From 6d91f27d9dce3f9fe72de5e527eff0ec9513217d Mon Sep 17 00:00:00 2001
From: krikit <krikit@naver.com>
Date: Thu, 27 Dec 2018 21:02:09 +0900
Subject: [PATCH 01/41] version from 0.2 to 0.3

---
 include/khaiii/khaiii_api.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/khaiii/khaiii_api.h b/include/khaiii/khaiii_api.h
index 0a99e1c..c064ed1 100644
--- a/include/khaiii/khaiii_api.h
+++ b/include/khaiii/khaiii_api.h
@@ -12,7 +12,7 @@
 // constants //
 ///////////////
 #define KHAIII_VERSION_MAJOR 0
-#define KHAIII_VERSION_MINOR 2
+#define KHAIII_VERSION_MINOR 3
 #define _MAC2STR(m) #m
 #define _JOIN_VER(x,y) _MAC2STR(x) "." _MAC2STR(y)    // NOLINT
 #define KHAIII_VERSION _JOIN_VER(KHAIII_VERSION_MAJOR,KHAIII_VERSION_MINOR)    // NOLINT

From 09aad0e8c4ea36b77ca6a58d9716ea505a981568 Mon Sep 17 00:00:00 2001
From: krikit <krikit@naver.com>
Date: Sat, 29 Dec 2018 12:09:14 +0900
Subject: [PATCH 02/41] =?UTF-8?q?doc=20=EB=94=94=EB=A0=89=ED=84=B0?=
 =?UTF-8?q?=EB=A6=AC=20=EC=95=84=EB=9E=98=20*.md=20=ED=8C=8C=EC=9D=BC?=
 =?UTF-8?q?=EB=93=A4=EC=9D=84=20=EC=9C=84=ED=82=A4=EB=A1=9C=20=EC=9D=B4?=
 =?UTF-8?q?=EB=8F=99=ED=95=98=EA=B3=A0=20img=20=EC=95=84=EB=9E=98=20?=
 =?UTF-8?q?=EA=B7=B8=EB=A6=BC=EB=93=A4=EC=9D=84=20.github=20=EC=95=84?=
 =?UTF-8?q?=EB=9E=98=EB=A1=9C=20=EC=9D=B4=EB=8F=99=20#42?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .github/img/network.png                       | Bin 0 -> 85450 bytes
 {doc => .github}/img/network.pptx             | Bin 59176 -> 59222 bytes
 .../img/pull-request-to-develop.png           | Bin
 {doc => .github}/img/win_emb_f.png            | Bin
 README.md                                     |   6 +-
 doc/cnn_model.md                              | 105 ------------------
 doc/corpus.md                                 |  66 -----------
 doc/img/network.png                           | Bin 85521 -> 0 bytes
 8 files changed, 3 insertions(+), 174 deletions(-)
 create mode 100644 .github/img/network.png
 rename {doc => .github}/img/network.pptx (50%)
 rename {doc => .github}/img/pull-request-to-develop.png (100%)
 rename {doc => .github}/img/win_emb_f.png (100%)
 delete mode 100644 doc/cnn_model.md
 delete mode 100644 doc/corpus.md
 delete mode 100644 doc/img/network.png

diff --git a/.github/img/network.png b/.github/img/network.png
new file mode 100644
index 0000000000000000000000000000000000000000..bfdaf7c2d584b033c122e68978133dc7885af0a0
GIT binary patch
literal 85450
zcmb@ucRba9A3uC3d+)tRA|pw%$1#$Gglrkfo>63vBRdIEa->LBvNs8pq=*z*m8@(s
z@7KGo-+h1Y`|sPM>-t`oZ|9uPd4FE-*Xz07pEv`3Z7OnhatsDTbw&qkgu#%IVlad!
zNQvP$jLhfG!#~KpbT0W~FckFYKLj5}Lv}G3Ud$P+nsH#x{AfU+vEvBwQqZ<w+8D!!
zAeuaf8667RTVcmK2}MOE@I3W*FM0{omy_d4i8Sp@ZHbuGl<8>IJ#W*JrbN-J;|Pp&
zcvBwq=s7j3n60`C8mBd$`LS=sGW0#?wC{r;;w+yBtIcIE9i+2VvH$&XQWKTO(EazP
zid&X64)>p*x1=0V|9uJ8DT+#+|Gz&c+KKu8`xdEec9ALnc?Vo~Lb;mGf8WDFYnWH?
zzi%Ne;6_XT-}lI(`~TrGQsb#r12#8X4!EqXt;@^HX^m8rl(^d3+6;;Zoh}_r_oR`M
zksbfCeMCQ}b^6Jxj;@om^l?}n;{RNki!yy$j{`lI`ir@tGW*TNQMqgLtncnC#P03}
zTh{pyegFQw@afZtHrrNM=#<P%5y4!od^IyORvGrhl<lKOkK%6KiYO>3NZIb`>A@u>
zF`KIh3u~E0n)vN4+7r^A5+T32_o{t##Pt93sZQT`bCiN)Kb2Zhl$gZhaHLy1I~^ZC
zGDt{Blvh?pR#lz6qboG+;~go!x4T<@ebot;m$c~j)KsMVmukLa$Mke=N=i!3f30KS
z<Kt7G4Z3mTa^C|bQgU*7qi-D@he)V7RaI5PQnr&1`dAYleB$AjC778~F#rEvg-Ykl
z?w?gVDOTRJ9;`N>OWN(4CsPwwY4XZ<37^i1+7`WF=K627gpd@?VB86^Be6B6*Ynvo
zd}uOl+oM0|^?U!P9%ceALwzj1zBECC=@#z|>vv$xiTTgJ522~dy?h6^_)cr_uxq-K
zsNayGK?uQv>TtZk-x5j9+WGQrKW1je*{+h1OW_)&p`jruHMN$rb72bQ-!h}Q1tnU4
zgn+tV&ajr|^w5XT|H=(da#@PCJoF%hQ{IpC%DbGl&Uh9>W8<jj&!vq`OxkUwY~SUa
z*3r{zoU0hSF(>X97awgY?Bw%kvfVXP-e2^?hY#bh!g(68?HwH&I?Sj}@H?GQ{i6&v
zJ6?ANt)pA)@63q&y@mm5bZN=T+aeK>k#rm!@pc_CZC}1{z#a7R@(P1HXe`P!y%MBw
zkuR;M++|PN?2B2I*UIBR@=<oJVFWs7&Zt>ib2o3#iea_2?cYCCE3c|zy8U;tXRIq=
zKW$=<n&vl{HqNmA`#gR3u|~Eu;dAX@>wGn}uQCg#ky+LIg_{(2?(glS-o72Nw&s_k
ze3M-%Xak$8eDk4E5w`2^CUm)-d0p;*>)-t>Q%_h&kiP%;z5C`RTEW4JPVaKyLrGaF
zi8<xGFhW8?`24?jVU$_>|K75D`WqV?TOB<6VGopoCflPame$r7{}ske_5ZWgU4$Pz
zc;H5jHo9SU7c84n%F}uu9YG^A(PvOyf47|rOSRi)!pOF;@Ixf+BO~U1oaoom_Lh<5
zS%X7E_E5!`FxR2dyho1?$I|d<UA%a)Ls=nUk&!96@REa#Qn>h)j43zor~kVP*ah{A
z7nzcHRFbB;lBd$nE4Pg`1sR%}9+788-y-X%ENQl(krDl?WS->NSr<&}kKXLml$0=y
zSZcfBayt=GQ9?|PLI8Ezk@H$wS~1<VSemz&K9XTl0`@2=wVK9d{;oaN9d>;i&%u^<
zRCKgj*^vG3v8JuHkLOeG-X+fO7m?@w`&n%8Sv}Au$^(}#KeK9J7SKv?9^N0ePo5Yw
zwdt(vQ;vtdOMr2Dn;Cvy`OjfgI<Wu6M5X%=RaSb%<QFKcX>m!6b<=aj{=Hj;R$hL7
zFJHznEDu~bpLgMjeqmu@>*I6T<$k}6roVrWOVUOc2kpAuNjhNBQtyEx8+uPz-w)+Y
z&CT2x0|NsL4-XHf^+Ung<#q~5+llt*Oqr{<QyQ@?reXFn7~Y@e_2Yv7zT!fBax%?>
z>p$2OSGonD7CY0$ObrYT`Tn<%bagWru4p~sj)9^J%ji43&(om#vbpdj_U&7}ne`iu
z+q1=(u+~;py<BB0V`IG2gd;}+15uE7(+~1eQ=9Ong5sI~&4`5<EWetuF=KmsdrE3*
zcz3r>e=7RIXp9WMo%T7sz4`l_*UoYmPge(x3T3+PV7WW&4}Lt|NwF+rYMUL>oQ8Wb
zIUoG|DmnS(|BV&H27!z>unWIT-z1JCfUQ9ujZeRTZi}II*HJBtl!5}GW5<rMpCpNs
zzisn`t+yeBdHcrw|8sx2&ieRBEAOpc$WgpbFPwJfUscYk%0-kD?M={x(sG3tYF;Xy
zr<?uCm2}r5<;-<WG4$qihQ9Kl`--pMys7Y>(h@v!#D3;|-qQLy6BqP!98MmS#4CvZ
zCN$DSBmbFLU3B6L+5cX358BPE1hzvQ92|_y%#+!k&29Hix_Yh8UhMBjuPnSjMXLUH
z!$1p?)!|Ql==<{>{JynM<sboymrqQL;`8UvGouZsh^g2G<@wS3R$bs+(fD^au>&zj
z8Ohgvf8&vCTBEP``$d@6b2wdB=-=I}vgiE22lfE9*`-T%Ycu^b{rM#G3k#ZEXrtn<
zB?yql6@DHVobzgWZ01M#_ge7Oy`7yzu(Ulp<P;PZr>Xvasy57a!^&F23lTi$L!&#c
zTi8lxVgJ4hY;1)m`uE$R{wh31<?r0Rdn@Vh1-i37VHLtV*%e=l(Dev^UC$v>+DK~Z
z%JZ)faJaX+u&Wu#iS%=o6Q4ah_Oi1x&!!~=jd-p^v^*oMXknfBQgz0E(-L9iLdC7T
z*#9lpDgz5kEQ?Jr-F=0Cn11J%6CJV3Jx6lKmnJ$2%gR`|lK;M*7T`zW=dIINRYEfI
z*S-e%WsLvZb=~vB!^5_3Gmi%@*8Lp+`7`$3J$AQm&0J^i9t&@?rTyRQN%fD9P%|<y
z*$osL8Wb6;&$2Kw+P=imwa>D^FSnK^+GlM5K6J!TMH&?`8nx8cvMw(#!!Mg|FllOP
zF0HOM%r!HUiv<oJ=55(k^f`N;@9V`c<YaNU5f%;wSF<nXYNgJ5Wp+xjTJ}SmM-4=t
z@nwjaHz{K*#dnFgORpJ-IH=G5v=WVOs5~?+eSBS#Dc$y<O=IR$Np{LM7VDt=b#(N&
z|Ew|UZD-D$Sz2BWw|w{UFF5pLXkn{9L`_W{c(AugM@!55XumsEpu)B-BFKfdDk>`K
z@G;9;+UDkF^;sGoWzKkJ@rat5nv`w$E*+Dd@e{v(9eor6gM*WkbMg9tYM1kLMs8R1
z2g<0rmumc}FKp%c@f+WLF3AeNWhGbX<E;&9sB&?yODA(7VR)f|H@rwodBpUaNy$NR
ziylv6w5JN8XJ62P+f{vyBOwu+SI$L0iO)~j7I12&aCLR%tBS*QD-w~A2=MZT<&&W4
zSP01R3qgSq8d^~6q$NDe%)-w?92^AI)zzau(I5z9lV`vyEp22LX_ja3b!bMDMa<Le
z=z_tefJ4ztbxY;To^5_u?1F~(ccDit+}Nw<=pIS4?MB=j=a?RR$~K!we8<>I>qsaI
zYt!tw0~>MCqjl|{VoW~Rm>ay^->L>(4~3~qXk)`p5h_~LsO=oAJGMI!@$&Pp3Z?b%
zHuTQ?Nl2hPOhlwsVp8KHii(mx4&aK534PpQx>6-OsZVW}1e^v0=nO5qn;x^WMUxmB
zXAPB?Y^|K+yX0VWI_KmAIqz)|%cd~tj%&lOW67`5Lyt53l9HR7OvNGF`r!jTjnW!z
z{k4S~U$p-2d;eob{RFtespSe9O=#<k(n%@2{Hqu9%i8;{XLFn+AsD~)keJV_^32xe
z&`_JRLH?NWEguc~1mXMTim-%&cP%Aiq4_^8F0SyOGYMYqWa)nWn*Lu4Gbi{*55!y5
z^VEXh+b>m6xWCwY{oP0{&O=nPB%)N1-pH)Owm!zEgPQf6zxoMi38}25rKN$hMNg!>
zC$VyJa^tgu#b%XGRLy^Wpl)~T-{2C*T}VL9yNWhikEmqRtM<|kpLi&XApOy&VlLP&
z!UWM_(U?5F0|pOkNnyJCaO-E95ki>OVBlkt5qjL?pziN~@>xkq^nj<not+(mh%PRo
z8!(b=LjDk)&5-C^?NgySb?W%s#$QPC2EkG&py)rtvF&H3!gE}I;qzru&d8Ksq>N?M
z)e$8B3u70vjYT^eJ99gX^6891-GXfj2VMB8EF$!THIliNXc-tH_jWd?Mygy9(jhZ8
zGK!d-yj*B3ivA3`z<&=64Nct-7*m749W^b@N~v@9Y=<__O~FUXbSHUG-_c|{w=Poj
zI<`2y2Emw;R}Ru@{8tr4&*#5TdlAH=7I#lL4R<<E7mK|NEoNfpcZ)VaMNu)a?5|hG
z8m|LEaQ*H;qs33ak5}jz8PPH$E_yLl^7XQvT{yz74$^T+>R5+Q4$^F&%4|a4>{OVS
zyn1@E*lU0HI$z!F>}+|^=3;{^>Cbof<6HLE%M3&q-#g2|$dAs>=4fncqGM)`0Th)F
z9XI98XFM-`^uac+uoHme?wu;ai~f$R+$)05E^Z*WUUjQdsHoP&3b*)R*`P=KZ`*G_
zud)~AJ@qQDx_TpNr}u$U!>9~tk>lRR<tH`;j~}<cdq=INrdE;z6TcCvgI7+D_0pwF
zRA1nFwG!E*TXw%onl)T2NwEZ81q7~cE+j54E)&?2pFEeSsHlOt`7Ky+%f>*upMAMq
zu(DBXC!EH=c(!N`m0DTwou=DkhMFC?EJ#mg`K&R;!aN4<^vry}jh`Stp4rsgD(Wr1
zL#hqOakA5^YJ-mJS97MG)&T1XTfW3FY<t-)ES9VDtYlLZpwqY6r)Z#Wd-?c8%%5`I
zJX2s47#K*ZtFxeX=@MJD=QnBqvSDGEmt9>QgM%0Sj`8EE&lV5Ly3i)EpCrF8?@ymO
zv)}#Q2mUzAu-Der@hK`c_V2;h8E(9O0xmMbudz-FS9p%h4@ZpI_1R;D=F@f+hcmz1
zR;azwnQI$Cr>0-AP<omab8-D>VW~=zPlrUNUhp?z{#jDkuc>Kik!q2|6D2ic>NOst
z<8yNql9FXwp(7v3wCOeS`)$Wt!+{KU_w=-7*o4HL%a+Z<<JJ9SV7`|CLd?p_djHEW
zv3du{Ck%GX`zPml{1(4n!ZtDsr0U2OoJ>_kvpxQKjbC(AwAk#df35K~eE>&zW`Ddb
zkLa97tH}b+xu&{mY`n?nO*>yW(b{A6I~><K9G<cueh_lx?Af#Mz4J`Vtnv8<3P&>D
z^cR~-Hm-d*``)E0MMn_rjU&0i5lumx_|j5+KP+5XL47?t5ZKml2m1lfY+AUxl6hJj
zd4MLre{Vp^E~BHW&Oh4?tOL)?G9Kv5X!owluk8%1nSRJ6YKc2es-mfRu>ph*A2v=a
zJ+Zt@VbN)z^Dt)2xK)^s&XDBz<(H(#EUKUN%Q#|bUPCYKh^6VGkZcTtHwPFyKia^;
z#>R#?P!)w~sbw3&+kofiN2(}53qXZXRkbd@qXnGpYpr*U6~|rNj9OLS2~&5=kPzk9
zuV2%0b5rN{<MOqW+sDSZq0r@grbz+n-V3fgGF#ZVpzihDmZ$V0%cRyhviV~<n%ev@
z8cuGDKC;#)#9$lN!%Ifac;UZkOH@Rb1dF_;{POVB_x!KrB(f_6DLM`Tp;B;VXL5sA
z$Vx5iBFf99oU6<Q1qEgPG^}M^c@UnSo*ou8z<pX?{_=-MXS(0K(Qt8biOt?l!WGt8
zwS;hG%et$3czASFKYYjwuRKglEc=HwTD{+V8Q1+5;CBKV0$vm9Xm|Z?+T_NJS9Ecs
zzW2~TgqD%<Fa-sLUbZY1?1{kbUr+ebeQ1yP2m2HeDsN24g)SWSDJ{q^KJo<1i%%+e
zs7T_!UgcG_qLy;u*^FGTx<plMN$_y!y{HUByI~5~KNuas-9>fm05=meGadjEpmq>Z
zboBK-d|H;Lx+>LlHha6eNR*WqJQp1K@t>ToUX6>0AQVa0g&JiuFJVqfO2XUIrDk0Q
z?sz9Xolio78nzbjd$-}m@$oQ_a2hv8TrpvVg(A-^YN#u(^dZ7~ZSm_-QE9_sa=(Od
zT@QlLSe<@D9PXYKOwMEJvgH>P{0pzzc*v4@CJElrF~-Qh^A+HiJ|)?_#cyO}Bo(-1
z3%flZrc25jPbH!8`L~~y3$Ndo^CrxYw0RXqaOh#bGwOo$x&=mdE7Lt*zP^!gsQ@nD
za&koVzWnrwIcQ_P4X+vRHQq|lxVQ0TX=$nTXKrXxZpb#7{i}pFXyJ$!5m7Mn%E&Ns
zobq7g$}SunF65%kTjY}`rZbH9uGwJX%G2yFpf10V5W#4|_4Lmftb@Lbt80btPkpSW
zrd?MO7k(;*?Kpela8uAJpeL#Dj><Q;W!^r=Vh8B5^n|CUr+I~hhyWyOVX-g8ceb{-
zt3#yt@xrJG6crUw(OWhLvoSF-VZw?oJW($(5l~d*Z0+T7$_?E=?dIl&)e-`1Q-4L8
zY)D_2-=%NB<>dVQywtUMvwp+ww>kwzipwt<baZvquU(V9wlES6JK{3XI1tPxJGHR`
zb_0Wh6Clxn5YjeSY)XK+(bU`nnt`E38?hY{7~HmgGl`izeERgM9ed*9XhZxh&41j+
zX?7x(=M`+srB@uD0EdkHIHd*Ny?YlpU(w|?(M}{NB$QWJNct`GK<Rc`+Ht=bgHkIU
z?7~PQ$0@JrZn{SW1u&VIF}C*hk?^+pAT~sNq@x>U<A^E=^POj)qw_x|l%Bzt{(8ds
zm$3XTDXlc>s#hn<Q(i62hC+L?g}XXj>n*Umw}+YuRKux17XC})ZI%pi=20yzDi4Ad
zPa_`L5b#?|O^x9B^XI5pT`PtU&Aqw7MoD}K)B5HO1ylu~mXscczn{(p(?Ae==DXvt
zxt!Y3sU8x37aPAAcFyKnCV7pM#D{kY^zZ581;`9NhKVAJSl1uxn)TeM1TA{t)wx{d
zCcD>_p}SK&TU{#q{nH^C5>_rUV$_Ddw{;6}iHWdohoN@gmsp%a1dBTplpny$30eX-
z)@BH3XlT%H7b(!toK3PG3#TZ^nwcNtX?tgdD!E=#Tk<!yB##^}!0Uy~H=Z!|OQv_&
zuBl#8q){alyCf-#SWEflSP+e)qhr8VZT{5O&d!2#`qa}GMAK)!)ckh;T~=21bB2T8
z<ph)>sEy;}<6duWLb1~UAS=9%)Z>zOy(2U%x+hXE+-rj>s1if}+a`pYyolbs%+6_)
z_JWR;ofyw<uCSe~V2Y~1+|^HIWKb-!05SXNaEoU=JS0N4CP*l`6<ML!@l&*^Yz@JG
zq~Uve%?=zrOhRH04G}?>lVS*i4PGDK(hwhdsBU^|^7R|9ib+N-?7+sw8IN&nn>9u?
zL;p~o;RtE(M`n>n9_6VtT6K0uToQj`bZmuwA4Md3R0~ukfLF<ljfykt+rpcFK!>Kf
za^<IqzSi175|=^@>|EQu?RDs9uRtrA`2HP{@2t4(?Hfp?gHG4o*GJ;z8`T8pyWcqq
z)WVz;T;ZR;i=i)?+Ww7?{z?3W6yF6wbYp(7ozr?``IS7<kjArtc565`E@{9?q+k)Q
zO});J#U=4lm>RiQoe5wN{_R9i2{=z&dDl3l%$`W>@9$r72UmD?>Q!Q7ZZ6Np{O8Cw
z8ItyYR=xuoZXIunwDa_o1P#mI06H*IjsZo(4+g``%slg<fMRlT($>{gOb+(usy#sH
zsHFg0d)A~tx1>t4xyH!>pR=fbJm&P4>F#{83@>PBx00u#=X%m>sH<cvz=QMY@aPsa
zfjdHH$gsar)6&-7P6!te+tj2KL&KBw=#gM;LxcVI*Xdi6@sdmHLpJ%Kb>Z`GXQk;D
z%#2it@e2re&%Ed3R1RiEV#4RjD|v%1#~r$9^{?t<d-eHKc~6}J*3IN~?OKdUv8J9L
zB_nzCj58K19s|6KzTN05RRBjQ<DH+pH8Ko#65gl?M(-xyWGywe#@=z26mb;o_Z|@}
zc)>8wWRH|&mq=(jh;BW}4Pggm(1VpNH9I?=Pa_6MG7%ExKm<E}W6jyjd}UIJ{ycPg
zz*)e(IZ&U27ff!i&jC9yjAxVc;swA>3oKinoX)U#{80~9P!CCO<LXz1^@P)q09A9P
z;6-O^7@60PJKX^<fA9j=dt_3aXCN{@J$)EcUR@niTr6gE;R46Tua8e?;O)HThmRW`
zkGtmp9jwB>>+<&6FI;jm3(T023Q#L{cUSv{YdmezHiNETuQ;uPb&3WpM~C&47*CXZ
zKx~bbNEsn&-TW?m(pi)Q1O&q^p`jhswY74)zYq2$mA7XEkR${%1cOmfP=HTb>>eH6
ztKLk{$mn?W>M_$tTEigNR}li*N#h_-zMoqM3W1A{kBpA~yYA{0YQxkZxL=0y-}>`2
zH+t}3vEldPmSbbVTp0{v03hnKrWY?x0Kz+EX=N2XP*z&{3K-SI<Rt&I5a=TDoEC6f
zN1D#Zb7tcTOP-c|vfKQ0$ji^0)~MxiJ$X>*wvC)z)w`d!qN~WRm8wqs$!3W#zBsq^
z`<(W~tl$qpYyT4JZ$eyAl}#ln+b8SmnRhIO`0LZIetbMpHDuiZA9eid2SLT{nfxt~
z87C$t!T}G#VBmv^3vL5GpMEq0=7L?%Z6Q(`ny!axk;Bu%{FtyFhxR|KKatqJ`m^tp
z+vetGtVwY+bn;99Snjrw<y?iu#Q?dUmP^)ERaNPnI~QG2lF4vY{yue4QBiv-@xU)X
zz=OD@?N6URsn5=>-<a`}M<NHde`rXJ??Uz4in%p8Uuh1LPxqKl?~l&SrKG14pDn(7
zV3{iRD054f#o6tSh{S7+kq2+y53+epZ%LL{Ft0{7>U*~KU-)U~c(hbm#WYSQD(l3R
z%b-F*tschc-IF2v_3PKxQU?tZtdATya?JdNj+ZWeCnGnPCJq|`dIiYV#7Z~TRKYk1
zT*>WzTLStukTMJ=T0I<$7=Vg)EaMDxRhn+Jf3>5n@kkn4Iy$~b%^(~wA2SOBB+9qm
z)Yw=K;1M(=@Tz|Oilb54r3D16J_~v|M?8xpvn<Irm}?yR{IqmghthhYB3pJP8Ks%(
zgI;L}d5&GZ@@(PRQhWCKj#J&mmy(0ZUae&&{VHprkZk$%$MWOgPdq+ErJia-t0j+S
zYJU9wmXx<*#BOK8+iz9Bse+eQ(nyyWPn>956ga=?n~sYD;#0HxI7cN^dF<w<)tx(c
zT6=pFKki>>VC3SW^7QmX?iny3xxgh}aDJq0TQPpWJ`hP~=>ZF&=M@nl1)*+ctVscs
zme;?qakzV}Fqr7*`a@4!_%I4x{A2J&TvPiOt+BL9rmec;UWoA!Pgg5h6BZD?^jU9j
z@1>26SI`JJ+%w*MZ@-(y>mqUO+BH^{5Ji}};n4LE#<r+-Q`6NYf2^0)zD^b=Uk+eF
z_=-|I%xsw9yy?0`k6@efY*C-Vx@%o`Kk6}X2bZLVb(+7i&g8&6YPovqovU+YGPjJW
zOP0zrhf_m2T&LZoN<|nW0LEmozzXN&a6S+$pzZ^25oo!LuwEvHgoH$dk=*W$Kds@L
z{<(ET!9Xt!FKgLnCyt!BdlPuL$|eAU2&4d?s|^&O=wYd8AAC)+x7QU1`mIy9?SR0a
z59fhG^LD-Sssu1vq+$4mkRa2*P7X&{MgH^YI%%~-K1{QdU^OZ3Z%vss-8kvv;}a!r
z1KM!l^SkDnfq`<M*BM;8)KlaeC!eyi;`8Y7<H_L)`@+&vVhr@33D~4w{!Eoo1ayW$
zuMi|4^Dt3PIpa`klohKxN@Z3wk5>AG?{ixfun_I?eXuWrZ2G<B1*?vczF}z&KO>nO
zk(hI24w&J`TJI^Z-7RmRJN(MZ+>s<S9zOT~v_+C=l$fX(P(`+3?(4_ix<w8GH6kJ`
zlGZO_n|s46kbYWhTDISGz91Ynf6U3^hbHkH%Vk4z(Swf_z&bvsuRw}{dldma`i`!n
z*b{y{Nh*vEpp{c=zdlN_Ct@(oFV(02PZj{s7HlQ}vZ@k4k9xpg?ic<20BQ6fs4s7P
z@f47kuRp0vZ>W%er+C7VCBmjOS4;mY@#R3D2RL&>wLe7DcOGjXhP&z2{b)D*vyoAC
zZfj5<RYZ1tIuetdzJI=ws|k=B!+Z278A!k976X2L5DL6;V{CKz+x4GZZ|^JA^y~nB
zm-79g7rZ$dK=U=&?G}8utO0iek(a?9*?PjDU{MsEqbYs-m>9l%47ff(Y-l++-lUdb
z-7VqGrB9sTOV@?XKu}^r0b8lR0yOz1lRwNiSEkR*e{134kacSXaK8+FpFI0R6mFyx
z)8z~ThfK&(rD&ZY>y$H&FosjJL`o!p_13M&+>fptm*L872QNMx*FE0|a8g7}tYJzS
zR8z22Y5aBr#;%k2{&-tAMcSU3$@<K?iQTH<+RcaTEkbF!74wy+%PZ)7`+G*#zhqZU
zoGHl%5drvNEkMAK-DT}Zj~?;DoevKW2WsE%?l<>|3Vfo`hJcoZZd+d?YisHzYsvYW
zZ=7Y+pL`@MI*08o9+ar`_4n^g;nOI;vF?IaDpT>gv{UcBXfOyq-BQK6=YVB5Kk%;B
z&?4fE<L9Lc8Ya*@-eMnFNP?Ex-v7bMj}e>{=HuRzFPY2uGTy+Syng5LA3tJ%^J#4!
z5f;9@`lGk^6R+S6TDaE0A6ZvYva;fUlc6d+*dA)3Wo5+y?C|0`!LJ4d;tfa^0Pfb^
zooCOU-TUfM3jY9Nz63faMN#O~E2_2G!LvH<{4UldpLqQO0=hu;lM32!gL}?`-mcu|
zyYAY;7nT!FZ(Bk0w{z?)dW?IEm~Ybe-|G5T<q^_+nj^!F7s#w4yiXvXMi$}{jmnD4
z==$~3l*Aoen}$WkpcabX1+BmJrQ!5|7FJ7(AI2tRApl3HyvtQdhMt3F4kQ!sQ*>EG
zpoReT<_ETU{rYtbMny&ConinB*dh?d7@jjWH-BA}6(^4~$t_fY_i_F7tPQ4-ROo)N
z+i*Dn;L$MMf_AtIh*?r{${}Z>1YDpIjAw&E&atn_#TWmGl!-?tj;MjY$M0g7zOJ=v
zSsF<fb~sEjI|_^Kh4P_>Sr#zb_xUqhUS8hTYVJV-%qSBbEx{y932hio$iD{(Y7(q<
zux1#z;tM;9i;C3D&DoPUPk)fwtI{n{cXK<Ilaq7!38(=)mrYiPVf=-Aw(KzjZ$H|+
zw_$P2rkOMQl;<mOm<ce|)zzS(yJTEY2nLr0I6(8l9u#kQV&XY5Aj>L@(n9hLo3RBS
zh(C!2A=$Z*$*kuQ?!8EgVL?~wEqNW@@MQDhA!85}ukLOwZ_QWs8CY8<h3swfNlDQ{
z>4Nzu7rZ3}9vQUsu&-Y&p~8^?6}Z}a>Nu=je}9~L4w#<w^z>2yl|>jE4dM!g9332N
zKR!MOhz;YRzrz7e8H{6%0wXegZLEWJRPJ?eFR#{k$z~if&|o$WHv~v-{Qi~<j$>TX
zMr+7kp8w=avbFio3>brd!Z&m>FFId*!%{#3+PU(z&u-g$eejFOFUXz|S>v(0tDC8(
z>Qu$8E$WwoS#nB)q3{_r)Q7M=p*ht&oATRQvI7CD6_lQ0lM=`lut?dpA3kPQj$t^W
zkJZylcsPJ9nY-582Hz1gKBJCzf(=CcPj^KM78ou3skkIfti#=r0f#o=ewn9zg+G;8
zK!j)a+UXm#7<_(yKM0#{pPv)uYbCZ98s-C?iHEmK0Ff4C8W{JpVMtiMa6v#`E4$)y
zCpp?A09+s;a46)@ca~#T^^6cipxRN~7`8_^is-Q*U8%9<%bxbV;p|H4su-7k@0pe;
z*E_g!<JQBx1?sV9>lMn<fp~9996wG6OWsfb8ZJaLUV@sROryMof%jY5-0Zr=BxYb^
zlMF?4M%5f&oe!!kDkAtb>N)^ga7ow!iCua1rJ9`k_m8`NI)-Ew@H(7)Q*b&DO<Za3
zeaUWuOcJ2xjs=)u`eM3J^1z77nl{+${pW}Q!h`ty^5sigqzdJeRl)aD{4<@_-Q67q
zY6Xbu^b8Ekz!|0x9LKaeNZYN?4z_DQ^MnP1U=^!kpj3j^l=vnK3yT0~^?;T)M=|?9
zvOUMYe~$)Pxvp-qI!InaJjrr|7(?pYuP$E0n1c%uQ_sh4f!f#eU49w~6>AOfl_$Uo
z0yDC1K^dBNbepZSnS=m8KOv|$n=d^047Hu4v%q0F1oj^I-4jrWt#`#O*sTI}um@!?
zUNC;CcJDya8IVQc&5Eij_wUA_XWcQ$s=3;G?_|oG-Yd5Xp{WZ<O42}*#jMJO31bV1
z5)nzszpc=1SV~SofkOWbmP-u`e|YPdBM$-z8HDv(!SJXE+Vn`^JWU@<&4t8ttTy9i
z!2c<_1>Nu7c?7ojgOUe+GIS!39XUW&r^W&g`P4NIe)@7oM<)!3H6nkZ`>VN`vhL>j
zJ2p4hFFIVgf-))L(b1hsJe#i)IRf|Ahaip;0o>wb!2Gp2vwi&m)~DmnzrqbN7V%T>
z-LWorqzJl8PPS@&dMw<?%kPpQ!%l_dIJmamKYgJdKDnTy5n}x0tvX~uTZc(qqpO9l
z)|JRWehs4Qh~K|>Ap>xVAJT9)cQ&F;ioe%KV=)qzwPAPeFd|3;L?4rK|Nf}Q6&$oP
zl9CTqt4mAHiPG+`K#DW8v`he!o&WG527}aWjEIPcs+HCK?m+B<Zb=^dkn61pj|DM#
z-RuEinSigkVR&J}hHOH)fE9Jkf3B>6IaBFQ`{TzCJlw`aiC3(^gNdm*1G~D|619-M
zoUFK_b5`4pH&tipdxj1ATY+;WHBuqFTQS8)A%+)!vDiedcyKbB#k#g^N?2IfXFLB}
zx42R?4BC{l=`#~bltyG6=iM*e(TcOBATl;uy%!yUD=dIcV*B~IW8k0f8OH+_ziOTM
zioAQz#${sAR{)HYU{Ea-oXvPR=%T49>~aAnsFcfq2y~M9;fh1hRfp?*j{yCb0@<*C
z85aeuQs?}6ZNTWa;^PT0K$w9af_LE+5WoU@4Q}yeaLjrYo{T{lm!vwtH!xdV{%77t
zhqrBJ>F2|gBCNKHucDHY=jV-Bm0bcDgC1KxAuXuc7#M(a6c)t3S28pQfRP4Jlup4Z
z4hiV(FeU*vqHQua0{F{1+PbXedsj!5>yQ`Bcu;E9h0{h?eWh3_%WG;{e(wE_1uqmF
z!fBuDX=MkYuCrBFbpc;n2Q8d@o2$a}yK$8ok$SK<Al3ji%L%nP$i;JDf#j6iLT8tz
z)6Sy^O5M*ce~0vX4ms;`63qL8j~M4WIUgP?*+Qo&{1d#^uZ^^<E}qZ?BsGF|6b1?o
zpnCSlbbGVKWwwrvMB_j1ENJD^0h6R<%P#B%64Yj&tP}%)0IGrn+8?CGbTUDh-`$!@
z#$dqVM@Ve6-aisvm6HI=NhEpS(+Nqp5YQ`W*st@RikT_Ix<4?_3r{N`nLS55ZKWb6
zuivmd*Zm!E*v<`06BFit3KjB)hChF{gNz^R@ymx{ShH^wmzBkW`x!;Xumrfc4Wf}D
zAt6A^HSO&9fOHWiYDGe33Dm(3a3z7EdKk}*27HS)DtZYMgpP%U1ezWiSKye|`u(Cu
z-?o~Bn7AF>GHBDybH0J#mggCZdhc!f&jZQTEdb%4{zelw06R#205;M}OTIass{PQU
znBZ%$<iUEy>C>mFc7nk@0dgJ%ZKJWV@qN*SBW|Ph>kq868-Te}gS5O1mpkz`>jauM
z-ap>5p?nqOOPlAWLLFt;QJM=)ObX^>wQ$q?x>f=@YzI2t$g(zXOdzF=u1a1^OpH_M
z1|!eSH9B~|Q~;JG7BzT?)kP9hwL@fZ88C(HJP#vd6l@26NUhqs;H$-Wc`<#8t9tqU
zwb&o8hK5*<A3x4|%7YWZUU*F<s7V7!UWe^Ytgfy`VtqD?M=y#dMe*0O5G3`Bl-5V(
zd`Khpo9u~GC@L=QfUuyH{VPg{wod(elH2^V>&N%+d9ac|@m`6QHAaBq<N$!AAFl~6
zB<NfoFIx1UXaMMf2wTL_3y&XyVh+JWx$7%}pPpTc1Q!CYoy-m0_!ab;M3uc?=OO<T
z(7|{S)*Jo+UU&&1rsFra{hpauk?(9QbijQ_nX20Q`pdsQ6o47g4o&#D<7-C9&yC*x
z-a^GDeHci&*$ZbnG+iMu)9oM)f54l7nBS>q(MuNgT}hhG&LThha+6WIEk}-yKJHQ8
zY56PfIFUOH12XfZD>H!e3YS4qG(N6gy*l%?Zk5X>yP=}IycPDB6wDpPo%u@iMF<Za
zO3lm+JRcg{au4n_HK$w{wCLs7&!N9&Pm}(NCV4;Tn*zv5!SUjpomfWux$D34Z^hwe
zgwu7SFI76(!9rldAjxQEZZ2E*yWFnB`~<sUtdW@+4s=I6jDVWGQ(6Po5a-O1_b1=q
z-#@p{&7*MbSr{6WoWs0sP8Agun@TRRIFGV4cZ{*mS%|fE3_cB(yDATo@(<K7ctW=L
zN^M%GVaRQ51^#tQw}S#sK*rG6*w_}p0U|<S*&l^AX9p?B$jHtq3m-W`0-+0`$LDj$
z%Fh)UPi`$w;Sv(uhJP$BCILiXgTdY4NR<szQt-|~4JHhfI6hffW>CSwQdQT|A_n&e
zg%5Xsoeyn?lu?D}H)UANmFhlwh^DbB-H=f}*!6{dI@Po}*6?#^9ArD~n>UrvZJC*w
z(Ms-xssQjf@%wi&wDg!FQ}_~gl#pLSBMD&8;?~I5c(B$kw?~mtb#1S_k>t=iITjC|
z(B$6cw=PtU;Ew?;-EN)@a4^xeSX-&vJ5U%<W(1k-`BQ3X%uTmJFGs+yGvO%rHQk+x
z<m&fLK>+~Y<#QkmJPw~P=RJ8C#!PH&ZEcWwb9kY6(X(eS`}4KE*Jcbs5k)r$X5qsb
zGq@WGTMIypWs0!G;6dgbRgQp|f*Y-D9)N?u$+#m3V6UzDA?LZ$skgoH`YJ6uJNftz
zmYv6s>k>e1g-B`)6$dRd^QZTmyMvH#qvPaEz~k{2++r|Ry9JVEMt;oAeOdYjrm!4P
z0*$oUmG5tAIPGByBzn*Ql|sRI*fAU@Uxio=O+iX8;Kot&x5+B|xcYkew>hU7U>Ffc
z=J90!k<fpj_}Tg3v2}J3^dFuX?<;)|Iza}fqo<ERArcxtN__RZi+qtS=1CM)#GJ>9
zHEk?TS`G^e^4EI~4i2JRJ(xe5+I+JqlT%ZvckV<{ut*?{0)$ZnF+q>UY75jCgTsr+
z`iT<^>+9<njH#(9tR<;FxRkw#C`Prq%Z2xPyhSRJ?wDUW!B;rVj2Ean3IwGSTow)w
zq%Xrxcva{7ll{{d1#pMbOt%Z3Jn4Ye4+cK#Nf+a^uFaJdm*;A%TwJ5h?&{-CzOK+5
z%xb;(e0+U5MmOhZ;tw|7YtWpm=VD@lkw!Cn@MC6%6jNAKbXD5^tb*?kO3Zk=sdXC^
zjnA7sP>bCB8j6N~Sfj8493$ugu>_gBiU+%^R02nTKf>b)p^E_8ZH0{IqU!*kVGw_+
z&ig@KGKR3P=7*^74B~T3*x$MR&3SQ5f<6Gvk5@r~0~nIX(W7kv-)_cEPdlKL$(uK4
z{jhioaJ%h5T&H6Bp({uq6%(UgJlNjXciu}L2r!vCdlU@dFc4<7k~rgl=*BR#V6o@=
zMcl`o7UJ=@Y*4s4;(h}zY2fj>Oeq3HyMVNMFHdR)1_fyp5ArK0?A*QC)YO!kkwJoi
zjDyCLk0<L!zkNHn>yFkIwiCdo<IviX5r6mYUDSrty0#W>JZmYPxY4kH-<7|-RFGlq
z;xlVfWIQr+?Lt}aC~z5Dh`a)1=>lnQxs8O!c6qWh5K;zGo@0#)%GUt5$pvj3V<gwi
zPmOO$(t>=203jja5~yNRkl<|m?Y3Z8ut@yO_!{2cfC9oJGfo9jJ%@MkJLl8h>r)*e
zLTn#9Pz{uj55^eu;o-_FG!j<zQNT5_dINl1Y@jt5n3xb(nI{epn?b6yZEVaMIc6?u
zuxJooM}!XsKS`Ee3=FT!5blJ6K|y6K1XE-sz6w!~UCUc~4(?Wg)z<piU`VNAR3U6A
zB~{*Y6Vz`|Y&o=g_D+H10Us6g^pc@rI5Z*x;7ov>Jcjqbxvo~3g5`n<gX01yCn|1a
z|Kn{Iy34Tr!2gIqG8c4!PG}~_udhsN&B^cIMYl4U`{tW#KzmWP4a@`Rt}@w2?ck6`
z`P8n9oU1%2f~4IAIZ}D~sjxC0H+==|9=r?NbF2#NHlZRx9xbjgF%=v&u&JO)crA^q
zJ=ny{o)u2B17j2gTR@RP8Cdi1loL)+#@Y~Zy7qf45tM3h)F%>8jkR)kHiiRejd7aL
zpi%OBQ%K1#eI8t4%g}OUFaSL9cu@iwU1;`^5G4bBcku=#tTo7WoQi?8$Xfi?e3N6*
z;<*EvuC6W^@V-M+h>n0gqO7c(SY2jTG0B~q7zkHlVq#L0P(43X#)BF64Bg>?p23&4
z7s@aRwF4eYWJANwp&>}LC?TeA-E?CD@DdTm%)-JRiV8yR^vH#WPhG0cjRWJe0~BD7
z&(>B}RDk$E_(gUjO-`B8@hm-^x))U2H@NhK%TEu7pAn>gc}dJJf{bZh*sN#kekyLp
z`jUkOsz0|EAu@DT$6yXa;wmml&FQVP4D1vR5WYyMs5C7tzl!ewR3z1bR)r^3aCHKa
z0l9O)D==X&g#@72eG6WtQr_RP58hqwLU<CWfu<%Hj;nBc6PK2}61bJv0gnt<yK^Ab
z67=ZHaG{&Qshu_l+olPu#|dHFq9`1kIjOkP#~loO62V9y2y@3_y=~py6BoYJOhI>2
z0}5LXan5-#;5U_fO(4*PgrHr(lYyXRf)KfcjupT_($Udrg;nft2=G+g{OSuSYgHJ?
z&0CWcAZa!B5RW!(`MVg_Std!g>^|@Jdx~%t0${MKCf-Q4yi8y_ac{2Zn%O0xE(0Tf
zHqknv`XDMgu2}QoUg13QbzxOLCIzPBpyE%?&nF;)jJK!l5B(pMnEBSCqOPH#*_B8A
z7~mdEBvs&A2z~tRx6e^d{fFk3<2uX`q((vLE(yNjDLqF>T%$Nr+b@uwtIwnf+kwim
zHN|r<TrLiZf|$uOBys-!BL^#n><4u0;?%cbhSuI+4<RQQkMcFAPSy19)5lHU{gI!Z
z&IE@s;z9PGno3rQfm8-)BeI@j2?U3TWkwE$goizoy&D!<#+rv>o*0WZIgYkJPU;_6
zMPk#t2TJiTUt%HhlE85ag4NV!H6C2nH+~5{v#g_st8$h51my=%VN-#7Reyi&i3&R1
z0)C4;Y($T`12`I@;pZna7QA{6@a1yzr-i-sleA4S_O5l-7ZToz#bPkt5U+fbZEcz+
z&s>Bh+#p7~S?MgCAp#zQ4w1nz4@xfiI|6kx0S(s{kZKtL2V(L|OQ|uoSFQjYkppi_
z^Y471h>{^ae_uWt*uZ7L$jJTz4G7WW&EGKkmy@9$R1Ii{*aZ0Nvp2tm^89W&PzLgg
z<^wvV*48!=o^<m1iUY#JkRgrKEr<XbgBb^>*V8PLrnXKZyx-Yvg%uVamDcw+84&qk
ze_R62TH;0!)E;sL##Xjq$l1QWbF{|o_fXk4@vho5;JRba#X&EF6c<&E5dE37K6^0M
z<si58ys~lvWIrH`m@vRSfMY5gdg$Q@9W&4^P=pgDoa#Qh`YMQ-mNrhU<$}sfrSfAj
z1B92GO&$wbZO*DFO)0}hH^X`;fJ5c)IfIb|z#2dp+diypu=FAQ6a=!Z+=D;4F~3!?
z1M<X_tcan)2BHB*z6BJQmWwM9Hh<fvPbPkfF0pGjl+gJtuy8S9Wo42t%*9~~*5-iq
zc9Ql3FQ>vaw~lRn<hpSqEXpoJ%ATl&<>Ys{LaB9V<;}Z}FA^`^w6sPO<KrnTEiFSf
zy&oA}%|Gg1?J;@_zQ;JEoSMs~{Gxv4c?RU&0tH%6PfzSgUc>22wKPoiM#F`iv{Pf}
z{W$60I!jMr9l3VJ8OzgF50lbml{vn-JJEaU+e=V$fSjVpC8|u62>BkghK`Wbd<)@#
zLr5rEQUWmqkXnEN67OvOMoPzrj~~6a*RCR&7eGd&dLAuT+6Wyu!5*Xac_7?^-whT%
z_&0Qh9D<M3<U)1>gF$u{Id<$N08)@qBOy8q*gHoBdk@7ZAZ10-(EbEo2`z)kTt!ya
z#CZ>fLf(2l9v?MW_VlSbu<{P)=he2J@8nU2ME!A8Y+}od99R2u$2Xw%F(DHX1Q9hz
ze4|Z?<Z01#a@0&TwFT?N;cyGtqZD0-nTh+DEaNT8*#N(F-J=iK|J*L$Tpt;cS3FhS
zQodb3;_{L*gysWtW2=4l%dyPe6_H0wW3BBh&071>VpN+KhX(S#Jj#f+vNf_QcJ<9I
z#ts~J|5{7er@V6v!o=J+X81r!MKm8sY$r^S3a4K7Bz`^@dtfgxp!$IfAh?c9W&c@5
z^!NMw3lr<Tr-&h&m+gk3W@|tB8OYqbs2sIy@TCED#i9^JEKf))c;#UyT|cS*3zM*(
zaH52n&|7S=-_hGkh0L2MapTc~vn<cRnTI2;>a!qkv7s?-Yx^)|8)*8!Oj@T5^c4IO
z^e|by!Qm0e&P*&WN<7+F2pl2GfANDH5}K82!+OYV%xKvqLY{<qNsRuJNa%QJX$&SM
z=A8_m2Oh&5^!p`);d^HVF(_P;uW8b|jm(`M*QJ#?!AJW86J`9%zYn5fYce<cvg{k|
zr&J!qwj}$;4@%t&1BW?Ctg+<=vwR0>Z|HV->fX1Gm`=~Op{dQ}`Vd+x;dd_omR8*S
z1u0B=h}%NOg2ZcevIDe_$=rk8BcR+b?eF|XLhZQ<z>B~pfkV-NfCb#VKeaHi2ZF6o
zsG#&5f@83Xdut#6R-=O8Jbh3cmJm!yhzmn67l8S!vb%H$`7>}ZjSEU&ab<P&Iw8@i
zC--1*y50SC=u+<b?+jy6cVAQHozL35l`~fuO{VcrOB=0Z4v*BMjdJ0!9+Hb8f@A0C
z7@d+rO9`@pzP>)VTcc?kE7rWJfMu>63Ob^|w_YI4<~Gl1`+L}js$L<r=aJWonuz!d
zZ}xAV&eoGNeyow&%*{rBn)~^M4lRjjJ<;g=`zI|qUq1UOkghxO^`|JEzaBcb+8Ta{
z<6b1j780+}KJ!8O<_TURfSgZcs(?1R-WhCw&yxZUhypz{7PUA|73U<yikw3r3BXMW
zYOlY01vyh7Z$8(*JU9~#1Co0pmlOCDC|{j$K%&C^i!3NFnAU@efZTAy^y@6GeQSaz
zel?|Pk!BK}345<jKO&Z=!-Qq*tS+pbslSDff6ZT7N`B2t?UG*!qm?RYQ!KqaG3gl{
zA*HN=o`KCsiQg+92=!JRZm;gVdfS`Bz$j675R^;eQgtsthpzu*(ewtXAGNl1@1<$u
z`|2<HRQuEAqabnvP~8qj0KS-(ZuEE-HX{6;==|WUfs2&=6O&lRD+>KpaP0|N`E@_e
z#efkET@h<($%!@=q`@+aHEBPy0wjj1#R?rClBD5jY3^S|v}-(J!pr&10AXgw>|Ps{
z0qghX_S(d?b6Bk3mncfo_*=J9?%s`oJiOi9rxG~fO}O;sz9GDi+^Zz6co50q7zf=t
zuA4Zju8X(q$1$H`eSr@RJdfzd4B!lv>9)I$b8M42*Xs}-dikIwLUKBPYLJB?y+Cp=
zJG*TA^9VSN-$Hi8iD|eK;N1}DiZG||U+C^LFIpV_KUsi2q+p<^;KwKD?cq!?Qps~g
z=35~}{{H=YbaWHsY+iZ!d*><X>H4`Ce1w4zs)62#;RWpk#N=1-i6BxQ!Bf+H=B24p
z$uV9mbT7rG%5sEDT;Avw#8+RD&h{!FXTb(;#55e*6no5(c2<ye+AUVeoNK;-e!yff
zz7!~v@Pp9(>wq5^(R)J2ExpllK391QEFLKk(~ucP9<8o=mcO10607K7Pr2h8W(Oyy
z%dpH~BejAj@*ZZ4#Fa4uuwC-twO~*&TS4K;)AAi^ha-6QNR(B$HcvHl;Rt_ys?;Vc
zr`T9{&eM;awDkV`PtI=U8DtZcloAF6WD~?;`RnPce0|~$$Gy9jLRg+Dj-R8YtBM07
z9Cb=naH8S9f}mO^?*tqXwc7oajFxdwG5}NGK36hrFOxOSp=>nVLiCoIVs>t>u91ff
z;KN^94b~7aLpb=`j^-dlK`u|ej2p7NoH<hV>J`Ld9Ug)O2Ti$g_U8VkAjHmj9aK?B
z><|W>G6o1t4_-330w_BP@|%?NdtSOKT=%QP4I5#fFRPc&SwS@$avUv|Y^__N)Ztb8
z7IrY<m13^<?!XjR$Z${Q>W6d50FBzfn;&_YxZdL+?cjFlkaq82M_)B9Yqno+o0y>m
zIXQWIezNi{R2fjAU^sy3j(0VrNWXh`5+a*ON`(CGBcr2Wt90KK`V|9(XD6M7$TCHb
zr<+?;zcb_QbJ$Np4yw)-W+()JBxkAYUem_S?YXikXs=y>qTFZSY4W=Sk!O+9Jy^!O
z!iD#9_{*7<lVQ9v0tmkjHb-!v16|J6+>2g5jHf6kdUXf5A=CqB*1?D377-VZj*N_)
z(e*pzRLO>p<G@setYpxibzn14y<T!UryOOo?JCRw2uKRrJo(smdE)gkELV66!2~+G
zgZ3^AuS}SRH#Qfkf`Wn^93Aa|!)?J;A~mmNZ^4s)7B0e2;f~=8s&g?nY8AHdv8E9-
z@&g2{wy(%83GtsREoWKny9^Q%e{t%Y@8GMp{0IWV0BiQu^92_=dj=;-sW1kRM;<I@
z0Z)7!hGH$-XL9&FIGlsL4fU1cr%zG(l%hwNe~#Wq$N@wlQn*imUMA+}ga7!SKYyN*
zRSE+DmS=x|v&9~cVl}#=2OD@*Y@0GV_i0F6<mFrP&=E4~PcJ+qCtj?T`i$<RkrFX`
z<7tq=z{>BrTrm#*3D(MrD~6iu7C<27J<<180i)4)JU`lIpv}VRE(R-r_wq%uE%l#<
zhjDPeTJOI6`mKra&>*-@JK1c>Z|Be{+WQ$jZtX>9g9tDcAxGaAS{^=yPvZYDO?ah}
zo{C;zLrga>@{Hf#pl7kGvvb&IH!XK}aYS+e9a>eq5W1Axd09<Uv-9M)IY_xV^*u-g
zh2K{W2$y=*pD+W73=PrrF)yXHgGN&$qYoyi$S)mZwtU)vtN-EBe-YRH+QEp;s9s;~
zQ9x)^t>&##4K5tLG^Nyyww`CIO)qgbZ_!i#l^r_zon-=n@gDd1DkpdP^aLm!W_3P7
za2zmuj=ExI1Lisk_wt199tTU7_`ZI?;Bf#3Mn_Gta29!aWY_Hxv}11WB<R;x0Vh6!
zXiJE>eEIUob9TDV<sS=nDvkt*zsulZf{tSzuJtQH=XVI+)qvzt6y>4Bv5eTLo$m$C
zKT^Q0!^yKk-Vx*iJ8(%^r5#A&DI4f%ALu;3cJj*48HJ00Uco&@`6DC-fm4V0)9cr|
z{e$rG(zD$Qq5uclns4qP*9;Fw+*an8<B4p7mx<6LCBX2(gL+eP+fz2%d-3b(;Em5$
zAO~1Idj(=u$YI17n3~4InFWlhxA$>&4h{m0MU6+5i!dF5m1X@Q%!-S&ozyR)8u;E8
zD__#0#=K<2433<%&23~9&eR-275-cvkT`6!DLC(=2IpA-nZo&Q{`$MUq5!~=Rw#7#
zuI<789+)bml$0+Yok?4QV*zyF1>nWtAf#1>C3XgQdjhQH1W?ZokZ&bnKO+YWcCa=a
zSpd(ud4GFO;WaWtfsT+jSto-cJTNfu@Yyp;Nckd17#@Ox^hIcuX1^K;AwjtWCeLKQ
zHV+X7-Ay>P)#cl#6bs;)5_u?)vhjukh$tlkO;j2}86|hm4`iMGQrDrekbEnKe|jiK
zGgBc~RhS=t_rl)3;DoZ~QNp(_-7mjtqX!)f2*Xu@ei#Y!6i|4-znclHAPS<T_0Y1h
z#X&Tt4J1%MApdw~I%!^G5k@5cDQ}G*LT(UHJB<q$7{Gp6ZaqYUJSW+J*j^<2BLEIT
z?+D$3`iTm}iG<HTNPt;m3;P?M3gW_1B{)0W{`Im!zd_Xe>ogcxpKG6Q69!%-)Es)K
zuSLXmn^Kt*J@mrxiFeb4Eie?oWm|p|=OBzlaqiqX6pYT#&!3TpFHwW-4x`1uBMy){
z^;c{$I|f!S#HUSmb^*P?69jBll`20cz@c|kEg%_aojKFP^|J|X2V$U7?jtftn1>)s
zL}lgbcOfk>;v{(*;E@n+U#h8))w|fdv#9V+82|0CT<nzDGWEdBYs1fGgaITx{pq+F
zSw@%2aurV{!ut(*q3=cd{R`7zIzegA)GT<zkdACvtfxT7*0Zy-XLQ@)xGwT?SWo*1
zI!J$J`h7HA7sNm^7tpx;u9)9cE&TXP;&)R(G4NO3bn^rLL}O71y&Q%zm4c{Baqwfv
z4uU@91tBPV(Ec!Cpu3_p8UPI~aIB$rF%SRecRag&7JHyAoX+=Boc!nOWZs6A!z58y
z*}{H<WPxZ44CcWsrp+BL6gKe@NKrzBbr(j2T)-kH!o{Fc4uAO)U-}7B^$a0<>!wI?
zgi6JP0nGGZg0<(;Ev6wwMCvj;2BT)Ziin7al(f7gKGLUdcjdkYG#nJh_4Qq4dXq%W
zC2wr3gr}B#E6xtr^6=pg#mU`RRf(!Gcuvkh{V`?=Y9_acu>ne(E?U3!nWOdebeM<S
z#Ji$pYR~$)$d&D&iDK^~B@1X_#m${-IdDJ>BvND#AxsGE!+KrNO1uC4`!Gmfqx}M>
zG%y$#N1#f3*Inf2pY?*}W#!>XhAFxBbg1QEKL}h&J`fQBgF($eysRCbp#sOvq`<ZW
zkL1O*CRy)FC9=5g8#1CSFL!97GaggKMq*`u1a}(V8Dvovs?^SHr2{NwbxqGYPW0*t
zKOY8S)+XMc|2iW91jCU^0F=K2h$%Yp(3^b<MQPwT29s6aEbO%k$Rx2QYGJ~jJUIeW
z^nK-(^Ti(VXm?t_*bak;?K0$>cbUq>ts7aP^oQ&H#bPMgd^&^RyQ45XsJ8y6bcO#R
zz<ws&4D+B&`jl}GoVbBYD^F7(^}w!$P9|~>7mKJ>y<j7mcAV+sL#b<j2<|Fp9=9h(
z0Yh<%>W2U_hZH5m#i_vhsrvkc;-K=@C6GK*;NdQiMQ=loi7+>XJ6`4pdp!^w_)>Oi
zLe-X#y`@hUo)bk+9w0|e4$bA+aXlx2538Z!pD$}=22t+jsB4<?*GC+hW_4bAWvzMi
z>7!nqjEp><*eLWxdr)Elg_n$Ptha<iUK>4csN3QBwU(+RpwSSjxmczkFFy%-G&&gY
z{uU^83srwECdR}NqeF@!A~W?}$a(<vSV^ggDjVS|5$k)Za!<d$|9I+*9~0iyNF!X0
z*mL-jL*X6SsfUf6|EKl4To_`fzSrvnClv@V8<3I#XQH}vkRJ}epkak#?$bYhSg*2D
z!=n-kPh4Fk$OCc@LMEuDA*1%v($W}AZ*Q;rMf7x%!>MP!<5thNEiGM5IU9Y<Lv?`P
zMMP?8lg}>NI96kY9J&SkL@C@@69MCn00UDFb}G17v|ndW!T$jmj5^X1bo>c0A0dqb
z$4wcmVBvM-_)I5|q7A_Uv<{u%8xUw~YhzT;ouiWV_=;n(4kQ5i2EaG7VKPoWW%jYO
z3<t-=6zj2ia=Otd%crzf;WLkA3~a;NI7k=^NRN9T4ThF7`xQVgNa{&q*2Q3}f@Xc(
zXIe+C1Owv<J^lx9apALPhcVC^AnwM?$43C}e0x1Ca21O>an$eT9SM-OD$wJ3%qj@`
z`};9ra0%_sGK3g-_RGgNksAE?FNW4OIIhzwswH0PE}L24GU4G#FNP%Jq(m|7yL^ZG
z#845`T{JP_mzRgUJi)*00sIfpG0;swv-puIoCQ`?RK#E)*fs-)cAyKOLspMy;H`oL
z9aIZViqVm#;FX>uTVO?_q`8cYOe=VlO*huti_4l{KOu>mNxELRmuDiE*JcoZds}g9
z<)}JLe|mIpaX1vC-TL##YJD4G)*#JxR!gomz=H|UlcbP){2t)B>*9UrLj#}5zRYMB
z8DEC(o(Yiy%)_FhxPWEa)QceUMAd7<_EUB+eJ~xCQh2{`@ZyB(KJgkgD>a%I{-^=~
z=fUH!khTS(OjVtI$jIgy6~V|oDwqrZlKDmUmn(9dg7{Cglgva#*js0-PCE|_#g%Fd
z{K^s+<~+7VWNt<DT<koqRKww-<85JwZc*D9O+O~Mq38$CLBDa>4niMyC2f)*C^Yct
zll5vQY_PVj*z?Sg%L?9_h;5fOC@m@i%P<%ND4g#PY%F*lOI_o&V2k3Cl1}tk5&-mq
zEueQJwfgJVuTj%OLqj&3eTCxCRa(CliIKnfnR>yW-B5WCcK7Uneg>zi{{^-I@*^*|
z3vi_ZSbHmmkgb&~wY8^#KNS{t-n)2qVw(R)qmji>ROi7kaR5tV%5MED;oz;ZbuWJ-
zi*;WiX|lz8bTrXgaUV}#udV&1SbFlvGtYs~YuEqo99TJdVqm|a2Rq*o_5E%G!J!qh
z_Ux@br@ptILCo<5AAedLtKJjDLf1i;XW>sCaQ%p?=+#p+=NbG(9oM7fEJ9f6byD6t
z3NlI;%;>6{*@CnVPorbZ@RVwGmLh#kLP^Z97SSAPhJkG2Uhp9=kI}1sxr~Q0yl4*J
zd)qx0xN%QI-9n`dJpcoKQWspICv;v$)-@p@VGx(>Av6qx$#affwi8r~KBdi^a{rGl
z@Td-y#ZdqI5Q|W-nF9C!{DgPl6Fh+OYzLL@N>pqtMbE)cm4i66pJ6Jgue9v0oRw3E
zzy7+@yiq3H%eW3-TF36V{%2{zItuHq@A+ZEpiLN~N>kRnWUxc#A(%AMPq3*Z&kdG>
z`pU;oT&+zfa<y!6nX4lY*2R{8UJ{_?o)s4i{a!<|{55ilmifJWv9a;ldllb|$Tt_Q
zabz@FjXj6xbtGC^T=uB5|KDHJ4p<h17UKRdroICn%l3W$F|tZVWF<4oN&}@KJf!WB
zkUc6g%C78e*_F1e6w0PDGZQU}ln_ZoMM9DBKQ8rtfB*aQen0Q`ZFuhIzOU=N&ht2r
z<2b#)moNidbv9{OMw8|pzPzMF|6OJnH&y&7&k4=6l%*$@Zh!c+ED8Bp_T#ox4gWNi
zrBC+IBpreo6T&x?Sga#Q#KOg${k;918|W%JIxM96P`wtex=y6Q5*w27c(a2`=6CT=
zdW)tP#e7=-sI|A`u>Ro5@xW)|>M8f=?i_6+VdU>+rv8MtwDE#;ALy+BGfKiw<Hy@;
z+NW>OcE29uXQr1zE5T-ImFkr9BtIfZD@L?X+g%viV^E!)MGVN@q<W@4zs$cV09|-X
z@aKTHJSN;HQemm8ntOBtnn2+qolMTo!xyx_EQvn<Cx<rh@kD;DZji_=Q}c`AEEGj(
z7>S%8mAn?{?GQShL%W=alY#G(z9L#Xrria5cv0M+^v&auNhnCTZgIWd`mA{_AJ7F{
z?sovE0kVuh#!9I|yC@bLaPC?IFAtBXsA#hM-G%Rbm>r1-bzT}c9Dz4cQWJuYV@DC|
z)YKH3-*B<h0u}~ov^i+A;2_Q4PHJCUDIF!Xp;_*j&SlQ_Wiko7b(E0@oWRkeokr8S
z-(vUa9EB<z3?LzJ(P9u1PM(gUil4oF*-%JHJoUvbnnl<i4?eEXE9G&h;G*3Tzx<TQ
zMuFabicxcz@3k?o!#6l$g2oL#nTSSZs7LzneeUeNDlUF0oJFdT_4>=)!0OH=9K|9x
zzIBqw;T(00fvxy3i)Y0i9Mztg`bp8(K5ng>UKN?6K^_QI`7yQ;NR|yT5hXP3yKTn(
z+9KlI2c86gug!B+KU6`1>q(Ak#Eqs4;S&eA%px*sHyac&l(g9<w;f8pGh&IT+rewN
zR+?vH(%JNPOIRfb`!7o(!6er^l&|#Zde2st`?$tu^w~|&rR={`>pfBthW?|ZloUOM
zu=PP-z3%tz?+R$|;59jP?3ffO3Wys5z^RsznogTIGY@XQyBoN^ZDDa;GjFANL_VZr
z@ssTB44KqW%XS@d6Ny6w#_q)<k&azrH*z<zMNrS(tc*x?Shx{9PU+V}3^gEr5FBI>
zeR@#YD3sHKDPeI52_CO=<xwKyaxio}1pB@^;&4EmT|lQMXdZ>n`vAdl3Vv{pIlK^S
zmEY~YDs^c}<G);gn=i8)t3~RtbNruV|I4(uNe}3dhQq7N+J`wJ2NlM`m8gUKj+tFB
zxZT{FJ&0-cgH|l$1K`4Rk2{!|nVA$C-<4HRcnk&}v`eHNU<3Is_Tj<<(0zBB^<9GW
zW>Luk`pEh~%|+K!RKE1|@kcqwn&ZzDJIuW>Yf-b~S^6ov;oVqm-0oZ9V*;Y33$}uI
z1RmVk+eO`>mVkd|?&)tJjcV!Ya+0K{wN?LN*{OTdOR%h#uQiV%*aE?V!LnZmWdq*Q
zM{p^2>}LD5Rf1c-xkM<2Z-nY`a7~ubp5CDz`BLuqeOoMCc+w}PbQ{06s8|LgGAWBy
zmB>ze?9%d@)1EQ*<n&~NsPz|*8Oh_wrAjB}V9vzINXU7oN2{58zkQ3wiXiq9Po6#F
zp*3W|285HBHw4sMV)F>jLfq4S(x$%YHMHj0HKoNTD<&=tUZW=7+$-~_T+J9rJ&G;I
z>js8r53c${q4?5=)$UJO-WmSHld;rYQ{!^HbMY^YWH*(6IRcX03w9IIEV=!_w}RnG
zi7o@grlAB%0Z)#wWQna1iB@2rfeUlL#P<)1M(4?|hcns?%~vGsPA3+esJ)k5Xq3zl
z=9+zwv$CQmWilw~;1%pk<$Y=H>KP^*<(6-l-^$fRb`gv7(lC8%llkup7#CvVdw`uR
zB;_I^OE^AqP#!w>E+7-&m|G)9GAcj44#+BzV`qJI`ZHf!Jud~%t_zJFdF)(fu_Q{q
zI_%CaeqH)P6W2fz)`fT}Tm18nD3Ax=0I1zmwpm@A>w<+@rq<!(0!Z-cc;Nd56AXqL
zrr`M&fOXJr5lt7d_W`cNOhIQ0I`@hm&j;U$m;rU)39rsABjpT+hV9BE`m`!brsllB
zf{NdVqK>$A5Cbb<exC*g(!dfxrh~o>e(Q99f+0jvGt2!_7e&sOP$mV}wJ-fHJhNI~
zCLzv|+(qgZeE-wn3Ok}_Nkk;jG_-Vf!ay+u)iHU=pxWVv_-Nl^*kO~<Rodk<&(F*p
zrfj!h%6I<>;7OepgH7+xBwycoj{p{@<Rl~*z+lisWkvQ-*+``jF&({pnGZQu4_I-K
zsFsxXUtWVs-iM(g#gpgf&tj^0FS{Pup+^c4OaTOel8M6sY-Sgv7nZO{R588w+5ldf
zVscS`)*_ZP9NQ=)cA3p0JRqZ1=HzPXiMHdE07mX$?{V%<Koc(rvkp~g{c!Fhg}>39
zi;lD4T_dn9;(0_N1P{{VBZ@H;(&9s$T-BELk3-Havr|@%$kAy1e$?=$M#q$YYF0iK
zVSDM*J<Gy3&5Kb2Z0ff-r;G*n0MIz~*gAjo_?Cr7=?S1_6diObjmCwV;Rj*JL+XBI
z{ZTe>cH*yHI|mRo<yJ6UjxxTjka-)GR{!1wFfd@zj$**0E7o>LcvUQ)@XE$x6gD>e
zbu)Zo4;^KNd`HZUrh^NhI*FsSonnqoib%xZx(BqphkPRXJATt&qzSc$B=Vy~H~=N`
z<)n=u3hfg+5x+X?xwdEjF880|!<?s7hTT<(keRfE$D*|*(AW+63^xSSnL0X^)s}!6
zUBa3k7(IJvxs`IEo`o+{L~xJXCKK1m(ka4+Tz@3IZ7f1H4N))8=-DQ}<9snZP9aor
zt#FfDgG8rFE*y#{SHfJ{8CoqzxX{8ys3!B^qb`ALgS&v=7#~QazNl$f$2iHxJ~W}U
z`tI6*2R3{y3E$sBMuKO>=D$`&Y)borz|jPy(cx!Phm=Nt@?1Q*G3B1IL)%O|hqTRG
zUk2JCy`anUqf4&|O{hG}ZGEGDQqfpGDWh=cvjFpAxlJy|RhIJIu2%1TEB4~x-@Bqp
z#>CNCkr(vpRo?N}Qee>YLSjS#LybcC0!X^$U{gSha9br(_v^^Si?_-P+V18MEgL!#
z_HcCQ%NJJLXa<FQ7E+Rrl5f$ozMhwqQyb9}VA}pWHE*UfRZsG{&n_cBPb=2Y?W!;B
zrxLh@C!6-rEmnN9jh@4QzJ)#ZQJh5-m;aTM36A3Q;#>Qy@~;r6?Et6?1nnc&LOyX&
zPY;LUA_}}%>LG+!2PGpAfQSrDDkFu(BeEtp{mbfC<~uH=&`q)x5ALRZ3#2=C<cTbW
zwfXs{C8r`2@k@z%r*gOF8OaG-3oLAQtWxehPXh2pdPz}7+Lrj*T)e=jt#g~JUanc9
zRm>M>;vtBNB0L8Yvew<>id9Sa3c0zYY!S3omHyU|I6R&RVwb9cmBy7%rK+^GZDS&T
z)xzx(gGTCy5|giSj-!hxI@P_5WW|yzK6}}EVv~_uzR|2%+LDp}My-~in7v`EZg|cP
zYP+T4m;8~BFIj#KF7hA9A*Vfb!W6XKiLgj3aye_GnsW~YRGHu-pvt^sTgu{u<^-|s
z0CxKG>RQPUj>?N0mm}8|t&`V7I3B8!JBBu%knY`VJ_4+_=o&_<Iz{LCI5{oLauBTR
zZ&yhzHE14rRBzq()86m-;fqQ<$>k|53_&|<U#@v(pF}cM_cDM7G)V2vpXU~vVrOSZ
z4^N{2rL^NB=3wPy2yW3_MRuYwHsQ4m7@eHsOCjJ2p!={s7g4kfXr2MW%=&z-f`YaY
z|JqI5T!%G}9#*<}V<t84QF}C##LW`_TU9;`Oi{i<srO&K7TT-hby`8xuEZ3D8SV-t
z9=66N!<%Pk&kU~<wIeW7_SieJL>|b(SP6+#1Jc;wPcMOv{>ydavwH%XajG*CADQqd
zo}c;$kz*)3bp)^6=Ec&;0dOA}2%fH(q#6jXs4;h(et2b-G_ec^Inath6)j}^bry8A
z)@7}F-n@I9MS07)oL6FJen0H3So>jz^XcTc8JF3cRa<1=h$>Q_o%vRMQTOZbADe^2
zycRe9iuU7rKbOyGMYmmJtZe4x-n&VL-qTA%Q9n+dOG8q&k|C&}rhVPpECKR8aJA-#
zpOGZcTgNJiXHVh`dxFc^&b(5^ddmp$np!2+S;6;g`Lao~8G%(EXZ}|a(Y8FM6XR;}
zxFUZH=(Mi6#p+Tv+V#z7pas>j>@1|GI~I-o<i6N?_F!{c#lVPWO_J^hxnoJYNT|kT
zKmV|9HU9o5gG6oZ7Uk2K4Dt7_N|r9A-7S+mPQ7_=rQYJ^_n+vlCU5HqADpb8POq}g
z-?8dJn0CK-=~A+waltzmcGqS*`7W~{c83&%(zhxZ*3$ZAie5$?Pm8xOqE{1nx!ix>
zP*=p2smTF@#z%1nDa9QtE4x?9&s{6qxun3pQrn3m$>fR1I-h{HvBHA&y&3m|VlTTz
zK00Qk$9SxAO6BgSq?dQYE~E)=H~)^mB&KyJ%HlsRfaY}Z?>~l_2|8!2Df(=y_TNgZ
z@2EX)_}St~Lq((Sk7q^e9<*K$@3;i(WS=VC7A2DP;et;v>WV>CVxsZTb0^MCKsnU4
zk0eukv$BdtC*60tT9?Poln3s?|6g_~>AE4~$&)T0{P$08{_V9zBGUqa`Q4(1-hsZe
z)eL>c{Xd0Mg;Sh$Yjo3h@@+ZbzN_4o(^u>cg?(SC0T$7}EA@Uod0*_oCH0~YhBhh~
zB&mba@#NOWxsLH4lRMmx7;c?xJ}ox&CG3EhvXq9=++2`!v$TSq(Oh@e`{qZdR<FE%
zbG4-ZDeYC<%ysvhekGY02G;Tpq+D!TciWc!Yu_P5;X=0qOyynm+qL#A9LqbzdtX&A
ztg!p{GBa5rUj&$!6X%n-pBdSr<n{1-kqjDulAPT}L+@6oWL@7)OWqc1*<MC96fY=M
zuwSb`JlyLvK<_=(R;w;v;LGkMKWRQAP{9yd-BoRf(wbbQ`_Kab%LI)wjWHNApXn2$
zF;|msx(_sv9-ZFhe_=9J$*cP3HvP&W?Q+?stW8V7TPSmTBf7?2ymN&QK9O9h?{V&c
z4o*~7Y>l9}5migbk!A5n;`s(0!Ltg1a}?g=;Bt4DhyaZ&>LD9e!SUjSaM@MB|AES=
zdMc>;UPI0erHm7Dyrd<(a4~)i$V(ucL?Q}-P##`!E>jbPWR4_({yJAn_W3=$gYBz$
z(@NFoi(l;`UoO=|DnTb6DAG(~X^rKH$2pyC87E|x{jYpr><!4{y?3~fiSR!FBgEgh
z5e2WIJ}4?s@3jbzda#`hdPPGhzT12~O7U7r$|jnvM1oF>+G<|E{Z`G_ujR|yD%oFa
z`rDx--=A)isN*7n#VuDb4X34PTY;%@MdjqmOLB+B7m6rKk8}OzTf9e;wJp{nr7K>C
z!jn!tt!J2@tzuVLSBZn=Mj`4#!;?Hs<;QeZvII#5i0L-f4u&1k>A%Gj8C2Qqaap(3
zDB{=!I<^Q?d|mFI2V%W;zc+`7y;-}+HB;kn4;(+m-lly0Tedlvq*>!`-WEIXks|>O
z--+S&NK#cMnkJ^OHpmpxQCwE`>dJm)R1p`ap~=m(D3;eO=)V_09u|m8y&&I^ehym5
zFmw?^g?^Z(rO9RQafrnu7UY+VdXP_+UP=sUwASV;mJZxV%1T^aT!rrfM-1)->810W
zKbdc2spx$57<ps``&Mw?PlfK_98OV#H_7eR-jlyzc*5e$8m8yzeZcYy`sC<pVSOTc
zbd)mb>%S3|elU7}{rvStNAw9-&CG5M`6RSY>f$AXE<L}*gr85FvU&%GTgGO*gqviC
z<1ZNc4NPfoXp$M~8R+v_JQck<H>2{~$Rn0R85+VD<7!LZEM;~*km|4T%*l5-mxtrA
zNC7?8Q{tlH@0~MsnL;$A<UUyEvpii01f|PXa2mfO-ks>%b74m=Xocv|UO<)tR`;Vp
zFRFW3>u%fA4uQJHNnw#~R|h|Hd7C7GawvcvVs(9XKZ6YAZNZ;S+l}X5I~|=<DSQ{)
zac#)t#fSR?lG}g#bO{d-FOuUYxB366d~H<j9-TLn7#2A?caFiJDfevT^*{2yUEjwo
zetef!>vlCaIzpK%`t7$vy!!e}zq4))Kbv(e7TX+}miH%Lr5|9bUJ-9ZbzkxVx<)l1
z&b=U9ilV`V3afqVF|-ApUGfvv{%~jV_@B0-YZa6hn^N;Hq(^Cbhi_s3edl2gHXz!d
z-fr&^xZ#5v8-RcGGJ~n&bd)IJrxoVmj%GOPcMt!JGdnYOpE(#U2;D8Jy$uxu_d0Wq
zHYxCaj9}eWGqxv^P0?YD(`xE6+fPOBAr`Bt`l*uNedGEwd0Ptija_?wCPU?WNMg~J
z?rC=+wzbFl4s1BRT|D*f(MuxQW^PKJpJ&29<ux|1HkrG%fvea8!0Te-?E~&aZ6CUK
zwg(Xp$yMr$*K{=~hyixpIe#i~r}|B+u7^8bO***x>nD30aw+#JInA$8(Cz_sg+0);
zcj=lOe?L&@1Ln-Mtqy!Q*5A0?#&(iJhqa8)KK+WcTlf<%BGm)!nJBq2<scv|jQL1`
z?B)K8yVJ?GY$g3<<N$)q!A@TrL{9o1-vl?qTik0jx>83S-l-#u_5{Q^2JTJeGCLdY
z7@%>RYW?K&*B53%O^^40uCa9YPOvQom6ep0iNJzryTDhF`*&6sFY~z=jo%F})zYMA
zJ!ijmhpi)z8H6iO?BS}x6c!bBy@Eq<N%7nGpA3TR?cKRi;T>x5Jo^3nw<Sn+0KmvD
zq;08@W{C(t5ft=_t8%~C>YV4*=^dsEG#0oZZlCtt_U4-od~MkGsp_}8Chs_M7e_2p
z*#4@V8zgqt&qb=)82#P|g5scZG<+oH5miW+8*T;~|IQP8=1s_wOZU;b892XmiR4V0
z<Fn0D8r_)!R!puRHm9s;E@TcJk3Qa#e_19g{^~)?HXm_&ligR@cBL$H7Cp&e?Xyg)
zT`YAA03DHJa6pMcmxcyCc&NRDrzc)jlW7kd1r_$}Ur4<KoOkF@-LIQ|-KJ$IT6(U<
z^eoeBF3R#JtG?^%Y`Zu4&ShJTqLtnJa$#?=NBko>rVho9;@p~R3N`fxspK45io%{1
zE)v+jMDgx5U2MomgfwWvlPi&Q<IzPAuD7Npx~NbVf4#Nsp6Ut$j=xs|y^KUAVi6zC
zCBX^Zbji%nQ@Wcs%jsgsQphO83&+2L>=cDVS&QR#nI>PVkId1e3g38V9jGt94AJ!j
zoOcKt6EdR5paD_a7Y2r;dTPj&k`YE-CCA+8+&{ZfZaH6Er7qsPZ3ng0%T?xpmBSmA
zzePS(#md%}2)3g{)Xe|&m+W_vcTf<BlKh6kJC4G7bQnD}j_p}C6b;^#{o$>j^bFA}
z&XCw}b>AjK7apZj*RM^7*(+$$i055(+kSNh!R}NbIy(A)>!!M4vEvR1zp%@g4Gq`v
z`9Ns>15G?QC~N8X;Wvo#gV;<o5?4EP%X&Y3l9YS|FQv8Rqk(xoU?yi}$q<jU#s|Ff
zos-+y3irgH8D;0c7A-%0ty;pXugg%{WYr_l2mF)vTTY)YDgN!i#2#7bTDfg&x$)eu
z)~b|>dzX@uOqL-1{Qe;BgW9UcUQ?U35M`>m3j3#K&!^tI<li1(IW8c`8dAsS`Ks*R
z*wt3Ap`r<+=@+~0ZnSEQITW}3Zty<2Bb&?3JoAmR_4gVjXV)UW^SRk4>&iI{s%i34
zkM38te13U%eddzJlx6u6f5JY52AAB8=TYU_eKm2=p6V`4h;W2X2H8DvQXzrB*mvTL
z9lD@-g?&j!N%%>95x}4!1eP)vq?8sG76_=b<IBKM$DkbvNq=-M$N(iWUtk9I2I4X@
zY>1Z{W9^W8Ln{o|Dz-+9Ou#Ut^f*V8v575@s$Q6fyq#;&82ecenfzu`viIJHaWj#8
zY{@P>my@m^qC{UXHl3UD*`#_rxc}s|J#**QKnR<}CAlNwmbSXGg{%<cQh%^f_x|m!
z>#?Gn%9%~O>Tk5adXTNDf6D*#nN8&5dN+xed7cf6yh8V-<J%YEkLC8c&7!Lu*LxoE
zk42{%>*MF2INepHf?Q^`%MNTk3)2-!6)|kFu0Q9n((tdb!~J{ORAGHkT+W?4H`@Ma
z?vTzL*mYcjf-=Zf!=D&9@33P4p$STu?cj{DHTpvqy$<9asGT)+=1fjjgup091nXX7
zj8EP;7}=P_R?}x8^8mO%o@HJM_SKcR8EBYxq_wa#-e0w{`vlAI%BTpAnC0jFzq>1!
zYK5{rRr6lIc$;;pRZUP@@TYmNnXRcGOKoPx^~I)aV{U4_tfW1(?dZ}NJ8<+!>c>=x
zBfcwwKfTV*cgw2{&u?T|%E5E~(a;5do3~dQS(ad(nEZ@YS?6m_U!z&y`TJ-4FJ&uZ
zd+r+gy3XH4c2gX}pBEm}X7}};pJoDvZtH!0aX9E!Gpnw%j!x@$V&`L=JsH{n>Zp~9
z7B{WIMv*`H^%q?;Gx-HlsHGw)QpF2+Nfm~x#mPliReZBMx973>fUMixU<*f7o}HIK
z>Nu+g0v<UyRIlckm0^w1!v%r!yc*=`8T#eNaZV3vlTA-n-}~DJs$!cR*RMZ!U+l4M
z9H}CO!Tu$DUT9lSb=|QieGw}Yp}%Ls`;QAz>O#||OYc6~+|4S^IO%9#uS`Aiv~{=i
zlBEJ`OPRRL8i)f12xHTs4e<Mi_p7o!iH{KdxgY>jTX*gAA8%VOVq&TnLy>MUa3t7j
zHX)%9<tr$+y+cFiz~UDNbp^tu-=u#EjSvrURfbQ4mf-@LovK`e!UIi~gsVSnB4aD&
z;|)36<!|x6=cMgCy8gQ8g(hmd=ui=O(7Kp!MRXJRBAV!UQivHU@id1YKKrRtr-<VR
z5g&u4+w&NA;x2(+9$IKnE%ThuZL<e67cfk`$HKjnJN;<a;@PUyZmBzJ1cSMjARR^+
z2!T-nX(NTQLRgrz7KphW9w9S@h?+ov0QCW`+dPLaBh;Tr1%DcD9-5&`xCmHNBDWb;
zRMM-3{OQTiwIvg)n~$ZudmHY1X8v(i^12n=+^j(0A!82(24dlteey#kX-stc&-;Ss
zM_ijIJs&@cXGvjM6vsiMf`l7{l#oww5Oq21n!5Dg0fC;(RsfLaHKo<TblTOW_4TC*
z$Ax#2rORn|#-WB6g&Pj(!%!?cijHSwWDv_9@G^dYM3x3U500JXofq*11}EEETP;CX
zY&Z2&g0BI5Je@4gWeAf(n_R{rg^63v*Gj3cDXdK|HC^$<f}=1*QzP~MBql-;&p&iG
zm?)s^@)BD=R2W42kEho&;y+*Z_B@_3_0%BboP1~q;kR0z<_d3iWrRYQW$|Jno5k5D
zl_y5~v#<cjcNVnW!XX(prz|OlB;{;5yQ0Z6RAnAA5Q;(t92^__vcH{Z2f-jAFcyw*
zh?Xx_g)V&339#-+a~Mt-F!YUF1VW@JTuc-}oFVsAGH_GV&8@KKHuW3Bzk|Lt3n>JN
z^B{$56~6;yh1!O8=M}tu;Oh1ztG(ZKvyZPYX74Z-kxf>RgQ^k#{y3KF58Rr_0zj@M
zY9yl60_zFz^g+{Ish+|AQc`;lq(VAuSHsc6K}jGBpym!Ch$kkpXI@mB113moxC9#4
zjT<+fKJvpU1K+$KchqKCDa367D$AuT5q=FxocA0->RfPC0;&f%zS=sUa5gCc#CDW2
z3U`K&;3bJWECiI|%___%$A{0eUDZbifp!PCPqYpHOn?4@l*QM$B(z0kWf#Mphz0I2
zWGaF6P-`Yww2Y(gxo=_Tp!N#9&Rxtg)7ZIt_ntlE+?Rl0F}D&1{(5VOTp9#NQb`|x
zO;>*zAE|zc7>_i6U|iMn^{d>&Z6v^9wdxb!DbSgkiyaoEy~v)3(|(TXiG>riu+aX#
zd2Rz-Pn<y&g7&v2D~X*VmBuQS)-sf|2XFVjEzn=kOKWlXP$-bl!N;Vf{}-lbaS$S5
zLqlYiPHPTMnpK#`0*$^32w;S=SNHg_xTGW_JeV8MiV89`KeGru9tO(4sd?&HGLQ!|
zsz{U9%l-6-vW85&i^#@QgE0*LB9o^mL_AYeRP?{o&SY+ebiR<i*_}9{jWC23G=%n{
zXg5N`OguNi3{YF?{WX8f^!7Qm!p7hqv$gl8uRh4Cth}k#j#pm1%)VHVIOC&j25U*;
z+P<3S+|!XqV@<N8^~I<xX>YMfw1^wlVDLdv|CcKQPzNb+w8J5{7f0!1Jo!<`P#GwY
z(BwfN`lNs68@|Qx?=KefJEmVemsh=?nNR0YtdxzmQpVA!V-l0*rp(1PtU!X?a8R+q
z%9fIC;9YBjein>*Ix^xto*dD;xWZp(f6S!D*f))vp9xzwBN0Y<@X5{^6r?JjT_z)2
zSXfvK;Dn9>GnaaJ%+SqvN`LkTopQepFodCp;;GhL*W=!%+T3m0<0AG`UlzY=f7RG*
zm5h)=$}>F*T0p1gtq=<jzq`2Rpz*}d3l}Z`h4Fd*Yp$?v5R991w{J_Kf!1JO!cEzo
zz5*rb9-*1*gj_iL3ZB9kLI#N;%s6U%(I7udc3SFEBu86Aqi(jQ+AKx^8MapKct6@)
z@_nCsqU+*Z-d-qMzWLA{v9mls=rjS`Kz8NeU{&6@-51n6`{ItFcaYX%&;uOY*V@}Z
z45Y)FEzW_zG_k$G{DY9jBc(H?(7h5*?_T62oNr??e0%>z$>6b9jMloNo4%GjT>9>i
zG`1V(fEv|ZK6Pmt0z6o|o%>>cj*nB}^#!9YQ-$ZGsju9C5sg^W3nMdXYYDp##8y!?
zH4#XI9|v)44+m!UT~lX#OXeUw32#(tcZ`|h>eF(!xCIpSZ&u}hOa#c{!H+SCphs`O
zOFV!6Jh4I{&J+BKmK6L7jt|(>#6Zj13&J4rcQE&xI@$sgW*lD7`TGz?xFvjcHo@ov
za8)2NHY5CgXpo7IAjxIm6P9|HyLhorAUP9^KRNEFyDz-8CDkUfNhU>w-?#+Zd*k+)
zYJUEWOn=7aqVJuv8GXZ8^K#8z@YAmkV5%6jMqqK08H9%Ar(hOB?Al;C3@;9i{ri`G
zulQw7n)jsN=r#FuO@<O4eoOW{2+ZJ!K~^?4hio?<U>&jU#efUquVrO5vpFbyA<Ejv
zPn2==HDA5AT=;$Ky`8>kISE4OHhvoB2?WSSYtenyD;l3b%zhB-o*b%Grw}1NQ7{lO
z<>@c!m!X|Immq6MQgQ+=EF2C;Hjrn4aDO*6Xr4<zfE8sm;{AP>4XKk6|2}s><XXVh
zzMRUu)AD>{k?THk{4^m?L*H+I%BUM##n43I(Xifc{EUX)Yw5DX9S>@15|ITb50877
z4WUVH6pX)lGX_pQ;_KIQ!;(Y3;{kcb85KAlHU0eNJU$S;u#Jrb0F913UfQ4fZ7d79
zMMl?Cxh=sLR(aDhoZek{MjaO_Hf9G)5(aRyr41`^#!{L)JAI~<aMM@z-ZmH^s)8f(
zkYW*U8Wb|#&wmY|rlWw&Vqx+|-2TCTS)}K;_E$lh4i#`v@GGV4=%T+ZIJ)?z@8n@Z
zR{YV8O7I6x8x2TzF$RJX0QEBu=ucu|>QpmC(A4+qVHSlCBa<TtK_(AQQKVB%ra?fy
z6;31)Fy$)!Trf;c<hsoNHv{w>-g{j2#l@8mXNFO|+IjFg-~Y+^_S}>E<@0lYP@y*v
zn<RKSqI{@nYvaX|ynLC<?7pJ_$_k1lY$G7{DlfZ<XXkhhMs0+5OAMs&8xY_0^Nt*V
ze4jWyp|6`9O8tg6OaXO%UQd)neS2HM36BiB)uD{sMwTTrUMZ9GSOE`EZS=rZfy}-n
z9)Dzz0ov`DaZ(GYnAnbj-D18S)YlaV5L$3vTy&p+g{2;lXA({Xk<dYfPSO<EYyK_$
zRERSJ=s_!C28B(<+Wv((cwuOv$8adx;`n;RUKXAyqhn)%kjf(mg{EEXa^#=vSHlyC
zBscPAMdmJ?ROAo-Jpap&->K13u?O41&fcE;+K1+5Q8|lNe{YjihgT`(McjPr-9kRK
zziCY=zxDHT`|)9#xt`d@M_<ps^)a(ai*qdXkP2d*+8hGH1={IM3~a82$?$sqB(y~8
zMul#mRN@u4$ZETql1ME($5*V4DuVmA=62IFr%RmIkWPO0Hdw=HIrvN|<tdy~ZMN|q
z$&C2o8eW+Bfd#?xw0(F2!b9)}b?!W0D=SNk;^TZw()BVGiJREf4RGCi`<4Za42PA(
zC*i?SCGsxrf?9pq{|$f?Xiw{SM+QE9vP7;xuo+~R;NntSCGAyIXBP+yb%dEoJP(nR
z6IKMuPckqIOOg<b@npy>QgD_+++k2t2QVGAIZ|nT{T|4PxGq^7J{*DC#@KG|;DK$c
z<Q_q@Pb4wm#>Elf0;MkK@B4Hrv!6VdsY=<_kmuI7U2Li?!T8438Cg%qx3Y5c`i~Dg
znmjTA6q3IDPHbZRo1{hkhOIsA%LelM?<CfYbycJWA9<bq`b}%^Z2Kjc3;cYk=(KOu
zcB4`8Pk}FF6O`Cb5A~l+<{6)DZ&m7v^*3P(Np6!Unux}eW=iN>Y4v>_i(9NZn&o9N
zbAf+Et}q#@*oyf^X)kNqbGPRk-)gB3mOBnjtzu+;xVW)hDC(hpmKLALfpt^b5;9p1
zEv<5&wZmdQdZ7=5%LE9p95hKYw;Q%WrAtQ0Se`r?3v-@c95;l(CbSWW2Ocgb1(OQa
z!A}ad7c}<-hp~x?dN_R{3a*h<L*yotl@`ddVEI}<y(Dkg^`pgr1Z!xm7nriYe*UZl
zw+A#RDajA;?jsWx3!HYvUU819Jhe$mZL5LK-1W6`>JOgi?!C57TW79j;KR<?D<`gu
z_N*40T71)+gXYq;&ad=}G;8;jpQpyZ3Wa#zjwpFQSFBbU6rN!u_Rf8{bF0~C_vMqK
zn@pcsTA4nQ9({36X`y)cR$RsNzS4hrZ%Vhum~Waz-R%PwJB<Qj<7O1*H(VHK+j1mV
zSE5s5`Q-f5?!+*y!}pyiYPL+FW}I5f>F88Cb#-}8?Py%&vaeG{cUR$R*#{aI>N*S7
zD{cI+{&MGfrS>IzwM^Z1KHtH}rMqLf`Wvf&skxq`m8Y4H>h@S|_HU@Bb-r{>403xp
zXu9=RT3U2PuNvEVvEFKl-v@rL*rJiUSYJ&rX^OpmHqwGcZ$4>Ia!Dd*R2j1h{^P{n
zt&jc6HR3whi3DFeivKLse>Lyk9Uk07j{mmXr$)Nhn#xN9CuZhYNrk=xK9nIy7d}F{
zuL-nN!J=|WypDe4@rYnDl6c{wNuc@<gPNfi1Qm^)1Zd1MfYi1D)mu{8HaKD6XVoE@
zRD{h5g}CIxZ^vbR&bL~DQ;eGprWyZT`c5A|e#BhxaEOUb<ECMUvdnRnlz3Jd5{+vr
zCu5}cU0dDYH8B*@(_;x~2HTY=9sd7s6gsVVje<g`sRHKv-4vZbVFjO+MQ|VYxEvkL
z4xIOCGzfi*AI(k6VlwNZ-CB|>7RIIvX3}ZSw05hkS6v!v8VMFYuGZ#K4~HU$_N9ot
zytn3%$I#2m%d6Wjqje=<UInq6g25KQvLinw8SHwy_MF73#_(V=Y=?^`!K--cTF0N9
zX-^@!eb?J9$t=|@b|xmK(TRx=fS=7Y@)7n_xXoas(xJy9+kL_@-v-)rK}@*|ymV>#
z-hKNP>FDUIknuItRvq?<!L-RrwUH^|eb)#7`+2Z&#lacbybeEWa241`>VI8p16wg6
z0q!9@w>*!oRP&2MjY>h>gwdnlcR$jSL^QW@bR<f8JG;Dn8qv#=rvANv=x_+k(7wa_
zWQM|oUr1;%nL8KWdJAphc%5J<q}cS31c*SXz6?K}Fsq0^J%C3|98W3P*=8ra!MM8b
z^qxsYMTHE2pb#g8&(BM&xs~?+?^P)&5^*!^+eQalB*=^*=uN`VRr&tmQ7p+-V4qns
zg0PzJGlhZI+;q!O(!{H7R8kUK9*NRkOlKB;dA*>|k((fre?$V<QVH>%h>eJ-#;0%D
zv;XsdTmUk47X0vbSJl+gF!q`bU46!{ovawu<k-Dre3&^(Sk`@Et|*S_0GRp`k385u
zfG^VnWi8Q7|3I}x9Ii-hSakd~afKo(M>J8W$c)K|j+8XqAXh~ZKKUc^2#p5OER{M9
zX1ly&FBl(3t(l#_v7)M-?{^=EWuJD>%o+8xe;cQ^7`Z1tObkZQ0Mbs6ePV@8P5?1h
zA}n8E7i3sHBNG$(se`Z@tArweEPUZa7TAJmLYCTc-^m28;L+}e{a4p7vhbhvNVlQt
z=Yu9#i>L@6#o+7LO%h_Zfg>cf|3)|*DEs*bMB45{jMfLb%S)IR+lIltd+!y+l;sDx
z+RVxW^KRd|b<ywdEpoM`g8P^XRV}f(O0b!SYcl-47au%$aPQUiTAQIGVTIQ&XfYTO
z%MHkhPkcRBz^l*6sAQI~vlFizn2Z$19)f-p9b(Xy)`7A<5W_{Q$6ySM)g?%_t*tH6
z3xhAeF@$yo7yt_k3z_&nAi`=5;OIB}_+`6J!q?5&#U(R@O7##D3RX(b%NrklA)Fl3
zUGz^MBDfMebR1;(&MWxKa`N(4HZ~0T4`-}3zJLEt2|)6r%(6=tqZ9&+<g{TB1^=Jm
z7|qrU;_L~#2{ONzj8#tQ<E*&&&nO__y9z321c|Cu+72%!hO4^DFfzQV=(xt6g?5T~
z>O!)Js_USD%DQ#xAQ9+47E$wCLK#FmJK9#lD;}XQ8BY)Mi&#+~F;gfITb{kZSy4^^
zdU^SM{e-Rr!{(5b6ft7YhxN8tI3Z(y4{xxNnDXTOWxhm3I-^n^yAJ)gd;f0J*41!e
zg+U<)X5?vO+a>fNoojAyH+<)h|58*^(t>G~{RXM{bE&D~8~X>Ose#Co=qWGMv)aeU
z$CK~exwF0-LyE{e>+bIE!FPqyU|M6^mMtAOcbmK;r-91VC0CRdH@{|xih^%O&mb-3
z;J+0su@3e`MJGS3gd27LV<1L%3r^xBL0yOz#6goES^Hcf=tb6UVmr<|a^aBKZXPyh
zUfJ}@QqGXa3tA!k0muUfH-r)gO(dxXu^`ls9XnQH^S!f#fy~oE9~7e{?AyK&bG@`Q
zOqi$G*z~!+#}Xjq0)h&nj6ONBRWa3zjNOBQUsL-A?A!2qYa#+8RVeHaqA~C;qjRS-
zJW79md+-4s(a8nw9$JLcPYligt0ATHC~-msV%L_nVAut{<B+#Abv_MUQkaHRA#PxV
z6}?b!+sg)K7WPGzf0houW*&UcXH;%9NIYPxR&9S68~${ltwTj=YTYMSzbmM`TXhBa
z`7^qD6gO<p1T4rdoh2Yn`79rmGW+4!dCmzfQ1Oa|1hEkK1q6D5Mg&Ggu=@D;z`mRO
z0}A!!r-x$J$jAiZD06Yi@9@AFSu#eFD0JbFcC5th8`RAB)}U}4DBdcSI;~-7$VaSM
zQDSn!LNsETbPX(PAUgps9sfWRLficPjN8vMaFk*KnV<M7p*WUl*Y5;Jn(WG>O&9IY
z?4`aegvlKVf0{u~T;c)jHA5=Cs?Ei+qut)=A4Ik=6&+{z`d->m0l~8QMevH6=T8S(
zJ09ptX_}vaXC%o1vkr^rvxw{+(^%$frat7qynNX;db5hR&BV+XVUz0YtO(P=Fo+r7
z=m?{SB?oaUf}LEu=vKd(O*glm2}C0{e4mBA`;sL~@}R^f!zMBO;`Bh`speb@HA)r!
zYF2hkPL2YRmLq==my=t9E#BlroD$>XxxiJoaGJ<)QN_>04e38l!VkrhLNl3OJf4LN
z$`ci)b=5LidHft*JuB1qnQ7+bdVXxKHm9a7Zatgl<clZwMMYMxNQzg&TwPolCy4nl
z@K1t16X(#54#6GA-87(qCvx6RVzk=NIA(A&#KfhEsTcU2Gfb>WaoOC`!nH}P?G)_5
z8c^FNm6=&sh+z5d1it66j+fSTXo|jNe5k01a`I$V_sZrLI!5EU;DwLH2d^X6&>AUK
zh--8dR3QY+0hDHGZ!hDlni~4Nv5^&i@nlvA;%UGurfn30bZBU&F^Pl3h`xmW)bvaL
zxgbdJcSv0j!yNCyD^w4rT+w^%s;X6MA6=aD*AC4){O{w!hvG2|7-)#=Q-6d0Zrgcf
zpWn_`u3xvow=&AJq$A@Su%WGi|FrR5&Zy3`tE~9CVjQ_RQO&?;To|vo4xnQ%yf-0!
zD!1P&PCVgWyx4!V6E5GF@Ubwxl9<(jxDHVi3VT{tjqG30fjsH`_OEWS;=2G*I&@+N
zEb+A-+rcjd-=GKQ)o=9ih~E;-%B%W5Jctw%r=<VnpkR8d-idNp>?1Gm&2)oMO+Zyu
zm7F;HZ)_sdq0G1G6n};tNf=f9%$HxBsiC3RC`c-dFU{F<RK9o9&-PusI5O_i)V)x0
zlylKE+p`YIQmW7?YyqI${~PRoi|HFAtA*1H86<-=axuZ}jGlc+k)F0V`s>RQ4vw9>
zc15_}T%qQ-0sz8VkNpx75<~<JV?a$9LBOqsY)u1`W2t<b@g9+$V&t%dJ5Q?WS(x#P
zi|a+^e%M0V4aR>kEl_;jy8Qp{Mg;@bd?wlqe}g*j{Abb$VZ{>f<RS9)JxPV|o*-^V
zh-_p48Qh=`*;uEBYGR#{*8#(z-kzQScv<kla34qkR-`(xpFN4~I7Gm6t%4mGnT0Vp
zG-TlBYJU%pJ_IctR~xnv`jV28aO6~2xC>}3ulPz@V(miapIHx!3kYmsT}7UllCE|B
z!%kl&x+*M5EbUqd!be{pQ0qkjtO0BMJco7#co*F6Oz?+YZ~yM9bB`&Oq#b}h4fOAF
za&q>c>A2b?j{f}QKhT3(@W-ddWu_oekiog`9v%eO?1bMb4pjZ`7M?yw-MA@$dsxs<
z0|M`QhL&Z3Pd%tiCo7IwN}QYlSn;jaGO4AJ3y&607wEgGv9WPTA0g5am;^xI#*^Kq
z@0t&;=4tY!igUq6wMp2p#po)M1xgj}HcLX428!$mhrhpnf(z^mO)+H*%w{yDv{9X;
z;OHo@u(T9~5k|s?M|N}lpFhJRj)52kr={6UY-6z>hx=O3&`{QTh2@b}$Bu<!Z=`{>
zT`Urn5%YI~`oh*5+KA)@<Gb~rmOxT73KQ*G1aO=}*_Yh7mE3XKDypdDUUG8+D<5<{
zl8oK$_n{tz=!dlPU>EHb^nc&eeiXt58eGib+l*Djg>yOxUr%*8Esgvt@)4^AdX~%C
z*x1<S_zXI90Bh?b`Jg_4(wW12oz6J@|B`hsT967n5r|PA)&rDH%Q0tC!Cf*{8`HG}
z;T{?W&K9CE=7ol_jU3P!mm!oI?0@ubC;>NbuJru+_AFLd8oX@P=4UoMd;WZC`oOq_
zBj+0iJija3eO1RNCKhETp{~NxRfi-wZW_OVVu|;jAEeAU8j~>(OlNj{+`SwwH-T;b
z)8lK1e;np$>da1yOUW!%nJ*10l%C_h<<~jO*jehz*DH3JF-s(NOS=TuWWyr$A+_+B
zhOX<hO|yT_oE=i5e){%j)0y#miW4K-g8eQ`2DaYLdfIe4^ytsZkeHnNq7xf7yc+Gl
zGorp{=apXxMNun$KkH^Q6VAkM6u2rn{ZJjWUwm->OQC6Hu0fFWft#YT8-zsmny=RJ
zJ$!b_;RCPNof@*U8X6er-DRxI<ekfP%kRX^C2It(@=tyrsA+w<cE`|bIZ{l{Dq~s(
ziLdZ_fS1zcW|wfm1WX6$t>sXo0s~QmxjsL9%7r5Ylabr=!qJG}?yByC0T};Kh3!ks
zMNlx3qz1vM4|~15AdUskcJcJ|^y$866_hLTI<v^E%YQr;!bCjoTXV?$f545b2F5vY
z(<&!At3Z%QFK6{~D)XCF#naAmO8k{o>NSyN`TWcDt&-y#itay9|De8YQRbtQ!(2v2
z*HdE43PmUUyIlD%DK8)V_C_;q`jO-IV1sP=-j5w-SEkig2JhSAFfQ05sWZDxZ6@@X
zR86_IZrhi^A<ID{l_JHH3OcicWvrw8N_5Zng`BbQ=0C%JGWyinkj!||iCitoo@lNd
z#glLJ<EGo^BYOV~$*ztY5Eqjb+`O+ZZ<8}usizB%)J}CeY_%5nu4*49B`H$)z_yog
zGStur6fKU(@cTJA844;4>SMuGHa0fIQ&XcMnTfP;v6ptja0gq0xXY7r3kPK|*4QNl
zvvW`b5L6cHvb5PHg0X?gYFlzz8Uv|khgx^VO_!FIVqD2i%(Y?6%uP!p6XR2Sy{&R$
z*4+JoVoZGqVZsK;wOv<fvE_sZW=3Y3CGiUiBzt%ZjsN2Oto_9Hz)Z-S$(l9}j*@WY
z$@_tA99#V8oK#s1=g0Hi?A}-J&zqKxy{WZK!s?o^zK5REKKGF>US+3{sd^##Z}QIW
z0u)&jtM2y7_BVCOev@_%wlX@tPMK1j-m1yP=~qQ3if>)>vB(xbOxw;~_+3J516yWE
zgzgTy4Gcm)_Z4-?ELU9mVRjA<X&^IjMo=jf_&%c0?kqpE49!hvXXgN_I88x)ck5f1
z6x)T+nRQ`|t<Ytwy=>05>1798mn8?#!%%Vc-XQRAa1G_4ZvTjaxrr$ml?>2OxVy3^
zyQh~UpE$}8pOH^y-7fq|NwZA4>Gv_Aw|~!%2&@TYk-yGTSU``4%-x6n^V><0iko5;
z4#}Vy!`ABd{zMt~C0Xjo+RmZ}g}D;M#te37gTupp;C~pEdCGVD{c)mjI%fK$mWJ+$
z!m+NT#~d4)S#h!^jpx>#JME_Y;;936w;f{6diFU>O;SFsT7PBGM14=t(;{vz+Ux}c
z9^wOzvn&v`5!r)@Wb7<{`U1d0%LtV_T)+rPJ|zIa6UHXUJ}|5B;@$D(P$+Ko|BghO
zL~(DJMO(MZn$7H(WyJKFlWSX2e$L_J7e=f7j0Ep<sX`y2WQiyWj!gi1`7mivaKGU-
zjKUC%k2>2VUn!kge7B<GwU0Km?!*aGi)<s$^P4e}C|2RHApr!(E7uHs9(iL$<G6sw
zMcO4&wJ$_#CB2k83apqLmZ5>q&|2kcw|mb2IL*fF;KPoJrNaAs=`N7W2L5A#u(GuQ
zYg=PyW;=$S;oJl9MbgMOL_<u+$x8|8_9hsW7~C@wbHAv17ySkNJe=`?tuLQH&o2Dw
z03HlHRim)ChS(P{G$d{>&|-c#IVib)?EJds6Q(8o)aye<oQeWmTnJT50OkfJCJLj?
zt$&{g>M%vFqoX4a|2cevW98O>n3$!{o;@>vXo1PopmHFzFV57T?wkd?0)QY31O!;F
zYNsASZ-o3?6W6zUM{(;bxBD1<$;a(Mnb-XKwPxh<=;3Y8keQ$wsK<I-lsS(sr2g*}
z*)Lnqyf}8`2q_qqm6e-YGJSS=dJOO;cbgtMb%K)c=%)V8=&xTLQHAg$sDy@U*G}<o
zMPR9~uSZ`oNOU4HG7_3AIh}7!q-H^ekMMFT<UKFjB6NdD9s<wDqD!y2+O}DzFf34@
zxDra{26?PEoRBcNbk^~3D3(?@3(2-xa?G|MuBT%8C@Lye!4oRsJ#d34{1(aDaPJ`r
z3A|C!(LbPJyja$B3$Fl=i_Aa;C}e46MTd{Wxsir^_w?`46xR>%WdfH29SEkCD_1J6
zUmu7j6$+?P6uwzUqGDsi5ZwqMlcLFtonr%=(9;kLm(0S3sCRfc63A*55)Syp6T@<h
zI|7<^$5Cb_)XgZ>@ef=DoBGhoi^b`*Cf4a)!74vg^sY*H1z09$F+aWr5B>J-+k>r{
zgpp|ejU!Mnoo}1%Qz9h9R0F#=_XJcuzq0uH`U1ZA2>M?Ci}dg1Sxu($X)ih{H|M>u
zap^Mk&#kTeO^?xZgDjc)Y!((DWZJ<c(gqG>WoO3%A_i(^F<2oqsr@P4jYoxBBaz?7
ztkS3Cg>nz}G8uSB(gc~@d5o`HXvBJtY&&vf^n(DElKYR!ruuO@?O=H0!$tWzfqu$q
z&8E5O&GF&kf)yUaZwNAg(w%HQlv(Fs{qpfgUjv0q`@)tJMSh2-TOCYgi6k2VfwY>C
z69P;q9`iwl4cGN5`~}j$+vwE611;%8*M2rUxAd7j3-|ykTqawO3>Wi*D?NE-1neXB
zjbvsD-Y6yD)2Ab__7lU{4iGp=YEN<~;LKuJk1+nH=z7(kPfL>kTHf>R8#{smQs9rk
zQnX+-kJ^j@&xu&WIdv3OgQ_IT9noBV#*IQaGkK2h){}X#B*DPUb-%eOTVjNdH@g$6
zWC{gPR6Pzv2Ef0>8I4p?MaP#RiM9XsGFmXS>1JWoyxLlnud#Si)pd1@xZiqWRDl9j
z0uUfFRZ5m}ZfoM*`4_tHI#0I@!+xCg-%MjNodAs*JlvoQ7Y-)aqrZvfg8*ub-<V*9
zR`>xBe3{X1z>w6SvuD+_XQyc2;3Fn5Ha3=L)3B5{l#|LCtAR93F-vC&WIDvh3cgt4
z?)v}LQONW=qZ_OqCawJO;d7796~Q8IRYdx<Ux1qlx`gHU(DU0h*Q<LW>FeK)XuQMA
zOBHjH^uF&F0VaWGZ}qLCn2aL&`t|F@_t+#qfI=dh>ZVZ2|9meai<xv^p~S{Emigs`
z|3=Ztns2lOKatGwfG(pN;**QfQ4o^i5VC=e9i!|7DLpVE)xa0{0ZL&n9vYBB36u{o
zo(+S4Odl$yLvMa#eqh491q7b>C0N9#bdFtt!4l~};(p*fkZzjeOvY)T|3Hj>z^nph
z&P4pH@e-Y34A{RumKZByk-{c$%5Zu^%H#Bcfv@<33lri1T9APW;}d&E+P0z#hjrbP
zc&fXaERZulKf4YbvNfk0dVa`bAXGJ18f9)h&ZbbT4YD@|J4-fDCI&=wOY4pR>?brW
zRJHUJpj3p>u|!Dup4XgZ#tZwF)J2}(J6Z8SiCi9N6v90f9*ejZ2Gir`YBWe@Z6JC>
za<LA5?%NQ)zl;u>vNF5e@|dLr8A3`=rfYoq^a(?x2TOi6Z-K?WDhiVCntsXdtmVY=
z6o+l6UN{9+LL@v9RzKp4tS25ClngnuO(>kh>*wcxwBcDmiR9wa1P8gv>SxcEkdXl>
zwk}#6H4H~WM4@1lk+=#*POgj^6Qv5gQe7h>7Sw(~#MU?fTNXm{*CZImAws%LXlIsq
zSHboPKK_O!3Q8vuw%NuCZ(+i02h6rQ>XsJ;PNeaD#|R`7Su5_mD5r0|es|~vyNCIK
zn>)J%Da7gwlsVeF6>=K@GXqtv-0hJ$H#Rz2h13HR>p9_9hZSOj)dKu@mgC<^tD~5?
zIpX1EPhsnLeqZD0fHTWXQs5rnLUunrWqR_PN$(_L6`WtIQHrD3KI;9BIUnn%|1n|*
z3vAL{ZBU_7C>Yz7AcW@A>5`E$jJ5>I((i|JS_31A(U&6eR#;fLza5Srl#AU-ieP-A
zNkPh#v7xlA-yj)KC{Qc8ln7!j8{En?@oFinN2Uz!%YA%w*x1+r&%}IatL;cv8E7Eg
z6=hAzKqp`Z2rdWR`m6Uk?>BuuB;fHmtJE#6p#1_L?SVE>Eo>Tt-Wi5X%tDbU^@)29
zh<yQQ;1NGIyryKly|<OV-X-3B+)AFN)<xHNF-a33I@T9@T*L<v-LoIRf7sxpPzNsR
zKQ|>!6v*!S3yEy?)j^Zg>mjn^8`&*e>T<>ykh;79?S`IGR9r0EX>MfnyQc4Hj_!FK
z$^OT(s!94X*`q7w%W|pWYfkkaq|wmbIChI1`~w7oPLEArf7D|;Y%#A~xsu_y_9;y1
z7VRi*Vcn6wOLJTt-~Q?6o<{?9k0R==!>Qs4Onygy4xRZ!m*sY-tK^4_Tb}A?k*dva
z`3c8>*iB+R60661n8)np<wa^^B*1KlDa7>pH^_KsYDl=7idW~W8%1N%N4`s+cK^OR
zxuzw7=+k_JQT<~QjInaba5un1D+z5BA;|pALiZrryzFy~#!k^xK`HJC%)|n6kd<1{
z{~M16g<^I=1?ZK4hpLeQ$!wj&;v+7Ke*~Y}9Y4;B^BC&^)9wdL{@|Iz&-RkZxjY%;
z2Pcr|U#;QmYPk26VE`bYS^tKlf+Xg6Rm{bf@qKas_-iJL^Uu%w5{>~X*m2~DB!bw@
z_W`*}hh9WCnIFhX3q`}G&U)S4%+CYp51<z$^C02kMNt`<EV3b3;rf9ygftt$kk;A!
z{nMvloDtSjWBo=r8_=^OqvRjf*%cqg<T?1@3K9Ai$WCiJ0h4^F$ABf$dB!i^xkQFS
zErc3k)I}Ukfu-ib!j3$DB=`@#e<@(;!>#ddxdzn?X(h*nn?2QOJnEGgq!eUv`&n0)
zFolpJG0!u(^FA6Z=s5t41I>rN?9_<Dg9?C<WG)SIFHF`oun}0r&5A7<33_NBz)YF<
z*yY)sFF`TFk=uX}lvQBy?&kry<;eRijvj8sfD{UaWK=lYUEjkj1GBvtDa2=)(t{qn
zY&*}(#>Q&2PR%E+^dxS*aGNY{HxJU0tRa>_IHq0SW4+rVMYh3AHKl6@yGm&9hId^>
zen(mTXk&h<5{R{UAfY(6Gs{;JZxcW*Yk!Gmc9Y~1No|u7^HaP5aSTSruuLr=cZ@na
z^?!}B)z}xHTS+LsWNH-}p*?VXKpu`!ACv%Bc}EIK6o^F_sVxBRYG@I%l6aO!L8_14
z@j({hpm6)cXqAz6D%r231&r`<7A;8_wMu{pi&>6nR*>n;09Z37&xBQ!G4gg!=NMq;
zpo_@Q&(E?fc{#d2%y6$HQ+@tcgB)l^5}=U$Krj%<7>OCYAp(9lV^}+xa!phUV6-~_
z`ff%h<`FEd<py5KPJjWun7fOX0}DnF5JP`rxqaitz496tWB{b+ibofW-b$MP{2Xx!
z1rka!JEGA+GXcpqamEh}3`AQk4D4cSh~oIS`)opXe2REbV>mdeZGbvNlkNbr0Ai#8
z<^>)4Z8Jcm$sY#h5(f$ec+YIeg_Ne<l0$bU4tzyo`GX*VNRbF+iH4;!Opr-G1iha_
zUi*9n{+=cj%O`PS#PgfzpkaUtQ*=gj)D7=4QOHODtn@32R`i5CjiI_^3?_L~QBhH+
zmAn~#f$V#zG&U$mQ(jUmOXRKSJp=|u;tmS3B-}wcC;;i}2#*u5O@u>ayME5EuG{;B
zkQ^R-@{n}EzP^1ZxN~ym)|<+Z^#TR!H`s3iuAb$8)AVsgk&tt-w{U33b2M71#d|?c
zBz>B@I_8I5)~S8UW`|9N1Y!YU@Rk33v4ui`_zP|2MGI4QKq#?-7p>9XsZ`7q6ej3b
z%%r+n>;MC$x8g<9<EHIf*8CayIXU}lTN|-I!ZdASz=pvF`Y|Sl4<|Upz{S?#iv!Mf
zQOYi=x;{Y+NKhq0D>Ps=WD>~iQH&RFnuS^f?4t;r+}P1k!dsR<e*Bn>NFuSFL<gX2
zhQ|c(?y%?*N^s7H65I%tGF;8i>dTOMc|ay|%Rffy2qWx6A%ajHgmAiMAOod7vhpBy
zA?1sIR)uCgll2&op%i~+(w!P+)$J%hgJHdSr$)@-Kv|6;q2$$-g#r^N$vGMKiJWke
zZYj?XV8^6|YKl^4Ok^u~Ce=(<O0UtTB&p5C%3X=D#YO!AjJwHKrnK%Vcp-;=m#-5*
z4PpOH_Ch1No4U2&b2LTzxbi-@u*W}M;jSnd1DNWCql5tiTOT6<bOtjAkD%Z8WfPuW
zA7~h4Mkq2zl%2HRLqVAERAM+nc)Q4(b^XZ5Q^Z2HU$Y<5dCtD-U)%=@59SfVFbj1?
zHT=8M1qRT5z)K-x2vJ8OHZTCtKlH)_+mTph!=LG$Wy|Z=#3(o4s+I{PMe=4IKW0)@
z?az%uXT6#P3j8sYIjI1}n|uYQ3ySn-WxQXUtG6b%f(^P8o_iQM2xm}>USH<6M8iW@
z?*>e}X(qNWqmW>M?ok*J<f9V$$h}4%3!NkcQgT89lejp5v8<*cu@T`J`PM;s%V)|k
znw6@*jC3kdsV+wB%Fv)?wmZv~xW8L%`?THCsnxuPplFAz^Z5}dynDBU--9H_TYgp6
z6>MVTJ=X*U1P)y*MlH`9<8b^q8ANH{hRJYQ?rP#5`ja<!BdCA~J>;eI(!Y1=4ouWF
zH*<!DhUV(mc6V340%9w;p>4Qsfs7lT)rh?muEWMm8P6LJ$aaxFyU?W~C)q#xl)gmj
z>k}V8_bF0#{JIomY1?_Yc`NP@cX$eUP%qX!Djq?!1wz1i7<7)HEhIx(o@d5ZbWV>X
z1nrmaYAL-MRQ31sG8CqIf;a$s29d~!MpIB`=9+ui+0b1r1}mtxs)~YD#*aahpz(#o
z>}l;;U@GC`+V&eB6!CS^7T<X(MESYQqXSD8&RZ_7ww3<@McP*c-cpz}?krGaQ%lK1
zz{t0HoN3LN$O9Au*{qUGIrBey8UE@X5E^RGpOEd!g^3qAeo*V0EzS%VB(u1P@gbgg
z!uK*H)MShnX!gZ%QA3efFl6cySShw{aXOUp1$udSxcT9;TlqS!d4Z;rs$JLJviw_V
z#|}BJ&!>_`%%iI(ex4%a6&+vtm0M2-pizkdB3Z9)|Leihphw*7_PJG0BQ;bWM<i0f
zu-mt#0Dnx@_+s#41OXC3$H4sEob+HSbzJ;kAEiY$;$>7#La^@miGJ2dDNl_Cu~gNp
zj0_g=nTXy1exM>awIw!h7JwD-lg37Mq?*g2RYX84e!<7f73OK?1m3^>rZe@aC%1t+
zmLgi5c@c>d877QY>*@wTF+^~IG#<6TX_Y&bJqe=@f$jD5_NJ@=#;}5%4+wiTFJk=H
zN{VMaeftj$xqK<^42OTcNEkiak<WL?lsI0vR|=B;I!=sgG_bMhLhxC}MRs&R9&r-y
z*t2J`px}`IUQn{ud4@8(PZT>SPW_yn>`O9Ty3Z142$(jjRSf<QYot#Xx8DlJ54>f~
zPsl+rg9No6S{ugQpiB&5*R6c=^5srIPk{7R8_fb^A%vO0H=uN(e}eWY?3n>7%iGPO
zZ>K)=^vHU`0@f?r-qOk?@!gMJ;YO*r&aCvV%;wfs&OATCqM~BPyGhq>gucxPqknT?
z=U)9A&Jzb;(UO1gXsyy}=KPsmkt2Jt(WNi<{plNVayq_O?@_o6U-Y?@D~=2=TEDqQ
zp<8o2NE&oH-X<=;Z%b1}wsU9r(Sp1z_B}iIIQ{pBlmL2RlZVM^$Jb)P(u6%g4Irxz
zR=QlOZ1H*+t%F<(ea+w)jn+Z1w!~5P5ajXA^3QzRu)VkVoipM<U26$`YxC=6pKmIF
zzEcbJqNLF*tb2(y7|Kmxt_Z)Va`8;S(;{^MVHDi+^YpwgT0DGVyF_VCUdeVpHr4RJ
z@(zn*tNUGKUg@nLRg}tq<G`?^DNX;@onzwP{Q`GPB=sEkTE3FH*sbvAr=17lrp5)V
z(}Jh8V!of4wClZf)?`<&v;7zIDLToFnNFV-@4e|eN^52$HmRuyHT?P(aOu9{i0r$P
z^)nBLSGX64lzbFdu?S7_`?E~dxTf``>}os3k-M=4-}^q+zWQt#wtQR5yL%6Ssbwrh
zcNIMb^j=iKN?ej#F)2vq3<KuVI64D(Qrp=%&((*(Eig0zsIn9~OLU@ZQ2CL;U=LmT
zh*25^XB_iI!_#xXVe<0xX*X@a_5e|=3a2F<1!8+LM4}lIR^!ukRF9|0>Omy;LvDJp
z>`qo$*%q9gVaPPET)j%Sx<aG9vvY;8symsw06aAmBOZDX>WOtKdJGsSK&iqbMtVm=
z4bF{(0n9sE1B{8#NPyoV>s>f40g4#hjYS6pX%ipPGcw5MZ<8q7xVQ#vKqLjs>O)9H
ztUv*<`%qaSI)OefRBB6$oZcJ6O{2^I7DFLWLdom+ZrHE^MI>3#S2vtkRlxSD=ag%F
zGcAM5>)Je{^vKBL*>RRDLT_uzBeI+02Nx}gzHZslE>e3tJ3(|}XW6ql>AT7`O$Jwf
z6+h_wnrAb+v8Z?h@2$!+mlwUAt`D<wIxeN-J1q8Eu<p-LpY6Jbqoe#htIt+MY$&}l
z!3%VZtL>nIf&whN$N;($re5qlC<WoA{loQ7Sa9$TP+ffgT&nzqO_z+YMNvHVj34=d
z4XQA^*SFb;T?QDUIKgFOIiM{eL%uPO4ga%2i}IoGA7#>UM-F$*zl!P+1!d-bOije&
z#1P)U3OMvVcu>hhoj`ibl%AosyMz#lyo`>5(F3Rq&|pwRE&yXMPUu^KMv-g@%rP9Q
zGKMD!rV8vCZD?kK6B1qnSd%Rvx&d!~Gh`zzZvPdeOi}o<FDbrmSh<YGZkP3%K5_m}
z?@WK4E)P#XE66^A1s~5_oBH!GE!y4OeQV2#XMb1;1$JnilM^9b^Z*yfUCmz>RYTg3
zniYpca(RCkc(afVE1n+hVY@8pgUW*w^E4@Fq*b8@!G-^iCDXqhF8lRlqzGa<8Hw`H
zZD1)*y-<WYpvq~XBJ#AZgKv12u(HxXYV}-#&I2}1|6}}JItmN{bT`L_ge*djGRtx7
z`OR%Y1eXK&+4Rl!$dO<`devZjl9ZaTL@DWpvw%FmEpQAn=<Gg}Z!sP`tZr~IOqMl;
z96n5Y#V=VVcNw{_{}!F|nSo^JRCY4t6Ts8O=V9OophvU+%*&>NA7r%j+O-GXHJ=th
z69SpijWKV$YdVJAA6{~E2bUbBkB`H7!G^{s=F*5X<>Ry<&=ZjJ$3O;?byrg;0L25+
z(^o-mVo}<SLKsKc#SfR2cs9?22`#da#IOhGNNOAyPHDbqC${Df-S;~?S@7ru6&bQ`
z4;~o;QUTJLDEuOE;v^W#VEeKVhy-?hz}PZZaImw7=t!c0H?XH&5QsSn@TO>wk#1oM
zW+lzP8d#O}k_=eEl%Tr0IwY(Vr0!%AI>-z$NK&zS7{J+i2rMAPki(#Ue&vV1$;b;e
z*f0I%Vec5UA4=Q#6q=;)2HVkl0ed7#ZbR#8M^;3}lYnK7B}ztsB2>@=D<Kop?;J=X
zU?KKH>8X*W7>IE>C;jf-yX$YKr2LTi5iwNZzjlKc3KxKJR%AH4vf~xkt%#GDqmdvK
zJimomOPs6D+K;-AD#=(m&COMD>*1YKF0x-%pdT9&qjBrV(blEP`?l0HWo&cbIDzbr
z%1-$(08mSrFzycA+*%DVh;Y$J#Y#~Gp9@D$4+0Yc>tk8@wY#ajZXO;XXV22XLy%}=
z(P~V=%CbX?hgev{kAjTE29O3(6}PGa7p=k#RVX;e7K8jUzh(8P)U|8I8V_x~ke-&g
zfAQdMb{VyRj7ua>HLCkSC!T6}`Pzl>QRYV}7cv)rHQBV7Ohpz{M8lbk(7LkalmetQ
z7Y*02L;Q|r4pDIdYDqU-3&aeaDsW6nm93X4k>7&|UL9`X;IJwJ)$PAebS3aG)dud!
z#4gC_<n)d2=3P)bDFJZ100>{A?WSS?Pyxpd+)+bBC-$JvsjDlh<CHqR8VLxxZy}}%
z4gZg^_m0c4{r|_Wq#>d-5G6^IGFloM(iWv9Nn2XlC@Cd{1}P~T8nlx%X`&L*PNgM@
zC>0G_zvq$r^Zoqs`~LC09{1z%e&6nNUDtV@$8o$~&-Frq|GiaICJ$xL@AbNFTL)rx
za&~c)+MfAf3|qv|W3&Qo+jh+`zpLR?&P$^2eDVnEhbM_P2TKm~rO2>Z#J8F`IIN+7
z%p*fFoJU#(pE`7|2c#e>8;++Ekv_mZ(I;_#{r1fqs;AvzVp*X<ic4&WfrZ!z`0^^<
zC3NQ4N<(p~XX2h&0GlA47mTn`gCv8XK1!GwKMn^`O7kx=G?6GiK~f>J>@q+KGi=r3
zo63OmHeEK_S%WEZE7L`9zb1_(q$X{Wz<w){@JbME3W@YxyY)Z_qMi@;TcwgZ&}-G6
zZ$%VJVNKjNcJ-Co*bVf`nq9YUpRYPAAapA^N{0@&l!+V^;wy-%3}DoWoc&ah5(3!B
z^#|Gl*m^47_VsNA=LU!O6?_6npn)OJe)Nuu3caSBWOq?4^hD=-<Fq-M>&iTf7Yv%9
zr^GhU`1H&%ASH(TdqEzcSAiv3(|YZlhB)8%!Y~&;pK3or`f#G70#ONKO?XK*UXt$5
zaO#0;DZky)P}Hs>ri4>BOO>Y@-~=GSA2>;bYoQ{<b{zyA5;<4z=CDG?n}r4f1-8&2
z+BKr^#e6;l7*|TlC3$HXM25f2Y96xt5%P@w{#7-(9oH=+@~^4w`gTFR{!3!PEjce^
z#(_R$$E6t%Q@VY7md>UG>_5;YC+`(m##58_0^tY)5=iD*;Aq;tTcB)*@}apEm5{@c
z$?S$&N0!JeYvfLvVI<nJS);MJd3}nfPaA$IL+)&(8t?hPPd~Ns5>^N}*xqvJ9SP`z
zsV|QU3TOzXgFM=_>R+fS%avStl1lmEA0P-37EXx^CdHt1EYTLDqw=%Ash<z^Ba&aL
zTTo$s`La>kT>x7snH@!`#4XI6OyN%YJ5<$Dp_3$VwdC@y-zkT26zHYDofo?w+<Ogs
z4K^8N4-*yBz15cM64O$c+Eqq75mc1qm5tW6{tGvAn+Ogl>>3k4BU;mXLN2Z;d0srp
znXY&o@(rBc`hOL0Y7hlBxXN8e_v%Ano{y^Q6K1l{0^Xvc0L-lc^KEWs<_8K3_0{V8
zH-k+BfffJojo#qc^H#B6OnKx&Y=VE5zq^UAJH!6K48sx@u<KD5xh5l&j5ZY>Jd*j~
z`6I#QUb9KF^r8$^zTp<D+mD%7p+tlNF=KB|Zf-ahd{s-!F+aWcwZ0P_H^C&3(b43{
zbVBP!rV*lj+W`#zN|=grF87i6P>Vx*+{GEqoXmvXx9}8v;5g@>YJP$>s&2w?-X&5%
z$kZ4Q!x7;l?zT-VWm!Ra1U10x_w_{~`}OR3>K|?jvN8a|u|%UI)D6guKvc!go~^$5
z<JD~LB8T(Th(4K~vO&b8656-`)SDJieGp(ex{y0#7<Z%pj67$KXRce!nhZm)r(4!Q
z#*s3-V<3S)n(Xx+k$Op(*6c&>2zh1DX5VAu-m`}*KRRFi!%@KWmPBK@cCA6DjfDj%
zUrDf%pH<Xg#e8Nb<bnM_KcT6|QO!lh^Fhu>5SRza%sS%CH(oY&?1j8jCrVBHX;X_G
zJ*D5Neh1#26iWYIY33}~{G7fVSJMiq4SF6iUII(=t%wNSK@Z^9ICZA%WKO~6L<H=3
zilZDcYe*R_9uQ~)$Ege$Ioh&-MB~)u!8yE^vRgu;hR9W+xTKs1P<x>bcXU;?ULVKk
zy>sYJ=h=L>Ony>Azje#cb#+4eN3pnnpCcoBw&HO0l`FB1>%%~2W8>AAwt@E@f;WuQ
z?sJ>LcjbpFX&n0Dib<CxPb>fe@)4y6^h%JPm_eM0!knZo=zDD$A;1tuctfPW4mb4q
z<R2oEhl3t@?+VuUpk#R$FSi=hh+)pL+hSK>765JmQL-7NBJc?v>*NA=hZe{GuoB2<
z<<t4)YOk-U7Ve}z^6OZmQlG|GHDzAE(=?l;IOl}5rn1S!jb{DlRQ7r^+|mQX#(L?i
z*ZGh9GRr#tcggy2^}Qv>^_S+p+1lEvWfb+^<_``J&u!j==y{?XLYYlskcR?~EZw<%
zI{?@l4*BPkE)=*4h(U>LR{(6BJv|Ao2o#7MBv=jGkR<^^Uf*k{sOvvfu|Tdavd4C#
z*KP!w9WC7l$JlO*m-Hy94XjT<^fnHn0sr+=)ZPwS33~fFvLSB)M3+9qiwm`Mgy&A>
zD=#SuO{}dX3-nH_xKB*8n&-`N3)cr^39Mr{w2F{5%xyRSbHjt(N4Kn#>;}{k)Dr#K
z@;`jopzv(#S@MjfCgwV756QIe-M(GH&^T$nTq5?0s3y8E64wb_=fR^#hRL<iGp@tN
z5d=Fk;hSY-p1F1-@c~n}zX=Wb-a>VU4gg}ogZn}JUxi+8+s>U?^@e{^LCXU<CsKfV
z2?!og=Olss&b|LkV*ECRhew?3H8Z-oZt5y_OtH3`^Y^Q@s#(9;>mvQd5bZ)zU^&}g
zj6GL=+qVh(VXPbQM26^I{J-e`6lGvLSn%aS_ce)T+&<;CWXH4u&6%w}RL5J0S)7bB
za~`PXkQ>8Zh(&r89ymN#nax{PQqT<|AVz(3OB5i}s^u*#9={f!egFlFU4k%Xmr>KL
z&3e551J2F+P%HLg$HyIsgCC%gIXLaQ>*{=GdASuu5$_);>*)TW!tL=m8n%<gJTJ9D
zkws4Pcm5_CMhP<i$n^DpI{}W$2gyw0(jdMFqN<?a2d`pAN|yS@t*0boui%e?ZYW}j
z<Xu|uGAVIQtqEJbW{tAC`YIxUVhrt?n}IsO4f6?3rG;R2*+!7E=!q0lhp{M#U5$`G
znGT!9Y?**KqdG9!b-{l1<|_p|@gQ34@%^DZrLOhYk>VF+E@OMfcJ+M>1Wg;k7hz(5
zw(ig57O@WdYm0@+54N?lN@I3}P8p_;`eX0nL3{F5Ksi(!pTr>7v_Y6R-E{p8it3Hh
z7GBJGdxg>FTO%VQA(%Zf>Md<BRSdF};LB+1)5AXKKfw}#rpw4$7Gi09WYOoomY^xI
z{=?BPL^%RLN2-BQ)E%2_;ScFNJYJXAA42E1Ya#n)d&3qKOL}fx2SSx?E@eOwmv;4|
z=CjMP7bIf88z)vU-U~n_<%Vq$PPf!)x*M2f0uchBh6-a%KzzjVh~--F?ihsGb#-bO
zRA5+o7BU_J2o_B5$T&hoZa^ZTGEVc+g4YUS6#be^=_Bx5na&P)#bms5OuCwtBfhE!
z*xK+g>D<iWb8Wk1LWa*bHxmpZ%F!^#qzXH0)G2KA`hT7&dKf*iBC?#sy^G@m2XQ2k
zncza*<QxPAsW)3H4p|5pZh>v1Extg4bU?pXTsuCq4GZ~!=RUEHd?x9&I-%!pPDn0l
zNQG%m)asZd_0sKE)~j$A-|F}7p-*F<=Nhpcx-ir+>rk_Da)tsOe*uttdU_fHwG8LZ
z3TQ*2>U+`My}8e4h7(X3tc95-cYa_`A*~r?;l$oZXoL?RHWF49&iP4MI2Zs^%%bu~
zK@ke*l~4lz$>?rHMWyO`;{hd76KoQVd6uf^B3`t$6?`lU0z5<ug&9~^<DuDAEPM(=
zQAsEOaK_*l)(_RAJ4IR!^l)<-@)}7(0IrPkENEgS92tn*E;cp^6+dDYbUTY2#>fmp
zXk~$D2%O6m_4x9-Vt4`<xCz{tk&#iNex?D@kl|1CM>1rJR6{vvU<QdA{0Wa6J^~OY
z3I)X}o#SY;emoiW0(L!J&v>9sPB~N`nI)L8n>`tI@7~-;OFYnwZf<TU`}Du{UL4sH
z9Us3o@ps0Hr2=T6(A8;>(F6LHhQm++tjD(Up#Hp;)*2wKgvZ?S3}-9albf*_e>JCY
z{?`Htv-e1E9Wn5|&BEN~fA{wZieG!pNg8b>J-4iouI+oL7~)?yGv9J&2PbeP43NIo
z&h^Oz<+LO>G76<Ms7utIL~TI)hc3hKIEi2Z)Dw#8q3P(?R>csK6JKsouh(i!hP3T^
zDu~$u{h)zBPZC`jb}nT|`(Xr!A9M=>Yz1?*{(05z%3jzDr|$%k6%?WG1IqUT$R2SK
zv+!#R*o)wd0A6|!cp7nQqOd3F6tft>OlGQ)^bUZe7kaPf7~gzth2Kb_u&}TYh6wCF
znQn=!E!Z$<`S~Al%&W8LJD^J>lewWc3<iLsq^e2>**!{zg!*#p_Y=fB4GA3SQt>cS
zrO6-hYyN{Bn}!6SAm!<W9~Ct)=};VI7HHlv4n!60A0<)4ZLaUvk3Pgjz{>~Gs8?Vc
zfEtp806|wW>^g*yh5Q_*cGK<hmPxTklPFHI5?27RmoK=BiHY;Jc`cGXs;B3q^*j(r
zs0B3xYe{#6QXdERD>4QYU1ot@x|<0ecl1$04%{NprQ#OTQFfjzWsMd&shc_W#u($8
zap0q>yXrcK11SJtD9&$UBEqwiQGIm;*N<@K!OBS-4aZxSd%-Z2CTb0buUs`^$&T$p
z%br$a8!X;$A-Clp4ta0)S&P+f)SDkzl8ReLj!XfB-pCxk|E`n^52X@I8hdLMrX~r=
z!l+IQRfwF;;#Ti-TpRRDmZ;|4-Z|{t_k(`TJ-;WxerGHUM4~j4Vor!ypRta4cs%B~
z#)gi$XE7y-h36ZbVp8td*u<Q;{^qb-$444AWkvSR81urutu-H+^p5sFzSx@^ojdu~
z?aQ3`XQ}h$zhvKkZECk9PD99;fP}qq-8VsiAE2=oFug#wyatU`b~pymAfc<S^yhmr
zSC_P7c+^PY&E@l3hX@s+;8R4YiiYyE<5e=w9i8p@(Si;#xdoM7QvIoXOy(gRJpyIT
zunx#Q1(JJm=+`wn4QML~dqaLN<Xfa~bN*#aG@lqWdE(mz#QdR_dh+gbN%VJKWET=B
z4^!kU<k%4ubUOY5EqS=`ktr0k1Ub_MK9WEHf?iuy%P;*of~VX9J*WPG=(GcAqhGiw
z?q3_Jpnk}8{`mg=&#>Vs7<;k6UINmFgMLz|<5~9wq)@?AE9E`ClPG@2#+-b1B8x~z
zN0Q%G@6ne9qvz-M{Pyw*oLR#2Gk8mtOA#U*@&~^xEG;)T^zApXh(tE;<W%R)8yW}$
zSK4NO_q!U4epJ5WzNgd0C5pc<w>C6xvOm3vjwi!s$5oz<T5L<q2bEjD8u`uNU-8*m
z^Ve#H66=ZBC*YrwusODWF^hq!U_DT{sQB&lD`4d^8v8&ZKropa!Euo?IR=>%n7{r4
zAyyzDhq&M{oXKiXq%a*~MqfcqA;ANbanJ$y32qUQcU`{)_Q6ea-rL&}n?u{|^=l*H
zVBC??Vg|Ky`xkEK{0j;x=Q{pJ3vj^cV{vzU2k<%9H}gGjJU-#%HNRtltrN1Ptd6bD
zXFDWxENu!Ys<~}WZcG%M!=S%(2BLK@%sBF67S$STgG9CfAWbF8zIXIF7z(8BZj+EG
zG(M&2$h4A*+*N$QQv-Q`Q1!aD5b|d8VK=7!HQ?O4CtVy?9Xaf_v4IV}o?VF2$E`05
z{59)#mih-67Jh2_azqpGXNLM|f$Ddke?DMh`wAh8#-+-@V2(YQX?%`Wz3A!fR~Cw>
zCIW}{-HQd=iQ=Kp^UPt$#Y~=gW>-(t)z!IbV)%&_iRDJOoBjL!QwPDPCr&HrN!aH?
z-UBx=AGpJn{a@Y+P!M8qGB6LjIJQdCt_zhhtLgPc?P@dZ6#qIn=+^zq>7?gl=??YI
zODDiuvgH`!R3utgGU*qKEktEd?qs)eThwtQ`4gO;8vA}kEuAPV&D#I<`-x`_RdYk%
zzFi|~O~euD7I0Xcllh5Im`Y-BfF^ks{!V}a(eCTybyz1bVJZm`%>V|<N2L^*NJJO$
z4z6u!EB||;jfuI9#Z`x@B<&`8(>Ppb&Wi~zzqkoIi_33{cDk?RHf*3<%abJ3yy-^X
zz6-k;e(!gF>Jsou*X1Y5^cv;J*SnW$Egu&|KU9>=$Uj#SV}Dp5szC~JlqT5(oI2c}
zc?edaaFS_fes%AL*=wtzvnd&RZ1QW4N#wV@ex3f&Em}_1`u#3?rIIg)ti2QUJww#&
z&Xv;sZ2eu{BW@J;J>|*wb;TDLuP4!mssy0vm-&+Wqe~)23t7<T%J{C;E7N#$cYGD{
z2w}+S6S#Wg6K%DuJB35r?eSY|EWv;Szq?k6yTqf(#8hB#3TE?zTbC-qhfE+b@Bum$
zLKs7c8-Q|zD(zVm2Y@ajd)v5%7C2E(b1fpq-e{cWkZ@lnpKnFQ=>~WnIKo}lzy3L=
zu^Ir_47Ia66)h%b>N-xpP#n6Rb2%U}qPurNna6G~BxmGcg~((RFw>~87&-64f(J@^
z`bb2?(jy-0Dl<Q&iHdbbn4OEZuu$XuAVlaVPMm=6^?7sc8z=UwSMT&~J+NcPtJ2dt
zG1$1MDG*3eD7%b5&X-IB&$LCbWJw<KGl}&d_DO`uA%BcQ0YoVvsf{o~oFn>wje|#Y
zJ7<0ama&KnHGKG^H2wOB2|u#&D;vK5G(bQ;;S@mCM`~CX<T}#P(S1dtKR&|80;1v<
z^4-<?`{I=Ei<GdV>8X9AwbOl`l(c(a61nM%94Gt21g@WxJq-j7@&j=gao>oK{+yB}
zxeeQiW%yk01v*5syaa=k9S#gfcB=Il+5Cleo-?@zf{-xjsFTf*I?_Ify%lEqCnBeM
zKRj0aK-AvJ@Z}aQ=xIAX^O_PuV9k!>#Jni`=kslS<IN=!R#sMtsKX=LP%KCT3*kGy
z03|urEF+`?5v}}5rNhIlE94*5&s={ncW(Q>4J@jQ)C{R&`M2Hn=2}h)2P7U<`&}Lz
z6VCO5T`B$Mm)j9Bl9SB8{<^<DkRB6v;^^SdOfSn-o_xOHvL_|B=!tygcnO>?stDqt
zH5+$Rq~pfI2B9mhr`@K14}n*xc$Q~T!Ki^=v%MDMdyhd8Ykbe?9M5$g&=z6XNBf~y
zfA!dW=~-uH%QqmIvWSzFo&I1s^zKG@xN5=ypDf9(s~H#`n^mp&(A7oMrxZU}>4?Rb
z=-!dLcibtV5RnYTfP+$QrE~tkDhh-h{+$kmz4IR-NdxdE^jT!C3$n9Q-&`Ll7i6Qb
zr}&|yr9#lklS}VGms&v1u74>W4}x{rmi_){oscr~$_v4^)+EJ^1heWbbxqIA&`}^&
zOg0H{Zip3U+j}F4zQDL)NIktbGczW@w{I>#!)Za`Z8jn|;C#}v9FB5vb*%;%O!@eH
z34Qho@DsKLU3QZD?XSaqPc*e?!8uTpT{X@oY`l@V`j?t3jMcQ25Rie2EVwk8o%8DD
z%YGEaIgfuUU|hj#{yhME84^r0gyDdC3EDDiNiEGD((6jJ`{*)+)fZs3u4*e-Jlq6!
zGEGdIBP5)YCOh{L?^GYhp4FCp_ZZh@i~Y3U`n-#7^}}^n6N+x|B8kv6IGn$HWr9uf
z^P*SJ50t6&8ES?~^YY~$iWgbU5w)+BcCCNt$DEq@1n6TVd(Tr#$E^j5gNzS3B1tum
zQPE+2&}O$oLyfTqyJ{oR?1N3fW^xGO4C`4~CZL{t0shX$t?f=(JamO9M6RNfAv5Pm
z2}8zoqu?Vs9>mCo8KEU^n2Eg{0vTm!stCXj$OeR~SAZvcpK-@cCbLm5PIsOL2QEiu
z34-G%B3&qTFyi1Kq!JW`Yy*;I3H*#2Pz^z|u$)QWGSqqKmIVzf02C76^Ta^VRvsh?
zuU)(L{Fj*1cK!N0GGd8yh6uxvgMeYO1h4`FKmtY1J>Tz}iKoT4<Jq~hL|aI*^HAkM
z4N2M<tvKm6>4(xC)2LK;!(mQnSu%!%JoIEE=DOy~Prriv8x&~FM9c|amFdNaL&e2v
zzs@oy)f+9^t^GSI7GixU6*3ztz=i3Sl;ZuC<3U?5RU*BgFjY|Yn~tazT*uR*kgiAK
zDsX&8w4Fuh2r6;j?5*p08XWFC`3rC~GNl}__xbb1R*V)XvO0ep65ObjIx9DHTI`(D
zt4(^!dcvO>HpLc=AN+Nv`QrG&2S=;r1y+_*H9yZ4WOIGWFs<}P;moC3!@42*)xQtj
z3>RZGE-^7k{+VewDavgay{e%3n~mGis<OwyX~da{#%tgje|#lUQd7bXhUyVx4MGq}
zW&z<7D3xH8A>3P@Hwu_M`QvQ&uFLF5KfuBqz#&caf^aFUqcqwT5O)f6_9g@Oap5jv
z%0l~bb)?V|oD0(Z9@jyV{z>!zIJwAHO8ID_odr=Xi8M%Tnm;T02EO<KRDd6f&TchD
zkP#?_0b)u)<M$ZT!dx{|VQ}~qxWTj%r1@@?dLTF>7?m=P`E&OJl&nFe6%HtnIBkv!
zMV%+>&*`HS+Op)2GY*|ghMtTD!_riKWE}%SKtFcXqv8c<Ig;)OM{iMB=DUI(;0;7P
z-JGIsRDN+kG1owsfMJ4(_@r#E;kYp5%)$e!tn(7h-v{m#uN{Kjw46K_^PCy(1uzC?
zXFI0m-`TvKsNEL%cFe44UM*5(Mevi~l6?8h7y`p^yyZ_NMdu*X*Jcj(8=H)S6g~)z
zIQi1@*hxVQ3sv<sJUd1s0<Cq7;Exe!xf3lrU6h9PNp0``%F9DFudSZQyc2+%?-dXg
zqDaI3qnO!da94cx91)&iHmeK7FBmBTuc~3!ERJm=WI<lm4H(Nv-jtZw)8=(!(3=im
zXRAd?eiJF7Xfva`PfQq`@|j{I>o+d0VRXORdJtN(&?!R`PyRND)Z@|g1x7oEYTjXS
zVnw9=zL4P-N0Gc2ijVOMNHgh6SJKk<dZv$zm~=y*Q;Uq7>)3&+G0sMM^K5%A4TTH7
z5wd@vYYBSv=yV_-->p_@<-Zo&POsIKdj*8JsgKvX@1vMZo%o2z_G__W7f(uDI`U^z
zn1s{$I?*p<{7gD=46dP((kqzT+@MXZ!PY(g(e1r@;ojBaj7b6rxqPY;ubJFL(%JO6
zPPI9zuT)H%WRL@tcg{h3F%IrIpgoylchU;s_Idy+Yi2>VtwftrSyS_{-Xqh?F;h=S
zS;b&s*d*Oxk*D<bSNc=WKm1gbpiA#1nu@QYorgMZX1=+iVYB*d5W8;ewhT#S#j8op
z7e9Y{GLXJX!R_>!JNKc0Px58~6#&=+R}LQOwMf%Qcfa~7O(!|gz)fN=X`c_<veYlc
zrC%Zz>93;d`z24Wqb<I2kZ<y~oU(PR5dYIA&sSB#CtRdPC5yZ5&2v9LO1?G{+vMQ)
z&S;Hzt~^i1LlVA+m4gD4EFQG1hTn>T6cb7tuU2J(-mOe*O*?2ecATP+sCnh(nc$HG
zh3NL;JHE_k{%m+GKV<06=tY}!)aC33o(xWve5z-ni+)NzBBiu))S>mqN7R&C{V)lH
z3n@0lV@6^kAYo#6z3{4=3Qx%Tl<lRG0OUy)1U%HOI>E79bXZBKh{$(A=71^+h63YB
z(A2Gf+O~Ot`&!z7Ile7@W$NZfW>4;7GHsBt{-3K|%7=Rb4)ome8RK|zJ@?bwDQerH
zCPlpYVgDwFmt!}+-#Q*idm@ly#maO@?x>SmzL7vkd_^;0AmkqGfrEi0bEJkTw(_p>
z_XCHa!I}J8$et-)+CwAaFW>EJj7mu={v<1dvg4`}JyT<K!N_Ja7IAT*5=QCSY2{C^
zJfvbKWMst`o%06y3$NMTe(mrwUQ?bY|0{#T=jpE;Qo&oTwk6piis3Ti^MMhIdax4a
zjS30!*msa~9Kt(=!RCbRRCh!klF8;Qb`WvN_~sw;u*Yqalr&7j>;M^^u6JKqBp6KB
zUe;CJmeC<Awn*=3zgmY$$0(NHV@s0CsEwO5b<T!|l`2j_DA9|b3NZKFCfgUJm6$}D
zhpS9pY4QDQs*OpS7IsRKn1<91Id2Qp*nq&uE(ZU}YJUDpllRif^YqhF!xUP%mCqmW
z=k8FG5DK=th40t<Mt)oJ_+C9<aXL;1la#PS%8acz5z(&7b3^5d1xBg|;8&;2O>p<k
zU_w$<;<0l827MF7Qkki!`T|O^v2dgSyKg}Hgg-)>s(X96U?1^svVEEV{=f|~`2|OG
zlr#stYCgBGxq^DI>kN%pV<M_`S;OWAa<&J|28BE@3bPXL7K)0PLX2<(!qBKqefY4t
zw)T+-h{^3cZoh`6hzw*`hhv>8MvD*EnJ4M0KV!A=WArZAzz?z>5{S|us125eHJ&uQ
zGF~?HZeBMhPa_jZy~Xh(4m@SXkHBwkMV}i^t)fv>0A4%-C!)9~P7zhih6J?Cr|yHP
zcu^WV{`VcmTaO(v2LuUQGA)G+R6ubTDaA&BFmYbvu^XDo;#OdlbXEW9o4u~$Q<2k3
zcIKlSF<!V!b3)6o+qqvM-IBR!7mkMki+AcPpA`gVmPNNBeBkLD$#<2uB8P6;r;s~r
z_ZA((V*qscsB~z;mZR)eu3VW27;ycg?HTbX8~Q<O)&g2I-@Q?YGv8~<4uF1#%U(Q|
zJUkJIFDpP$05w`SWaJ9z39&_=*_QE0#A@q{ws|Y5P6_NgUk`j{ePro7!WZnnCdboQ
zs?&A9a?%K9Bh+<Rzj$3Z=2Tyc+7`MkGJ6^x%Mh8cGM7QxuA;N+FxHcVaKWq<rn`c`
zwP<2(7s}w^MQ?`T5CJ57@vrF_gwVV|4~wKRR&wy5vpPRkGE6ibU}lW759{3~1p-7#
zJuV?AvM>+e`#)PKv-;skX$4Dh3>bi6_{n}Xy(4VV?ES`zhl2fsDbdd9gG}PKBA)T|
z?$`V}Evm~SPnqrAoEoOwYK9S*^SAyA7oM7B-$WDrJ<AEa3ZkO?Aq0ht@uSNRET-$w
zM}Wzqf!NL?a{M9k1<1dFqac}Z;H_QIxDdZTY}|1A9F6-jfwU_W595HGJI<U;nw#=n
zn(+m|cL?#Qfct=@gjA!yF+&M)G_#F3)bLw&Xlsz2SlJ^GPlK+W-pV*_!+}{GzVChA
zncJve@8>tWOw+q|m)Z8B+n`kfI#jx8P|5>ulZr1_9g$LI^nYbh1&mGYwcKjOn~EaG
z+Amyv{#+JH@A<;;?^9mjPSe)1=sO|ojzmKPWxPQ=a;UXXJXZl))P1xcFu{R8pIHc!
zqWyh+PdMCb|D~oABQe*7&&-rcsO*~%eM+7+5;n63<Eeb<{mAkOPC;VYE@>3UmoQA(
zwE|rlIo8sXnZ?hbs(HQNcLl|`xV>lZ8RgIz6TQTMv!e7&E(gBT;|W*iyq0E~Z?sd&
z*?P6eaQtQbXEz1k;;yz)*N)!aUhvX2Xo<`v?g?`urlpSz7lC)ZWb_x&s(pTa8)MhA
zveSlO4kRT!RvL+?0J>f;=Oo7A0z4k~Y$)fweq3Ig%fzQc=)n72v>+KLBz^L&{G4>p
z&JQ0H9vrP~82z%opCgonBuF{;`+2;D8~Zx29_fTAG6O<cx!FA6@g3Q}Uy~_*fa&#T
zUoB_^r+A3VuadT878BY~oM{hjPtwDeuijB-%gL;uEUNXaD9cm+q^vvzFs_Om)8A?Y
z>sIcUrumK0X@&3HFzqFTOO68xGcR#O3pf0MMDhieBjFy$XJH}=xOtNa1a%TXi~m9r
zf?>s3L1z2{s9lSe0^-CMUZ!ZPAg4j+h770k@Bbi30dqcxf`n2D%Uu()@i0M$$j-p%
zBPz2BiX*c;qx!fLad}1Md}yQK*LXa2_qCNj+_b3FLx|u76VZVSBMe*#GB%M@A{FlF
z%T6;RyyYJuUX4Bw4+7?06ct=PsCOIB5U?B)=Jh=fD!Z&x6}?`J=wjI@{TR%puCl!4
zWO%!#Q;Qj3w?BlI5ofV-lv@3PwG?_(Y*MRq23baWxjGm`wq*)1`r6EI7=4~_Q+hpg
zh3dYn#{;9Y8x+rZpShDzWVDm$)X*{np`x{fv2KWK39Tu3{WE9`lbe6xnyv{{x{M(9
zDqK)P5Tbh}ts}97pg$xAepByJal$yEE1-6E%N>I=npgm_G-o1DK57|2djOd(yxbJb
z^I(Dl4AOgYzaS}Bgu9*WNjO$N0ZX6#njB6^G`2izIM;sM7YII>)>k+q#^Ed^F$SQ<
zs{jELVKhlGrTD>Mn2c*6AR<x&@FWzMi;#w>k^@b1gH)1B9(4qKbt2;uK=iKyf>aHv
zmP|PPgl)+m1DtWlxIp?2Spu%A(n#9!1FEMAh9-&cm6b%5)gDE@x~7l&wXi>ghH=w>
z?Vsa)XT_zEa+)7|Vw;Ti*xEXA(*UiT4~{9wi`_oY_Y8yuIo83g4M0y%1nE<G#Nk68
z4p`f;JKSa%2KC5^fh#LNPu^QfA;Sy$A!$Q_TxGRak?-~Zi9ClCkx)KFL6e%AYE<mJ
z4UcvZ%3A}OT<u3MK>dv)b%jW{36BfBj95t!u5$6uQ1r(At|=b#5E>J1m!vV^Gyw=`
zR#0Q0itrcWeFfG)O~DZE)yVWDxAMY;3&e#8=O^iQ&|V^A0AfoZqq2b1Nku*VT~SXT
zVBEo6(=x=ZF;IxmjU+S>z?t~<Avpi?E>74*<KF|SN%)Ld));{ogNgynEPvNW?%hnz
z5shgLx~IP`tzd}J9T8BT(sr^rpyasqRWj8=qt>Xf#}-N@&Oj<+H!I?r1JtveYKLJm
zT}O-cK><NAPiF5xrvd#o@vf^RH<d2MbJNexR4g%~AO;%neDUvhJbfhfKQn3q0rUJg
zB&qRDp-U!sB1G)B#ed;?{vRgfNVdTs@c64$W{*X3jj+dIG9R&}o$wXCp5)Pe`H8g(
zp!*w-AMdtzd#r^kz_ewH$!M&yzBL?(XkQ`!GY(()m9V4?!9k8`*>#9z2tv7@Z&@^!
zR9b_EOI*E>!8?6BC|(w8VK={^t|KEG4Y1RGMLaAS>qSmQ3V}$1lY$`}oGTnYS?pX@
z8lHeWZ(=A$hq<D-MD$b)b~D)YEWptaYXc+(M|0G5*Z}tWb92M?OZwUJ#lQ9>`vZwH
zn>KAC(E+F%w$Ur4p4v*eFYjA!Y?+w}?&NWi=+duOAPMLxSVJnH7_cScsp-~uKQooR
zmC@4fKH_2b*fu6I)FSO0R&Jp(uH~aD&E>ziskCd}$PG0U<H)mZ;(^WB5@xgWBk7MC
z(JyM$Zk6+5B>`d<7Ccwia$W~5H2}{Mn@vsB#PoE|@1;yTpqC_Za}n9iSzy<ZOa8pz
zfsV1USFPKji7)aOto9zqa~mCc_*K-Z$oi#*ly|RRQ#rdmG`9wMR$#wgKM)L!u5QHC
z+`pVG`hKn}SjKvES~lCSGtxAs-A@fu(tG5(lG9}8l&|`C?4;Yk%VFmT(tZ?xMqfnE
zI0n<1*J8U7Hpr^n_g>ag?h-H!220UX<spP$M`Sqh#*yJk#9`iOD<&S%1O63jnpinf
z!xPYFQ!s$UI6F_g0gj0zflpXTY+U>9HGmIv2<_lq;UqF817$dH4sOVbfR03bMlb20
zyX=;4+rCUA=wu5?uv!>SN^fjP{mMKq8@V^C^;Yo6OTkXY`_V(&tG|XQ<Oj66jM;s9
zdF}i_;0BUjflf3TxMC6KiFUJJGc!e8GAMfkzeIOB)Bqd9$hvVTh5E4t6Dwiwg<;N%
zkiCdUoM1iV4&^@D9HLP@;l!SdO$EnUVw3GSC=h*nhr{`Kd5U^^tb_$Y+ZpZcBag}k
zpmlZCgGJwjqc6nFDk+4d-(~z%0j*18_6bQe;QAXoaKAN9zc2`Fvb|!OtBM){2~$KK
zpYrhGCwN`S-6mckOq85k(5=tAd-y{^;s#cRGaTV~gmD+_3XCKEF7L!MByrJhy>`ma
zz$U%xOfz)mkFsJmtyb0(mvEJ-Q{GRCxBJtZxhB8w@Ze1{K4*OA#;+$A>Deo5ns;CN
z%!Lh#Ol3m_*5fkGO(CJmgmJ*dm>9{-t#*(gKpCt~m$e_|hs_}$=N5wTkw=;gPw}!i
zAD|?<EQ`iOMdCXRwMzJZL23(X+n?gRe1PRoay?tqk1=oAf_>bxzi(+#5^hek5O%2|
zO@xf@IpKyQ2pMRase>fb5ed(u!wCTMG4aRrn<WldctN&cb9m<Mbf7nko3gHG$}c-q
z%=bN?!*k~&n(LQMs%><PgHAMP`F6)L7T@$>rF9crVZviZ2Y_rOLLm5jUG9d$`*XB{
zC8zsSSNT=AHHh}9>YcjJ?B=3<hpuXctgU-MyY8QOH{JDz@~U-xp+O^)z7R^Xn=Jgn
zrNur-{s(Z>kSI8EXhdlQLd*057jWX!D_~#1jSs;a40{7ojDH?au*V-`m#9R5dk8l2
zgQ=wpS)ZPGS<f-{YJ-d-LjCAdmp0vC-=zd+0YxFR%>sHVJfY)qp=DFwMc%l6w>Rv}
ztDi(mM50WH90UE=I?DLOL>1^vGJeG5`T2>kQ~XV3v0zw{UNk19!n9J%zWqmhyf3!<
z;Ly<Q-2uD7gb?K~#Qw<i3B<XEi34Ntc6utTE1yg4Ub)t8sE4k@e~-(vZPC998J0Ut
zt$FoADi|7G0s~D1h@XB6CnCzfjxPT6pAQkSi8_L$*zrzbya-AEs($g}VDn!<G!d<$
z*hJwQ<t?z+&=@<u@olPe!7u;Y+4s<N#`1>NZ(3*j?2}oS#1`%DwUXYuencI16O=od
z$JH%4=5N1mgZfA`>jv{$(eM2~JkLsA{%NR_$JLi3Pg!^F%;NkF53j7==bOjgg%wQB
zduE#?hPem5a!~C3*)i^B_s3^Q!##EClg|{2jwr4Kj9P(TL2qCdPv0IS`F#HgbJaxW
z=>&7!84~x7!Jc+;n=pQj=etyy>8XJd=!AsjOk;?E%py<kj04;>q{H#|UqO@_n1xv=
zUt!?k09Y2X<E)AeE2(VqOFnLl(J~k?ib4Nr32EjKVX@Z0LsA8?W1?>Q^L4)y|3?b|
z3jl>eKxy9Y&0!oB+YlY=#aLpp7XKhh3+E@1EDUjlV*&s~TlP=?;cCh=VSx0>EMZg_
zDz2_ic&Dbk<lw7VqgvZla?4W^wGn#8n^wzz%qh*=8NVB}C1P!^sBCY@k_@OiI&|;+
zl)mnCgHMFN$@U%EO07?6N>%!es}7|M@tgaO<euL-9`o!SgI&m)RNDIszgDzzcwZa&
z?lSgy&xg#^3#Xc*Ib6QZT1%xl;z6+6{b~Fguk8=J@=I@Wb}=s+x{ddb1gOc}F)$Aq
z$dxPQSdC;Nbf{{a$y-8Ijt$W&d(^b^>aa3%{H=&-C1a^j0gt0B>A1jd$AP03)1!V$
zsif*w^_EZf-UEK>`r_Yo3}V<~^n=`XaVpn$o2;x!#9!jjCb0E4zG)A00C1#OMie=Z
z`U3VAGY~P@vQl}cia-+)_Uy@#$g^k~AfStA%~`eo)uv*d2&z;qnp3rAqzt>T?ytM*
zbcwqaTk3$;7pg1bH||L5#<fVL@0l)(ZhW4a%R9ABUW_A||8US*KM-W1vW0AJOMB>d
zQK!blRP37AdElZOum43mPwN#dJ3s8GxO(vh)#zKNuaeG8yB~il9-Rvp|2m^zDo>PT
zzyyf)oTNN~GC~1Hnhrf7{gPCpYj(QmiqJ%Qcer2$5tlm2xrXRWS`D+`@X}2$)}aL?
z6%py?Hf`oM>|CdOiO*lnPN-y8l8?6nU)_VqP7v;CHFezm%*k{~p=0CXnW<q1r&<BJ
z6D1M48<O-3q6AOoWt@v}L}$GTUElT;_$_f>`*SCv6grQ>c@J1EQtv|WfLU^&)m1^)
zk>FP>3liw3TMM`uIGFLvXG6-Zwo$Ra)k<7wDbX8tWE_zdZ)oCoVAXRybLTE^#x9Mz
ze(Sv3(zI=sTQ81NnVuh@`O#gp4)nX9akdI9Rv^6AN=lC1)+$iVY)eA;Adz?g_^X$j
z#!j~#f-a11B-%)X<?KNjPT&IK&eGG1R8C)x5J;0C;B%;f&KWKTHoFgk;>)GZL1o_;
zV4}tK@z=*^W@=$lN(oastcHb)K4^#m@*cv#lgG*&O)r_r01Heo2}Z%WU4v%pKg+sZ
z3JDDpH-1;x&?s!hEB$xsjx;t%@x90W*RmU>pyPBmL^K=%Tl_03mK<xzC>NDvW(b3*
zDOZAmUi9>oY0R-Nmp3{uUXt?H&Y#=)A^xB(P1AyEx_+Ot-kt3Lc9WhxDJlv=--PIL
z&+a`~yGY}i^7_7qC_=R!8owuV&j1_ruTRGR@Rk6(>$khKo@?^PWo~*P(xusNEVhNc
zcBN1`KNcX6quvg1>;x4e<KeW>YT+1duO%r>*ryI6`~dwS#Sfz%QIHT}t1gMrN$Muv
z_%mk&aX{}##~j!PE#G)C*+%?dxOIeHx_Z!d!*hwPDFOC{>{U+d)Bw2X3gVsXpHNZJ
zOG<CA`{6uu8$Q;{q?JZfPSV`}W4EMaWVQ<mzAg<RBEts;nN$?Q0KK)@$K~~Po1i5Q
zah)BlcH-#bFvf4Y`T{=N*rUU(YJ^l!R*ulEcWCT=Br|QgxT-dH^lDJo29Pw+-w5&R
z?|gZyA~qm9MBBEC+;wg|<+y+^3F<~}sV|bb!1*J<+MhrEJW22*y^qI5tV%*bZx<5M
zsBXnu=xx$r-A2Cx(#;*|3HR>3iSiBRScUQ-O(zf2)t?u$bJ@ug{8a1nDK}-t%SD!}
zSC&3iaebHYA^ORx9^*BGQ<_sXY<cuSOLl9oZi;Jc{viOAImzx=y54!fO(Y9>1qB)q
zv{i=RK0uZB1TZss-CC)fjwH{vl#psxpwx$Sr(J^mNw=-X==k*H1586c5*^6$%^M5a
z%H!}(+Lhd`UxUqnYo}e-gct#xOEB6Z4bH&jri6F-ITKRvtfit{7GX9S^GmyY>js1D
z=Z@7YR^ERj9TMXIMpu;wL1g<%2NaVxE4|eQFThiJ7SJRK8N!Z?W&zhjS$RETmJ8Bg
zVkHqxYHGoRZ@IGm;I^DE{EB|SMBDM?C3*K-3}15&51ZXLSRr-6M;`nFe!V|ERj-YA
zI<&l4E*nvPC_pBTFTW$b!h6OE@R4pr7v6!9218bU5L*Jwd;-@o<ZcH62*D}h=?62r
zO;q%~bmr@}HvLael|wagu^(5i*jmf}SbtnaMCTw|`%Vt*T!HqR)03B@P@CjnjV0AA
zW_$8ZIXXB1!6#2JhGjffG&+djmWgkV63FO1bRQHy0AQcM^Vgu9a-N@c!T48}513T3
zI1%n^h9Ut?4jJ;{Yql+eY8iXvr%)6{ZmoZ`Eolx<N_FfZP0oPWGzUdG)f?+wb6y_7
zH*$Xu^o;B#5GAOOY+z@{gh6#(v}I`z`s2bmCJ{zOrqz-OOx*NDB7l&=D^TRaTET(M
zpLh`O-@gymAt;E;iypGU8q^UfVcM+%DH}|buRUP0OJ#{o2iir9)JLx#^<S#rH~ws1
zTI(@UT%{R6FGC!;;1;6ER)v>(Fk>}GGnh28eUhpG`c7<dOe`$a6ilfjxu=MwB+WHA
z_(TZRA(w!}I9h7I&S0YCNMd9scb#~1X^1FRsQ$?f8skCp(r;JvTj?m#mPa$`41NY~
z1B|b^bdquT{lpI=E{_(%SJ|ccUw8w8a4S#PEO{)?`Na~fmTs@DRgDON1l><lYW}{G
zR9`^at-Ud!#in-|M1v^l$w#+G3Pze$$g<c_(eFOPWbS&41Z!ps!(sc%;<8Het^fII
ze*5U{FGD#4aV@{cnj}6|zWQw&{*=%q)znx?n0E)B%^PkGuM^S$+AY2}#((YH=a}8A
z*~KK&nay4Aklz`Hcs)^R3<mLl?;EcO`kyIeD=On-u^w+;$&HogBUq7idfZvSN+{h@
zXZy_^Umu^k#A#Et!uK}EExrx(CdQG$9G8j9Nk(Qa%o1|X%P1J<p($;*Br}Qigvx7|
zPXKY6e+_O~M_2y#y$r{klX<CwT09d2`fls0CaG|rXsqHj>OcKNl@(@=)?gB(j?of^
zQ8kM`!Xa=MIRz<_Mfs0rZCwdze6_6@$nL`<0y9<s9E>2rhF+IiUcQ3EyBDXqIf(&7
z?ia?8+5fnZ{^gqgnstAss_(e<^}CLpIlq5p<$FQp5u?`1z`Jq%W=E@AEZ&g)QwR#=
zc3N{NRH1PZ3ONqg4eJKUp>#wPi)YjWPHXZ4QjZhO=uhbuLVp;S76ewfrM0?@#Npu^
zynW<A75XE%^i~kr2A;ASY{(1<k-f%Z&;twHWAI0S*;7?EU%{3<i=i{HwU0Mzine|D
zr8|mqyb0t)P;=p6W7Mc~eoM7>`Yma8ac5I@83G@*6@%h@fTRN=hbHqv&R@7-4$)Sl
z?cz*%35l{ss1)=B(Yyb-HPsJAN>iAd$~4~XzkEMpjTL*U24|`%1V2?sMJDs39J&jM
zERMvv?S`HMjX9auc&_L6!#p6>>*Bpe1Oe<s%N#&?L=5)?4iMO+Z+O$5_vu>2-FK%x
zB>WZL$$kT11<t=z6mtM5SV(RFj1{2>WQI=c;D`~qgUWhG$f3p<0ZKG=Ge6b8Jth4V
znFmfdAY=}ZkwAHF!`cM6M2A?7vfwIY5zsqmaOP_}BC`ceyfUs^3DChsDWI8D2===c
z(Nt#pvxRe7Lv>b~<{(q?PG0>6#(-6a`0`u#;8R6nrNEo*3}hS4#wQ}l*?ji7&cMIy
zbsjsf<!VO_J+h8n^NUyQ)bBjGbZgTkNy^;(T;Gxkpb+G+GL}65bm|vstEhN;@FLKF
zkU1%}wZl0DAWNKw>e;{>LW0z5I5yIrBMfUi6v7I47!kMa-7z8lC$@1k{d3`0w!JHo
zxNl1{h%-s9;rsEnMR6Owv$MV?*~Z@enVltiY*Hi>fjd}Vl+6dToj&;T8}F=aYUw=M
zEtrzNuF~O4&=K-1rkNAD1V>8B{XvQ!cZa~W+pVIwJb;%_f+(goRMndxOP%!4u;nn6
znIdq`(^CcgI;oGougTuD&xVW-{!r>6Op=bk?_+T(rR#kHH~sS7+45BsIC7LAw<ie2
zj~^yE?ZC(ah+6BPxYR7l-tY^K76S5NwY=e6LW4=5XGp);08$bs2`~=~=8_$!tW>Nd
zRQ1yQu>4&)ZwWHl7$q2HeO3S6LPnpFiv8(RF^ZqSb&@qgvSV-&jB53##2>`ge&N>F
z&uI#IDa*NrJuyZ}c_5SrG6iOJ@@es=s(2@&NWuY%^hA{rqnTG$h5f1O-rHDGk8gQx
zK?P<QGFaYpa#|_7%M+ZRmcNLXW&f*hM}vj%;AnC_GTjg63;Z$0oOcL?zQ^r@QH_kC
zhrkX>NE&$CiL;Bu=o3*0=&cnL=z{SN$w9<1KuS-rBqTqXSnJPYY6sY!P!Jb9$YjUg
zA<~qr0{lua1+>Ij0Dh<~<K(6?)3c8A3vs~Bp#KGl1Rn`O^>ITifP4eh=Y?;RTv&<%
z3e`2V2TOc@z&x-A5F^$sHw=E{6vZKaVD860Qjg>QT!wbAz?6k2LM6T_B(7wKKmo-#
zbqyz{-tK!PXLm4*J0`1yD&Y$f5(lKQI_GeCoig*Y;n2$gaiP3Z;tPu;uvgt~8EEPe
zB{gA{dVTqR@5d!>8$TzBy;;?!ig78R56Nik%!}{9B|_H}N|-NLM4obO#mDX{bi5I`
zdV>WxrJ?DaNTz!Q*mB!-9x@&inkY+b7e)nkYy@h?7DECCG39uO;|h=#C45t86@1<o
zWNHE?!X`h9+_cB1-*aJ}9u-=a^3hq;^XBjF@5jMi3nNq38z)9i=~l1pyCfvoF$eol
z$*-~L*rPq5+L>saa<wEE%JzSqGITroIX()R#3adL3nK1E$apIB7uzK0H~ZlUB-bxV
z`OQPrW3uCw)Ddl>f`X*`N}0lxD<}%NxJ1q@;ODU3)pRTN`g5ZdW%|q^=n?5P1Zn|!
zEc}mz67u2VL^njTP=~@Y+w92bC}NO7=$C$ux+JD5yY_{H2&1Bqe9g)RU>neFkqA)K
zoq-0KG?a^=f(h4%yrr=cPscfw?O04q8#Wl4ABopVRx!sV`B3iT9k~20N1t6BT%!%f
zFPx6?a%i>bh}5~rs9OQI_MBB<N<Cu`HKIA8QSt7aNp+Fu_wP&-vulD2W~aRWJR2-K
zqWYpqKVWc4E<oq!CC22cl@x)dO^u#)_ab~2q<zj{IMcFu8lBQZ%y^`&MA`vv5*qfX
zluM8|6ZynSI?en@V(Y{`HFWCBEoEh9z6|U8b^!skh`*%0B+*-)p)aBBJ|2B6OSH>N
zK8*go>VuntsbRXg*x{8}51AA@KhPN5w|6he>%{Fe@JfE}aHJeCVzdkw3(K*w8SvJZ
zk@o_5&eR7Y=ZB--67b=9WIU4ePU2em7iiod*OB80q=cG+o0(H>`M<^s%t*!@;;eA}
zA@K)Ymf(0NYSdZpd>IHAKGC>8v7+_XSY2Dsrj`w5t1A@Fs*hBRe^%ByG0XOR;d{=X
z#dDT(1uy;%Su%a@%WAhXDWdT2^4qs_<B}SE@BB}P@0J5`k1Tuq(`|FI%l@8g{t~ll
zYRu=q35vn7$mrWliP55kGPby+B(Gd~>ZFMYZzgD_&r_+fW#{=F)*4u0Gy}jAQWj%r
zK{#14I7{3T2TZ4jMQ*>Y=Pb$-tf;-A#g{oQw-+e9l(aNSJ|m${g!iF577_KzEjVkv
zH-@m&;+urKqaSC~yY`HEo3b9F_dog|{diQz6}xc4XLhsPd!Mb~Rkh>(d+F<UoE%fa
zJWluvR95Lw^KVd9AGwjVhC`KKO8S|G?kT?O&Mr@a?q&WR3;20!Go?tE#e|K=s>mca
zNuat@kx`OA$T-uNKS^NxC*5*AiHdT|Pi-b<aUN>gHGbe107m|lZvX!gCu3)(pJZq6
zBK%MQ-MUZ2Y>0kw9Pk!|yjeANGE!0sxgCb*!|((d$efLN5>pTDd-nKUr~%+8M3OCi
zY6RoR-x`AlHITMZC}@(E;zdq;QsynN(KsH6!tTFU59tqewy#G$hvCq$n5}>p3!blx
z-F(nPK*XISkc31uGGDzM3)R&U=LO&+I+#pqxx9bGKY2d;=>~;J-pJ^Z_Rau`T}mO?
zIf@yiH$!2$$MYXn8_@;b+ZQ-iXmtjp2six-v>R6;;&KhoMhXRrE)?BAITHd>1_wW^
zr+PQA`?%JJgeR-`++O|k{qL&r@O&arN9c8x<SNj>d#vi1$d)KSKTeGgu?LFN?U){{
z8n&eh%jpU>Swi|@ebj=tBk4l~X!?&^hleTz?2tJUy=I;p8o&_8gj_EiY$TKwX+uIY
zvkUE9kfwmSy#ZnNIv73_u#aOPq*;&Z_)mBWu92x`L}HBdir7nGoqJdfLw^^^QX)yh
zjo5sMZjqcls1~7XB_8SG;$m+vxilvzD-YQ;v3(xdJ8}B2sqz^aSg!1j-+CRgKA*<v
zx{Rp$e<L>$Z17^!73EfLJcs^08#=Z>n~iPN87YII_!Bt6mpCd(bwzxpq>`cAdXcC{
z%pqAA|NVO<1$)OQaQn;C>H%t#lNhFTLTHi_6h>sDgd&H&vij+FLf!X%8d&Gkp8|u%
zF!x*kk>~M-iVmml!d>{}R21-nwRo?a_wR?`Ol7cyrFf!xwP-)0c1S;rayz>Fz)x@^
z<GH>It8qUGrcI<~BwAr%Zl)HgZV#Gl$G&4@Nxdo9R3%EgVkMyFD%<iIIl=TrVTqW$
zeTG3VO*mcG+~4z0vV^Kw|30aTs{?;n3N-dXK5CzyEJlwj+OC3_cMdFmB7GMX&1IMy
z4!nAmNDu`C1S$#4-exBpS0ZUyvu0Va;Z-D3d)x-XC`?;#c62i+@I-2F>b@B`0>TYX
z`LKbh2?SC9)OeTQy*Zj*&|u|+f+6O4CMM_ZVYdIPklBasC)p^6?S`+Z#>&&2^FCE~
zD6rKf?GEd`w;yh09=z4LVnYdiXh?!L&PJRNHAiwgURdolhxrGbh@fD)yXIydyYr;n
z^7`pzF2wL#(oDE0DEeSxs|&{#Ot*!lu+QVZO0ptA?h$UChIWH=z2K0LtY#x<WU##%
zoXO&*r)OX&eWb-JmrcF&gCjqp!e%Xd%r56UH@^O`r4Fpy^lbV3Zk>&3urb=n4jqV6
z#9!Kg2hHnjSae>HcCF2O85p7xVP>Lv$Fz;du3kgS=0{!2gYn!iW*byD3;Y-HlC(^l
zCka>}-y0tvOb|Mne^}tTuJ0cJA3Oh<v<AK-zERq8PtNC)eiW+9Im`frcz9NQd&`b;
z{z0SFWfqc4@b5Jf0!`Wz#6&?DMoaQHz<^oFEh0L78wS)?TPZWXd39gj{PHFURUaRN
zeiXXNNOVY@Zx5ns(N%pgcho}S;OW9j9X^tk_3!0QqNHh{%6&oFI9N>FN4wa7S(C``
zZ5}(5XlpI<xcAR}@fz7>l37^0g;~dX89~e$txgYFAiuPKXY{}x{uAyVPXezuDE_zm
zBzf1!(c4#ns3mnJ30<pf@Q4A56$}fu#{G~q^QL<mi3a<r{VD4TDnM+dQLXkk8~)p0
zQ_2Sj7(~7K9U7e1#%|rdpKx~halkJwJ@x#B?rLZm!-3Wr`(nD>8QX&A%<6h4LFBrx
z+Ts^q_5L7})UhkUV<-0&tW)aQdWDAl$B_j#+GFX<T_x#rf`fUnMHHApfJbIepu~RA
z@tug~A&%7GypdKFaf1Q<vCj6!u3zRU8y<2OeU=V0>A7s%7Y~p@MnL?8bB~#S_;vZ5
zWlRERLYr8xOF9j@Il@UpQbU`ndkq>7(o!&sUr6G9aj?AX@zlJANvR8iX)I;AKe!>J
zV^_$%=Sa|Im3`8-6yME@bz~K{-h`}zi_8nYbg7(o3VY~{ezV0aG#c*w&%1SRU|>&5
zm;!##chGFmU%>J77VSAkx|YJKlcUs|OX6DdzMfrUv)Xf^ZOhXQG{wTjJO}U@bPgu#
zrY9O?4t)D2sO|%v42s+3Qe2!B3gE}cDO*Ww-&%YuYiOzQ&u=feo$RYI14zI_7R!(?
zH_)DS@9Nyg`;FE`iPO2kNORze6}8N!e~HG7p_raQCL%*;si`Xu1FKM!#{WW-4tV^D
zXaIUN7<2ItJcKJ1DHM_<_UVf9RA0R2W_TKt3RG1;2(=P8R3$Ch_iXz1#$9XJS9=bB
zyXh$QLYuWQ>VF?CW9S<h0Z#_R^^Bh0jxQ6;bJml&NBhBn4TisNbF)|lPmb_<Y=P99
zx(2ZJag}`z0F=9s3WEzrUV8R;<}LA^%UYR(M+!*COF8w3Nx)hHz$9Y2%$`!b@@Wcv
zcL`_Mxc!@khW^-q5G7Mrv7lbvzx}U8Zeh|ncFeAMB4E|}JZ2i%FIKCQUseB~LP3YK
zpK~Kmh9uUuLb^H~|K@YWT^BUaoX6gb{Fp!c=l6!<*o2)Pk$i7dSU>!i3xyB8{rT#H
z%2O(xlSzs>x7PD-OM3Y5<x%>85>`6hB8AgC?~vhS|6V%Tq-7rols(D*<a#srhNy8j
zhA8}J@jX>ej!jyltIN_?;xDOjKGM1U`jC}ialPDsfAd!4@-F(IH6<uHbn|@N!R6xl
zpz`4mZu;P9pY<9s4AA|3Qfl9mzFk}c=9lgA=V{X=B&1A@gqE-DL4WQ_i##>tZIjnx
zb@*8pow9zXpP^WSFRUa8lpZS}6UZ_8{Mo=vWP-~NV;*3x(2etWWU0Kc*ppmY%rdYp
zP{Agz^8Ho$tLvG?caR(c($r(~8N5u#r2y21jD!c6P=zuBT2R<GT`#otZPl{b_xH#W
z$I|=WxLVbdJ#$-q-)rsde66x#l?;^<{i<yWv^P2Zlh`IQ{QN)Zdpmj;T(>{8D<?5K
zD#I~*SAi~#z5Z`qVVOE@qbN@J3dEUstxlc(voP?!Rg7u{-~QvK?`s28x&FM1`8h1N
zE#K;Z+q;RqPlK{+2X)SzNk=3>Xe*|JQ!df>#!xk;e8`gBNwxdYF*Q-^4K^dQCDdJB
zlgC0wUGg0a?tL?A=q+lf^2iv!ujite9^uFps61&SM0v`GO12TYumZGoXhYP&qqG37
z@_-*_Xm~g?Eln*dD#{DKC~@!UAJTEZN`~W)!GUU(A9Eq6H_tK%EPH4x#!pZ#3G`mW
zhnbCEY)bZcI4i7QDy(OGJWU~ispVFzczA8LQiqrRdF?u$XT5i)PM68e*tFzPD6~NO
zh%l35pN&@8{hyGz=0^WA$w7IuF?$_;>pYVf{k_XzkIMdzWL^&r-c45Zk=*B#*U^F7
zr$E|Xerc||Pe@kwquK(HT7Z>D%*}1?!~k7xg;*mCQIy+>tRXTX;cSWo<!--YzWw<Q
z^~qmr9d63$XfkO24zE~DP^#`~oh$jCRQy<BtH}A2j!GN@7aDkWgs(nU8^4&b^PK$~
z`FwY?jy(-?Ekc2TE5$A@7_(C9UjF?bEx?_~k*GjTRt_s0T1y@EP46@D(J7QlDUK@*
zSGq^<bdNqPr{W5K)Yd`g*Zy$Z4A*b9$znOzXAQFDUY~Du58mo-8GkQqaO1OMQQDL0
z?1rVcO(iXClWI#9<t4)#!Uym^<n^$L+kqAr+yfM+@PPxpY72lM?9eTPirJ^T1mQ+I
z^*Y%^a24C9mZpm;z9D2n4>(I`t1mjq-;D>?3G(sPs4YN0oL@EFmgXE~Z~a%vW9>El
zKF@>Jc}5&5Ybjzos1z4tU59x+Mmt}tEm+@>{Qj`p_n~$V$MA)QL6_G<f{lzt7iCy3
zEkqRk?(az-=`1^$+`4qc_vBOknIje2dpNr{r&pN0wwumevHYQHnz&b*y}jl}r*CpA
zH-ASb0;duKpBOnzKDm3GZ}akxPzv&6*HRLv>ZP8kMzFHnv6WgyF}eF@v~@%0cYVF8
zUYCDAWxHt0v|{%stqDUR0eRt5qP`yV68RT}D6MVfE%!WT-)g7EnRE$w&xosXg{!9K
z1Wca4|1kOAU+w2v{;SU*{Hp$2Z<n`8o0_tLeZwst`JW1LyI!|f-FJLAD5m}@M~v-R
zvtIxFTX)uff6-*6H`5AR_n~*XJ~0U*>LXlt%%m78Cii>Z7=~#Ze#^Jq`nB>~w%yj&
zLj#LC4=u$v9eO7gRkBX~$?Ny*r={_wqF|;eA(TXB+uPF#2<Q^OG=JW5ii(nbujP!*
z6#Fp`s#bZ9D-@2CUrH8Sy-lt&j69@xZ*Fm+Bx>_e%LqzX_^nOoOwZ<zw!8TL+hk$q
z;o7}9btQ+b#{wd<>;}b+J?-gQZc!+QBU`?;0;MT{Uz=0*{82<dp*eko$38bHiD!SD
zwjAiNSUGRcxOeB;^bKE2)}3J2q^6S(Z}!)SJK$psn!pA(%s3Tmt;9s6$FtiJ7ti(!
zuA8m9slnMl7ia|dWSekZig|syPd;b~<c~7+RC!imni;8W7#n7#m**B#|B>$Z&o5@G
z%O9u}k~ohyyNc3)`;7EeROFP72ziWsJdSrlw#>1kM{i<;$>$;cnIgfR2H)zvSw4^2
z+r05#XLB~XqN3tf7ZQeqrKF4?^%oKp+@mWG2NZBa3Ej;K3rULfiA_`VKOc>muX<9~
zsOm%2zqd0>f^ke%QMhFT+4@G<IdX$5+dEl>J;(IYy%TtE_&-;9?KEm7y`GY|Y7X{M
zF^Gud7Jj#oN2C#UjXRouj~@1c>?-0_>u(>*)b+s0&5aNhXZrXe;MND;)T_;u!#CP2
zhF-1ovq=v<TO5(3DqQm7u0i*f`8y~2#OZvpE_$(2oXxqygF2S*_$aU_1w5t-<&rDy
znn7>dRIxZEdhiVX+~?Y=BS*u>FFmx}$ks2ew#}4f$a2&|n&ICgC&COc>IudYSf4)q
z9pb0uXIPA0)IRMS{lA|mUL!`oxeh$*5Qc|<n>Umx2mNyLT#twJrTO9*Grw(I;kE87
z|NStwD2^+P1o>`-TJPLwXTB;#p^Z4*PPA=N*Vb-F_8J=-o3+%A#^NWd!_RaMRD84Q
z9k{WM*3TxbI7vaIsKq^mR_BZK&ge%U+}Hg3rL8O+SEhczXxnsfom|Q~`T3&|C>SF@
z1d(oO5SD62AgZkm6+DptjoJ)iImxGI(h}#hG=y(xsZ#lQ&3GtYHP~|D#3`#>NzT;X
zsM6`ox`IQ0qTUQm6l;O%CNBlK@x7dhsWD<_A`ldw2u-qjME59YIJt+2A4$W>gegQ<
zSyqJR7W15X#1Q_*E-&5Qb}Lyqy?-vyy&iHI`PwPU%GP>V$FVdfOixtfB%R)mE)n&~
zq8)V>*aH4tAiuu~0gvNpXlaFZ?>-7^dLtUh4B!qKXv(Ew_kgF!{)amKQ`%bB8TANh
znwg03=!5|=%F!jI1)jE|9=0wwyMRKU5+%dwQvGt%#fO#)+}M1CJRNWI(sW&|>u7Ir
z=K1fcZNQyK%*xWjor=cbxCd}3>vQ@-E&m%*O?#xG)3#8Bz8Q7V7pJ$ZW?_6p<#&~3
z^B#rHz!>Rc^QncWvs=fVI%2}yJ2g%Q>HW~a8+@RB)4^^p#f&$Iy^XjCPWekFzz6r~
z`hub1kn?(q0fk98-rsa|bf~Sx5ADTAsO;z<OQ{{>BU5l$p?Ud+SY2}ayrw_5%-EhR
ze=DEidit8b&x<w*?k@K&6x-eGHx6OLJ9FW0cfE-0!;pOM*de;@^(I03Ke~j}86NNd
zeEPfNzmJgtBNJF+{R~8BPUpg&Z`Y9%haJ#i=LxKN=tpu04IwC~zBVY?PLmrqI^b~?
zEWj$hh%6xqa@eeP@bQhr+UGRknYNbZXiVGtZtNECt1DB_>eMK<i}Rz39F68#h4J%x
z3BG08%v79hHIi@Lvox)4Tj}^Y-kf)EdH4cXByrVuXhZ7j>)mli+ST%F-Q^B?VJ;B$
z+4Gsl=Q?I6?F9GjI|dnf3+{I-CQ0t!vKDa+csMq`p`$1q!w0(ze&i6)_%^JP!>Uvv
z6X#{^23a0fPr8k)uV2GT$v!xI)5v;*)yTbw+@VRS!?h(=<%L|6PcYkVy~FZ~j+M%6
z%K+gEQE?jRo3Wm#hzN__ZAB$sGh^9MVSR_L8qZh@I`YW)_#E6rVEuaVy-LR7uC|c<
zmt1^m=`elY_qNA(8IT9OiTB*b?dmuU82L+NaiUTvM-4HGSQjLl2c!>w{{9^){hnj5
zoBs<Qlt=KQ*g+Z9RpRytkqbF+vXRnjci6cfRAN%G0TkoSXYZA%xHS%WetTJR@alT9
z{^U)ZoN^=_J{%pKJ(oX`Q~uYEhEB3!-Mv3^bN0wcEnQliYlC;gxE5|JF;^i9<*hmG
zHrz|<H=;6`+5GbQfthdJBkk&YYOVGCo@X;J?_!Cogarj*75n}a(+qIRczkVC1ofI*
zDl08*Ecwwn0NFKb9%gOMESO%mXZXvP5pceS08l<VOGVzii9n)WX&TB?6WKTTgkp)m
z&mQQA|1sry!6btBS$NTaa-Yz-t%eWW;&x8z2G8(z?xcFS9Y2_OEj=qs&)7H?-zIT#
z()IXG_8mKR{D0M*iC4~h_xG=a%p9}GaAe95DMN+^q@rjZG?+DzP)S6`OcEKQ43*NP
z5_L+03`Zd}7j+URNhL%XBK=;Uob#+_t!MrIfTwk@``pK^>-rA+vxoP7@Ar;mI7KCM
z+JV<E_R)KPSW#ii*lB)iOGT?aqy4fooP&@D4@G>tpg|&u*s9z~s&JNXRzc`zQkFgs
z&xh$|TV6KWmlZAN`*z^k(YasDJ#(URx0(AbyjeW=*9-B<s?6i1CeTJ8lV0RU|NUe@
zKootV!TKe%-rfWQ^*TVS;_Vx;n4>t(HjcVqVeVWT$_}&Dt3zwfNF;@WNe*huW8;b=
zuKqaJyuf#9#4yDRQVIIcrwwzp>vJhBf2zWk{$}DboF<<-IeCWG=c^k7_v|cg-@(0U
zdVnYo^p^1HbDQ6c)G?%+c#3O4pz>~{>;->+L`VIYqzn}@dEhqiNeVH^Q^$^}@98DU
z4apteercS7bwAbPTbJr@4ZG4mux5s2`H`ZPCX3xeQucl}Xr7+xGIreoh4Lzgk_T3M
zt|na5-a2<v%!K~FKdyKZ+iscn=K_?i(HGx5V+j3Z!<8%l+4iO0@j;49O^Db)LpD!g
zVL+Cux3HZg$VQ~Ne>)|?TSCatG4y73m7$Hzh4^Ntq`>C=b*bg=RDIRXZq?>$N;|qN
zy&Ct-LP?U_*WDt$?|`!*XRgkdn|)nxS=r9uklic$z3Mm3p-1>b<B%;MKmHjhKi$f$
zVqm{FR+rWu``P-*QsdC(oPM@DoaYrOPB3-aS^4uvhHb0i1v5-$@!82McGq;$(}X5P
zN6(;c4Hjd`u3l`SaHYSEFhWt62vJBt*=e8cW3F9O@{v_}XxD-c*}xdJ^-m{XY66^g
zNMPrR+j;-q5B|+<pB{^s1U-x2{w>P3uv2=&Ba^5J%f@;qZu;WjQ&VQ)d($Jl_-Dnm
zi2+|CSKGyzuUl|qzn+JV&yG*gC;exsG(WM({QPiYT!cvNEbsNU#=7mtD=L;iFs|Bs
zLz&dFU{gRC@&)z;0dd|@<5xIvH<*^b>CDRBlH4D;JYV3H2#}WfP=mB;x8mR^`i!is
z&K+-t$dnI#s_)>z)K+u=65(E~TBO6mEl2nhna{P^YL>3bNNaffx<%JS0yuY?nrpYU
zHQgl<dYd#3#g{JV_bp&b<COHKka&@;Ony|gSzJpqu(~m;Bk_e0r9pb{yCkfI4>ebT
zI*pZ)!fT1xg|wTlj*bDmXqbC?3)@Rxq~QT0rdK>3`c}5TV&JJ3%7@6#2{M64`E<#i
z9z$nGK(nr)#o`xA!Z}_=<<#pJYe+B)ko8e!tgilf`Z8R+5$+6PMKH0aws_e_6Rt@w
ztGy)1RGe(@F6+)-``CjsdZ%}{a14N-if=&4P}`K&tXZSu2P0mEHvFpS1&PWmx|3xp
z5^+l5uxXg`)O1-Wefg4<#M!3p^K`e~D3x(NRaR*lwB#UXM*O^J&3^pY#MJaO0F^ST
z(A@`4R-ZO_d~4x!9tt(vz0Az0(9rqhyDFK<R+Dg#&+6TzJ4;EvP8Tj<f!7>!GneyY
zeX~~mn<;+1Kr>PN3X7yaHZ)9;3paru8`t!eBb>;EBk5-goZqE;_YJRBOC&qIJC5-3
zpv+smu<d$+M1($hZ{=}+v!<jUkX?rKI1p_7^(Hg)l>GMCL}`j^G~M_r=NW~J!W865
zBr)B(YqtGz?pyc2^K17FwYak1bV`FZ&!|hcZpuc(z79<)i($Ck4h1FgkZ6H9sP=wR
z$;QAUq)$$sK7E<##u^$MSkW<p(-^FiAvW3WHZQ*82&-IF#mx4|rtIu&69a=4q0_vY
z?_KZS_MLV+C-mRfIA~&Y(%GusU8ipSFnVslUZ#m@Luq14r{&RRE*QT4ADZs2w3smk
z6BfGUf`Yn$&U&&r;;@fTReXm5s}n6p>F~m|YW+{2zP5W{uI7jJ(l2B-ZU6h4<(12;
z#Fr&|(PXlRHOTDpUi_AW+$Z)ckC&GZJ$|C4SN{%QhA0FsssFx1CE<j}AJZFFhy$4b
zJz~x3)!KfP=qgwqNsO%i{uCYzYnLx6;}YswfPu7Wmrk7~%XQT|=j-dcep=v8(w@PQ
z7ybRi(P_-la*&R&P4nBZnw>`hPU%g>Yd<d>u$0hqscPZR%S$t~M-ll`AD6hui_2^k
z4T<Ks)@DUsnYHuZ@Q0pMM5#{|yhsO}{YUrNADfzrX#WsGU)9%dP?epu2e(h5xb=DV
z^=sFV;+uIdD=T~Lb(NJ$67@w|R4hWz$THEvByp9Q&9*S7v<9toR-*@?(Yebdc)<)4
zcdLC>)vrD$S8Y(aeno&ZUR}S#bBV<T&VeK-i<}zZ%(BYvYcXuN2=h8iSw*qF#v;a)
zg&o=~r8$<7arSJ2-4JF9%{)9(#JPmIZjVsSBjxszAIm9kZqa2#btVj1Uo*KnRJm@?
zSHFAze)&jIgWoEx<ZIRY;88*_03W$orsyyz$;qhC8HCZHT@elFU0{&giO~*G0yCc*
zW1VU$zE4wgH-;YX`9>mnJd8>!AB`u5JRZ4p3%Y&MxUWL$a3fQ1lkz+N=(W`$OGV#0
z5iuq>Jt}e?Zmog+w3M|?K9z2%=U*=$?}uJ4TrLI{7PBx^b@Qu(b6bCG@nj^ewlpys
zGSHI?5fZOOnNwxFvNuoI8aMl6#HzpxADdQ=>%%PsMW(zCa8zbDD6Ky|W~MJ!N+R}f
z0)<UO%L!0gbC_FLyil`GE>g9w@1I+<wT`D72y<&9Z6>leuk1V5WxN)>{+#SOVe3Wt
zaif3y2(**t!zaII?6!#MFN8$IF{7}tgea?$S?*e5EspE4UbnUi>2TpdnPKz`_41P7
zcJ8tbjk&j)sOZ>yCd*7vbFLehR5mfXsPz5=o?2&ofdAXN-K$7E#@`>a`1l`h<2v(w
z%4=AL^YFog<0L-tB@&rUx6CSsgNI9;2Sk;K%tk%DIj5v$I5}CR)dypf|LH=7c;@JV
z(GOvqrDY5gEGiS$t#G;IP7hbcD{odM6JJOs&Wi0N2u$51J5W$iILUb-hW+I-A;`vk
zS$&r;zJ}qwPP0**8YhK;3p3t^EL(?Y6ULFETKZq|20c911D;W&U<Pqs9;}QXJ9DN5
zbRk3N9*Y;wn`&ZV5y^XwOh_<h#<WY%o@Tnbhe?s_W=Kn-h@29vEFS_PPIUG4XRDLS
z`pL>hF}qrj=jD+%DdKu*2iMbCl~d$8a22;Yu-4h0I!ISrJC~+6=CrgcD@Pz^NB*i^
z4<9+AEZ6z*xum3|oj*TagwSCv-mG{d%;-7bK?Q6=nCaWU|60}2l3SBS`H1@@X-uyf
zsJ>^_HUSv~*?($kG9gom&G_k2U?8!&<-4<S<g@kD3Z~@lh5%J_czkRwLle-lMo{w9
zKx7n-ovpi(t=}%TPLcafvV7pIqG?J<fKKZqpj?S^KwWIAKC|ftIy#4H#;QlV0SCiV
zyiuHO*CQm@JjS%CreL8+mSYfzT1Na4Z?nc+;x*c9&|aUnI|6SD1C-?Ua=Q6#t!HE2
z!dg~&$KH4?>oZKv24bcP3CnbSGzU8NZZ~lmATDgZON5R4l?@S0$m4L8<hsV7vUif9
zFPG=p5D%hYU3QR>nZ8m%;x$eL*|XX8evco9%Mz}=UcX|Vz8G9{q>cyd{3h5I#j2B7
z3d6Zv|Hw|=+tk$bGGFOHb8)lk`D@SC%OWTLc!07Fc3_$29#4@7B~hFbl8C^g9VB+@
z;yZM9KAUbC*wC}7(-Omq&-4Cu2@9D=zCwnoI3p7iQc}8hytmZE!0f7gLQMkGxKUxA
zm6equU!b!st1nX+o0eupol^o~E&*Uign@rEVqaf>{{b6EN~<rrJ93tnDp{ggzu3*E
zW-X%F<N-Q&zP$MCO4W9`QH-fNh~{%BNwXbWdunro_q;T|wGZH;mc69L8XA9y+K#fm
z7LoI|(w^?711v4_sVQ=OBujlsq>{(_XGQj@?x9VlBk%?B6_LXIB3i8uQ(`HBJZ2RW
zk8{|ssJvMzsH!IK(g6l`m=;uQdRJcNGUSDyAR+>e6fupPKnGaZ6g4b+DJRQymn_W@
zUlZ~7Lq(IC*rmz4PHO&ezbkCPPJd_X_lmj;+uz-IaRx{8{6~k{JZw>Kp=4(-JlUu*
zV<4AWKo4p&Z*7T}$66IRiF$#*(Kf5z*JavXvmHZWJoo}6<f2Z`v5H^0kCDD;F+&F7
zFo0RsMZ??qqseOC+FX*st57X4@ao9;?(}cpz7@@Q*WjLj)Z~W0lj9vS^d#z^Zg<m^
z3@D6TH|tdLr`1W{B8;3Bsjm64+KteiXH$_`O6Ad6E?5-tWNCiV_E02+YRxj5Pl<dG
zhJxO?83r_4_xF+3TefCRm;hU_vZ#;{&3}_bAkM`LrDX+0`sCTOk?8G0`O!ag=#VIf
z#QEom$vhg~-dV%!M?{7^09M+1)44|C8kEOyK0bc_Y)bYzN(aX+a1`FLjmlaC4|&|V
zn5y2r_5rg=xv;J};q965G&V30zO(kptoNZa51K|p|3QsDKop!0AL`cseZy$i*So_S
zlQLTL7(t{rF%1vkER|MP(Tz&jfs@#!bLT0Bt)$e=_Y`-oxU!SfSqNNn`UXe0-XG~3
zL9o)Qt*)x7VcfzLQ0I8_jSDAaH=Wv3wLrY4wfEOvr_-QCxQ$ai?dKOFruV?3jFvAC
zGL5wwr<3{fwRAuVHe`<kr-&HxU{LnLs*7U+U#|AD0Pw&Yh~7mL(k`MjMG5K#$Lq5A
zCDJKb6^E0Q$6lX;G@f>YOQZ+>n0eZR*&f`jT~QMP?vB=u>G6$qB{}9TG0Vbk|5h^Y
z{+L2-Pw{Bg9sylO3bWaTO{#zdmv7FV*j%mF8cDv=^!Aw%po1Zl6e{86stURC=Vxxw
zg89sY$wS6RU=iN;R;u>T4HYXHORTo|Fhd+I$}Vy)-Nmo&zLw;k6qW4H4+qo+#<Uh(
z{CTputyP!1xSZ*R0Ln`&Xp{~~W)bhIF|gu?miS6ubrYodLHeDQtx@AXS>#~C0*tDv
zd&}%^?fcr??7$vmSu*EsStzHIa_ak>=^|M=*rrI_l_YmJzz_Jr&C+=G&AvIi?m`|A
zLm_OYwezI>aB)PeHHntda!GCe?^v+0diF;kmO8~<-@M^82$<!}nX#$mPOAMS3A)AJ
zJ%xiH;+qf#$pbTGv(nSj$jTOM=@oMUe@B#WdN@Ar4aBp_ax!*#lwJM$_6=n>;Tt9f
z@|eCCIWe0C&LVpi(_?C<w$HN*o;}o;u~Ehu%{ZT<?YTwr%~P237#(WUGw7pjoY60L
zv#5yR#R(J^zEC6Dkq$_lv-|owIgv8*SZF&29PKQLY4>kcVma3PUI+<~BRhv5w&iI@
zjt`;lmU+#`O9%Is3)7{MW1{27-74ds`3712jtM)a1?oZl6p@-KZo<E^aM6r5Cck*T
zPvX_Q05pUjul{c{e%56qF!R*(CT|`XT}kq$O5%puD&V!sS9NvUEmY3whd-otF*lE3
zsX*1MS7V7nY~d<R-=2i{ERL&xOKzayyu1gBtDA17J!-mLwRupqPxcom3ywSpzfdw|
zkObWa4<0DTOG%4b9lLknaHzU}2;yJuR1Hs5yQNk&xVwCOr8BY#0o_*(?neqcF3rI8
zitEhX-SRa{&6Lbq-kv2=mhmeW#8e`{I;qa*S@;DHT}XB=G5&Olw2&9`MNpt6+iC6v
z|E!4>);SN?UJW;*@QR%=Nx-$pP`TdgXJ&`pU89RRQYyxVKP4w|NS^?7!#cC$cV8|A
z=#ZUhQD9_jEFWGF3BmrEwkJ*rcQZ0ks!-AUfSa1QF3H4VGP7`)=v|L;QUlNgTQgnP
zAy9+um^yNR)ur9;NJ)3JD*ooU>bgb3>Z5&^+qHXn=B%NHMlj9p$L(*e89#C2se&3~
zuAyeEq;@%XC(J=!90lP~y}eI7UOV;Nqa(;%VH^To%dBMqy%jd7_SI_1oF?PMIEn0l
z|19w~;Wht-3MwG1QVb(1iA-OJEV+Hd{WYOSQ|P#-S8^|I2FDYLF{(`Z^y$;oW1VhQ
zH9M_+373y7FN}m;JWc3_`4qXY-ebNaB+5%B#6DkqZsU#qK~p|_7eJh{DE=Nkew_Y-
zEhhhTd<>WMX<R2Cn3~oC`)<Nxzk&Hg4e<33lm`TyGY6u!C=m)Ns5n}7u!21x?1}b#
z!EaJQ79ko>7YmU%lA6n_g8K0f4+PQ{XNjlnd{*R+Lx=9Xy^H8bv?(NFH&Ph}9PP%!
zl(*|AnqO5HBJ=wCNBO!}n{O=6YPWRBTenmJeTwX^h`3$^F3x3-Q&P%zQI^=9HeZ==
z7U?I$1q=42rF~VLSNYEu`U(=e=b}(-dABvgh*Sj7EEpl@`!y~K+#G|z$4PQw{$C*g
zihMy}`=XJ9n>ur}N1+Z7-@h54cC%aOJlbyKM~&K-vF?xlfbkJL_FB>qvi{YH#pk3k
z&k>X(wzl&I7qwHLdMi3~?eFw%P#(V$>7!-F$>RtWOc9U3>N5WP)X9@}7z)*ZRbsXX
z(V`H7PMU|nl6bB54Ei@0#qA7Zad(7m#?JkQ(yk{c8Y{S)U&$h-CClj5_ipWC_Le(=
zuy7(9(H;O@p^Gu^<x7KCjeGy&0=)D&*G&>I_w>LGqfx!0PFbhg(J~l=#6&WGHPHO+
zk?BbJ^kkN!vdk!Zg#SFyK56{NTQQ2%MI?&oKJz*fTxtxZ?f+QRXULE%Jk7BqMreow
zOG&e2OD`|jt0L5cW7i~>*c8wX6x*af-|gV_enrO>Q}bf2X9@Q+ZCbHQ-vlM1*Oe;q
zqLhuOK$=OA3n?-_9UU}zxfd?$Pp9%IaufkRL(5Txb31Z$he;KTJ~&#h?4@3(f(TLf
zK}z$fs!Fe~^F{Fe6xbJSlZP*zBX+g^oS@MysHn2YXlfptgFLZNVU5km%g`i{`S)Gy
z=X@;yNu;o;mb*>|K*tb~rlKhn*f*?VJLbp~7SX8yA~LO81B709nScuF5Hi&Lg_J1q
zp{L}L;$I>c&8~L~JSro}Y?yLrZ3zA4hmRg9%Sp$uxhJcTO!dE)FJ3e>;u2;ho6smQ
zE7M#~lg)#Aq^h#=5=gDKAG5@q;8FC26qJ@$@fSrTLXKfV;N#0gX5h(3E)yh=57Lrm
z;y;#_M=hnY!uavC-V+*f1E>l|ckDJWh#qLJOSc3~v`$O9Pyc(lswkwqvf0`-dqa7O
zNk&WkMp2cB2OJz6OmK>@)M{L^tN8?uZIlju$7A$1`DDFJt?_2Y#-WrLi^&M+c%CkI
zDgiHx-nm5Tg&}2KQ+?T&G=Jne1UZMOH9@UQkjabVv1gIZN)Yq50y(7m+mx#dj3`bU
zI#TJlOMWJa6ijp;OKf*pi)JSXqA>>sw4@D0S(j5xConqm3n|dXw1)lQMC&H2SVkW{
z+-1YFYYQlo#ks#?8ZB;_fxv<x?Q?^~Y>vp4E!pkGN{$GioN)r+oj}_Vlw9x+_0B5o
zTM}orAI|w=cBJId?o&n<3b6XT_#>XkFWBv?6bZ8n^p-lpp}#yQ_kKl1Msd8EwX4Z0
zE$OJ4fQH4f*5h?R@`^<4{q46EGx~X*ep1<hu#3j#)bfW<o)ilj1p74vu7!1S!Y_|(
zr=EVPtndi$cOl<DMedTtrW;o1zxcmPuI!Nx;Kf0IbgMm&D7W@}+{{jAijy|l3Q&jI
zRscC*xR#wsE)agC6(DM8(X)dYok#^#U!H0v{0UrGphdlf9lSi|?ddg<RR<^TzRV>P
z8h7BX2HQY@tlBE+mNyI4@~P5dc~WPON*69qv=rwHt4&zR=NPbsN5C6Qfe1{1q%(MC
zr=WW;4Z2UX*|8&SF5^O7oiy7z@i7_{=<KhHg!%;Bb(k<$L#JPKyLjQwHUdu9(<|b(
zY5RcUgcm%!A1Mj?Iho6|p%)3<S_G~HY9M~>61k~FCb%$~2UVMEh($EQhDWzZ?DB~U
zZ{4=X_@~qOQ&3d&C+ahAcfFgG@_6M=dH;AO9dafM=r*P;!wYKKq<X&ere~_?Y31iL
zG7aiRio9q`FZJs{;Fsw)MbHP&@*QRda3PTNbLS>7^GLJpMTlLmkbB-xr9by;yDH_Q
zWWvIMlUALO2UOCd?|=B{(P_BZodfx$$LD4mu|2gIZ*(#+Fq(A%c5H^%)z$UmBFU)1
z`Mo}cXdLXX2z*;=)^?{cxMmk5{i1oomoEjKy8qyT0YL~Ac~Q0#WwQZTBc}g9P`@yV
zvZRe>lth5Dro>>;pa{7~5*GFAdM%>?Dq0YuP{ASH=iv6kd(-HTpFY5=f3`!S!nA3o
z0DV+FNP-0ja4)N!S66X)?^E^m<|Lh+EG{r%Gys%H_?SO0LZxxvfdhd@_x2a1t$<IC
zmesof?^me*E}1-1Ny(6%y!G1@D>8Ov0Dc2Xl^*0RkCxqQ*Fb8c4C{H@wuU<4{e!Bu
zGrR{4XXfLJY8j;s<f!G_s?RsWKZrhx*J$}?Kt`_7k~#M(7AH@D79z4MbSAD0c1qR|
z@Z?A-^%Ewbt_T2XPx4>H9}ycx)du)uInJi#l8I1C`_*Ca(h~bS$yy3$>|_s5E#R~%
zjU53AVFR@>Nze0{a8v{nn}b$b$NZ2RoL`YAbHbw`E$QR~{@V9ycdNsU^ZFcSN3uE5
zHGpa!%CyFVm(C(00wr?1?H2r(``!`J#{<oO`s^V9yT%)CJF_XS4Fx+lbBhcAd)(y7
zxz2{lGL*l?bh}^&4AL>gdPTzJ%4P^>C5c@ric6O`qc#{@DLhBvG#RaNeBd`qM2O5I
zgD!#Qn5ODmSe&(*WEPV>L;6-vjsNO|cX)EJchMz}uFrV~O@~Q&P4+5W;^K`f^>ki*
zx0>WH#LL#KU8^g|GAbxg;@L~YciRvXR48w_`o+W(Gwei6L0QBU%dXm5Um6BPS@UO^
z)2f-y(3BSbVyamtgK_H^BIM^_)xS@lkvgi|KFW93v^w^9tr2dUEkEH%MmYenr&)BD
zsIh=SGy%df%o>6`fLoN?4I+>@L%)#K`K-@SI@wo!(O=Os3caay1V4qH2_D|(-|ZyP
z5PY%}hCy8)tCVOv$R!CXpZcdQXixeKoTACH0zvXJH7(wjb%DW!{G{v)QUj*)fc4Hy
z)&SUm51{Uy&$x$dmMDuzk96`8<QFN0lH*|Y^k9vH9)hk<+}w47<WczPf=#l2QQb9x
zPx11NRXQ!ZTQ9<FZW<zVW8sH!X2iR~i$zb$&dG_JvXe2PYvvOYpG=xx_sT`}6c7wS
zy*T+u&#ZRpmCq2ZO?*%KFcBM^XJ30Ut1{Bn!bIZ3w$ECzEe)awe$zo!SA_xXLaN*d
zIwA{5(Bi&)nqE{U+EPJH&E@ZRn$>@&`vLf-PhT;*Ve4@orGQ*v+7+p82_Cm4HEjE9
z^EHX)M+(2VTPaMOXhIx5-rW5*aSN*!tz}>>SFAP~eE0GyDfMX4v!WxKmf1kQGoJib
zt)*tkV6~H>R8<_GV*4x9LEE<7;4#xFkLK;4pS{hHRJn+Dsz}b@hE8){DXAtDHHkS@
zD7Hsx2CAU`YjttQ$3x6Vl?|`n7}fUTBsVMT>c}S|WWUSm_!T)65k%GxFY382QTt2J
zM*oZcXFI-JV%iswK1;iPKa2Gu0T%_(cv_w^>tT(HtbPYcjDK8Q+`E*zNkIC>E!WHH
zxlf91&oIK|;0L0!xBQlu_tEV%;+kj&KP)R-MqSc{J;R_TMK`vrBQdpysUtgOW&7>V
z+I8IYe%_@Aj~>y#v4x>{=KOh4P+q#YyvcpG;JML8?K^1DHj8R!qCUkwp|YyJ-n#xP
zc?rte!EEYoKKv_6?q6Vd(OI#Bd!T(UU5DSYL&%E-LrHXN<VjVu>EW9aguIhICAn6>
zT<{K)GdGc!rhj}mzGLr}GJKUs-W1jm6jK%s9?{(^``-9Dh29vfja#ZU2-_)nvM$gU
z2T@DZ*|v(lm+02eWdwiCBd?8${oWEUAa$^87F)&$4jb><^4e;hqNiBSL#L#==^5Yj
ze7YH3UVOaV$sZFMV7p^PYI$t+h3)!&g&Dhc(GV`*`MrKX`a~G6u#!yd>>`j|ip3j_
zBeBGv-3%>jiMDOg<@|~4hIrP*te1B4=Ypy5F3FkzCeAn|zwC%_FKRi^)>u`aGsCs^
zq}ZMzH*@oR?gw$zJaVFF`dajtDBFij?PQ=x`DJWjG0r#3$n^?%Hu55cX#X^U2aZ)%
zHqK~md7tR{-L0iDT&)Og2+Iv`nKNw4+4a+C>))FWh=LGmQ~HNHT>~KWiM(LN-Tfvp
zo9f+C9aqBwiL-1~MX;8*5+U+QFj@m-98(#v5$zlA=cf@+(Pjku8sJczXgWas_CVgj
zrNnw`(Kk0m>l9<)#u8{+;*dj%1e&lYB_Im<mp2shU};5+03{+pD;nd7)&4cA<xyL@
z@^ysx8o@^43Yb1zZ}5E_K+20dM0O34QVC`?rSK>nweQbzsHkHpofJ^!C{_N3CXw#0
z=7NijtlLViH3u_vEzp}@6B&L%5GiV{ro(I!?SczCKI({@;)q|QKe;`xj#lv-3O6H^
zJ`q@x#|7!+1FZ{c0F_RpzF@5<aJ4|rx_<Q0qmV4D5tK{D^c?5F=JpE}{Tl-US0lXn
zg5e^2l_;l_Ii>nM+s*%s=@lo*6A6UqhMReAB+H}+trPX<qX>C2fE$?;(nW!UGsa;$
ze!c34ZgB`ljQT7(YUWrSKrzbACFgTQ{mERmiKp7;SNEgTk^xN&UDGH$YoA>Y2UptZ
z?k>Js;vkfV0hq6gHW3dnK|ig_{5`TWH-~aoqvQrds*;Zk4Gk@AfP!W*p=i4QmoHyR
z043O_6w2Y6w%%L^2X-pd(*JiFg{#izYyo?^Usje{`<*Tupob1y1xz7oMfr56IcINQ
zvu1Q#!z#!P`KePs0sU(kK`UaE)x_9XaDC|pb~-(9!5+HcF`FJaC6%b6Ruf)a$@wvj
zHUU&O$+}tQ_~E<5EVoEJ@Tj8oHo6t>CYqZimiPC*xA}Qx+l>|abJ*<33bB)mp}Tf{
z{rr7ATS6s4u9q|pj;<ttsvN~9JMi`qe?392O!hERhYWao;wD-ur8MAsK0$uLOedeS
z5yc#)gTEYN;xZS9i7)WlkM|MDiXG|d<{MV9kn_nPcwJX5)oyj=FBE%^(#l)?7A^}N
ztZ29Oh@&7iB)VhrHD#V%M%klELH1-)d3us{6bpi6tAE4>pSf_sCcGN}g&!5~Fq2uk
z3=EEQD7SV_YI|s2nTd!Fh`NrBEWckM&OBQBd2EuFsCR!feEx~e1b6hsEozbS7Svu5
z{wUTwel%o_(h)5e%5z~gD0;NKU8iP5Ji4&$^&ru0d7AwmXt;2s@9YrVm;r5Tz*`@u
zl$y(*id*x1cZP~G5n!JHk+}BJnP5cp(;M>_DNLH=lv^fJRmrWMQ<y@91?~9Z#qGA1
zpa9YI26oph>>6EUSs&e4|J<XZ^`~0a&+6yf@dM|f*fl<ynqyr5rP_H$0t=eyVj25v
zXP))si4&LTywvVOcVDokg*U-%7cp9BRi8#bCZ+z*vPR!sKjsN;u2V+K=dRN_P3-S2
zROM>m3kj1--0^KQ9@yHo$Rq`qOt$#dp*o|m%yV2;n;S{)%c7VRl$!Mf{H8v2Ib_kH
z1tW4(Jd+gf0JFzIjE<)20asp1?bWY$@6FHa1j!rnNNrhC99XNO`A9W2TSi(<5YBT@
z9-T7=kQE*l?5rV!Z%DU26tPRf<h$jtv@885rcaqK_1yLQ;JBOJm$~S7*td_i{`$fc
zz#dVI!=?n!e;Sd<<>hXs=F>Y$gr=LPs5*(&3r`kZX7ZX?_tkGR7Fjq>JYI>UdknoU
z$|8|R05TJPm$ap|{yrT?wEYG3KCC{BWzxoxo{r!U;aTCSi;Pupge18~82M=Xn|8ZJ
zZK~o}96{aU2T&f4d5vRMhb$LYwbishN&-hj#5Vw3=oR9(yhc~zYq0}UZrn9XZmS(S
zg~~Dp(Sl|!8YYY<DXM;^I2HasN;YS}X3|l;^QtDI=~sKQl5J19;ystxub;;nxqzLY
z0xJea_e@y*cZS#L106>kVX1(yV3lnjSN~V}DzWZ0k`Jcu{m!tB8>Z3cGr|;goCzZv
zo97F%Xv;5C`}SZ-BiX05%iDkR&>A$!T_46nzn^H6maM<3E)enqAU}?L^vI?^P9OO`
zAXn+X$1#}%E26lP0<M-)Vr00+A~EHxcX6&YMFt7+u>?>HYs?jt-~3YHC->P*2Q5Sa
z_KH@dOi&F|4{c9%xjIEZGBTIh>R4oNuN^d(R8<R6p+9j6t09N!F!Ei3BwYy91+7SW
z1x;imDNm;EkrK4UyjRKPPNLuClXAH!?8<T5vZpj91mCOsZ*BQN?j7naLXZxzEec>z
zA*oD{98zGD#2P~ji)W*(b1ECg!kxy@TxD$mkYKE4*SOfYtS8|)KNZ4B9XJv_$_S&Q
z-e3VDKlAYLNI&<<&?y_RzBZw8e1Jd2rx;ZMZSXO@PS2E2cPI>Sz1`B)?xtO@e|GO|
z@U|U0dh}(3VRmzd`ig~MTGj>4hQ;?=bK>a}i>{`?@0T{n>r5|qIU!JMJZHxkaZV(5
z1M%^QrR)=Ij~J^PhWO#bhYsZtA7V&7&K^C`{{`7X3K^bYaU>_t0OhU$)Qy?;niua6
z>IRQ4`slXVLm!`?Wro)i90ozBCV0us-H}|ZN?S+p9-w;_K_Wtn_uSSScNhnDpg}3p
zS}2vGlG{+KOc8ACvyIJjbl?OHhM2A(FF#*IGg=v9g3ac04jKr{lFHoz36_Njyb*2|
zf!_$YKRQ!vyB<DV9WSa9+lpKa*`itcLf3IgYXn=FW~d;&(Wkckyg|L-BMvJP*h>&t
zShCo6;6V6iS~YNaiZaEN2#8YTrWTUV2jrae1OB%>)9+_Xs$lGKF3eJGD_jcn9#T}P
zzNb;QCaKKvytCNLf=9f8Jvra)PK-Sz&oMA3h2%dJWBxy`CqHeeqnJ3!<K!P!l_@Cr
z90yE|P5byTY859kvwDrHBp_DQw4}d2?=o6{T3*ck5W|<$g{!qh;>2VVp0MUsB#ic`
z&WWaUUQlA&QD}-L6GbKDG~@oGbU=NK92}zPbg!5($kI<>7EoN95wX8QY<(kgCpw*r
z*pOxg>Bq0Jeu5~bw#~i}RVVT(@uZ8$vdb|vK{7y1D>(ICO0<@c{xPJvV~9YD0Heb~
zLq*lrOl*jOP>$W0_{<?1<5^EpJUhmvo03Hq&U&AAiIYV;F`-uaDMS-|Wg*yb5TebX
zFSbN-f!eEVdA(Rr{%I!o|M-L)hhigS76%XpL7)U_Z=x@40YM{3(4`nr(7Q0(f{uDJ
zjXO_qttKU|w0)4QFQkUzlN$XtW|MbsAqv+79~f%5eEEw}>lPM#`m4{9H(0jW`t?_A
zJey32jdqZFn@L<F=(^G2n@5o4sjS+`QSsS=6CkH~ShacF|C*Pt)@l7|ol<pr8PviU
ze;W^x5?Eec{s$1HG<maN(idf>l)sZz$+%dBCY%`wk_t-KqI)W;oO>$`NO%;#Rd0&_
zUIJg2&RHB!tOkMilSR$%Z|I_8A4_s7j*yQ3Pjfs&#4zPfn{3eC@&Ytz;+D=F%?}m8
z)ytl{>^`WPS#8?KmeMF1OUCKT+iXO+Cnz3q(HIy+5Q3p)nr^?|H=-wpi$&)?P%uUQ
zK?;Lw@PrJs$z;z5wf{Xkzb`v}cFRZq3}a|Z_#|r)dHj@#CE_I_(-Sexjf_ck`{ye^
z`CD&}&U$xDUO<#+-I9fQhJt2BKzuf-WT8tUfdg9P<1<s{aU6>f|BMiYlz-9iw`5#!
z{9=S&kQ|jjJ{Ybb_458j3D{hvxGRzh-_)@eb&D~67}4;{6Z&V5syK~F3YG`F6#sLB
z^>ToH8bwLSQGoa5XksK56zDEqT)1TC$36nZY0kF&X;{5QYvIB=++PuRbtEK~S~S|D
zKc_a!(FmbDRlxh?Cw+Mg0M5EVUGR{!&_QnqWB%{}z|}}J%vbGi2#d&WsXM*QDNn%L
zqK82(87Z3t>jwAu=5nBupG%EzKaPF#?ePYM4$wa=Q(Pi1cz(8^>@G?EVOYHTCvB3b
zxg)-LiZn<|YroZ`k_9;ka9M}946F^P!~JKRjITB-N&KW}S|==p!N*m;bKIU;w7u@N
zi{az|=*J1(UXKtBB->1;1;*5Ve;=?LMrC#S6QoWWTT4^F1dWYu?fK%pbzk%LBj&4D
z_UK|a+`m`;!!w<Y)U_0p9}V8^->WtE(c;M`;{sGowJP^1FU&MQbtrAG`!K0ZniC8=
z8!Vo8FQ~6p`xOK7TUu_vMf=+9Y15wy?@PK>&OVUyZf4H41NW*woy)Ozbi7MSu=myQ
z+zThQb#?O(<~(>H)UWow3M$1jv%PVg!*@!!6&pADvfgP}mwlRF4gLGgW-Hv-xznbZ
zDo>|;mXbt7Mb*B#GVIa4dkeH}lvPw-k|x<XId#mv*}J#hl6*wDF3}Nksb{FE5A2!w
zR#tz{JHH6=suKegk6~WHF)`gqG-SsY<5NF~)el63C2S(-I-tt~4aqUxFtN{tojeat
zED2p={YELHF!~r{1s{@G<0(y)MuEle!>j|Jyzfu`S~Avt6pKqXY=|+?)x2n#OD9jP
zPg%cVLmn=)eTNP&36;q%HGaA8(0O6ev)~rVWf4+=CjMX|?JnDI*qDVxYMj=&NtNvy
zqmRcqUeOKPy?FFpIy=fX($<dEw#m*?M23D%fVsQD)$@Aq8Z%~$(B=`$ZcC+KOQsHX
zTNG*R@ODAi+7&DQ0N<Ni^_#)g-h&2>Q~lh~@JD?76d0^SzYc8rZem^hVJF(jA(hzg
z3np<aSIx?#<16Rm;}a1Rvo9=c)M}-(9}?-iN^+A2%|@l%<t=mnG+#UE$an($^<=+<
z7jiQbCQNwo{=F{<iKJKg)eFatN!TDj!$w+CTUXZrti1}v?dtXG+atEz*p=9KPAeTl
z6ny{qL$(4Tu-!>X)1eL^&hhQ(_u@aJmZn>NS#hG;ch96x<A)EIaB}HK{ALqKw_I|I
zw=6aL@AqAvD28P18KtEonsMER&Ng=`IeRlPh!%=(M8w!fY%9W*&SN(P)XD?w9KQ49
zNe@D>5CH}r+D8npgoPTS+C=~ADJyS<iWA+`K0;9W=GT9^ioHCv(C>WQW}Owb#Ak=+
zN-36K9W-sJF`CuNZ=zMBxn;}sTr&%cL84cms7UK&J$mdWCMM2Xwrn4w=JZQ%#bZ8u
zA{HQWM+`*-RBSdd=XX5J3m>~!8H_xxonjSJnw!h4@d;<S6LRP8xg9aH9fNoJt=!l%
zgVw+IQ9V7q1zB2JTD}RT(~!J_R@|va4<W_~b!7S4wU_fYy1UN+zLr6p9x_YcSzF)D
zcXS+&8*j+1zT-GevA?UyRhjf*VTR-6dv})aeJI!M%RB1`V+Di2KRr9AQYMDvU4A|<
zp|bkOpVhn0%ykW}`vLjq7U14-G}?Hc(5pts%eN<TNM>Xl>~^TLW_~T`(Vfxp4$Z~8
zKQ}ckV9%382o&<NSTwVr^;i(BmOD5c{&S{`7gbbyo%tHZVyi8c##cs3mM&d7Jf)zX
ztA;_|>*qje3F@h0eGeT4#99v8ROW|k-JBbLJ}=DeC%{_X)2E%dJqAMzY~|jclOCRy
z31;_Wa@+cy-#t2lnY^I*h=`7s7IJBBH}191yDmhqB*@H-vASx;Lpk^6DXX{)Bl|Lb
ztg&oapxuJEq@2#|K70A}_HC)m%uKdDYU}81+dSrjgfk%2uE?VILVgUMLpD0{PvDp{
zN^82n+B-y85xY|Oh7~JTNV@HswA!7GXc8}CK-Vo=dSQkyc|cgg9uAx&2pbTKEw_K3
zu)o`&-~Sql9^w0&bSD%%dPv~&b&w>drgm+Sx<xTCe-ryTu;Y@{oY*&0GZ^;u0>7I#
z$$#GH6Tz1+cce1ub2I1Lg4Hu7OwpB%ijK}JEZhsG>|NRv(;+5yU(98jX1RX-Mx~5=
z3!Ww&WYM$Y!2=z0huHo@hxRZsGBWVo@SjP~y0obN2T-~D_U&CCXe5FZuc9L&n-lr-
z2G8@&>>j^Nc%GlO{2hk_10JBu;=`)M<y3?`=y@d5LQa-lk?8Li=iVKeuY;VvL};6a
z)GrPQ2>3vm@4b>IDFnml6FSPPGIi*Mx9KG8c_1Vb`jQEw7ho$xtL^P*g6=hp+>3xr
zgthhatCuJ(wB>t;vQ`Ese9+f-Ky1yxW~=)js7^Z{n!kJ0qSJOK-?DFbu-M3inFi~%
z+{6=}8Ed+%o!<GsQ{5Xvif<Ma6fC$H%teUhCTQJ=Uk++fRgl0VKm57@--~UH^_UxJ
z+-1v_PcomwyL!I%KXs~~p#9>dSE4>FR=2$g7+*c$_)F&wBkY&%n!yfubvvVy-A?P*
z?<EuzXSrVK5c9m#GCj`$0nK|B#gNfRf*M|`+T{+<>%Ai=-8d5(*dcs(aBOT30$TW{
z8pDwpyiG5srpBUp>zXLHoz~WJP8BzOpn=(Wrm<Yt!O>Br+5f?6-^LO8ZQ?kM%5)c2
zqK<^c>_e3f)`0~1EOy1;DJ_*{v-{*JQ#5>srr2Ii4sX}Pbli%S@-?boJ3DW3-DfpP
zCimyABC+fUuNxAhdhmh_2smEs4z5tF$%5?@bmPV#s_Wb2H_HDM<mZc3ZuA^S&YU@j
zENwo{6q@4lHETxc>R&<YPDoKf{}c<Y(8zp&qkW1rDOb(D$7f^QY%W~5aP{idTbL$G
z+-IkuSR&R`L5C`KXCdE_fm7O%r%hH;>K$b~r)#z5{s9P!L(p8w{s?hhWKp@W4#ciy
zwEOh+^|kf%`u($N)v9<#NQHJVX?drvUH4c`8t_>ut%CAmPlvAkm($3Ih>KfDYW*91
zza6avb$2@3Qy8NnNe~|epFmSrw+qO5Ps!xz(-(>zNbkzK;=fs_rO9^ji4(gkon7ow
zZu0x_;cOabAX5X4Hjk#-eU!!K%a+ys@OU#TLu3uqaKoohA27_b@wnE^)lU?Ezq@?*
zG3AK&o72>$G)+znnpK=-I_5i5qy||jpBzxh!TIo>>>M2O?X8*%@0X1EbG&uY6JRKr
zpc@O-*LkWaD+h&z?b~{(O~EQ}%ScwoQaX%9ql~7>RoL=E>Nhbl(az3J-N>j<X=$n0
zBcIoy_{q}%oe#_EcO8=*t#)wM^k_EC`clWd;0TtaEbN<C5OVX+kPW+(o1WW!8CKC6
zH~;tBB-3KI`jFY~y@Yf+j!Mf%b{{G2imH>A^XIpx`M95d6HDhMr5o2idH7IMORFO%
zcJEX0XEap@!au!BFx=;c8Rl7)tovb|-lMjfhYD6T>ndx|uUMz{=eC`kobEq<yny~p
z@lAUC$1Yy<C4Cvu@Nq`Yz6%?v{PwU3MeI^wuacCV;*!D0={D7s+`V&0^I-Q*ozy)%
zRK?m&u__k!+-<;M*2JN?JJDkoG~ktVneIP%qQ&-5iX8>rUQ}slO`Au|UF+m@lyM#*
zj~LS6i5ZA}p$1k~N0?f<`{+?O*iZ`z#$tLXU?JR=8M}VYMg)1pDcuH&PPenecbxpD
zf?_R7o-ctsW&2f1c**mQN7(!&JFHRLa@n#x_*IY2&l<p{P%_WR*t2H`5tZjx%}KOK
zbDtvAjiLVvg#o*qN00sj+KgxKqrta4bVXQ7_z#*qf58iOCLW`<C=C|&mF~*?obux0
zj-+<1NePOI(yiW<_~)ORqfgxa`|lysH$E?nIL?GXPqfVT@9&tt@{iUf=2R<{K<TxO
zDw5JZp`bIM=@KNpgwYD1pisJ`cz{IgRAf)-8M_?Kf0L7DV}(;s4s{Kwz6~5KRPP7{
z#2HuF?eCw@F*Y^5^VeVP*)U^0cdMRWcVf-UH*Y%4nw1)^x|M%vARgGil`F~P(V&IC
z8vqMGHGA@tW1Qx2!$|RYU`jIJz;|x$DMbzp?N=#M7+Ia_qq4HHOU=w=SlWk+_o?$H
zZ(_Z{z!}cV!#iIK2#_r-F8=3~nayw3C55k9bbB=^X~3xIO<~zTQnpn07&2rCdWpgD
zOI9vCXFGpMPow(K+fSbkL{E|4*?`h+23?zZ(Iw`mZ*KdvmQU)^S?j-;H|$ho<T!%x
zZYUK#?O&3BBz_KiKK$>pvhVRezcjNBHfDLBZVZXPc1@F|mi`Ykf;>AgnBuQbSs;Yz
zAj@#NS_`X>?%6Ys|C{}pRET2n0ys*oa<RDmF~_yp`PjqOwnG2J9`)~<NSIelI<ZHx
z+}^(P^yx`5H^b3wd{I?ZR8@Ne*1>IfUD;GbH}4f*l9!Pxe_aBZd{bTVL|l^Fw;gRa
zx#)<WDfEsRp3Mmnk9#Esp}lt%DNH1zf+He2iPfDp9yOy7a<7{+$<yIgU7zm}!u~jD
z)Iwc7;p%jw=GP=@=N1=_KCGyyc)(a;UwC-+s=?B*9L0~p*22=#!^Op>k1nbWD=I1y
ztI)sn9_6tZOBY*w`S6kCfq)fL8d<bHcX2O2Q;bdceX%JLMu+sROC`H^F5RA0`la?M
zUEdjOMxm{Nj_-^a11UIBM~&+hU6<S#c|2~(6^}GIv|O!<HhigQ=}me<h6nIqU`r6c
zMZ>lMUS8+;i4!N1|7*aE(%<^&*LAY~dJpaK6?Dp%MjD@)y-kVvf$nqULk5UFeyHhn
zKBe0Mbs;ui9@;L`d(Nt4nGyCyoGFI7&YV5_lFu#rTGgWLs4>287qR~e2X*z@wOVCU
z|4lVUBW)`@nte5w*;0@p&YOkjl_bQiw?Y&gO7#>b$Hu4*{dItT4;f&M1}Q#~l@ohj
zyjVW3=K<}d=H>@E-;V-Ry0SrV|Ink)se<A`E8N}PUl1}x#t?j9{(QN~8rnTwUfhxr
z%TgiY$T1)3^ya0@EUHJO+irnlA3J{hkJ?%Xp=b{patov0_~zw*&wzjCZ8_XovzBD7
zFIR?Ieb~b1S5GIDvA&p*%VE_1zdd>KgsNDJwV-eb_TVu1%eOF9IA`<uLkh8r*{K-G
z|M_|I>{j~`u2ErO9ca4jK7U@$alYCJ=0aLOfR;&17VGH+8!I5L%m(v4_-L`oNaFnI
zwVrmEA+d4r=H#|q<=1~To?Es1*qcE@*p+riCGL;da3DN<3_Ur1f7P4d|NCCL?(h7s
zU;6cML!9=ijBHzl^xws${5HDN|F3^=yj8mM$9A88FCG~xkQto?`txHom+$@`;qmTU

literal 0
HcmV?d00001

diff --git a/doc/img/network.pptx b/.github/img/network.pptx
similarity index 50%
rename from doc/img/network.pptx
rename to .github/img/network.pptx
index ef999e5b44f2de5488e5771ed6407b8d66f45e39..e78394680c38c1bcf4aeb040d09d0db63a844c34 100644
GIT binary patch
delta 21120
zcmYJZV{m5Q6D=Iuwrv{|+jcUs%}Mga_7mH-?POw`6HRRUo!|exb#Lttr%vy#+I>Ep
z)oZWS<ul+bGvMgTa^MgcAW$GMARr*5AOkxDodXT%G+=;nhd~yUu<MjJq{KQ**16P$
zsQ!Y<3V791@b+;K2g1!p#L%{D3Q%#CH8|7`@y6fZ!g;{kBuz^+ct?a$7fNu|`>J7Z
z+{%baksm@nN-0#g)uwDeMP*>dFc?moXJ>VLDxl5_bT<MJauz+4H=JNtn|c>WSQoUY
zEReD+esBWd`_@6MdUWmQlx3h7H1zjHHfo?xni8FM4j{VN(HNnJyJklzE5HU1#Pc2&
z^+O4d>FdA2OjgQ!ISwZme)Nt8)YIdYw)`P=#VIqzt?3S&uzM1sTV(YZYuWyhBECYM
ze!6qhgi92-YChW2h$E#^L%0mZaJJC!Wl=_;zI+Y1M<K(E`<eVQY0>bq_+UDE<coH0
zT-`r|Z9$m2=2GNg*GPGtlWxH=;&dUVEDS4Bc#2nIkBXUu=SYs@`a&CGY|<UuzvlqC
zet*LzZQp8PmIt=8t=>Ox*<;qD_9)m?f!CpfXc7B!MfC5dQRz?3-zM^DlxCfDq>Jvb
z8F-WeW}UR8YE2>XZIot}8j~>Yl{o}Qr{mGwzTofdQ+4B!orUu-F$Y<*O%jsLLS73A
zTm9q_E-s4?fU-|`pb9=l6D$M=c+)MG8$?q#-4YnkVhg3o9bXC=-sil;E*cpGWKIPX
z=!B06G=QcDFz|{YNOf)C2=T0|IY41WAW{S%LxCNdK?uDhe2_}L6EED8NXRbfb&*zd
zMu^#z5^FsA8Eh{b+<!rb7Zp5L8_7d?w=M>HTK_(`-0tQcd-CAR6H<&g^#Za@`<$}O
zzVYW^f65#O|LXe~c|W)uuRrc6YZ~|HTwG_KynR{$tQ2?c5|;oy@12ERzc;6&RdhS+
zMs_sor1QE|rG^+KD>m>-f~yyWjP(D$`g(fKT}pXB8!ydHlqM7k%ge3ajei*nWTbNS
z%3B{43OkQ1DKT@L=AP?3&(~}DDUQZR`}cKD<oapH5XTZ@JoW+JFGThbeH0A|ndPJu
z8FFd>PsEq@kE<6tehu#Xw>_ukoqp|lF_lWI+}BQJ2O-~uZ(SOnE5Zyt4)><dFRNEy
z)N)ssSHI=1b}}{VIkl|aGvc11w~L(^^PVnW(trl)i%h*=Z;rVa&5T{bv}N`JG*J9G
z%t;}TY>E&vTbQUEC=?N)l&}X>7JRU{T1J`xrQnS#tWVu5;^_<MCEA==sBN;m!>*&Q
zZQmt2fQ^`<0f{br<;DwSk#2>uE?B{n42UQiy@I@N9(?6s0yI>woy7izz-RTZhUF)Z
z(tZh@V)(nxJkLDEZU)b@>gHW5&vCWbyQ*1h&PNUUYH60dQyLPH;lPOgkYOic>B4;g
zk9;~xwg_c4Ys96>PS*IiUtf0i=<B^d>kKrfVnA!5Pb!ip9|6wlDBZy33iln4gOz=)
z#-?7E;Ccf(7rb%Bes!xPiz;4+RqvK6T|t8^Ev8y?!}LLr?(2+!q;&>YVJBNn@+r`y
zuXW*Wy#CWEwVR)Y>S*Z08*!aFZe{~;{uXzZD0LKV@?Esk!>mp5Nn#wOY|Vj6S~je0
z8)+UHNKu%5b~8EQk401kD7Uex`l!cw$V5W)F%f#morQ453a~Q+)AABmxAP_YVYMH;
z67r5(B!H6e;sTZL>qWm-fe1#k4Aq?9vY!fSwi0X2UDiVn@3`wOuSRY~KA{VUJ~Byo
zHdNPA+OnJuJbiLfT^X$Gi#ZooWIlD8MdaUZJn_N=T0S@Bdu@=>Z`YRSv~Px4a}qqV
z-<=l4{VMG?0su0By8>=Qi^QIMqdF<8_!vj5)Tz%qxT)8%yX*A3Rapkyy-W>xk3_{1
zZ%&#{dUMr(S_?}c9hYnB;IaXajTf4N<DKu`^A9B*HK?O(28%G|uw-%w!>Fh#{RHSd
zPj?z)q4QL1Q+a2dV|lx|wS?l?=7AiD?u^Us>uWb>RUSc=Ia)OuD^??ql@(6}h#(B*
z{pZ*t)ya1|=UU8;`P=B`bzMG?Kuwp43o4RlkVO7?{XHw(bf}irF?)dfuZOg2?B>oD
z8_<bROkJy<C!5!IZv0{_aO3tYcaHnC77|fEEL5RBQdEaElHUlx1u)0sQd@`)>c?T1
z(ulKD86kp)q=EX}P{pVo>b-ia&J2p=)}}w9n{@D2mI=>shF{IR=|@X=1wUNe47sMg
zK_zI_Qiq!9&-rEjpO=AO@PvTQihGyJ!xqh&m7ofnKzvL1s$xu({tmSBE*h@oi9BvA
zxBX<@`k*L{g8}-gqrN&r`#9a|Bn_{t!ug6w&=%(coo<J(U--_32<%3T7j}P#W_%=u
zW<-Z0IjBfhruA}?Lb}*Q6Xs%X)LpoD_wa^hE^RU)$$6U3SGke-<-`zE_AJA424v5Y
z8-<)Lf%yoy0B6fm9N+}svH{suH<4gr{txGYnGE<ko%rIGR9EYZ(g$e}%q2R%-17!Q
z!F_dSmh6nlhm+Q4M+WDLow2va?%9x7fvUu{=C*<j{;ncu1&pj#Z-uU=u6E`3HI9(`
z(TfeSOVh=<%bk9~EUUsb4uUdni#lc-;;gWP2dFD7fE4;jDQP|TO*u#=4HsfyPoSm5
zoOpvZPt#8Ny-2yA-3FE8R!0-{)ErX1HP27S9bAcFPA){}9&c!M)YUbP;m5y(&XPx_
zi%z7z;K<_#tW4q<L4<0|OV)@NzGsk4g%#NCw*yus(r6tUepFyTKWIqQ=Sk+aWY}BN
zeF_XJ05x$;u0DDd*dfkEXc1J8K`Hw6_b$vJAi&E=uc@AZ4s4cZu)l`=Tkz+H&3TFz
zdXbgFa+X;hU7yuv?$lw>Me+BBRae%jn{UTwEfSBjHbSqL&Bc@zPFa;&%i9tK`N}BL
z_@?aXi{P-Jp+O9hT)|vTq`_mOW=SHVKMMYk0D|@h_RJ93aC=f<Fi&1lFv~GxRmN#7
zoEdob0wKWyLC|oF&}zY88y12<hj26wYFO$wXkZn<;<A31H-jvXmg6W}vq*-r(#5El
z#K4)Vd-6z9Dtg=5b^Nxtnj=6Bb<qmBqY4a(W^~-L_?H(FAS&W%x@ZT5jd^`R;_~|^
zKwTP!27^Gy`77{WRg$GxJ9_X{Z~MBqH*;|2(#1NMQYb>Kfl9E*`c(nvFc`VHnYp=W
znebH_gg-h%$C5cx70A`(@xl^mW8_WD;K<c*GCGG6X(Qp!Xp+0=F(4MgBA`vxr4eZ`
zC?J|~n{frCLR;hiD<W0=cI#HetOf8>9Uk_}ScCmTC)LTpltkc$aehU^PG3w~UX0QV
zL`XW0wbFt#h26GzrPLdDf7(<%h@L%G1vdZRx8f8H;$V)Jvq-v!o2pN<+PYa?CwVE;
zSg~i(6UAtfa^@>1pv+m!S+MB-JG3=^U*tC5%gI>h5(X>y6$Z<8v~D@t7*K%|lK@KH
z5EbxJ5D}q5O1=V7a3l{ECL3eSIaw`=I}+Q{<~AW~D_fK34k0J&{)VZcsv;?Y-j5ch
z5)I0LU~B>c53a2RV{Fn6{mfk8#Ucj&7pX6&1Kfqig(eD`7lYR=vg)!F5hEQclpu=c
z2a?3U)k7B!b~Fgb=ysn3Ks{6z2mLSE<wS#L2?PhpIIBo<tRMpHQ2F1QtC~7O{0BtX
z|1V;jJ=rWtlR#i9`+sw0j|`~O8wqCC6T%3h2%`2y{+%HhEG8k5pMC1uN=qJ27GCI;
zAeJbW=2~9RKfqgfe9<8VCjlpMO9BWb3uP)(u7Oljr3=G|BKh%O7_NscGzerSr9S{I
z2!s~n-+I6k@wOR&^VvD|{;w_!<p1h*L28j@TaAy#65CbGWo%8bJiqW|{N5GN6EYu#
ztAJ649HtUQKM#@qsW2oyEYe>S@7-!a-fW`2VNq~0>fIW^rod|tK(lAH|51L=X5RuX
zM71#}u9t2Fk8W>L$e7LlXAATgiN{kryhH4n<rjxT_GW#j0H+?y;`-pca4(<FYmVo`
zDzX39D)QA?Q+VDl`MQ&fz^DM5BkJKa6={Zu1PH!A6sUA7vQ$*PZmBz(40?;WKMszW
zB5YLwrPQBROuQ{yfF%bGoa5-6n?lAE$*R#v5C8Jw8f-uPxi#K);-~vWErScUu;6Ov
z;6=;Y62_<vJ)6Z@y6XVk3Vi>b85YjqDY7IB5PoDC1=~Wm#$ACWSFKC-p=vBWa@uDd
z75T-(@K1xXCL7iHW^Wz#vc0UgtFu}It=)@(Q4^l$fu}EBz*|?>r-OG&xTVq(=kfaO
zP?1O~_9Pq)vLz??Gk1@pj{)u)h6nau{H(?K<jrp__a~$&EurwShLS9ZDrxZ<NX|?e
z867f$*f1&|_LqB)@Zj6)(v@b2NPRXWu!0EXZaRz+Seo9ir{(!P|7y*7i~I*3BNg#X
znt2X5np;0604(~T=TM<+#7c~0W(Gyzi%mqekJe<hp-5Iu*7W#Qr<2B(RprHc<PNXX
z>dTYDh@__V!%X;jF~h>Zjz8{Soxcb)%$aX12;IwXOU5{sf_1*E9v<3Q;K$a;?=Tlx
zpn3*SYq7M|MD=30utf6-_5JO<G9);+V&l^3R#*~}0NbiB;(j6U896DVVfOu3tXl@f
zx2ee+2F+!6EV9Fmf<ULs{EM$seajj|SO{^FeQlYV-It|h<!N3G%ea1o>c$nr0Tan#
zvY)e(7v}o54VASmvB_Brqo67pmnIYxtV`k{xR`3hL35)&HT2KDja{9BStZQK8|&gP
zWZ=&!0qMCl1i~Jzn1|uH+Uv_dE9}z8B7P9y?=u*pqYr0NpardEvT)6thi}Ou8DcJ$
zBknJa{TisJZw7%Zlq^F@YE;W@M2V7Hbm&%2y+7T{XYbeHna+d$%}{4N&*5mZMSUQ;
zb5BNi?bax8?%7>P5ccY1;c5Dv?<Ubq$^FcX3?N??#;$EHgDx7N`5wyOX`K&=+fbfo
za9avX`U@_Jl=n<9Brq8copN{XC1qe~aHY>;BXeshUYa^w6vuBgc>o6AR>rVZeWnNR
zo*z7i9fF^g5E0eh==3b5cDf~-LF2x@t<K%#M2{J~GT6bHoY(4PnttTDG$^cy+ymW#
z52$JE(WrNlvGiYW`xbgT26Pu)_kF$oPOIyCJ#Bi=`!(~WwYY%`j>&?KqLrfLdqCH^
zrMj<noE^%MEOew*|3tspj<TFZJq!2FtGe2+3oWN>Jtff)_T_+!Vf`EwFeI_FberDm
zv|6~SCI(`V3uKx_o8T#hWuE>=1REAl0jUz_DNNhqiAc|l9a*)(C+GX=dEut(7iHQ5
zB`D=sq6`T%aHd*a#>4eARD4M7np<T-yL@pyWXx3zHr1K(9H}dVnXpL7Ew*GB)n<!1
zEbPX56`KkTSG&pMr~I@iYcLnw;1}zI<4ce!^j)%O>@$qfktMq5xG*F_PLeq0fQ>MY
zMh;fT5lBdI42(MG>T;F2ZaxctPk!(UYjCq8D*YzkX8P+OriM|y8(HY}mwdOrI4s|V
zQNvxUk-oF|+=pL+mWq9bs*jijWJ)3>>lPhBmPU5vR>r<)ui5D2kko<6+1isFV>I$1
zKK#n46MT*>88te`6@xOx+DLF4fRRCdZk+?OI%%m~{MiZrz0VzVdRFr+$R<mR-pa&t
zn0*wtewgJ3$}o6|YL?8ov5*EA^fH$A3#w_`^w;&c+@lRH2O<{RnCe-}DO^GA$&}tA
z$!m`F$-G~0_Uk^Mx9umZy&H`{PZ48G_-!*`kVF(z9fc2ZnIvf@@e^$h-ox96!);pp
z-q-`re2xIs1un~)RHD9fA-pFWH9{=nvIC{uWj5Co-hG(!yiv>_nR@Y#6>nF$@5DeA
zRm>)28y-;LsPhsK*@g|M=kTYA$(|7uSicVkv~`^LUy$GW50J^88i;HL2~@M21y1av
zG)XuGf&vi_;ecMgA59LvP@q65+a*NlAG%D(5Fj9|C?FtYzyfCs;G8x*5XlS+lY<>k
zMs7u9RrQRD)Z?vIR#z|9AqB{5Mhob(V&ZfCJN5d_!QScNmsGOAQ+rlC2Hn`1Kc<=;
zE|Gj#1^zk{g^Zw6X`LV?qI-Dz!!X|V^=*Hdes#h>M^bLUhqVyxK5_Nan50fc_vu#;
zpMG>xz9G#7kL7ikL<(dwXVtQyI^=WuHN&Nd1ZfY0v`Yt}QP1qlPagTfHAcXo3f6J6
zv+9WT51o%kUDcJGwmZM>6b*fL?w=3i2`4_|nv-Z^T4<T0nnX47W-@D9rcaV7*g9~7
zM;u9aB&LK-)4`|ln+q4~PY-E%C8uAW3wuM37V)0Q#KbU*o%$8~>8p3Q;Tj$KJCz~f
z_y>9^$>Lk4=Q39<%)0FJQ@;U)3;dkEEmcsT<Mk6><7^ZPP;BTfIl{lXnRAtC?f8bg
z9os$b5jLHs`G-1J73TQ4XpJ<`x3VS!w_Oye!j2V&oBmzn8&y&e4&Qr`?WgZqZ*&li
zM@IXf>S28&5RlJ$!BM^YrK*aFH5=wGipouuC>>97C`~k$MiQ7pBvt@Hi?_dvHJ#}>
z<DZ074G!OWAraKrK}FV}!bCChPlPxywbTgoyyvKbMz&IZV)l4!+k^3moVirMT(77a
zvJABM#`;TP#g=z7tm^L2SSqAO(V3`oy|fhJsi|Zff64D568==PrjQU<EAU7HmCh&K
zakb|Z(?B%K)@*k+csi9DW-4?SdSlmQGxrt;=S32-Cx?HtR!BSb!Hbq0CnhV8!rdh_
zm(ldig=t+3xTf)hzqz&7fm7|I*l<R=V&IND*>}4}{Wb=cSdan=)s%_+3Kz~lFW2v}
z7xq@dAed<j<3t5=+NtU}%|O6Sh}YJZ0v)ZDTToOq6#Bfxd11<6%=Z-j2G5GnvHjuy
zvFsi)PYLagG-s_@XP6=6&9`aFOeC0d>7d`k_N(2!n|J%;Q{2_ov-Nee``G9od@q$5
z31(f|6GIaisP+TwFL--{P*~h|A)-PVjdid<g&HHQwa3`q@ApCDapoS=>?lR}F6JEK
z`r8^A6;BpndLcVWyI3QZb)}sZ;Me;lu^^~MiC74Ypj!(zhe?R$BR})IBE&&st&>Ll
z8~G~tv&iWv1hBhQb5y8fgFjA+%cA8$O+{4qf*__td0hYuk^>S%+i;bCo_PBb0HReM
z;^$w;i#2(M#C8PGSVpSGg0sp5lg@$1cy!9BTmhjd5ViL0%8PczrqBC<gbS6VPthvM
z^X+H(pIL~n0&`#S6bVJ-Z?|=eIcZHf*o)U~C8zgN;Yv!`3mZq~V0>^ZAkCPV#<2=^
zq~S)i1qXn4ys+Ur=3Lk(M-S8FoJh#-h@PoHEEPKs&iP2Bh-lV&jziFW>rT_AnrYYN
zqMm9+AkB=9+xIxulg&XqvZ4Kl_k30wO{Z|XBD0EZMUy@7$8Y1!jtM4<HFIqCNou?s
z&X5LE{M29pMT3fc?h4bKYbgVrt>~&o>EXJ%%4X(2vBAl^rE8FL;>C+?fwn{p)D{`X
zVJX&+GRivy@SSV=F-THwDcL8ZZVrQKJD<Q?D;z|qw8rX&f0(8IPiy>tm_=lS36R5!
zb)fQ?Yf0{Gh|KA&+R)K!aUh&_SV06`<TJM0K-+*8UP}mywx(SQ`LUg(1pIyR#_uMH
zUFQhOKaTYe)-rirXFdR~lAWG_1S(NtAlpn6uS!Ou1gNFTPGr+E@9n%Sh`b7ob-Mc?
z#q3RXd!XFy*T=(B&(BtQKKMOJz)lfZQ({WfUUs{rH7mW0d2`reefBtfc7x}3b}^~B
zLL;5K^h^4>B0B6uBU6b<cI33L3-Hyu4PR<)vdA!sT^FB}2bT7}Pd_@ku|;!hSXabp
zd`i8Exl^1Q4vTRUr;`mO)P~@+iAW;bvOsE5Mrv#@**(Va0{Q#J0Eg%sKoQfb7}j?M
zUpDKHb4ieUm8~u|`7_jb%2wxFO_c##uDJ0gzfU}*eTb1ORlLG`9-%5zr>{~vt+cLn
z*~w#$l9qXW1}2y{5w?9G5AAvgoKD=`lH)3N1*;Ahzvgu{CUpa{*fsjD+~&A841Cim
zGj<POlebZMVWos9@_3jUAOb=-jf?`LjYzUC(tn!>309sd!~up~`KBhMH*ZNEM4u>h
z-ds=nh8eiXL*|{~c`Bb=hDgikB(Dd_x#7Z{*B(|nM3N9lD~7p-<q}B9gT<mg%Z<3Y
z$hqvlaRGAQ!B|C8O!bHAPBXGmd#vbDb<Cv{1G^RL%1*q7iaELk0L6m02*G`DmZo9e
zJ&g$E_UAAMv3E?Xx|(ljy|{SNjeJ0(=ui9v;$X7&b%}E(fKbbel2mx=ocHCUfF=ai
zpnQV|JU-R-^z`G#DvQ0#oq?nAcI7#%D0p@7G@iPs=vd&x2faOh*=1pkv4V1cTff{=
zU!RY1aP#1Eeo|`?aC%h=a!tcfdY;v2sS|^41gRK+Qm=jjGStkm>VkorxpQc;w-7C=
z_%pTfY-@NMKZbfYj3oCV1Czji{8B!ytYIcCF*6rWXcwQb0ekaT(a`j1#eCED{q)|-
zE_P;Dcu`H?s35UCg-s-Z|MO3`W<ajkW1P#$kCSzjv;H(cKr5d(XuZa?Ij67_ry8fm
zP=-J_;i$nYV>E?)Cu88F`K?NlcG`B0Msw@LOm+=lzluE!Lx`_&M<6J{qLZ82v%8q(
ze5+z8#v;#%NcRT0SA!Hitm8<na_N$SQ+m#Jq+2en@eB-HHfla6yJXy}wsWF;swkRJ
z_Ytf&v2F`HKsO$Md-e77WOgr5e-Z~~d!BjZg#|ViNOS|%T2E<91%okpuhH7Np9O?&
zm6K7p;O<FKme;k~rFQHE-}p|1_`IOnes%S_@o~GcdAaXp@$K;I^xR`xzOq3ROJ%07
zM9`Pg3fLz{UJHu}2=v9Jtt_Nb&JjSrxIPkL?!kfw@I>nW71c{ye=adSftnM1f;6$M
zM0PV`*AFk@+Etrz(a9Jl%x1k7l7;)#^g2uJ$n~A$c2^noWK_0XRmr);8&@K>YIdq`
z@{D!I5bWoCjt9emNxI5vul3059vW1u`kT*D(jZZIw0P{`>{vMw|K`nP@F#mBET&kC
zZlG-^<>-!!S?cTGdNxv{kddE#gVBtmH%?WxLF5L=gvfghZGfTsZ_xid6gZD-H~b#R
ze;9@ZR1qTsit?xfZa;Ed|MI)`X7C=k9&;~><IXom-P=io%4Js0F-#TlY-+328Qun#
z&|$(i(n#OM)RyZttIpQt3cA*6*Ctspu1$o??{PKS@M^0-$6!w8C}JX6;i=1ci?9o=
z%uaF`ck4K~6|%Jt*q6ybTW8M81oNojf4ZQ4cJIcEZA=aUOk~+L@oBBmSmOHdh9#DD
z>9HC#lbiP~il0oR?W+FlXmWf+KqXd654a3><#$f_A+G|Z$AgkRyWc*?58sVaZk65-
zi1WHPc2f^DoJ)$;mm258%^M8$=}I<2=Sx&0$xTXCwD_lgkW*!?OKbAgRX8>;8+$hy
zl@=bJ7ueMT$_4E7_BT=e+<>$u+qg46Ljxn^W<!CVA#P<k`Rv>kH<eJ&rTpdYmUmX3
zACt3D+LW+nH9r}PO>pbVx&4TFhyM{NZ$#tif&SZiMBidPA~rwl(^19(#Mq<yVm1z9
zDaMifTS@Ij9)^3m>!uPGC4B#8jsw)t_Q>)8@5M!c;MTn#vQ@1EIQ~A76TuRs!S5^7
zMXKJ@)3!{^Kk1vQKJ(TD4g!wkDwww3UOlRf*;29)tw>wyg>j`c(IR5>SWO{ZE|2-u
zQ>m|GIC^}|(u4w5HmwzU;X{a=V37r!&f-`ir4GGB(;FTQQ(QR$8)#>%UgJeLW_u*Z
zdN;ZNdya}VzD?Qk>UgxZA5m7{Y}mEE+A6Q|j9R5&g+gW85%Ej(9Se^pcxA7a4HN*h
z2=Ko?NJ1UebRvT*W?2@0&U_F9YT}-+6t4t@+CSB;r73ld?3IXmm63Ya^Defr7neyZ
z%b1G8Yb+Pm+nO{Tb$&S=+nRjr^xmDFULCFjT#t+`^+d{&hYU8I>yO_0TL_?e3<gZ%
zZ<MsI9~o935+mdDhJv`j5=eI5m2uH0N5j6j=)?-G=XBF2zK4w{5hpmkXGU~}mHbpv
zbcf(SY=FA150ctd;Z*{W<id(*>nNZ_4ytrTBzySYO+qKy*$th)DVpqAF>CeM`Ext~
zr9)j(n)3*>R8(lT^(gwRmofvGIQZ>v2u#8XBbkg4*t*m(TE=#N_Hp&gvy@fsNV`Sv
zif`Z&5DnSFS4C5`TNcB$0_<+^*t=nL?xJ}7ydcM*>UhIfhu;P_FU4>FOSBEs3NY~P
z%b_p?Z5*(j-Q$e#=6lWp@wV%jWB{~y3HtH=vhe*LFpEdt@?-Jt_IsK-oSE@;^?i>Q
zo6h+S-0Yk7+2Ma5uzzCzx|=t1^b~rWSF(L`&>lNIg=x<X1aV=7h%+bsgFY>qD<iOY
z7bpA}nNBCe2b1KV1-4xM{+mF$_iQJ#d@PK@AWxX_TwH0@$3yFwH+M&(e-y9=-g=(3
zWxmMDFvKoD)Uhn2Q52wpRkjq*9Y)#1W5yP8J#0`Sn~^m~lGTf}P8T20Qv9R2nB<qW
zByXI`9a%IJq&bPlW!FW*_uxNje!h4;$}fBq{v=>>OG{Z@f@*hPOn@@o-%qx7M63YK
z=S-kPj@N+T^$~K})M#;WJOs!$YqpblXO4XDN_c|EFqZ%3-1^1l|MK(h=JjRM!2ILl
zY9Ltbn%hN|cdt`$_E7-e?H3-{+;3kgPkX{OA>tc-7(W-V?DC|OK^DOztP`FYHf(pq
z>4rd4JSr9+oM5JTL1V9BH$wkrZ$TcxNjiV6#h9XZqqzFCO@RN6%R8X%K&abn{_D~9
zfYKK6oZvSX3ZMxb`X}+9T>GF}M-+eWQ=G22Wp{U_`r3dOg_YZx5pAtQZY!O`h7Vr2
zmM||i{-CiH`E4%a&*ap#4VQ^p_XeJB`3a@q{e9_$Xr<B9y@^^;<xZ@eNSNKtzJV$L
zI0(r1H#o@unNt8V|AK~lcC7!K`j1*-BL5GT?b)FL+csO07{9L({h<pCXKgr>B&#UT
z^Ji7hMyV8uek3*<j=WWJbglnuq&489rHlnVEF)q1?di#)3M@;E-kJIqN1i8B(sz6<
z@G|F&z4N{&Q&YV6-_y9KGuzGt&P625gB+)~KdYu)MfOYl)9s#K!jTE)_Bz12BCjQ_
z(H>p^GRileTB&JQX4_<UTH7*w!vq=Rw4Dy|#Kzg~kOenTA{SUS9Pns&wRwLVqGi~3
zF+R5ZME}N-gSKu;M{Kj~o!`EO$0)sj&cRnN;g5<0Rmt3uxUk;#;e_bSPo?^e+365r
znBj4~`K8@Zkk6OK6WtTAPe>Kh-qBMyIFjE1I)-Cqq+aigpC;)YJnjU!-`|=)PBGbO
z8lk;W=5*QKIA*yTAG2e-xOHbc=F0Au4!jHeRk9i^*nVQ}ISz4WRc;bK->0XVv9~+N
zQ)eoisO9?&;J6wtpHBbk8Q9=#-RbWg)wRMxVPcdms?6@fL$VvJXbW~gcQtG2kCht&
z2wdDM5klidU5@pWa+0(m;^MScgF?#^;#bDOQa;4QVX3T6f|pKO`VXZ=)LP+%8|GqJ
z6SsA9)b3f~kh?^=FjBkU)2eDHxQnAoXWD?tBogLlRYD?E)aby_C7I&Qo)MA=aZ^T%
zmV?)lJ0{aKE-!7;wNp^as+iF4F)`2qIIS}fZAzUYit<~l3>znWvuqJ=)}%jS4GvLh
zb;>ql;kpB<Rr>L8{aCnN3Wrx%kPaT4$8MtZef-;0Z(jX!UCFyAHblUP9Ml&MQr2%3
zJAS$K`VB5O&aZGU-rXR2er@Y`e*}dY*@HbMGEz@%qv4T#Wx!YL74qS4bpe9`caOB5
zz1&BC>?On-Fw_?1jW?x1lNhQ(kSNf9CKM0FnTy~uq9x7|&nh@T6CTg@m9%U((FZUv
zPa9sO{XQI-@%QF1Fr>@7HgK3}1pw+gJk=e(xvpPleaBsU4O4$Zr@baQ&>9t1NVnRM
z=G%BckHPK91vi2+6#tM((=f0B42_hL5Ci2t>CqzbA1h3gJLn#FVch&&gd447I8saH
zpm@djATk*?f3Op}6!TRE@{LK-k-@e(PmF*f9rVL;hkMs^4a*P4|4gI*X(t(&rN8Cm
z6-qN_?M!CptZX}-NG`<2ga+82-hL<c7X3o3nSiXJ7M^+X_xI%LL8$fx$OUo|%DDX#
zSqQE8`<2ZJPte=~D%Ls~2U|@yC-ws`54y>oCJV9-NF4Kf2IOCblLDA}(>wq<guS%L
zN|wR|PED))$M2SD8Ojc{=y4KEgC4)3Qf<ic=S*wFhM+1d^hzz73CBIKW+7fb+qq-v
zM!=ZJ{v}J<i;-J`RDb|<8WG%*CP5o9mpR6W)>6F^C!=T{f<wGK$!RYxAL5mE6u<S|
z?tV(#A1|DY*l@as6_8qaNb_Njmo?d0)6EgyWgOBRij8ydb<(aUcgiQ%sY>w)$8<9I
zp&jbrHTLm07zA^>kfY%Tq0mJ|9qRsxs_kgQ$UyAmEsnWk-m=5CAHE@rMir1>Nl_`L
zU+jR5?Rm{Hc>Uib4xHl!c(n6B7RiPFKe*~<R0C{zjUHsM?{Xhc-L7FXIg~|m#7)r6
zH1$^3(llb2ix7>jYZ~?XPlJasXtd;o`X^bc80|#Tz3y-PCN`katfJ6=_PkSJWR0oT
zeSE*GuZ><dtb+?vIn$!|Mm4`!Ct+6WvuHcgWM=rxLf}{MQC5*=emszO_?;$&n}-bL
zi~`{GW|1wbcK9(tb8CwQ;Eai`6r+y|(o9wvF&;koBwgl*oNSFdr%0lluOVoH(l9j5
zfX0`hG54RqD@A^$0(_AwTnl9gc=7t(_}(q%o~UB3PSEVP*nL>**kNeNvfeGc$kL9G
zO4*<u8ouXWQaLm=lt1j?^e$_BE?<gd+yRmpBbv9<a2@$SHV*9PJs2SsOzW)>!_tsX
z@e#Jge4aaDedK_5)tP^gO2NC_HX!%lDZzB=&X|f-aG#I#a-l2!`rvp~L)DA2<}+~x
zLMQHtZa0yVwY^%UR706lw7)W3aFiy2cb7UV|J4vG*9sD_hg>*~QK{TdRVDh%2$WLY
z3~Yr%ak0xsjR~meVt}Ybg~ZMMYx3#Q1;+5V7DMM@BV==mpQ`ewAhLtHYASmvPXnU!
zJrQD7H$zxWZBS=7Rq?X(R9@2(GbAWz8*qc=6to98%*qZJ#DvYHhaYdi31w=$i86Zg
z1rkVIUl(cG*RGTaD_wm)y^UNo8sG9_GPrY$JnCbUAa?XiYD1U82?hvatH^6u*-os(
zd-PW;&^%Qo8m*|TM?L}KD1M_}LaT#i0OT3{G-ZL7o)*tW47Q^DWtjn)P@KrR{cdX#
z&TFluwS@o>DonGrT+`FqXHdm5887Q(BD(0m2;LKh7GlvfOAzYRd?ZMt{rh@B?U@d0
z7WTD8Gd<O^Kc8GI=5wM#dp16gfm&VPxh>w2h3_@<x0E(DXEdC!&4?DqbQVZ|()sEI
z3PF{`a5k6f?G9a*G3kWL`BuC*mN8$a%Ng#pKfZWWQZ&6xg-tq6rOCzTF>f`t&>P&3
z^bPQRowYNPjb28M8WOOxV#bHC6d=>w7TxDVk9R(=b>IIra(oQP5wH1ILtvgu5Euq{
z6S-?LDClZayPG2zU=I;C3!i-y*?R!dmJ-+QSI2KcuMCjKT?lx4JO)F6-`zu3-Yx<6
z0i@2n7bq+?FIIOtnL=8mSb|-99GR_y(rSWfx?q*(<LyrvGZgv{WK-sFdt*%YiHSQl
z)?gZC6J~fk2-2Thhxsa6?N!)ZU*Y=2iwXz8zP#=i>%0Bm0AQa`nxem#@0(rij}^`2
zlvWuH8$D@F`6_09C};EFCk$Mbk-?=alaYL_*$FbdE&Sius6c(846*6H<aBPKtU5YI
z%k`}REDbxgR`<%uW+|BDI2&R*83j)}>8nqoOpALkLrjW#KK~SoucsHIKJQ;2*HwzU
z`>Mow%f6^JfT%|<6l8uV3=S9?DIY5JS*B$(Ej3tYyCIh)AxpooG4GY&n{Dm9k~~Q7
z)Uf?IR9ehi$o4hmh62(k>yi4TZFXB;^Qs|n;R)e8Hfv90k&3Z?`g+a$sg3ZJE26(_
zQ|9TD`A4t*2`qEvcr6gCNfmLra!b0RHn0_xx{UKS0N5Hz6kjQo^owBjml8_E#S_nx
ze1{Y%Ac|3IcahuF6%ri<UeetSGHMPE9VJzL?GWYbb|8#Fv^(d_oUYadU+PI=SZz61
z>dAgsElhx%+j3ZKgsGgHjF?t`ueg@&$yr~%9De8iiAiMb&KxH|WYy{?DNVPyH&Q#_
zX*P5&fG{dwJ%UTPUB8_&iKJHW#>10Hv!NsAHyIA}j^@wfca4yM2;TEJk{>?`w~45Q
z4!~Gl8iB9oHV<u|be)+K44qt<cK$HJb$p0}i(Op@zpPb>;yCtwtq=VZ?j0p|IR9#i
zakmU7i@y;12!;IF9nG6St!gay{44oKqwHoR1Ay<Z66<H;pZ@Gm#rBt%<|p$mScJ2V
zbvP^qm{X-bye2fQW^!eDA}hXe2kqffhG(_FL;u;7+yYo9oTeR(4vhY}u(1sm8{p8#
z8RwP_65?xYGCDv3xPG+;8j0em=NR9|SQ2{_F!FRjmnzn1#Y++W@8QcEL6s|_h^T-A
z03edAYQo;q!-MpZ?GyZkb9N8oxk%Nxy5~Gh3(@DXA0a*F(MkY99fH)YMbwb?VH}?D
z4ZFv=LUyc3A4QeI<E$p^3XJj2U>-Juc7%<&)9$V$7@@iRm#`DCwIp}`fA>cbTvCj$
z9ykxVQUABS9<ar70_gPZDoAk6p4<(ZnK5rvoRAa|5-bU%Zk-HOnXdKL79wm`%@AJb
zNRn{b_VvA;4}Uqcyvz%QhHOD!`O>Mc(sb`_FyyzIC&Iv8Ls_XKY&QN)weYeb68+5m
z$PBZBq+5><;<_s0qjM?G=|@<6Oh0YmOzW?r0#MQOkJtNZ1FxC>{?Ai`Gw&affB8rJ
zh0y-jNri;e0p^x}r!q2J|N6QLG^H`smHeMh%8NTw{;!jo2t_od@V>rr65}+gH;~{%
z$mB}6-ubAHryVmJwyRHRjvJF|&q~SfyP7TAu~65sX-?6}V%ADpD^DLvB;WR$7?y$_
zOuI<nCJ)#Pd1lf`Yg8mbIa7HaOd|j5q~XpufDUgS*`k;hV6qAtivyqq4_iZ6@0PR)
z(_H=bpM2w(IlKaY#n_d@ywgrA7i-A-7ma4tb#>^-jKo}I0`gIK^Zgl6{Y)fx^4OG9
z+D2EKqC#P_{&L@_&6GM}F_&_yN;?A1D#NX<*%_x;V{LIzTG-9$TT%zrWIM!$Nx6#v
zu(ntH%c_Un7nLAkD-Wk<!jlqK75)1ptZIjKL$)u##Y$yQ!*)x>Q~6r))64R1Q%&YV
z#dJ(`#0My8E<n9|>t|-0s}wM68&J*#+A1aghuqFf`k64~)PNOlVZA;kv)z~V<TdC)
zTK<2E3~^%yixNMN&!`s<>FFTtR5w7ZzU+)jCC~;9{{`zd@Q4h$z0(m-x+GWRp=%q-
zDeJxQR`itMO5pmRAOr91Rh9Jr3o?jXfs*(NDp*y5YOHdFiglNEbc^;^EXKATLfY8T
zTYt>S5iEdFj7^UeHh7^g)>Aaz6b`WqJkyS&h=wvGGuJf)zTO(CK_QI79G5iUt7)5a
zO|Rou7N38{zj`cHvAO2o9E$uOR5!j&>UR8Swj8-d*47xx<fm1A$abb7uR79w-wVHv
zDu!o)(?4hdTVEu&aIPz*6W4zZ)Afyw6<vOf6$ElbHFGZpK_N<q2Ju`$*rJ#oT>hKp
z`J{#2>ig~~7hK~9+Z%xghHpMHU9WA`=Gp0YEAQ^|ZDnu0-^c!40lw`JtBnIJ(z~-v
zd=+s$)X&f--MlgvVVwv?0|$WkEcG_VG4j$J*XV437{w`;P`)?feYFtnEYXkloo!wO
zR^GbEwSiJZMus_wel9PvRv1e>d_FSqV4~8Ki@s%WDq@-OOTtt2-fc1eXuIx~`m;=|
z*1%Xs2zgNfW4_**&lP(c<@mFu-c2J}XwYcwY)SuaHoBq_R9jImGLI5Eb44`B^#P59
z>3bLe4m+ASI~tp1>6kM<K2zg>sxp+8l9#AZn5NH<^QJs%y}sU^O&)}555vF_zFYSk
z@TyIs8$qOM>kJeWIE~du1ni8Nb#Q(&bGeJvqI@%zb+Px{=wX@s&k%|eXyj<<J|&`_
zYT};sk7ja(T*oWNYeI{lTs5i$C115*y#6>qaA4F++L~x5lrKVoNaW=@l+Z7#7yi!)
z8bPe+XWuMgrZnEM!b*EEy)wRhCer?edg$`9)3|B1tmDC8eND<+l8JWrvZTR(KG)-V
zRczlZpEsw(_we<>@)F;f-DdF^s~2DRN34v#8g#0X_ZFvo#D89wpojtJxXM}k&5&q0
z{oSUD&Y|-?H;g*}&8W!WwLZ#edY>7>0%UuMtxe4j3mrQoX5d{_zBU23*X;-oks>)A
z#6r!Py2Q)Q_ci-B^MB|B;ivx+1e6o=17uhKKFC<zefjnAruROv{OC~r5X^BtJg<*9
zQ_e{lf#=+WUNf`o!1OC%Iho<PsFZkAg+*%<rg3BWbuz7%qDf~>|Au?tu+}M-ChPIo
z=2Ae(;{36C9;wrx3!PeIIKE}S`TW?I+n|P94G?SVImD-`p?m$7Q?5CqM4dt02GGrs
z{4*SD$Q^MfN`WXUe0*zpshFf;GfrV6gJ-5MIU|Il+M9T@{MYvO?(I&uyWOR#bZo#n
z&vR%)zzASUYe$#3dqfz#s(1bMrSoJl?t%wVi22wj4=g$Kp@TL$H&RMbRFh|-UOqbt
zbV=WeeQQ(^=J0E=d22eaX{K-e0LapuyK?5%Bi`y^`wjTiazxFHmbc9t3Ru_`T@tZ(
z8s_8+d8js2lC>92iY&W?DS1Ev1k{FG5(62`yfIaoFQ=BTyoee(R>}Jh7=NfogG^FF
z-Jn=7EePV9_-IXqxwR%jvC$fF(zlZcCe1mNC64NBO<3YNw^0;<#T=t00oc;Wrub{4
z?9Plat)@glPzyoTmgSAvtg^UV>K!^gC+Yi5Yg=6~A6*8HCaUDk#znW<b7)nj*eTKU
zC>#Ldi3$x08<_tr3xo;h5s`HqgTJP(Q}qSxRKV~Gs6-EzP37f$$z`@!6mo^47biCO
zQ!_53w*9|h>Aiz(Xj}j;z|>udFFC@YBALG#4(;sirD5A*tfM<cdjYfEo(Dp2uZr3S
zi&4V62!sO@Ch5_A@b>!<w9<2HE<Gj6&+G8%y#wxv9i_UkE@TPykt58iw6d!B!K6h0
zbRK=wHqUD*(yu=$cUP}7HzRW9j=$7I)Fy~IGE8(az_){``qPb#0jI5qG|RN1xls}3
zdBOm)V=@wvX&si3WD)HjT>A-X<>&IuEM_qHC<?L#`%<wUQ+Xitp=OZ9cF$sCWwZvu
z1_>6+k(pb~5^8V{wYSv3>}3R9?w9@xeg`@6EoAy-wr34l5e_tp%(&sTNcH8SO3_O7
z5poT1G~5A-$^=3U0H$gQ9d3nawU~F25(T{!Q8K1(Ni`SMBnpr<TT|02rO}Nd8l~ti
zDCd7KZ|T*FH^aF6ivR7{QP6WxE3LCOidH3aj+#}J=FysqR$0YwqvuT&Nn!b{l`<RQ
z3HWks9YJ3S;|Xnd`&xTB{_^##f1C~6zAPK6>5~)$SbBJ?c6|JKPaE3Jjnxy#RgLx(
zkX8>Nxj|S3Un7UPu_7TEYfzAK)kynu-|Qu#yosz=GJyT}mu#?9W@W`&f#^ZNfayfU
zsMPnh`{)B(wO_p|sajOSrS+1sAb7jB8t-Jcszv7+NU?6EVu2iKU|y9fsx)Bpeo;VL
zt}=QM@I;}mR!;uf?FHB#WWm@nT4Ka@pOn~kQEAH_xj%hB$Fi42+-+tI)}lu#pDO|p
z-3J|NA!R|hdJMB+x5@dn?J{>bo+1Aw*fN<uZnVTie56h6{9)GK#XiHDqZ)^JrBmrp
z`2&3!p4EJBkRw22)|rM~5Z(6JsGZ}!&<?i>Fl=&b^VlOW$=!!8x(b(dWDu?Ko23Dg
z!H&!rWgl$l$<3H`yD+9xy=-OsOu}PvHn@8cv3tgb84>GOhx}gTnnt68LwF$ZWuM5*
zu{YFO4A^{3P)r7@i-6>~#8UTFS+AFL_8}54VvSwi%5fjkKf&tR#+5+ZPnX_a(rt(o
zKxfS48mJh7BtKLicy2t&QD5Xb#0N0e^Y+m%Koq_Ux}li!mqd&CyhcU-A`tLr_x*6W
znT%ueat2(_2Lg=jWHb8Xs@SGpcm<KdLSg#a7=uCnI%&D;OKEStvL}X4+xA>FX!{<v
z44!sRNm6Mrf@||FgSx|jmWCPpFi>a%Fgn=Rf<N~#!1H&L9nJfp!T;)deT#0(`p#|h
z9#?oYub)V>OIgxM6-k_K&xwAE=7D+!i?``tWFc)Gk;m@}>V!xtTDqMXFPGbIAyL?m
z`*4F>M(ge+zRQgGj_t=Y$74j-pW}?uBx#~xLfQ>G<@Rq}OY^z~It$pXr@$1$M;8>4
z{=8gn+kgEZZH1IYT3+D*?#VC#0~OH$omdjvYY{+bW6C(G7KaiZbR}?bNuWdm24s}S
zd?=#X{Gio*D5}{WmUkut`++Boj0{nm2WVM6xZ`m24C(CN_ifDIneA<|G>dc&9gR3M
z`}_$sCLrDP#AV{gjR*pPjtvc+TCx@;(YWi@nCGg{*NLK_-JKiZbQNvN8lE|TU7Xxo
z)LKwInJVKC7-w+hDCIBiTIoO<{3Lry$`e$*t}#pl4aSlmI<%f&hjd@dUfGm9)k)TY
zavMydPH{BReHmu%i*=et<}I0j!WhZBzj4uWRvGj%=LsB>U-eT%XfqmQ6bwz*sr9_+
zQp%0dC)1##cL90?e$<8>kKjQ7*6WleC7B)#M>G?FW2;Ot!_ikxUbP!DA3H09Wnt0B
zq+i!{;oem`Mz^^X>1CW*=~}ihrck(NMRGH6+=eT|PXA|?e4$)*ia+?#R;fT0S9|p7
zi67r3tMrGLILSHUrC4T?9ffe5={9!}PEh(Ih`Jxl2TUi)3L^EwCC?lH(Kh~<?FtL2
zcKMyo#$t&MHU?v3h`Iu2X>c(eABWFidOF_4@S%xNoEZO8!Fdo%t5|-xGNc2FHOYh6
z?hPV4m+!ASeo!pe-6)1+rah*6!^l#@x}+RQI{N>{95An@3N?UjkB0&(IR@{lZkoTh
zOn!KN(&gMM^qSSAiw^<13-sJ_O<Bud#X>~V(fAk1=qu~k0=42^F1IWCkDtw9kK7Dm
zoMDL^S{4>tM&_H2?u``lh~#qYd>_mJ-Y=4TpI>I)58z6Df)GSSYl=FdB@f~s+}M({
zBm0|alSSaC+qVOE%eJQvBfZ7Ds^X&j9Qj0${cKYpKd{gGBYy(G#Y=J9OHXkvARLJj
zlN*!X5ho#zOo?#1@^yc4)<I^lR_T%H%_@tnvittQ&qdOQA@?BtiT@IJsTYT-(|UCM
zn8G4~u~^;j+$u=vWhxsQaxIa&guyY!N;gY~Vqgl#11_@&opgnyIo5y!iEZ8)3DfzV
z>-$ZtyN6rzTbL1G`0;jS_AO99$_Tk~Ur4f42xH}w54swL^lJ^vNmLLiXYF3>9DJ#n
zm6EZ+NC~wd{bBB@cNOFknqolUUGUe}@BN$UyPL=T-{97_XX~f4)W;bCZOO1hp;((E
z5Q->L{gLY}9t>9C+bf*lm4Ud>JJaj~LP*yN(Q*xQj)kcJ$WF_OUU=Wl#z0I_7=guz
z0L%=JKTQi*p&L4kws#C8l+jnZvTK{4!ouEN->VAP%<~UMsXc~;t18FuwI$F<^v8HH
zpa=%|UJd3!YaAp9zQ5sPBGj~OZ^+#3?rfHpm&xVN=+Jj}GB*Ob#dzNbR@KE8)i;~@
z0`)pI=UvQ5T3`0+xu}iueN3rterB->fU#=07V9mhz|Qu3H#IzQ|L>;z?HISX{?GbK
zhx^}6_j?Bt2vAf2ppKIh-DTec_~vrDsv4D3Lt4EJ6x2<NtKcdY_$I4@Bj0XBzyvg`
zackN1moPKRZ5Jv%J%zq66xki3?Xc_4A=O@RHTXx--DkAfSDLwUfM{(+ourrm?kej@
zT%!eW#-bNi;VuRawN=8JVegWsuLr8HN4wNAipoU^8M+k!p{PtCZhT+Pu>ybg%GP0X
zSW+r6dCxZynjW{nrW9l?l7@ssk(>j|5@D`4A0u1IASJimj-bX2z^CK&QoYwY3pyEv
zwJOzU!D8b=?bOM!Ln%c79r_`za;5L2ZQQ&i8U2#@3VH}5UW;8fDsfdrM?sxwRab!1
zgq*&#IENL`Q(t=DWBXKk|B|`A;S}g`s+y>Qy|wH%MVd2>ES&f0Q|RmE=koMdofqI=
z9icj&?xDve-j@F=?w3SC)p!6A6fz64gbG0uCL}FC8Lhw#t1!unwiCv#NSk=(n}V2(
zJA-LHDNy3BH@iktGf9*f$iY+z`b$M(Sm|OIUY?nwKEOl@wcfxAY*!BK&xt>pWNO{H
ziOUjp@Na<_Hu##^9_B$3+0BcUQ<ExnPHY9ME$jAepuaLUAZ*&jrQnAWj9vksUbGzX
ztqR_4yVZJf4Lliktk)DzKHFyh0=aNK$`8$c^iUQk|MYecZzdN3Y7c#ALd15<bkx)|
zig9GD<O|01<OG)*WNdT8mEFH+#$PHof1R|0>#nQdnAU3bfp7xZcg3<;SP)uz3u{PP
z{rI$DysGj50L8>QX3B4xj`Bu7`P<KL7r8P;@@NiqaeSUnPu&{EnmQwg-0!=+%Ug5f
zX(bHpZ5Vrg7lc1WIG|SjC=epUj;r7pVYw#Wo$l}bGdPzKpALbv<Zb~TH=i*7O*8ff
zhBR6BIU0U?Kb^VV?~AW@d;5GHY);B+=h>j&qc%$bzFqczKo`Iw1|-;mFCq>NbC5wx
zoIGXx0f0PX9kpnux6MzAvOBpQf*<lRIg!DL80)dt1zv_lw6*fEwf0mq-o>a3bjB|X
z?hyCA?|<GBw{3SHvV4dj_>Ai)5sLPjh7ma)L&<~0_+mpung$QIk{~2vhiEdC+ah?S
z@Hun~;<_t*rq7km93iI`JpW;I9&mM4fIsW(`@G`sWGciQlb3LQ(Vla3YHCaNx}D@9
zD&XJ*J9Ri8!Ag9;kpe2J;=om88+F1#fq>{@|L>1XQpEzpvj$eG3V<bt0#8(FKqY{v
zY6t(!fJZe+Xc(vb!cApYpr|?%=r+(<9RpBn9e+ms!_=WxAX+oWDaTtaQoul(@7a}9
z(W*q-q-yYzw9=;0cB8681480twf&5Wdo=!a4DTFn-Yo|kxdM;!;>(6!JWzSQl5Rpf
z$!6G$k$(2nj+F!5F(bBDuC6)EVe6Fgfw5cf72*D8gEo?%1F~ob(%5}h8x7BQ<_v)A
zsTbP#X}CZ6^VS)c(5;UcQpJV&&l0jhDeE)3(`716yGcX;ra!_JgC;1(ebHr$%jd||
zftIKO?aG@+iSLu*+NXcnmj7UB6`Z<4&K*M5lvm?#S50dk{<pshuZwbfLsd0*kN75@
zrV;_z$-OAJD4A=;4=FGtC^7+7$zy;@6Hzm|v4(oz+y}(&JsdVcR}CGl!^}?cRTTAS
zHS9?|GvECsiO$kLaIA>#v;A5QgG1WoN<Z;II8(^$t<W|>$|K|YWr1r@KNKD*?7;J>
zgkb;0Dsn>XL6(-32Dw0Nl7RJ?fBwHNt^=xxr3;hL!9Wm#F$9p_dl3XFB1L+WYAAv<
zMHCQeCMft2kh)0kRVfODB1jViq)U?$2vP-+C!hi8f6V)z`kZs-?9QEU@0{7QXXfsm
z@7|1W4tEJkF6Cd431Pj-T1N7{o`FuD*kDYh_pWJugVZGQ{E?5n&w}Re_t@KmsW3{p
zNNI-K7t9msg@w~x)~woSpchVQk@HZT036=EmyGufSIdVCVho(Uuy3D@`m~GFd%mrO
zxmxk&7I+YFyYZJzhpt%{uE^*{L+dBOvuc!;+B=DTi}`!wtDUK_Y5ryt!v`jOB`D0;
z+P?46w$M9JP01nCX{pg4jU|v~>0x(d=Sa%H-0iAyh0)vr^BoojMmMK*=!Kru5k=)Q
zInf%BO}Ry)@ac{9xPbI#!LCcU7oeq6LQz|)W7qFAk_X$}^fHK)Ysk6knZ{aqC(5vR
z4L-qI6{0SBLceT5-h=t6@Zv#7f+ZU)9Ue)m-w_gW#+m%8B0};`G~W#;2q0<6hskDg
zR^U;SmMKC;$A9JCzS)r4YgcL%pkdf<(b)Cg(ny=F1wCP%%3BtwPAkF_`qXoVA_Ypn
z;u&^7SBOT?vskQ@>ZXSObM2%^wi)hH`2OYDemd3bIzesL%Vs+JoYLyJ{i&*>{gugA
zTa)QSv9@h1gsH!mo7n}Lb28lOE@L$1AYk8-`1=}h;71yhj*7KBlWEhnap<Frw;Epp
zRnH}`3Vc>FD!#@lTj6}t!NZE&ylhq8_VSGo&&ybSVNX9>{79AZi4}xl{=Oye=ffWR
z>?ZqwApCrXME~8vBD?7lYu8~X%Ic@ADbn6#sB&2c1uM#!k~7Y!4UR(cPADDW{Pwc^
z%#liQ1CH<3X%q)+m7-&B*j{oyG3LJhExfv+8CKu<TDnuT%2nfq<91;>9)8Pjs9sP>
z6#qn8%krI6SRp=`Is2HJL>eSik(p8u7YcbeX`#SbnPaMaHWg<x7QXJ{ob2lxIH+`I
z^<-3Il00symkZgAM&i)bwdSwxmC{-nbbFumc%_aE(5rz+l%i10!K=DeDU|E5yF7m{
z(RDql(!BfZeg9YIO1nWDl!Pz#p~5CNx}3Q<H~s6i1`Ds^zt6|8Q{5>jspe=vn~NNd
zSRX7Hc6ng}qF;{DHwj9pnM&O9;lR8n0A&~ST-5bUB^D%ygU%HM^UO{1qu1UC8pOM}
zXM43yNe?#GAgZ6dh3s~2^n0~yzv|f-FMqwwymYu0`kK?L&>(4A_q<VG-cho9Rob*p
z;abg6wD~3uToU$lte3tp^@`R##b=lg%@zhT9Rx!RA~JrK>Ds+#o%D5F=(VeZwu4VB
zDkl+NY8%F5D?RUII+d-Bfd~;AniiV3v`5r0pKEsw3T-(uf0LPTzl`J)A|!k4q#HQ9
z8v-Hx*`Pcd3M&RFHD<@<GfnP1`kXXgI;U)(2}FIp{EGW@fi@l6Ub*bRykYaZ$o2#-
z*bv6wfJ^AQa9iUGjsBuQX7X@JHaMg$#Qg&7OtJgbru!$%LK%CMh)yA`RLv&+mwk;T
z7n91~l)p+n+0kg4SD`tl#i}XSk>aa&bqUSkUx>PGV%S-?JAsh7vGTZkp2^_jjm+or
z@{$*<BuZ$Wl_aqOHKK7qNAOG}j;(E<finODnRz-Ez02cY+<KxaSZ<a^S@8>X#fhiu
zcW!zwd=5fS!S6YrS!Rxj_#TfIy4!%GKa@OrpFGNVg-i0T(*~hGZJug3@=6fX%O)F&
zI0?q>y1#IPjGWEJy1Ef#vFYu_UQgPy8UNak?YOOZKQf;=)$vKH{`RqAl*j<<VENWj
z2L`{fB(ikN)l8kZGW&%Zx4S?k*|tKNhC;?Z+;}hQq9l)%qZz{%ON?tW4em_@`klNU
z=V(<m;c!lUN;XPpG(IQES<3NT(Zdc=QI&`W-IXE{(+>wv`zh`93oXAOjx7N@PRH7e
zJSp?tgZ{Aa3EnI+q0J5WU`SZsV*ycZf2{D8(;yD}o=0@>O2LqGGJ=IJ8fJrYgtAP|
zcn_<0s~5|YyZ}d(07wOhKp{`8tJ*Ka9O&sT($h%;6{w4JgwJL=K{5L8v}b`MlrP90
z@UfEysRBiI(v*AnmIG3}6)B?1_g)VW=S_z->byyf9mA}i0dR{@tQ9&OL>sLV!|udK
zM&^X=%2DUWh}jrHy8;{!V_Bqau=8tuyytSAmlVk#;*HEQ9PX4)36c+1b3*(*wuhHL
zeKD*}bC-D0tJGN-v45rMQppi~BRdS=xsoY9I=}ZFQ(`wRZ4s7cn6`6@5i@7?{F=M7
zmmF6ux2C9WrTAPORO@q0eUW(7906`mmD4ToNOUg?W3>5>lJmrOqc$%uw5?u^Ldz@8
zPs!CqKs%awBOPLP4)v_#VByeo-xP1_Z>A4DEN-KRH>5i1+u9(zxjwMJ=)7mdQS<_G
zs(mf98KMpVk3@F`j;PgGl+Vwp(pK*c_`WiW+9BW3J0??JC}*UboV595GnSNzG{)K5
zvr2JeKs9(gxWEtTdVaS6ArU^V6pRQJ5P({5VZJo1_=;SssmoHRUL2m229P?Hxgs^*
zeLQ^Lw&(YpWYb=z)@ibs!A&$-$bPG!-s$f=uj`UeqEIV-W(Go{e{XLlT@6goSsbOF
zJ0`n-se6MEenhT)O!j8nUK1l&P;+tXc^407h%y8!!g~K|NP-UQW{i<p8$otxW4|?|
zl7pk77YS^;?Wz<nt)U*mR0hsAMYAl{x}Z}FRs-egCcZ`Kl;up@_daBmo1QUb0Rd|^
zEG8E5Vh>qN)F}U7IgiH=5Aq@f$$r!qQk$vYqi*|L58A*UzA*)1yZReJScRbr<g{No
zCh5{ZWMsO)<H6Gyv+kQXNodolx2n)afrRMJ8p>YKgS;jRH&AK4%%auRRtRz`{*b`#
z856x2$L1Q;8+S#u`h5lY#Zml|V83J^D#O4+6lx1I_5#P}ni6K#=1<M13f>Cu@2|Tf
z$bi5fX=#x$(3@=W7aBXC?x8AcRpKO){6FTyBHFB6g(ER%!L|J9A9uO27nD>6x?c3{
z&lu``UtPjK&?DEO;5ix5rA3|buFdk@1q$92^qx%uQ%Y7hs(a{Z$t3QSK=rD0utq0}
zdN|y<WpneJ;C#|8xAPX&Q=gchHpkN>#O7o<B*U}ZT6Us>yRp4COB(hzVq3dA&l6G1
zzD;h`4|a?(H=_}gLlMTivnHhpBV&-c>}1P0+KA_tTwQqbXNhgCiCy9oWt#WJ@?Y7P
zEAS**XWbPR3_3{ZA27?^Fd3h!tX<sk$x|^M@s^e}_46cTRjWE8=v`OICu*cVGCNRu
zwj2x*E$iom0V>Djs}iQ!BJ93~76TK9J*vjz)nlHl&;&bCK=vTij^>amAR~snX9FVt
z4<ny8%v!y{2~0+2fDIkx24Rdw)mWO-zjN|8Vb}A@bY1k`o@{&l#Gv*=zzzF4=2tTP
zeJ>vf$R<GNJ7^*hI@=Dsa0Y(wjaG@gnd1-Fd;RV{OgU4G{?hcq9v}R5d8n`)txuIB
zo1Eu9c0EQ@?Ys1cyzy%J#<BV?)}g;5LOM*<My9Ad#p-vy48<hS3bw0bm=mE5?wXyv
zQMKiVG<*8v6`~a<b;n$(G{&&k;NF}T1;-_fZ4+N(Mk*Y8Luk&V=n1QrnqNVt9K#?5
z**nE&o7O1^S3X^8zZ$#qx?G8&Mu6*{_~o3LhXSMGOen*Il>NROuAAxmx*n$y!!9bP
zv_3uZUA%709>G*oT)#r&hWVWH$Rdhr=5w#1UvL)sT+en(G(M|v7Gt_kg;<yD2#n;5
z_v2K)(8F6WlR}*{6X0RbwWp6b_#mXKHtr7%d0BEbkiTf%Ph|O$30z}}V0}VN_ASJ^
z3{+T3fg(?`M~B~j*mvI=8RAvIo!0Vt&C6<$awHBnRZXRToB#3UVN6d>>+QT~Z|q&y
z$?=h)Ii~u9CZB-YhU+Eft*fd<QkazNT8d-dB%Ayh#?@}T*qgOF=-~F?@IurU!zFoT
zT}%f45Uy$bGSdIOS8O``Ti_#Rhnp1`$El;d@Zp21%x{ZMevKO97|2Ly`>MQ1^<(xo
zXD!}GSzdgeZk^jb4WNcz@B6+?hJ@>AdsWR0-raZVjB62f>qfV|k!KbDmKc>B$g{af
zCnL!*H_5xMrLP`hn~T^KRv&+-0{eJi@#Z4q1F4UJT(eZ^OMbp<(P;~#Y#pVXAGba#
z+u8n2+Eyqh9(z*;Vp%76L71=vb-i{;BFp%}Z0;hyaMkxl?wZ;)0cvuA$htU{IyWGZ
z_c?CnK6NZdBh5Y(EmHTz!FK$UG1U9gQOnN465o&mmzftsCtdm8qiXQt{rPiD9#h)M
zK<=z&)*XF5(J{I`H(FN6e0%64*Tn1OPWdvDh9LwxpeoVzo~(~}J_c^9Lf(C&^o%(Z
z-h(Gqq}C-;rt`}r%PlwsyXfo6FvY+0a<k>Ms?GQtO<WlH%PDBP;vP^eiSmm*XtAc&
zno^}NS)>+vXzQ{eU=t2$s>lO<RrXq1rxV)~_#WtsHbIS@D#?IRvuM*68@^<juN-*&
zxW;$@*Q|^#%a&gw_+bc*0YS56c0AL!u2G(C&5++@2f!s|DncJKrPs7NMyBa>43p%(
zDG6GW`<OJl>rcUkKj+VcHTT8&cl57}BR?&c+zXUHou_a@XI!AbirM?(rP#l^&R#PM
zBFtW>U;AM1v=ZKQ2|Dbj7COM_B=)E@W(Q+=K>IKK8*~yZM)pUhI-q090<;GpaU8%s
z90Dv(4IHy_1FvwxV0{|mEgU$;$x`SNRRHX#Ie~_GDDZ3=0nVcbI;Q!F8PO16KL+5z
zv;gwAoM>b)Vj{d>rJo(JzzG1i<`Lj1F<=hQ3`lygku?1f#f$%6!~{4mAV{cWiJv0d
zn{<G~JQBDu#{kZh{3+(Jrvhr`VI*{c6iICC^-E`o^iQOcI~|bm;U5JRGC##Rb8N&|
zaRj(d7FeFQ0<+5h?13!Gf{|x`Hg5*g5nl$BG!nPMPqEe<0{Nea@?X;>-U%uafgHr(
zaw_2c5)}ELFwXw~jZ{cL(sb}p<gZw%ABhBiFKQYjVEYV=;s<jdP?=!`DFg9%PQZ5t
zMoD7F0vppX5EJl$DC3)fAxT6)Vn)J4xj%KiHUrWYWRSmzKz{=4u8;;hiH9M7oX1-b
zf5f~w{scX7B!L`>Akx_kky3;`@h3>wjRdM7f_`~p{{&^;AcAC&znnvV0%~rNv>499
zkiUKm{{$WRkwBTV|9lnd_lQkqpkk2=XeH(gC|E*(wb4N95)xou;sv)n1QeI}iIyz_
zoK*zuEb=3N+O|ad_SElho=om3(3f)>NL+#fUHwq{Uw!=mXqgvK7=jX~{cnjF1VWbi
zkUwo9qWhWY-#*f{cNkD#=3)D}2NOBX#DoODl|-PG*iTwAiN;h3ai9LL0x{rdnU@Mb
I_2ZHM0dr*w^8f$<

delta 21219
zcmYhib8ux%)HfP)Voz+_ww;M>J2~;>#Li5T6Wg|J+cqb*lgXXueZPCZTm9Ewy{h)w
zUDdt%hrK)vSvn1gp&}0ng$V`&1_uTPMhezY)?d_Lhd~1YoN(-CL6v+?dP9kC!D5|F
zo{#Fy|E)nRQ$)N=1=$L)pWwp$+YN$3mE6;a7^O<MZ1bF^F&nV~P}SWG8#8FJk#P3N
z3B=_^wGu!b7CfulpwIA8ky9mx1Ss?+D7}AFE@jx(5|tMrjn4A^lrBRGM`b^!M?%+n
z8y}q9oBdV_wA|yy2dmPmK>AL1D$*V$FCf$L2=<c;ecCs3*Inwa$V8;`oq=Tx3zdi(
z<`iOT$vTZ$_V^wPCd43~r3xA)Z2-WRJSWGPq(dk~w}8ukxNd0NoBqusb0*{-Uwv%c
zr!wsd$tYtF&C9@YlRla?%J#C9%UCsK_$?#L6+!(ID1<_)wnrls*<+6#Ei5i<sXra9
z4QmFi)Xj2UJCqFDWCmjNz8>T#pbax44tkS?Lswj(aftpCx9El6mpY(K3Wa#!0EMhf
zKbjrhB5K>fN&b#`b5#uS%kVMbQEBUWve@>}QZN-MY7H^%8JHHeMi|a&kQ{ef9-^>9
zX>OY+7<XD6qA-Qdwv`-LrpZXYSf=EfSYl+@xre`+W8&YYw1^ceWlXU1A*g3AaUL8U
zb_hcxnh44${^Gwh0t3Y}V>ZA+aX=#K4PMR!qJV)RsDXoiBV&Qy3F(2UnWSL_Bf~d+
zW7)3k#w1*EU_!WXaUispUGcl{Z<j!snP5y>)tRMg*>a}+xb#zLjh{wZo`99k_w8Hj
zKZH-wV<a;kd^cQ|UZ0g2&oFD-N+$;;@T`PRYQ;LtgvUC-lKW2K)6VDJl^!3@PC>jU
zfoFHFH6>5JcMIO_?R8+f`}2E&`<;i-mDLrn!<$u3{<&#&;|%LGo*uxlX_{DGrOG$T
zcX>Yq$e!`^gx&H_^hA;Rth7g<eH&<WXUw@k83x(d9GoX|F}-1EDSSSNfrwhO0Xg4%
zmrl$-8vi_3R9xJAntqsnl$L&dz7I%;a{uA=26QqI^47UGuLFUNElWlX18=71lLx*F
z6YlR`Ew~yv)lS>DN+a-J-jALYZ&g5jpRMDOtJ9jT&)><L8(Y6qHoG{hjP0uD&)5jB
zQQBqqbcHTAPMNJbN{bErUYCRg)@<~ioYj;zyi~q<@a0g$BbsNzCA89$S(C`1BFdu=
zD^9wjvsHJOW&>fGwiw@<Hic8yP)gnS@R8ajg~mK5S~^}!HU2J!WwwhnVyidZx=A&u
z<uyWyU4p<uC^XYkUxaYfLrGD-`OQTQHhI74JeyWrUMd7e*UR7>KC)c<bKTZ&mNm@#
z7i>`~bB@;tl`Ji4Rn?O9`4txgK%v2sL13T{{3MBtIsmemh&jU~)ee(1OM5#}Q9pnB
zxPCv8{xr=*vCRZ?kOZV6d-6@%GG@KFZ@fqmlw4L;HQD;uHCw#{hRwklG%}Zl;iTlq
zajH`E<S03$Xp2mvxQe}XRp__dCTJKJq|~@Alh}T8U93py3gI^sPse$}4i)-k5-&yk
zJYKErss>({5_zi$5kCq45q!y%oJWBuhD(ydKy#IgX-d{k#+4x(FGKDrowVAkdkZfw
zZXSP*7Ogp*&BZwuzqY1|h`uo+$7t;*&o@25*n1H7n@x8XHcf<uKsTDc0ij$&@ihs8
z6|DzgH)?LO`AxR7ULH4kTI^kR{Iaoya*)ums{;xk-)-NYou$23(z0w3DUIuDKc={R
zj+{~wUbU%~=wTDqklkIFU}NSq^~$x_*<fEao9M<$ea85G6dq*L5-;xvPmEm?b<Ca4
z@Dm<Xj-Mk%-6f(=z&asKcn&*VCS5O1RpT3_tj_rSo+tY$K;Oz}t?8^cHy_k;wyXs`
z?E}>DsK_?o`s}oRo8Mf9GfS(sfm(!$r-a#$h@dh41xMihN_W_Aje&Qp3@|v9_AIf3
zMYGZ&Q5?{h2N-a1b-tBn<=Wy{8|V~ZtyC1R%X5I$=0emMVK%6Z!!q~y-y*CF{&;Gy
z&|<NnESYqKa796FyC22+l^kfJVi&!s*<wl!eD@cvu&ct()z{7obtE_fXif<CtdTU~
zf(>j{2G^jVokd67{K1NJ9)RJsOijdY)#ltXEDN%F9CR&?I6IXUB6vvXtIZBoit4i3
ztF>-VqeyIS*bUvFL$tPvLs2sP00olagCBvi$OwTx8Yf5kV|G=#RnR|{Ah@ToHD&2Z
z0^BH<Jyq??lX@Iy{|wSL8AD@@2IG$eYmU(kdoH1_Kqaj2&NKYs{_2X=WezKM!fxlf
z&$H{w&$A=lm-vo>dQR1&B-?L*M=JAY6N;`agTo~@@7zOnItYzmUGhfjw}HHP(jLxm
zb)OvYEU{jw&+b$~%H|X#Vk;k#&ss+YK?ZHcpe_&}-=o@RSBwi9=UXx$_v#`N%+D2Z
z>0?S$snPje*p%#Mb5Z;t1BSIo_cQIQ-cWE~-GwDH?fl_SHFqGx?S$b3*-Ph4NNi_m
z{Awe2-nwrGBb*{;MzgnKM~hyc%KIuu$o(+T{!{$YY+?3tyEit+I)9aepoH6UfW?+L
zBkbS-#*tSVW2l(4miwj?Je`IMDX=RLR%uqE&W5L9J5^k&^r!tgmD6Th%gEv^a;*)I
zsnfQo^dKh}l1rC&i9XuOD#zgCaTZAG$ZP?13g4}p2*ru?7t3-eJkFUpjEB%(kj@-1
z1ab4lxKWKFQvXs21Dw?#6SdqL)%vOsPfwwBwt7YrHkr4pA_TF&RTW+W<lHC6v}%Ee
z0~Rp$L?B?P-y4IN>H^NrWIP-Sa9y7{>8p}h%CYEI?AY?xtZ&B{2%Zb~uF-hr5Ow@?
zd+aZGv*OI*_SBtEU}#rdsyMsKS5X1dz!k>U;Y}U}h6IR-CQD?9W~{}Dnw#B?kdWr8
zIHQW*?mE}Qyhd-!gdjZlMIbCj4VRi@Ht=9)-t<BQih#k#FvDvFhG^UL1(?K?-)iD&
zJ0yni`{XsvLa5^xn2Hqj;bIM!vPgxp(#5E59pEPMoYd7F5oNSVyGdY<|3m|hpF^QD
zHFtVs{7XcKK!DzgRcyf1SMO{I9u<sMCH=1m^PyK*WHNmGG+i$}#>%sETZ>+C`b5%r
zOAP|imq92YN&)98h-z{bFCSYUA31BLN=*RRDKgoQd2Oy(1y&nesbUtk(0JMy1rzf#
zBI#lZrqTkHs~Qzl8GJG#V2mm$XRb;d>a69g<qZMWy3l_`=%5d|&A`SlMX54Uyaa1Z
z-B{`Lx5WnD^tMRoY1O$Pk+LEs#eu(&@D7)zhpJCk#_MJrV@1>u6`<><*~G|!Y2X(0
z|G(c|?n_W$`%9-l+&5=1SnS{O>fahSB9wuL289H01p~Q)=ja-LHvYs!iGm#phMcgK
zuEm{{uGMDuuQG&^A2JU4#uKRlHHac95rcqNM+L`d#m1s;3RR(Fi8W5+SlLzZLsR#M
zZEPjEwp#gpO=&_D8(lhBGE0o^H)9hpL`ZGi-gZ$i!~wxpApB4;-gg$ym#s;{IPrLK
zRe{mG(LA8Mz5};IZZ8MTp^5~!E6u<0N0Aqk21nC6FC!&}B%BQ4zvOsvr~W*cUT3fk
z0!}exNz8xoe`DL6$?3o2N96xT_VtT7#R3m;e~A6RIetvne0_;W&SPO5hbo6Em?;0f
z0-Au14&q3`s2-?@$BW5}ITTC~MbMrpjs%T7jn-v8BdA8BMiT`~i%ILw8*0m#nxMv7
zr~(E4UsUZ+EG;IKb2Su9A}W+j-hZP7Zu|$L=In&h4EevR^1c46IK{10BcaV-zX|uv
z^3J|4R6q~s#O~GN-u^WakG_yh1Pe6-PO%aLNEXS7gn<SN-Xe6omL5ABrfQQBny)q8
z0G((vZUeH8wt*BgNVcgUykVWgrdscp&`Y&O#BkW#1V2IM@$79I;F~0DA6Wyr9BhUy
z9ks0mrA=ci{e3(di@xpz^?cpj1O1(~as>V^n)ygtp{PTt(+iV!$B3DVhEU9cQ%9-*
z>onO|Z<`t(noZZzr@_7pdy49V-q;1trQOELj*2}B`gn{kVsISFICMFc8t=s^9lHK3
z1PxF3{p`v4-FaqO!cJfwi=Ro<p4?1h>r|#FT+s<>S?Kuk$ThI|jV)Rx9sGigg0m`e
ze$4+~WyYXs=39fVeQe>Tl_aYoi1^tPm^EE4E4grNF_jtSbk<SP;bHKz0Tw@eZyt8z
z-R|`I>Z#*MktkbGbtTiP1m!o3jv)hmu0ZzE=H|{r_bI@__{XY2sIX8*74rbG(b5_}
zi_t3q*+{Y@dL<beFV3QMK7~c3GspMzKaaaJ8-&Qiz*OA@bPCro9=H%PvD;!5Af{FB
z%iTte^4~kNrYJ7RqLJ4qvU!QfVpWmIYd6_c)NXui4Y5W`j`>Ln_(5)Q^Q8`5%SVKD
z`z?DVkCxDieW@Q7%t{Z_fgQswPIO%Las{5e`qUQj$WKobEHi9YBHnY0Etb1g%Ox-a
zJm=m=_jL^e^|(CBzN^y4J>P4WfPjFl)c!MY{>1oQG$eS$m<$k95SnmeOh}}qs(EAc
zSqj(_w6h`K8(B8mjAw+E%d(YWm{nbQbptBt3NliN9jwK(@(G`vgA54?1low4?dN()
zb(rmRS-UforEyZz6AUst88(_Pt7*-FZnOrV%As&GkqbGZ+!lqLG%}A0NH$rT`i2_9
zg<Y|)I|G>!j2}Zwh9Cda*IZdwo4N17{^+Hwp^8NKX>hz6Ti#+oLJDes71ioCJ(Xmc
zA|7>cIixPKn*UJK^`<;k$^rF6YMVaHAP<@bm6I<tE-~vqF;qoA-Qys%K2sS5)&Oj}
z-{vnHkEJ(sbg}o310j4I;N7n%_s)f}yULy0C|BI$fwZmI&)|8ix#Z2w9)TG3Fa6ee
zA%0vWsn;Bs$L3fn_r<U%%e(@vL=IvW*B8HC++!AEBK7f_142L+pv|qXlyb-jHBA@&
zH{Ycn<8OL~&Flu}hH@vy#Xw{+BUDFwF*3US@%}Ah#UeL)i~L0Z(DO%EdZ=C*ZX@<|
zRo+KuDE<ILPYL#46Q2vg@h$75%c`JIdA|FJ?$7J7p0ATEzptek;cl<jDdil2k43kq
zp7iBu+|q!wfYx%qeSEFDqh$0Pz7vMRya(Ezz^0wm;n31;0dLjSnfP{?Wj^hBIkW00
z&IiwlT$v6nD1z!cV6oE3Lud!~R`z#@KX;h`?_`x>T179L<*k|METt9(-5U2(swnr?
zl&LLue2E?Gkq<V42}|K21g21tCREcv302&Vb!hT1N!B@+-llyatdD7z3oLIP3stiv
z^ZdF->}3V#+)4wCJY&j?tB^GC4Y@7HD8N^-g3^v`V6T)hu!qQa23x6^sZ@VWBO%qr
zQAdGcDScD~>^6|SpJsupFuo|#o1BRTH#gTS)=V-s=y-|szGDMVMiHNco_jC-sKCG*
z$32pBJ5Y;nbQrei9niQ7;(B`1rT?o4e5l+x=XHC#OelAP!}3O^o8p?l5_mTimxMe3
zjh<xU((=Fq3l9f%i$$eO$QCmKqg0l?E2iwH_>{4x$%b52W|^?JJI20K`%3cNZF*@w
zAzuH^){dmzU`Q2Yzt@P>mqrcGg_J7&#mK5pV;VXaB1ILGl))3)2jI5Z?w@4#cwks*
zcDWDGF*E-uWH$Y$Pq+F(kkhAR%+UYvl-uI`abzK1%jn&9LTMX1P=mo8^a&;*CSlHb
zi$N(%)`psAHWxRt0P;Mse_{Hy6F3hhC%ci^r0erh{cgTmp$O3SOZ_FxS7~uf=QfD8
zEF4Q8z|oSut@gMj_ze7Cge?H;4%HBXtcB80i?|4ZL<PpU;|dK1#)a{ZN&&HQVS+@I
z41xNURs#HP$M46PyhQK)Lzf)!!>T$%-peUVSdz*445I@)8)p+MHQv2(MdeQov&wf2
zXA}IYlt%jRF9nE0wn-u9f0cyFDZ}D5D*TOzmH~C^=Nf%1a-X)8%=s~&+@lkVHERxS
zONjDcR0B#NB6ZAp`A;6u#_f^n@JX;J0Z>ATC}y6f0zZyT?jwE32<agIY&NPO7t2DO
z6_?cqDD$Y)S6_9&7ili*PCrs9nzW*Q+$V)DmzGl>b~?~CmIIVo=uoCYF;pTJ$oXib
zZdV^_VvF)(#0TnP;}0dN#dLCXtI4^H#_K}0t^A$yz-U?1R`+rGwsThg_v|TG0q8x3
zEwO$z5z|8Ge|g(%-H{wr%&Of8GcXbUfE0E`+U2_i)m0?xt*Ph?yAWu=cNugOiXK?0
z1;Cjvjd{OgoC#KU6Z|Hi38|fjtAdl!8$9ZqceREZzv0&?ry0zKzUhbU@%_gbxZwg1
z%{?9QBE4vKGsLO;FpY+bJT1z7vW#rmrHCFh952@_oK)|k0XD=^M9??o`&$1F18{6N
zttbW`hkl1+LJW<%Y*=nO+6lZLnX6%~-oL{X4Y3C_wL}dJG3L*Bj=vgARiMGaOF-7D
zr{IDhH8pl%`30;Nc7hozf`R!SR?ovL74NU@9qFod{VHW(_U8NLWAv6q-@Y52_L+6m
zc^i|+ch|t!T4V(*A2460{M7Q<<-`V@ivXPrl?yHMp2~uHjLR<q*<a-AD5X$rz%*T-
zszj=ptI4Y#?d1>;B~EZoBwM1!fYaLO{aOb_14(e80^Mk>;K;neHDsm|=w_ZqP#t>;
z?=G=xG7MX_6uwl+U4$SeaHs{!g>$QTa$R3K%U369@M3Q=pC1d+x9NR`y2j&6(C!{x
z@>!?gKz>noN!1qJ8h>(HW0Ihk6=3Ww0ht}f(p+Y;HhNS}ukNP;sJyi<r@zZk)j6zt
z`6%F_tGXHX?{{5@mx{=z->+(=Ha#D2Yv+2|uVtg{xrv@IBW32o1BElDBQTv7pXVda
zbiHpkJrV!U-ht{4&aMY3$ua}=>^7w^M|P0@B8&Ybvg1sUs-Qg2wW9@?VNer`#9MYn
z0nV*G4jkYZsvw7p8S{R!4221_rY4FhvMe&ba(GWx_lR4$-|{sf%APa&uVzJ0PiC@z
z?PSJZZEne#nncK=S=eYLrs(DeERY};xEIh1`RxWa(_k(tiwjSyZDIkXTT!o>n3*=m
z_$3LdUkdC+pt+=V-9M^VOE|7&)X*MW_HpEOe*Ji=Oo<;Xsg&g^jC)aD-IQhM(?pbw
zBrY-)%=R07sa%PMxG?V{>7%WpBIS9az3=%W?L9Y{3>>xR?SsxJCiv`@cTa4vZjk;h
zna4pUzMME@f$dpK6fOolqpmc2yEDF<pf_>7>tcC*bAMZ?n9m(0Mvd`iB=ncNd-RJ<
zJh(zyi9ikY%l7G{SwgXGs82(lh<a$*!&h9owttlAmv6@F*VHmkkA3MGe80iQ3Myw$
zeT4a3va=8F_Kj^?WFRrnO_bK)_2eX24f){&yHh+LpA^K!gd_u8#LDp@YUEh?5jbKi
z;-EuUtG@(Bq`uCyD2Pw@2alMsKuN^8Y{JT9ASI^SI$Z*MEUHpEC5WMH&l=avb=;Pb
z+Vh@v9w7c~jz<&b2r2nP(0IKalA*|g2hKLiN^j<j&{&Au0xwh@R$qFXM)i<hr&b5f
zW;k)bU$fspTiq6T$9BD<Fc&f0+FU`DoV3QZf;-Ddk0|XIFI~yN|EZ)u3*d`#SsQ&d
zI)yQjvqw&#86%RWG_0GMxweE6>S0FkSI`>gQtBIcl=Ep9;N4EUBFb{gfornwe4O&~
zd6DAzy0Z8A2nsiHfOxLcS4wZA<dpbe$B`S51aO6TQo{llR!aI`x3vRWu+XOF><4NC
z&&iNz%FEgL;>IIHqu@7wL27U)OXNG-S-SH7;Q{_-thvlK05^9rxiba&OcGsODE}Ug
zMqD`YP00XV?(EoI5h}I-Gaf&zx)<?`m`g@rQ(n~^X^rZuXsfSCa`7GUOTQmX1e8M)
z6p9Y11AYK)`d&+N<Lc!<>9He;pIgl`+vpy5;M@g0gp=p$tqBIy89c%i;J#@90!@4`
zWSRhoP4Ut0gXoS2iOQ_8=ogaskU?6ojfpn}t5{JX>#R=;(fGfZ*rC_ugBTi!_0vg5
z{21C~1~_wGo`vs1KJay$5OtOPGcP_ruI&Brv>$=V5O%&%7qnpr>$8smMHN&mSbxCv
z+!VFK7-s-YuaFxBg=9|?fyy*gNR8a7v|rNbvqSImS2dsk?uCJcF%;PT!^i2@ziSp$
zje6rDC$-VFOCEs*8*tVi+8ML1fsPr!9JCpgUAkV``-FtAkP__b0`5mGxHwcN*jDjG
z_(Xue${|2Ut^xXrEsV{u)#OHp^<=-qGTE*jDMGslzi!7iZgcLH3@zsT+BkfYn$r0u
z*iaic7_ufVmUIB$2vdoC74JBn^3C-e7Buyh445jk%FMEer3|tK`zU7e@6MGRmv)gp
zQ-{JxLeh)Lo3LB+?T#6xEwh}E>CNs!b82vq<vUkRzoB$ojPuSXolb^~Z2L_!eL??6
zp(6=H>MD2r>*(s(|F@$ff_%i)fwd>sne01h_fs!1Gg&(d7#0!}bkhx3SWGtMVht?6
z_wCIjyL^UVlgrf^M)OI=*xU0R=i~%?e4ZxTOxxu0n~dp+*+Mr)%a4pD30X85Joz+u
za&1%?kD#Ysw1t@$0&Xqa<|4;;VOp}N2n%hbsFC&+93CblAMX8yk!_e!fZ?=L+75+<
zvb{WiZBcPF?5eZ}byer1GUC_6N$aYDeD&sT$iGXJlR;0QT?d`fJqzk9lTjk7ZQq_f
zQ-iL78pFbuzkz;twm`jKT0=<*YN``b#K}WGE@g?fI8rHcl)pmKqzo=OAY)VQEy;CG
zJBz#LG=q<({L`WmYQpQ=fcbZ|4w0VH4|9seApk3$b0k=t-*HH=+i3Gy9zoNah5Z%Z
zH#4qsm$SwSwr-h)=}nU;;+mqG;=(LC^S2*<9-_{S^qz1AaKb{43~=4VJ9>{k7!=~S
zgm7^r4<wAw`Vll?1`p$#C?3>KR&>rJ_X{|@%+1G0WA0PsLmVqA1D~B|o&|H;+L7!d
ze0yC~aNG$yq!lt=ic5BEH3Q5X<xeF`Tv~5CV<j{ykM<tr#$1VL(N_7Ke?;=7=QF?{
zRcTM6ua7PT*R$yzlbhL}&XNIca?93=9SM=S7vU(vmIu*vpwil3eK_nkhS)aT@2urZ
zjSeV(o5qC(XFB!S0h{IuP0#J=^Xi4<j3H_BKPU8C-%KTs*qijz5F*|(O-T*aY7XUA
z!fi5-bA}Q=WJ!=y&v{=C>&v0w^*|2kffeVRV9k8n9sF(2FS}1aib#{Je>Mqvzlia^
zSOV)kjTE%POS9YKh?=uyt#+n$yDTpIvT8zaec#6`N;*JRz?r1rEC=W;HDphRk-p?}
z$Q4(DU3chk!G0ryu9XJ;1!*cFe`LgUIFTT)YjFz&27LUtjnZ$C#iDj5bYB?2Bo3OD
zLtY#!G~q`{4O=AU8|93}^aAdFT|xaLS%O!WgwZfKTmVQ74;p@FG~-cpWItv}kSfg~
zy3BbieeJXn$WFK|SRZ96!^#@39UQtFWuuHx-MiYmWaa#7L$_>C+B(Z8+6rL?i8He+
zwiS|KinxcBKpm<TV1JXq+lgI2Se_-+31J?*lDa23nvfH*5dPm(x|^3`C3EMN5@rq9
z+HRAj7xe1T^M>p`772dm6M<oHYKdD9eB4NK$0SfKdSxM{!D{%KSmEiVcMBFYvSBr&
zaCI@sJ<{^Uv$*Lk(Du1edD7KDylE+Xz)Y<8<Z<$Y__MM2IxI9Z`&Ym@9vlgd_^viK
zgwPzrVcH-&<MqproQSf7!eElq7Fw05cHm^-VbxVtWR$(k{Nw9Vz0!kwm6azT`{2|7
z${rXaXxz~~X23-_z$P%jk*Rm7!HIxt;Uf3xWu`qJElwr=qO_?|p4bMDRDXIIS9J%c
zEsqfLAu7*4?h=!LqvK9X16gOcV-fy=)0{x7m*da9{TgW0cmf}-Ol!gzYn(xS6krVl
zgWbmjwBehLJn;HH7h%zRKy9Cs#Azl>3<IPru#jH|%5Lt2Hd2p5nZ4EQZG&zHa6sWR
zBRf}g2UeaaL<H?U@ju4lS1!G)Pu6lb7El9(j``c|&U>F;15U*LfG~C?k|c1#PU9Oz
zMqHBHO6{zFArUOF1kMfMx<_7t50Dq1-xN%p_nqo!-a2tk@w9&(!TGj*p1?A<d;$5c
z2IR3Xje58Y>z8Y%UF;!m`T3nmp1`lxsQ$dc>=82)nyu8eS6{&|W1;qp@ol%^o=1Ja
zFh&a`3rpuMqbgfUqpqcwZNOC4%%kh+DXldzVTY&4zS^kh6Uh49j`0ZogR5|IEBWkb
zd%o}7UXM=y<ms<3NU;ASSA(*dVWAlq&T*eWF8aPe6HiAg%jtioaT=acduKA$JP%z3
z&&;toPr<laV(2T$wC3^Q`HG+MsM=7FJt~Da_SD$<KZupx<F%*8sz(_dYAi_18u>e3
zqoFUeb6lIXf<~kDdyyz72uF2zBg{JuCI)8$%R&XF<#p92pG2cb?)zZc<|DIkAy9~5
z*yIp+yhjtuwv4PMCEMHjcga>?{nL6VJVrB7btv4%r>6{EkGZw!!lT>kpqPQv^(b{k
z#5+5Kh>04K7E#Sm^{b%^_%-5lT^C{~ykVpJkEg+tBuj}EW{y5GEwkq{z++BQ6F?1i
z&$1}8`!kqs%H_y5uD7N2X3x^W#qcBkZYU18d~ANz0JyD*Xt$1)HPORUvKv?$GBY30
zxW=_9$`Rq^D61{kxGLo><u38>WPiOZqgXCgHP18ck}F`#23@}h)pO2J_nz>IsAqs}
zkW<~DaS0T@Blka4|3Nyn*OrPk>c)NW6ZIJ^u^~wIPC{8zNG%~G5H#D&igUlhwA}`h
zuKoakN<}y$k^G`xdR^lo^|!c@3e{#v-nJZ}Tc>}nYv;MccqTt*mqvvObZW@zObb#H
zZ17!)<Ubb6L@#BRrBbR=uW)gtow9l2)^hR_S4PV^-ZYkdvjjykHat?;B80x!Ez-16
zJI2oQOgRlB8OoA8T~`tKd21vGg)j<p9MM3or0A_QypCO2e(sL_?Xprk;Zl8bEbxC&
zTmLet+5Sm0>4YQtXH}s@VM)!0Y<mj&sx8&;J^$}f+2AvX+K3jY|3!j_09Y>WzurAe
z)iyR(a_tz)3*?Au<Q@ouL6Qj#SiOR>Lp3|NRNO)Sq=nz?`RW=KT;+$@8$twyZ#*(N
zt!`H5*z0$y?CkLAXKudV$L=qK-nNMs(!VBIytxNnGTAOl7pap^Z&*q&cKJdef?>Ih
zjhA|H)8eeRsm+17S*a!nfCu5h61Y09&=;4k4mVs=S3}6UU`2dg?cC@9hZ_-lv}q15
zcZmoXL4~=soFzyqV)(I3!c+C>E7BfGu4P?qfFj~X?=Ut56)6!5q0x!QITvTCG{AP}
zlDQ~2K$1a;T-XpVX>LD)lSu%rLoJDoIx^^Dmu~d94Hg)K2QT(loc+gP>n|jnHDg+u
z;0o$ZQets>!Q98S8Bmpx(X|5sj3$3guW8mx`*h%v1Ih~#>~f<599(GmxldTM!pTJ-
zVSPR2)48fjed$Hs&l$;midla#^oQW^q2L|r--gS-4Hf*?<Edu2pIe@h*+pk)F-EDo
z%S4fkNBV#wk{=Y!an<0&;?^przRiI%+3<grdf4QC5A6$D-H!~h!Rn|tDH@-v!!A-n
zX@7qLd*s-v^R~X%_d=a0ry!ERTe(sfTdj!g5%O)ieb#}kp*qcZ>FP=UhkG3|FnM86
zPU3o4Bkj(5?`WSp!r)!(2V85ZOX*2?<p^s&Sk6SFK0MU1y*K?}0&QWpd1PL)5XlJd
zQ({E<U|ZZ3o3KV1qKEcu)cmn9qYq%1h44lOY|0S=-Rb$T(3+ojejfjy8zV>U^0Foj
z7?>;3|Nq7~W1;~hpdk8fWe7aDzCCxkq}N*M$eTx*uf&jc<K!t=P9(7%oKkA!)9Pru
z0N4Cczxg}eA6L3h-+_G8R_lb>s6AoQ(!{NAC&IU_pEKWnIj5rOULdV%r9q1S&)b?p
zCKQ_Ra@0a~&i$5q)3(AVXZ04+xr~NQ*STlDF0XKwx&J267Vtd~gWG3|WE?S`1MV|(
zB&UgAR<n!}R7z<R=Yq%kZY=XE)vuLNGc<8NVRySfUX@M;Tk>-o$?o3oGq$L<Ou|3K
z^P8bl*ZaHGS_b8W4Qhnu;uo7@a>98TfnOC+P=imAwJlS<A_aEb#FhJOLh5t&VX6Fd
zrn+Umdi4%SpI14v<kinL_|^UgaoMd%)tyVKEB0xh)ePU^Id$&w4C=8C6s3XHM`(at
zp3Lr7&;D|dMphmB&75?iBF@Lgs@h}ZKCY;Z2%N|XJYN?|n~60boW(8vY7K6o!<4JA
zA(kUf_O)E65BfAd#DhPUV|rgxXd$A55)aE*ASDi%*oJAc?V@4+1GYlM3>zW>6)Cmc
z5&|T;(HI{LSm@F-1SfK9N~$PLL1fZ0v>=nm-XFeHh&0Wnjb9~mc_)vuk8+r=4mdkB
zQTauWJU;w~jX1+B3@M8>O+bM&qA_?atQ>~3JlQ4$*XH|BI7ml^y5Y4Qi<^7!0joQy
z_O$_MlFWRBZ$jdUeybc7e@C!$1>oO$(UcCtf9uC~9N9g`=Nv-0hi)U5*gaN5Qxg(i
z?GpC8ad$dzQ-dIV@?Rwo@-Mjfe6VeOSCP=rH>~sVZhxa;z&j0(9Kdq%efG!PM=_+m
z&c6O;)T#c|%#FoS^dSnuE0$!P6|{zkfrkMmJ8lH1WvTrsi_m_0y{C3%f<Rx)gcg^<
z3xPH#N9GM}pMHQ4532>AE5xMpU$h?ifUi}P(G(f?5DxK6pMP9{w$aQg9aGffC=$8;
z4rCkCnaTwOqfB-mG)oMA)~YeO85Lnlkc;M@%EV;JXb6_m70PJ{sBx0SqbcMgtf>Mu
zNHxk6;^_wY61m?b>Rk^-b@DXEZO!VxvEmKbScGcQ%1hV7vC=WG{?9|j@3Au|<;=PR
z^M5lFQuShPcZr1L^o?>lW(vu+Z)s{&jZsglg12P?KIVDh4sUlweFM@12L8K$3-%VK
zCQ`zln@2oh^}YyJxBU4w-|QBFZk8$3NfRjDK;Dnbwbr2n+9t1&U<pY$h#6<W0Kp)z
z-!KqC1RL%gS^#(<<!BtJrq5c9?|vrTB^+LbW$ER_bR5XI0>v@beh*)URtn4m>#+1#
zY(=;m*-}VL>;ci7Nez?q7R43FBIQMAnC+P8FKu*Kfe$T$!c)PGh3zyzdu1`sP`-qj
z=;sy2Z{mHQl3T8-E&7X$IDiQqft>Ch$GPIg8^sMauO^5<1VIE63L+MLe29TvL#I-%
zg5|aqMp(t{%(26rA007iACi}Ms=pr5MwG)k_zwscXpdHRvy0A(5ja%4P=8=EDL`sJ
z5`?oYu-AF~4V7nS9#5_V$`N4KXzolj_}%l9x>wceY&U5GzG3L~_ESN?kTS8@E5l+h
zFpihX18ldd49&Se1xER|de1AL=8fx8CNQ%j3@Fc!J#b&<{^B3{Wm^SI(V*T5bFhp3
z8lh>6NMHF}#6gJ!ZIYJcC0#DOeXR67U0)kI{DQ1dwkw*7aO9ByLQh&S&BD(G+Te0I
z6WXW2c(^-ipP<;&EB~f%+><s$&4yZ^kqN5%65Q|02_A6@`)!2mzorVqe043zcXJ-C
zNQ<%Rq7jBCTr$#Ei@NvRY=PQKaAo`J94?w4NL63!U9}>J9sRpglP4=>ZM3l50!+Sy
zFBh4#cd-_+wBocGI8-$S*zHcSFbUTx6lhmAHF4rL0^WMu@7!}Sx0Q8&(|k>+O8Cbv
z(A49#xTylsMs^K%Rgt{DQw=t!+L&pnq=>jC@@@0t0c1O|3uQOsbCmLQgI1o!4KZ$W
z(6kE_3M_&U+ok3Z%##O)_BL2Phn2D4(V6|(euXo3pa;5tV|C|{sO);v6!`jI1{4}{
z|2>EuBxub9RCL;uz#X}k@q?Ynq&D0*Xwl$K+JPYz!{EGx#l2XojNHMUacoeJC)-tN
zf`ji16eoKGixMHoFAooRv+=ob?1p4F<HwhYQ%h7AtUS3)OFK#Xaw77*-rS;hwUDv~
zX&lKct-zFvjTg!_xLOqo`2_@2&>;<2MBSHDkOP!}-yg5u!ihJ_!WFX}cfC0+5R3}!
z!@LTWa4W*omnksRb}g&QWyeN_u^3qn6x$1mFopN+aM~=t!8B{`R;qWrKJeXF-15lF
zs8cQkB~}`QBoMG)0oljSijAter)@bYk&L9K|CY$x8{S*%47=9r9J^`67rPJh0{~gl
zIrb;OGPV66Plg;~CjT!7myaw0>oeFyR8@K_;su|M?fm(_+xF3PI&{aL_v%&ne<sZv
zGRLBxYEJooVxc?_H=<GsBUm^Ux(*%hmqiNIk0cgeDrk0%ncpL9n4=q+R6XKb<8WcA
zHccBKkNg?1`7MezLQc45t$jV8Xi5IaD)vt!8-jxvv;#wifD6L)0mCcC$45p$TG+;E
z0Qjb=kZHV9T%OKf%12@Ah*cDSZ)hr{<Kyt@Z%yU97{qiGl8O(8_~T6?rUVrOha7-2
z%oh<v!veV=X1ROeN!|PNN**tC8G2K4n|^K)U!#75#$qT5y}!9XORv#3sy<tvqP)Ql
z2y0;hm?VU;XD*#@TRn6$99I&U#BZKD=8ccl6)40{q_5*=tj~WV19=^3Z*FfP`1u%V
zI8Sq#XC&_KdC&{$84(EU2z?%1Y+w|D6dSvFwTZgarn`j=h`vbrolrzyXj&FD9c={B
zzNSLg-0EK5UhiI1yxm+|{chEGeLZr4xYx8~@fc~d5#q)fnsB+WNV~(MP<0le*%q(G
zSOXi<6cz$U{zK$_ui8Ar%%$6gQ~08mllRFn%iOT>CmG;y3Ois#A*?;9qu_z(P|bED
z&1SIgKdBU0$r<S3#yyG??Um9JZ>mOrqz8ra_8ob{n4ar{dn(QW{<1G?j=B&7DRC^?
zLutFelKsAjb@zCysD&A`Ki`fYcFAgweGw|JM@$MwSZ8BZV6!0&{yN+5O2I_wAHv3e
zBD=1q&Lyuhm!z_Z`C7lXaf4m*Q}poxa|A!lTpnz_A6#z_^6I^A4=+v@q7{0d6X6$v
zxpsyi93xD3$^5Gf@cDY3ziov8gJ;F^fbR-7zXyC5_T?C2HbyP-_BMihAp_0$dw-EY
zbz2Pg{$iy5=4BgA{wa{c>7AUA;Oj|Fd-2$pL)<0wa^4I|VR&PnFl5#_FGYly;Sx(J
zaU&!HBcA~ms%H#f!02aj{+x}01~efv{}|z~kv1?8Bx|f@d%dhe;k&V9=Q&~a`)Pdk
z)?0U>w3}R=)B=&_;KI{6d1}?VohjS-VoC5@#bXKLblKg4%=G%pzVZv{KY9|XIXE{Q
zRAcW0OeK}f63=$~s|?%e=Ba@_o-oT*QxRz!<NmMd>RNdNSX7CHWCBX)7zZ@MqT9(Z
z0E)VgSJNjsTQ<WE@t6C6E@On90^1R<{AoP9;q>KVH?@5ab?(CbYfDZ>)`Q;<A`>+J
zlg~_uQ;Ps;dKbg8F5#|7pkT%kZRDA|YC;qckfEYxJbd7E;=O>}!O6W=l~;JQA#>dm
zKDdlD^Q0PBA+8Z`={|<@4G?aXtrV?L<~)YOb)8LjBHXq4;X$|4I0Y#ek*YDlY{TyG
zROa5=vRNXR9f{;4t#NMjV0l-kJ0AIj@*HduHBy1cI5c{mTS-NiV#oNGvrbA=LAC=p
z^kB-Fb!}jlF@0cH*|Z0CIiE*X!PHuDnkvqhLF_N|{wVQ!^LBVKRp$ElLmRCqf%UT4
zKEhG-KJtx1R>6D(9tu7Oyp$YH<=fA=)MT6tNA&b$H;PUqt4uYrrB^z98um>3@nmng
z^XBY2RoxUpGDsVJb;zgusEC}^NSsulp)vTtFEZnfS*Y$5sNT6xiWml?m4U-jW|-gJ
zxE9#DIle|AqKVCGRSN@}WwuP|v#krxc60Fb!<O5iB-XzMKWEtch7_QNLs|IXVa#~j
zRC1|9s#LP7hVp*D(BRy*TCXKmA(9b)zvVs%7`KO%|NYS`2Brh9jM<O-VzLhbJZu}i
z&0KXMbm!R3Hj&acOrQ|fR;h0>Ruda<64Fof)&XY72fRx6Wek`gb$8d$wHkNarK#`k
zB~okD*OIiZCk%VELTt;he*4QVou6r->K&61O}8<Ex%0(b;(o?aFC#A57`Ol|Sz4ce
zdpkOQWhUMxGJ;wDkO*9gooyxuj(`b8p#{Kv6GfHya2FIY`Gc2SqSu0Q5i|E|^&0!%
z?u505-T4i6`6UBEAj9MP^nSdQV^yQX&8(C-l~Hmc+}$-9@>s^jN-V1Pz)X$oj+#CT
z2G_>Lvc;mHK<=s%;+O;P582v6Q!~`x8Ec509#}w5OtRt<h#?9Gu@U7BM6&0W=xiE_
zue-Exh_@=-pZl4J{q*-`3+MXxSL@|5W%wQp>mN!hK>*zAFFc>lEPNJlmOBGfieE_R
z8$m1*hN!xAxz((@8fP#2Q>AuiBOzP>CT8uh?ckO{=$R=_#?te@PaW5P+9Nbc*s+H8
z%tU+ornc4ghfof@LR`zxkYqfIaj3|i_oUkHasEf%L^Ap~ltcK319fr!KOA_73VLux
zgGi4D!MX@QU><-ZU1-21K#nd4;LRX4S1DNcT@lF!QWsE;D-&?agC$Fe^P92GD0mnH
zKiK*x8>)M*)M&vQ{A_9l{Y<_FwZ7H1I+-#~!t|$p1colH=4!dKZ$x?eMf#JiBbF%I
z8)tR<SnREtQ2lUvK8E7KkT_jl$d-|Vq!x*~jD{qhyjpB(rNt3%7Gs#@K4Zh2=V{XN
zY%G6QNO!0weGbsXkO6xlZ{|{zW(C~a{}~(l#Asn{c>5V_G8;Fwo>+*fAo%LJfJE*J
zI`b`Wg5%g9qIwG*nwtUJYvXrc)(~BQsefaaW_8c%&kwm$+9BPY>cp9C_>P_H#0}+D
zca6aEvLD$dAUx$KuW33YE9sq1TA&ZDpI4Jcw?FxXV<iyoCTJdq3{nguBe6RH**$_H
z7emf`iG`F6>rh);NcjzB+xc5sc3XKRt*fGe0Xbaxds(yv1YQmgAYKYnpR_pd3h!lH
z(F+%`N!A`Rx7Za5ThSCBjB^f6+Xj6Dth6jU!~tw^klN&5Y)p0)Mnv(F;=eq9^k#@m
z5YYv&+JB_-F-ttb`E~x!m}Q%yQ>8;PvnG=fofbD$M!^X1%FeNqhWLy1ink#@S4wr5
zR4F3LRa7z*Rc`=YN@V4<-`Y4Nz~l_K5q_C+YA$as?4*aPzD4Iy{jAL-C6vp2BZ{3N
zK}AjW)5G69Nu5JSllCn}ja<U9xtCi-^kNPuNXh<M`WMjda$G>QjoHirRFAaPdVCYg
z>iLP_E>dL+Y>&aU<paJa-UZ|@{WTB0;V5?V%Dmu2<}77h5)2)HT-W9giF18(M@E6O
zK8Pdf4$JJ{15c_VWOG`;IpU#pBrr|Ntf%?<(WN=U-+OZOv|OlNfJnDoY;)RdId_V|
zfdsiBhRx!R2#I0W8oSP?j`a!B0}x{AabsenMu7+ZibHar_;<O6lV2T%0>uI-ff;D#
zX@3XOSdXk@GI8oW7pC-|bbpTHyMAxp7i{kK;ah)s9GD)bFj+#B)5gPVUtd-T0Bk|%
z;+q(wLW37wHd-w8I>605j;gs7=$T`g^_+F+>{%h%DDzDCA=rG0Q^j@FFrN4P&hDbm
z>B9SLrHnDGxpchRaM?#3kMz*f0IDy{23qp5IlC8vZRK6m5^TiVleI7wW(iZ@&30`6
z{#7^x|31B7lLn2%kjBRXPZbcwxn)@qg;ZOGK2bPt*>WVvI57}Vy~oiGl<eQG`FLD=
zPii?~H2j$Nd$vCts!9h?32KT&+Xm8-!_F>6ferN0IjcjHPAq<tufsc6lffgjB)rW0
z1+cm6otdZD-HEU$quTC8DhoMd%L7I%K30u#A}6z2#tY*6AT=;=Jq!mlL{CeuncP+<
zvbG2PB4v0y&UpC{{$F=Qry&yz2-VXMSj&%-b3er^wBzr4@sBhb5IOa=IUbo)Lt-lB
z9E#R;+4#MjY}?|*30^^~b0y`#9YO9?*1TuYR56jlY#@u>klVHIIM}7XhpXEBhiR$Z
zJDXwoT&LnZ)Ug?r+HkN2u+QP{)~&E|cDdL!Qa6zMOFd~>eC4-o5s)t$xOw^lWUN7J
z@yWwEf|fahcAP$#Y4=aiupVQumOZl|kn&LuPjsRX*cf!$(ygJLXSlpa($+gpYtamD
zMsFkC!R>x^c(K;P``gaB>Yv{U{K@iiP}Mk=pjUPhsq-IoqzS?=bMIR2iKp||6G=jC
z^6}2&08NxKQ$M0?u9TN`v~B}s#aZw8^DCK`y!N;TJq<3r&)eqdxbzOTG@svCo2#nb
zVgOxyxk4_bqvdc{)_kRMRxkEWH!C?&tv}e}k{D{l&Wir3UHUgfGECW^&mOXis7HfM
zkiuA#O%fDxJA|A<?jfC;(m$HJblYlC1V6}_i1#E7>uiqU=(=dL6_NqJ*U(Z}C`zV2
z14`G_olJL8Eg=g;aaj~gAr3hdE4kxLY;Q!pVZA1oXc|gW;8W7Hl13AfS^?ZY@>8|s
zFq>&MNApYn{bArJTe&#{S<^_*G(PsiMXdKf%#n^j{%K~_!qPB|6x_Ka7OBS4ylH(h
zv}=88jh>6IxYy1eIj|-Vzp0(|3x1oHC78iziW@?C`W>7B($+UCW0U7SJBLe(7%$>A
z#Eh)F&}cmC>PbvErst9J5mF0oatqh*XG3wTXgAr4A(UecHu&L0@MqQD)BQj=b7F#;
z>Epri&8gne%}MayUo#hd)!h9*5|-K)RG?4b^nO&s(6y%b7tqw{<u1v!wV^B*QO>AC
zwA6$hE*0suFT%Rl)|tx=yMZuc?R&`C+Cs{CeuULedqMt4&QTTy9eb865{L95%rhAk
z`TA5A!i9cQZ0sBf14{P=ELvy(#p596;`s{C*sluNn%^vXN^`j7AD>ep;j0wcV6?Y4
zxHADElQGl`(hZ?L-@%E9$0HDM>u3V_*lCI^#edA~-BxB1$J=Vpo=Q|DD}%axsc8(Q
z8n-qlK4<OT?79{^wkIs$g(ExUu#Y`<+vfh93L70`86*A^;DSL&ei8_`7e8cte<{pL
zUxW1g+`*kdvwm#gMj*(aY%o4MJs<yz&k>Y;Ghb?<{)zu6AQ4`2P?QiR@SLkXZB=wd
z?QEFJ<E`4(%3-M;z|pbxf6YlZ!AP{NsmWrBaQ|yg8sEudZ}<9%xGLNHV&ch^&g(Y)
z0d$jU_XNfb>J#sgjn6|Hrek;`7nv+a&Obu`ol8ggmQ61c=MlssKL;T!0SyGYcpMuF
z84YF?8K(jt;iMf3hx}@Lolp~I5XI0Ob17(!>j{b;bhrm(MXWJu5D5tcikm^&agwSS
z2_$l_mO|ZQAC()iVuzL{%%ZeqvBsFNA#X<Z65eyV6Hca#Bc-{P<!d+wWCNK}^4wvS
zNM6<7!`~aF_i=3g?AdPa&6DI29z}oN68E?r9OM97UC+%w&V@YRF2Y?;Zro1_zv~{_
z;gV{~zvt}xaT@g{g~EPfO-d44*3}InP6b<PFDblhNVanP?5cvf4`x!*mh}w7XSh-C
zC4I-yCjO4fRLMUe>(pB3H?`_K&|%p)%S4i?mnf;h|Ieq?hE%eK+ia3VE)Q*EaJFY1
z2;2w0*H`%nK?zlX)2ySe5^)4F(}z)?cazhv|KQS_KNgRnix)a?td(8I3Ix#%>cj)a
z^iwI3Xc?W=_29YI*|>9f+RMkfnIXgzv5eDZ!IPabM6~p0W(l=^V4d4EKrWs%SLUOp
z2V{)uQAYXEB61@#O_i!Ctm?Klpf#mePI3dk_sU3^F2+<4YbK3xapT4XW&y(N<7=#S
z05JD1qMBnYQX1q|z$E-sy1F&_e2RCz@|!3~k5q2<n^*B{MALfpCRGHYs^xjm(Cr+T
zna6pmow-@nwe(S+-POYg=<W|MV$U|zu~}0li;KK<q!%~NI37-l68!U%SC8Lo6hCl_
z(B+OWbjVpi++7jmx=-8tL<&_*wjsi;u*B*hn~3E3R?Bkg%`||fr~_~`y+Wb_If<jn
zzbzslbenq=k1R->Ovp@4$K+fmNk8;G8G7@iXVK<dW;^|A`P)6aDag?BH~qnnCBBx@
z^<yjg&iYFre4OxTHrZk&!2-!EUmB>Dl83?|39z2JCtOnbCfQOlV-oJF<tI-Lzwan-
z{8Ps-kz%>X(+;buXqMfIma7&^9)>F9KP3vbId3C9&pW+(&RVYoB1k5YL|FIQq;Zox
zDWt_Tg&uu>kVa<aN}^r{i{+R{HecA=F^J%?ZlCP35{k#mYXaj=sIlCt5lHmKw*7f|
zIrc}O_U0FaolUt-2Nl$8vcc|~yU7oAq+R*qb6R(YG4^5<`;^`ERVN>CD!7)}Zsl!1
z*v2n%`1Li#_WP5UU4ZwE?Z>?kw}0EucF#RFjcZ%<;9_dIC^!N=P{5%mj&?g-AhIBu
zToi?DJ=;&JKPwZ#hOUg*cR&u;G;+OknH^>agN4B3d2Tp)4S`(m@~c&jChuzZ<cTPr
zv)n|VXH~V;j5*<)PP|Y4%R>#U%NC^yEe*TU28SOY!r)@J$Zfq@PQcr6RavhV798P5
zEy0l4;t_hM9N}dr9>DtvcTXU?{I|HQeeXFeYwT1&7{h7Q8iM`;C&g_tX~Q^Icrxo>
zveA@xEJJtJsPG?vqi?4^f6U>cyFR}gexUphGtf=)&kw-ID}+^u7hCwheo7E3|4%K-
zsw%0LU!YiCTA=X7w4~K^Chu2{rPNBb@}0tW>O!RnQ<4hAC<aPN^<kSjzPRg3sNTe)
z^e{;RON;n#hmVEHHOz)xu2p9%I-Qu+Ddcpl#FASmuy@~!G<sGl3YsWT&O__v5T5L{
z69ft?Wi@?S;jdYDyq6a+p`Y(q7!N%JymM}koKvMDwSZCu0<_i#Arv|?Iv<a^x=UFt
zXFoeND9Y0CS>McM+`CLGEF$yhX{XRTBDARLDK5rR1yzqElY^{zql?HUbw=#(jo%KM
zIeeQ1KaZ?z+P|{gUv4ti4z>nHVEY@ZHDwZRbXot#sZT61k_l|$?vSZ{cl%vM-L@iR
zJtfhA8x54ZUc6bV{{BQQo^YRFA(Q^ioa&J8>-x6aT&$?{H0HQ#ZheZ~zJT-V+qogP
z==Ex$pN68d6O;6LvKt$#fD>7EQMyNbtjxbff(<<W;~yZ>YYRq;4DstX1FH4x11GsN
zX9V%h9Nm>P1zP$~OwWt57o@Ix4l)Y^<Wn*AV@4nkAyB~<t4ZzN`dWM`*bsHKx6vKZ
zID){KP8&Q5%{qDiH0qoNWnH2+wL`5kf2F!ZQG{B?fH7%qOP5ZL3U{St?s_yIEh%M5
zcLM*Ya!gHDP0AwJh`*c_E6Mw626LN(!<;s@5cRygILBOMAe}bb%>nR3L(w{0v(?4m
z@l<+{_5dLC#;(a`@z3bX3oq<Ij`*&lm~!fq93eIIC7}|6w?lYjW9Xla(A*GrP2(wl
zbL*gkz}`-=E(?Ce!2RpbzWX&A&;<0#PYUe+p+eM^KY#vtxpt2;zqcG#XrLxx0R0a0
z#IXBk7K%y&+&Ru7pm&*bDNI`faIlYdM3PCC7|NZ7A_QADySNQrZXyE3%v<Ufpw!^=
zIQ}rUutwy2^XJsw^WZFLZJ_F4>h$=UI$f`$VQS}C@asEZF;D38Y|Y;3ooDwZ0RnhM
zvV(*P7ia;6tzvHuxzWrDm;3*8aUEbyEL}T<5<0mwmEJ*mlO|F{sY;O|y$DE&bU}nH
zML<viUAoi=DAEmZMG%lKpb&a5kt!kxNRfbl$p3j2=Xo}pob&EGv%8a-Gc)JRBe8+#
z24{;?bU{5~WsNsPD9{cuxBW;Ok`X={IhQuX`e?=;sh12#$GML65z$tiv*(^C?xvxh
zO^c2+_~Jx-p6<;9ZMOb<g?H%li+P_+vKevD)s7~A&U|y7p82xi1)Cd`X;e>MzRIbQ
zyc`d-3B4A=T5(f&<!sUH3s!x;T=n&*<kXLq@9)}u>@obL!@CA;)i!i{*+oz$9)?l4
zdy}93)qwyQOJ`k>eN2*UD^iM*!ay6gHsjg36utxS5pl9(5`e6)^_gq+X~;$~gB(?*
zkYH@mu!ICdMll!r&3F_vA03NW-V`ZVBm=_ZWj$f1dPxrC>ZH1QTQ%zLzwH_FEK&I4
z?r2YJPseS)+6NmQp&4<odb(b8`R?_s0w3dw<s(q+%TPNfXZ`rz7D`^i$luO8%N8|k
zyrjT9Yz-oogzM6{6Ks0EKvuc>xHZb0#>KZ41CryW^Wy^=I0H=YNN}5`$6d8Eb(49=
z8gr)6ppq(=)aSVu@rbg;l^*F{8WEXI9g8V8syCz74{6Q)O*5C<{ckm=PKSP$I6C&P
zyack%P=d>nICQrd84)-$O9^@^bb*oz)L@s)!T-*J=xIs3l4hC`y)4;Bgl?04Oax-Y
z2T660PCc>#2TdZNNnB`)Ye_^z1L!qVFDSsh7bEZTkaIJFby!z=bjLU0{Dw5q=SB+V
zaaoJ{qaX7tfvSy)x7w**p}s-s#eGGCq_i@`8DhA`yxFyeC$r1AJ0$O#t$+U12JhrD
z+v9AvrJZOJc1cQJaCRg!%5An~mL=C{ef?hjS(&%FgeQ)!R{QzWc0oY(&bNX2wgf={
z2o=x5ZR~B2S%;UD3W*@VmxDElC23}Mx2dsRWNT-oQE^d4Q!R1o<=M$vvajXiC3pmC
zL!@IefU{CwA#!%6mLFf$MvA>0E%aQymg<%O7h)e$x%MRE>uBQS!3Eh5i=_3!+V%lB
zVIADsw?^dF&fbi9Rzq8Xic0Z5k0leZis0h7O)Yd2ZWzC<|03Jw<@>TidhBTQ!i&(4
zKd2vXmB8leDpmAUsWwGv+VXgfg|SrLm3fYunv4T+4*Yel!@gZ+F|o)bVBWd^D@TO!
z<8X8PQnmDIy3(l&J<e&}lBw$5TZdc@xkl&QDa!aw*sD`sRNot<H`c4NL=DjcmU>)!
z@hHhLnCpg~9!nu+T~~cdI6;8B!Sl}Lm9#1LYud7W3v--2N>bN!_3dq9^fv2=|9Qi_
z>YkZbT{apZiqWF6F+zx9+GuS4fut_K#Gg7+mY+nIg@wkJ4fhQJ>Zx-6ZI~8%o40u$
zM-gS1Rt>|T*iAO?fPO_f<?{fedJ!Cf^se>R<5Rc3JG=H(xx~@$OoEmqZGCC$$K7Ib
zqJR8`+110sF4kSR64FlUIDbQMt9x+S>%i3iT=I2)gzzX1j<gM6x%WIUuw%DBUzyFV
zb-%y!kZWlsuHK`kJo=V<?Ah3j^V=ch?;0PeCi9v(8?E;|YAU`m_drG-z@b<UnsJir
zhx#{q{0v)d8%W`wUiH&5Xp_udt7R^sH7*e{HR5%WYw|VeERtEbxU(ESl%AST>d#a%
zYWNENdKd^Vy55O<yxQi_+RX^x<CTzr${D!SIl=feCJ}Eh9MvSrr4TbuBotfFzDF%2
zidxdX-(29z2qIHFC+mO%&V^Z9@wb+XLU3{S{nA8Qht6OttjubAM-uJmednr}VlqEj
zt6yPBB(8Vlt%=4NCq7$+k0^d*nz!Ul7ReB|x7I{%aV;sBCk~L!v+#@*gji`Jbvlm$
z!Nk5M5~X&(^sF+SrV+IRBZ)H?32GN&-<I4}U~?^6<D-u_K&5m(^}W(rR?_|LB(m&L
z@l%)1)h#;g`}gTOA89Y?GQSVl`qp-t6I+~94KQ;s4qmZ1aD7*1Ikev%yq2qV*YJkD
zrbXlJz6kcG+MhgNV$(B&ja(MzZ@#+)>9K*WI9JOIZ3iS~m&-nFE7;T0I@Ch!U%0hD
zQv;Rz334kk;hPU>P-K=8Z7&Cfg6s2?;DgX6G^#L<2>l^yhKm9c7G`YC4}n;t^<Sv-
z00*_zLQCPIYfjmMs&506+6{xFMkgB5newzCcXlG@%4uvZt*2~Oh6fDzzwf&S4OXgp
zq!yjCYvEwncse@8P%xfpXGiy4Z+ua=57Zg3#m=qcM3{O?MISBZ%e873#G3I>%3;Z;
z9*@YjVM0`%QOypX;rm9KWAgq;3G!?%2cXz9SX=LQyZ!vsuG69MPEvo6S2L2q;iaoy
zwL)Vi@Y2rRs@sdwErPRZPu=O1iRjq}9-8TV%EZ-sy+xk}hvrawp??3qdm-)XLz=_)
z6VkK{jftrjoRiDB{xQ|v2RVye0hTfY^5(40i`eK{z8ZKI$LisMZRb;ytD9)L77w&M
zTxyh!)Y|y5<P~XnV;)PTa`kZ6hBnHxre>>yIWxmrz}P94S(g8H+4Xxj%Jf_t^<6-7
z2DxQaoD-!GD(ufg@oo-DklZF#ulOx(ms<gU*7US=&sAd`wu~^G>c&|`6^;sg0dM|<
z>}9Mj7ci#0$Tl4X=AAZR;V4qaK2bw@9q6>w`_DnPg&8WGL2zc?mf<r_Pmmut1qX3&
zJt|Po$J%!*QJE<{y#umY5|?TYI3Bw(C9yu(_eQnXJuPYulMeOJh!gJTuRz5KLH>6T
z{~S5Q{rX)l^(}qxfsdcZ-Hpw~_ob+KJ9J`?9uY{f07p9`ZDF9OA|-m_Lp>2NzM_DP
zCW1gDi6M}G@Ghv{-u@D|-JAmreZ9Q>#Dfr?4~-43wn$RH<^9I_a`Ji|QIzSb05{1g
zzC^RvF@uQ6alSZo=|p+Q;W+F|g_Wmj<F2P`&5PoGt4OUP)1LNcVRv%xT`A~t+DWgU
zRF<Y=H$N*;9g&708goR3o|C}<md%ST!54ESu~H(LGLP4v3FTrH`pFoFq!A)tg{o?E
zEE~A7FSiBN;XjyC<Gvs6JV9nT8#keuiz$;<wMCOAAKwwpe;5A2y4gjP3sW^En6fCF
z=^nOVF7KGP_BD2b-@%QPsoa@AFtJ9B$7fvQ2U+@-_vQ2TbR9%n2_CU>0K$!h46G}h
z&!LpH5Ot=EF%<2tGQZqrfU+yqce=E+(!9IQ38&R53*Zj^l3`70+nm@Tkk%V?uX<!M
zP=ARp!>3Ove|txNd}cRY+bUgeVMwZuMkhoscO@V89YrH*_xM2GKh(jif%iPSfg3DM
zZ7sg=!gvd3M^+`nzcGV%_?~C=IwV~PtlI|q71sjXUgrWU6z<4eKy3m(F*jD}pA&jS
zbv)ofr}k<PQ?T3S2RJ|>5Pfv<AP*7X!Jy9AKy?)3u>L{7)QJWZ<1d)E_vCWBkWeS_
ze%8e${fyJ6qPQsI`uk!p_!1*6gcZqmX7d__>h|dxrVw(mG#YERJ9pPZx4tbtO}{Kp
z_eNzUyx%(Rm1*aK7_>Q=<J8(K{C!@ZsJJhuAJmfy?K0!6s$Q9Fyu7MQmP-u-3=B1v
zUkynK2#wX;KHul86yqG8oN~>^Me-WLs>X6M3bE4gIYM$LSjDbtt|@Gz$%;x^UBp!M
zd(Y3f;QmJk0|y_FVZpQr-4>j{dc@X^$nmZFDydxqS1Q%k7%`<>8@1ikKjsEs7w4yl
zX}b3;G;2?B2kf(4vvIt*rX|h-1z1?MQCaz_f<Jsu(@3;jja;6e*j`h>9f9K_H>*A}
zdWbOlS8dHo8831p_9ltfN=yz|Fs}Wx5?9w4`-FFu;UlK{5^Ysx#e2%ylrq|^;#+wU
zM+#z#9%?J1>cvf1bh%7f;+RYU$LmYNsC`!DLzI@SBu!<pb8!cwYqP~!K#Yd|3v66S
zFjzjm!B)SQ?8fGSw{L*XT1iq<8`Fa{)rWykBLI4n(iKe125yfT8)bF>vfwB0pyX`s
z*h`Tc9c@$pJ|>ARzMpz{%lA{{G-8fMQsfIzbra~<bVrU~2|AKu;dt`C+$boxX4ECR
zrh*u7c9y?$alSD9OvcPzY3mE2obmC`Fyt%C2L9Xow?tB?_q!wFzL4o}S0?5?x2_+&
zt8Q{jh8QhUpPB!N)?{|hu(K>ECD*j->b_KCzXOV9`k@~hel_yYUqLC#Q9QwJ^RsL1
zo?>rd$`<Un(ve1^mbh-C)kgV<fQL`iuQp3EcB~5^V+sP7m7>&STu{*<AD^S_bjH`%
zp=z7)u37O|SOQ0V+6FhytU1t|)l+TWzF+U)=UY2lJJxIiXRJY!MV=|$<H7`GL+cKW
zM|YQ92EF_fUD-{f^r|8_GQzc7G2e_jO}eVX^Dt*m3pNGU>Rj&{;0X2?2MSJG9g#m7
zbGh<DIi&BBbgI{f)6XBOqNQe=b|)+%oTU7yPc>h%WW`)Mvt^myke1Kb%BBgWPeP~3
zQcS<1eOLvf>ben=)cnC%;l7!J!PU|i??k&Z#&lvA4;ez_U%wGAABVq8V}db8Nj<;v
zngo^iv5C6c@S{~|{gLQ&%7Zq*{ip%=!}g=9rnR3IseDtW9EJgec9kfdUCfePCEFiR
zCcoe7%Cs#nWI@beRPS~%g1%^X&PBsRrqvcH^Cnv}c{1DDI{t?cWk|mp{7LygiUs+4
zk;9>wLEA$!A$fM(z-S@z0ofG06aW3L_+7w!o9b6Q01;FS!bp2^2Rm?U+>Gb|)Scj9
z`(pz%2ot^q|D7aY;O3+tcyodU`iU9Th4FwZvvAOo98UckLv$LvNzMaK%)mhl3R>FV
z3ViOTL9R(D@E+}-cym^;9M5Sq$pS4C1{<iD!8R<M?L=VspB#_~F|g8s1N6Z%u^tD+
zrr@WWnee-{ai0LmfJQUNvE)=B50;7TL`>QV9JvC4Y+#OwjSw1vCsU#Sn~X#S{Dzfb
zJK?8%f<;ji&%#S6H^D;<&w|E3k&{LGs6)uA^b-@?u@lM|zI#lg{2Qcpg#g0gK_`~O
zPh@9d3}U=_LBn}C+wr0o34YC7?)2YGc@_kwjCm%u-?Q=%2+s+od0PS~a*mnpc=~Gz
zUvMe--=IM!0!U_#8O)hwfgaoh9cKCP<4^%mQjP>HpQVHnxf4X)Vw8mXaEyX^5JWSx
zEclD4;bJGUrST*vH)N>r)r9RYB_o2MoRXyakBs#eK}0kCJ4<daf~a|#1#FZ3vpNZS
zgQYVT&}bim9zDgvMi}+t_i8*m_xCe;=nGDn@PfCoaKfDl-6dQjGAj!e55a%6fXapv
zB<pxTgvbRxs1gz^S`YyB7FeJd0CZgtVEfM`#+$`>w}n4aMJ{Oc;tWW#2*=w<EDY-H
z|8EKT-*iSeOAtPft1vk60gnG${z|o=#3CQt-xky_Ed&zK3sx-hf~mcK-mx$S^jhR+
t{jVp-2Vdgf;UCEp{JO}?N@#iUcG*jOZ+~o;{Z7Q7=n@}U#rU!0{{Uw44}<^!

diff --git a/doc/img/pull-request-to-develop.png b/.github/img/pull-request-to-develop.png
similarity index 100%
rename from doc/img/pull-request-to-develop.png
rename to .github/img/pull-request-to-develop.png
diff --git a/doc/img/win_emb_f.png b/.github/img/win_emb_f.png
similarity index 100%
rename from doc/img/win_emb_f.png
rename to .github/img/win_emb_f.png
diff --git a/README.md b/README.md
index 6898b8d..f395999 100644
--- a/README.md
+++ b/README.md
@@ -9,14 +9,14 @@ khaiii는 "Kakao Hangul Analyzer III"의 첫 글자들만 모아 만든 이름
 ----
 기존 버전이 사전과 규칙에 기반해 분석을 하는 데 반해 khaiii는 데이터(혹은 기계학습) 기반의 알고리즘을 이용하여 분석을 합니다. 학습에 사용한 코퍼스는 국립국어원에서 배포한 [21세기 세종계획 최종 성과물](https://ithub.korean.go.kr/user/noticeView.do?boardSeq=1&articleSeq=16)을 저희 카카오에서 오류를 수정하고 내용을 일부 추가하기도 한 것입니다.
 
-전처리 과정에서 오류가 발생하는 문장을 제외하고 약 85만 문장, 천만 어절의 코퍼스를 사용하여 학습을 했습니다. 코퍼스와 품사 체계에 대한 자세한 내용은 [코퍼스 문서](doc/corpus.md)를 참고하시기 바랍니다.
+전처리 과정에서 오류가 발생하는 문장을 제외하고 약 85만 문장, 천만 어절의 코퍼스를 사용하여 학습을 했습니다. 코퍼스와 품사 체계에 대한 자세한 내용은 [코퍼스](https://github.com/kakao/khaiii/wiki/%EC%BD%94%ED%8D%BC%EC%8A%A4) 문서를 참고하시기 바랍니다.
 
 
 알고리즘
 ----
 기계학습에 사용한 알고리즘은 신경망 알고리즘들 중에서 Convolutional Neural Network(CNN)을 사용하였습니다. 한국어에서 형태소분석은 자연어처리를 위한 가장 기본적인 전처리 과정이므로 속도가 매우 중요한 요소라고 생각합니다. 따라서 자연어처리에 많이 사용하는 Long-Short Term Memory(LSTM)와 같은 Recurrent Neural Network(RNN) 알고리즘은 속도 면에서 활용도가 떨어질 것으로 예상하여 고려 대상에서 제외하였습니다.
 
-CNN 모델에 대한 상세한 내용은 [CNN 모델 문서](doc/cnn_model.md)를 참고하시기 바랍니다.
+CNN 모델에 대한 상세한 내용은 [CNN 모델](https://github.com/kakao/khaiii/wiki/CNN-%EB%AA%A8%EB%8D%B8) 문서를 참고하시기 바랍니다.
 
 
 성능
@@ -24,7 +24,7 @@ CNN 모델에 대한 상세한 내용은 [CNN 모델 문서](doc/cnn_model.md)
 ### 정확도
 CNN 모델의 주요 하이퍼 파라미터는 분류하려는 음절의 좌/우 문맥의 크기를 나타내는 win 값과, 음절 임베딩의 차원을 나타내는 emb 값입니다. win 값은 {2, 3, 4, 5, 7, 10}의 값을 가지며, emb 값은 {20, 30, 40, 50, 70, 100, 150, 200, 300, 500}의 값을 가집니다. 따라서 이 두 가지 값의 조합은 6 x 10으로 총 60가지를 실험하였고 아래와 같은 성능을 보였습니다. 성능 지표는 정확률과 재현율의 조화 평균값인 F-Score입니다.
 
-![](doc/img/win_emb_f.png)
+![](.github/img/win_emb_f.png)
 
 win 파라미터의 경우 3 혹은 4에서 가장 좋은 성능을 보이며 그 이상에서는 오히려 성능이 떨어집니다. emb 파라미터의 경우 150까지는 성능도 같이 높아지다가 그 이상에서는 별 차이가 없습니다. 최 상위 5위 중 비교적 작은 모델은 win=3, emb=150으로 F-Score 값은 97.11입니다. 이 모델을 large 모델이라 명명합니다.
 
diff --git a/doc/cnn_model.md b/doc/cnn_model.md
deleted file mode 100644
index 3a67f29..0000000
--- a/doc/cnn_model.md
+++ /dev/null
@@ -1,105 +0,0 @@
-CNN 모델
-====
-
-음절 기반 모델
-----
-한국어 형태소 분석 결과는 원형 복원, 불규칙 활용 등의 이유로 입력 문자와는 형태와 길이가 달라지게 됩니다. 예를 들어, `져줄래`와 같은 입력 어절의 분석 결과는 `지/VV + 어/EC + 주/VX + ㄹ래/EF`와 같이 출력의 길이와 형태 모두 쉽게 예측이 가능하지 않습니다. 이러한 길이의 불일치 문제로 인해 기계학습 분류기에 기반한 모델의 출력을 설계하는 부분이 중요합니다.
-
-기존에 고전적인 HMM, CRF 등의 방법에서는 TRIE 사전을 이용하여 들쑥날쑥한 형태의 lattice 격자를 Viterbi 알고리즘을 통해 최적의 경로를 탐색하는 형식으로 많이 접근해 왔습니다.
-
-최근의 딥러닝 방법으로는 번역에서 많이 사용하는 sequence to sequence(seq2seq) 방식을 맨 먼저 생각해 볼 수 있습니다. `져, 줄, 래`라는 입력 원문에 대해 인코더를 통해 latent 벡터를 생성하고 `지/VV, 어/EC, 주/VX, ㄹ래/EF`라는 출력을 디코더를 통해 차례로 생성하는 것입니다. 여기에 당연히 attention 메커니즘을 적용할 수 있겠습니다. 그러나 seq2seq 방식에서 주로 사용하는 Recurrent Neural Network(RNN)은 속도가 느리고, 입력 음절과 출력 형태소 간의 연결 정보가 끊어져 형태소 분석 결과가 입력의 어느 부분으로부터 나왔는지 정보가 소실되는 단점이 있습니다.
-
-이에 khaiii는 입력된 각 음절에 대해 하나의 출력 태그를 결정하는 분류 문제로 접근합니다.
-
-
-음절과 형태소의 정렬
-----
-입력의 경우 각각의 음절이 분류 대상입니다. 형태소 분석 결과는 다시 형태소 각각의 음절 별로 나누어 IOB1 방식으로 표현하면 다음과 같습니다. \[참고: Tjong Kim Sang 1999\]
-
-```
-입력: 심사숙고했겠지만
-분석: 심사/NNG + 숙고/NNG + 하/VX + 였/EP + 겠/EP + 지만/EC
-태그: 심/I-NNG, 사/I-NNG, 숙/B-NNG, 고/I-NNG, 하/I-VX, 였/I-EP, 겠/B-EP, 지/I-EC, 만/I-EC
-```
-
-각각의 음절들은 아래와 같이 정렬이 됩니다.
-
-음절 (입력) | 분석 결과 | 태그 (출력)
-----------|-----|---------
-심 | 심/I-NNG | I-NNG
-사 | 사/I-NNG | I-NNG
-숙 | 숙/B-NNG | B-NNG
-고 | 고/I-NNG | I-NNG
-했 | 하/I-VX, 였/I-EP | I-VX:I-EP:0
-겠 | 겠/B-EP | B-EP
-지 | 지/I-EC | I-EC
-만 | 만/I-EC | I-EC
-
-`했`은 `I-VX:I-EP:0`라는 복합 태그를 갖는 반면, 나머지 음절들은 단순 태그를 갖습니다. 복합 태그와 단순 태그의 차이는 원형복원 사전의 사용 여부와 관련이 있습니다. 음절과 복합 태그를 합친 `했/I-VX:I-EP:0`를 key로 하여 원형복원 사전을 검색하면 `하/I-VX, 였/I-EP`이라는 복원 정보를 얻게 됩니다.
-
-이렇게 음절과 형태소를 정렬하고 태그로부터 원형을 복원하는 방식은 심광섭 교수님의 논문에서 기본적인 아이디어를 차용했습니다. \[참고: 심광섭 2013\]
-
-코퍼스에서 모든 음절에 대해 형태소 분석 결과와 정렬을 수행하고 필요한 경우 복합 태그를 `I-VX:I-EP:1`, `I-VX:I-EP:2`와 같이 순차적으로 생성한 다음 자동으로 원형 복원 사전도 생성하게 됩니다. 아래는 이렇게 생성된 학습 데이터의 예시입니다.
-
-```
-프랑스의	I-NNP I-NNP I-NNP I-JKG
-세계적인	I-NNG I-NNG I-XSN I-VCP:I-ETM:0
-의상	I-NNG I-NNG
-디자이너	I-NNG I-NNG I-NNG I-NNG
-엠마누엘	I-NNP I-NNP I-NNP I-NNP
-웅가로가	I-NNP I-NNP I-NNP I-JKS
-실내	I-NNG I-NNG
-장식용	I-NNG I-NNG I-XSN
-직물	I-NNG I-NNG
-디자이너로	I-NNG I-NNG I-NNG I-NNG I-JKB
-나섰다.	I-VV I-VV:I-EP:0 I-EF I-SF
-```
-
-그리고 아래는 동시에 자동으로 생성된 원형복원 사전입니다.
-
-```
-인/I-VCP:I-ETM:0	이/I-VCP ㄴ/I-ETM
-섰/I-VV:I-EP:0	서/I-VV 었/I-EP
-```
-
-이렇게 모든 코퍼스 내 어절의 정렬을 마치고 나면 92개의 고정된 단순 태그와 400여 개의 복합 태그가 생성됩니다. 이때 정렬은 수작업으로 작성한 규칙과 매핑 사전을 사용해 자동으로 정렬을 수행하고, 정렬에 실패한 문장은 학습에서 제외했습니다. 그러면 비로소 각각의 입력 음절에 대해 500여 개의 출력 태그를 판단하는 분류 문제로 접근할 수 있게 됩니다.
-
-
-윈도우와 문맥
-----
-하나의 음절에 대해 태그를 판단하기 위해 윈도우 크기만큼 좌/우로 확장한 문맥을 사용합니다. 예를 들어 `프랑스의 세계적인 의상 디자이너 엠마누엘 ...`라는 입력에 대해 `세`라는 음절의 태그를 판단하기 위해 윈도우 크기가 7인 문맥은 아래와 같습니다.
-
--7 | -6 | -5 | -4 | -3 | -2 | -1 | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7
----|----|----|----|----|----|----|---|---|---|---|---|---|---|---
-\<p> | \<s> | 프 | \<u> | 스 | 의 | \<w> | __세__ | \<u> | 적 | 인 | \</w> | 의 | 상 | \</w>
-
-실질적인 음절 이외에 가상의 음절들도 문맥으로 사용하는데요 아래와 같습니다.
-
-가상 음절 | 의미
---------|----
-\<u> | Out of Vocabulary 음절
-\<w> | 왼쪽 어절 경계 (공백)
-\</w> | 오른쪽 어절 경계 (공백)
-\<s> | 문장의 처음
-\</s> | 문장의 끝
-\<p> | 패딩 (문맥이 모자랄 경우)
-
-
-네트워크 구조
-----
-![](img/network.png)
-
-위 네트워크는 윈도우가 7이고 음절의 임베딩 크기는 5, 커널의 크기가 3인 4개의 필터를 사용한 convolution입니다. 크기가 \[15, 5\]인 문맥을 하나의 필터를 거치면 길이가 13인 하나의 벡터가 생성되고 전체에 대해 max pooling을 적용하면 하나의 스칼라 값이 됩니다. 4개의 필터를 사용했으므로 최종적으로 길이가 4인 벡터가 나오게 됩니다.
-
-이런 방식으로 커널의 크기가 {2, 3, 4, 5}에 대해 각각 길이가 4인 벡터를 연결하여 길이가 16인 벡터를 생성합니다. 이것을 hidden 레이어와 출력 레이어를 거쳐 최종적으로 태그를 결정하게 됩니다.
-
-실제로 base 모델의 경우 윈도우 크기가 3, 임베딩 크기는 30입니다. 필터의 출력 차원은 임베딩과 같이 30을 사용했고, {2, 3, 4, 5} 네 종류의 커널을 거치고 나면 길이가 120인 벡터가 생성됩니다. 최종 출력 태그 개수는 500이며 hidden 레이어의 차원은 120과 500의 중간인 310입니다.
-
-이러한 convolution 방식은 김윤님의 문장 분류에 관한 논문을 참고하였습니다. 문장을 문맥에, 단어를 음절에 대입하여 적용하면 됩니다. \[참고: Yoon Kim 2014\]
-
-
-참고 문헌
----
-* \[[Tjong Kim Sang 1999](http://www.aclweb.org/anthology/E99-1023)\] Representing Text Chunks
-* \[[심광섭 2013](http://kiise.or.kr/e_journal/2013/3/SA/pdf/06.pdf)\] 음절 단위의 한국어 품사 태깅에서 원형 복원
-* \[[Yoon Kim 2014](http://www.aclweb.org/anthology/D14-1181)\] Convolutional Neural Networks for Sentence Classification
diff --git a/doc/corpus.md b/doc/corpus.md
deleted file mode 100644
index b83b2fc..0000000
--- a/doc/corpus.md
+++ /dev/null
@@ -1,66 +0,0 @@
-코퍼스
-====
-
-세종 코퍼스
-----
-세종 코퍼스는 국립국어원에서 1998년부터 2007년까지 10년간 진행한 "21세기 세종계획" 사업의 결과물 중 코퍼스 부분을 말합니다. 여기에는 다양한 코퍼스가 있지만 그중에서 형태 분석 말뭉치가 바로 khaiii의 학습 데이터입니다. 세종 코퍼스에 관한 자세한 내용은 황용주 님이 새국어생활에 2016년에 게재한 글을 참고하시기 바랍니다. \[참고: 황용주 2016\]
-
-세종 결과물 배포 이후 이를 활용한 여러 논문이 발표되고, 여러 차례 시스템 경진대회의 개최와 오픈소스 바람으로 인해 세종 코퍼스와 품사 집합은 사실상 표준으로 자리 잡고 있습니다. 그러나 천만 어절이라는 방대한 양에 걸맞게 오류 또한 많이 포함하고 있습니다.
-
-
-문종 프로젝트
-----
-저희 카카오에서는 이러한 오류 중 약 30만 개 이상의 어절을 수정하였고 여전히 발견되는 오류를 수정하고 있습니다. 내부적으로 저희는 "문종 프로젝트"라는 이름으로 진행하고 있습니다. 그리고 문종 프로젝트의 결과물을 공개하여 협력을 통해 발전해 가고자 [제안](https://ithub.korean.go.kr/user/member/memberQnaView.do?boardSeq=7&articleSeq=94)을 드렸지만, 아쉽게도 저작권 문제로 그렇게 할 수 없게 되었습니다. \[참고: 한경은 2017\]
-
-학습 코퍼스
-----
-세종 코퍼스를 수정한 천만 어절에 더해 저희가 자체적으로 구축한 6만 어절의 코퍼스를 합하여 학습에 사용했습니다. [CNN 모델 문서](cnn_model.md)에서 설명한 음절과 형태소의 정렬을 거치고 나면 최종적으로 약 85만 문장, 1,003만 어절이 전체 학습 코퍼스입니다. 이 중 1만 문장을 제외하고 학습을 하고, 1만 문장은 다시 5천 문장씩 나눠 각각 dev, test 코퍼스로 활용했습니다.
-
-품사 집합
-----
-분류 | 품사 | 설명 | 분류 | 품사 | 설명
-----|----|----|----|----|----
-체언 | NNG | 일반 명사 | 의존 형태 | EP | 선어말 어미
-체언 | NNP | 고유 명사 | 의존 형태 | EF | 종결 어미
-체언 | NNB | 의존 명사 | 의존 형태 | EC | 연결 어미
-체언 | NP | 대명사 | 의존 형태 | ETN | 명사형 전성 어미
-체언 | NR | 수사 | 의존 형태 | ETM | 관형형 전성 어미
-용언 | VV | 동사 | 의존 형태 | XPN | 체언 접두사
-용언 | VA | 형용사 | 의존 형태 | XSN | 명사 파생 접미사
-용언 | VX | 보조 용언 | 의존 형태 | XSV | 동사 파생 접미사
-용언 | VCP | 긍정 지정사 | 의존 형태 | XSA | 형용사 파생 접미사
-용언 | VCN | 부정 지정사 | 의존 형태 | XR | 어근
-수식언 | MM | 관형사 | 기호 | SF | 마침표, 물음표, 느낌표
-수식언 | MAG | 일반 부사 | 기호 | SP | 쉼표, 가운뎃점, 콜론, 빗금
-수식언 | MAJ | 접속 부사 | 기호 | SS | 따옴표, 괄호표, 줄표
-독립언 | IC | 감탄사 | 기호 | SE | 줄임표
-관계언 | JKS | 주격 조사 | 기호 | SO | 붙임표(물결, 숨김, 빠짐)
-관계언 | JKC | 보격 조사 | 기호 | SL | 외국어
-관계언 | JKG | 관형격 조사 | 기호 | SH | 한자
-관계언 | JKO | 목적격 조사 | 기호 | SW | 기타 기호(논리, 수학 기호, 화폐 기호 등)
-관계언 | JKB | 부사격 조사 | 기호 | __SWK__ | 한글 자소
-관계언 | JKV | 호격 조사 | 기호 | SN | 숫자
-관계언 | JKQ | 인용격 조사 | 추정 | __ZN__ | 분석 불능(명사 추정)
-관계언 | JX | 보조사 | 추정 | __ZV__ | 분석 불능(용언 추정)
-관계언 | JC | 접속 조사 | 추정 | __ZZ__ | 분석 불능(기타)
-
-세종 코퍼스의 품사 집합을 대부분 그대로 따르고 있지만, SWK, ZN, ZV, ZZ 4가지만 원본 품사 집합과 다릅니다. ZN, ZV, ZZ는 세종 품사 집합에서 각각 NF, NV, NA와 동일합니다. SWK의 경우 한글 자모만으로 이뤄진 형태소에 한해 사용했고 SW에 완전히 포함되는 하위 품사입니다.
-
-NF, NV의 경우 품사는 정의되어 있지만 세종 코퍼스에 한 번도 나타나지 않습니다. 추정 범주에 해당하는 품사는 NA만 나타나고 있는데, 저희는 한글 자모가 나타나거나 띄어쓰기 오타에 한해 제한적으로 사용하였습니다. 아래는 그러한 예시들입니다.
-
-어절 | 형태소 분석
-----|---------
-생ㄱ가한다 | 생ㄱ가/ZN + 하/XSV + ㄴ다/EC
-최그ㅓㄴ | 최그ㅓㄴ/ZN
-가운ㄷㅔ과거 | 가운ㄷㅔ/ZN + 과거/NNG
-쓴ㄷ | 쓴ㄷ/ZV
-씃다 | 씃다/ZV
-규모ㅂ니다 | 규모/NNG + ㅂ니다/ZZ
-가ㄴㅡ데 | 가/VV + ㄴㅡ데/ZZ
-증권가에서는ㄴ | 증권가/NNG + 에서/JKB + 는/JX + ㄴ/SWK
-
-
-참고 문헌
-----
-* \[[황용주 2016](http://www.korean.go.kr/nkview/nklife/2016_2/26_0204.pdf)\] 21세기 세종 말뭉치 제대로 살펴보기
-* \[[한경은 2017](http://www.hclt.kr/dwn/?v=bG5iOmNvbmZlcmVuY2U7aWR4OjQ=)\] 공개와 협업을 통한 세종 형태 분석 말뭉치 오류 개선 방법
diff --git a/doc/img/network.png b/doc/img/network.png
deleted file mode 100644
index ee509fc6080e61b5dd31bb2e98704b01228f9593..0000000000000000000000000000000000000000
GIT binary patch
literal 0
HcmV?d00001

literal 85521
zcmdRWhd<T*`~R``IAmlyAt5_E^B74)LiVa;laQTxY?5R|Mj@FcyNs-&$d;LgtZXuW
z*Sq`ve7~RH<M$W*&ZEzL-ya?C^M1cx*Xz2T*Yo+jPV{A6O)7F0auf<hbwLYjfI^Xw
zqEPq}q{Q%(6O5Pi;9q3!TGwu)P!#mY|L{;5S<EOD7wQ65#n3lvY0}5n@cKCMs{bBO
z$`t))f0|r-)S1ut;TW!ql_4=PvGY-7*BK^o!Y}aioK*Bj8BFe}C+QJ}_hs@B;`8K2
z(h?LTDRZ4qN$&{=uK4yvtLT1XdiP<2aM=gh+dmX`D8;`k{0y{;?zR`dhx_l>I8pID
z3j5!$LBCNh9?XBe95w$x{=v6Af+XMI<NfK`xj6zsK|$^<2?+@_3k!L*>7?)?r@S|%
zjg8In>!bdBO|jWuVym}lBC#0s{r_5|qVvg*bof%{X&<gE43&&l+L2ImD#l6N{P^U_
z6I7b8g{rzb(esiL8&PJ_+dp)8PM@aeF6?)pMWgZJHBJ=-nVPcL+S*#XxSTBra&#21
zwX=&DPK=GEAAZ7M@piF`i)S-Yr94D`_0<;B|NW^pVHFX*kHiQ{j7md|O4JPvPv8>}
z%q}leNlQ!1$jCGg4>M{%rSq(-ZOOvJ!yB!zA;`<i3m@Fs*l7FFn~6rFxeMMTCepli
z`4U-CA<Dhg)YMdY`=@SrcsObC@9*Cun*4V#{QO#4ho_Hj@@VL+5QsEzZEkXM^AN!A
z{`=DRZ#|=7Hku`S+gg;FE2SH&$?ce8P-Fi!xtk`x=qsD!f}l<5D+c!H{|%|GybJ0i
z4m6`|JoVMjozA&G@NXfc|Co9`HbL!wqs)E{38`6g20T<qW0er|C5@I_|9c6X5Njbb
zFJ^Y>n{dTd**pUuh6skAGcKG$a&&+1E4}#b;j$YpKAu4;^V4iw6q=qstgr8KcvKWc
zw=(h}UZ<UD>GU&cznzH9S8qu6=S`tfMFKj{yTe-4`cwA#(YL`$y4ecR8$Ung*VZzB
z|Ngz``SZ{g>*k(xiR7%TfKI!n!|gBg4`^xXbb1Od@Twb&-uldAZ*SlH=@U&DF;!?<
zT3V9O-{+}mCFVv~lN}`_^OrZoDgR{d*Z=z@MGWkU0nCZ#x0zMd)X<!qi7jCyZ6hO0
zuCA``o}QtlrDqKV7{0AH$VYIebXNrQ+K!Hoy>egw@kd_$iQDuWt&11cf`Sy^9PK&1
z>*&Y}K0d0bsbRYRccJHNVD<8(RL&UZm+0lqGylDh!S_<mqakZ^op!@7uI1$`(DJ2_
zJ(jxzq0^py-aqN#!^oZ<oh%s-{K*C%O<00lgJNtaa?bO-b$bF>|6T+BQ9E@xMT;lT
zPQl%=nk70q`jM0i3&a&{e}&I)%Lfl1#w2JU?|p|3NB)Jp&bIgWql0(bIQyFXeX`_z
zsHdi;@=pETsY(^Cf8K)~{M^{oG{*M#1Es=XI~-G<FA8!cYuS`mpxGn-Su%B(iCQ#q
z7~bLDR#<6i>AVc`%Y)SJRH_mKeSIO1<$dr`W#)CEpFbP0$+)A2hKAzzoxi-kWT=ow
zg88|w@2pTwSg8G3U6UKxZH%KnyhFzG8)*WYY^a{n;fV(CW!1pIrg;<Oy7sf9T#~2G
zoH@fRaf8y)(Gi7Wm2y@#FreH1J&kM@5*R4=A3Y~^9*N>zyLJsTM>i>Le2E05d^D6I
z%Dfdi0B0Ot)ok$^VM<49iSP^|mgSvg77eRoH*>g8og&c4*Y{FD?xgCnGi^q6y2MS}
zmQW&-YDb1hYWDHk!?jM%PN(vw$e-E%)S1$5#3-G|att{^duKTOm&9I}R%;vD=k5Qn
z{!V_=NpC)0F6*tgsTUh^Jxm_s_obyckGYQ7UxhguQPk|0uU(7YS{$+K&r|RE^ofZb
zxf`4iEo=A<TZo58qBi)*HJvAFD?R+x)Cf>dDk^A$f`U*X>amQo?Mznh8t-n11|9fv
z-;E9`=lW@aQeH{bX}R!sUvQ>TIXRr!f!o&<_U7{;4m+P-OoI<RBgcp=z%{PAd6bur
zz4%GKOBvyvoWX4SJHOo~-uUmh+TXmXRz5OUJz`0KA|xTPd;j1}ULM_fTE*>L8Pi${
zV<(YW$6j6;(}vgtWJ_SPzw*$LLB52gz4zjU5LQ$3-@jqRK@glu`_Nygm(%e@PhpLa
z<L)<fk>T?Wm&5%XCUpG#z55rwulmHbrhhoub@l_6Y4P5_L5QwqsWVU5-QB(MTY{NV
zG+pO1|B2G9n7<1|RLFeJZg_aOqN>XAGaFA8kMy&Xy>I=<_lO_QNd9w6E^3gr75A4U
z*q~VOrCj)D2gKFbR2({vR-=?~3Ce?c&Q}vieU@RLXu1j-#UK7%Hd55b?DMfPG2}1I
zYN(5gi<u=IXb%n#)jR+0phPR2puuPDe%n}*eQPb=fQG+6#^8rsNa3^j)|FiG@}-)-
zKKjk!w%H5wI=VCwD@OKs<ib6Knza8ULNq!Z9gPi5V3)@T2#~)$Sf@Z@cc4)3)albC
zH0L*I_jZ1Z$YK6QsR!Hr|0H1h$P?s5dz-()(yW3`A~~YlJ`_p`d+gt<qyFpaVyvvJ
zam-@)ll6DP^7HfmIgs|mGNDPh9*w{-YGbF1*iNp$H}04+F){JvGJi%>c~MbO>fghm
zZ`J|HF+IYm`swbkcOx=MZ_)p=fEdh6kAqzgA=6i+CC23%*RG}id)$fLWAriN6x)tW
zMc<5yp1+N}|9r!}W{dvcBGV;gXl$&OdwJ$vG`&pVo+KR|U0;Kbn+e4KKba}ioe(~g
z2-ZcpkQvoHCYb)8lbW6w9!}V}`T64fXyvImjZ@TYG9gG>AfaYcR#OX`_eD0G2gMqW
zwU@%r%Hz667{O3iG-Mq8Z|Sm^T3T8X_57x#aD1>qN=m8>Mg6t&SZD+VQ>(QoE<qFd
z{nEHAada`Uv0?$c?nQ<I+*=eB6k=ZUmu<FU;fMBJv^HBJ7WLtfVAP5Y$&0NF3_{<%
z!@>`(-o$<VYD-8+IN@O>7(vyz6ngox2jkuAo#<a>E)n5a?2q=3EsU?XrIG|qkDV-c
zR4PZ0)L2fCn{|DUF)F7)rNstCq%{znAWu))`@{5p=GOvYosP~SMf=mymu^b+UcWE}
zijb`Wv$L}wN#3ISm?azb;K7OAKYvt&fBr%eX5Mz>G>S}DS9f<NjE45a37&$Z3Y#`j
zpOuLQN80M}f&!kNOlg{<qa)P?nv>#$P&=c_%gdAY!boU3LWwA5mX`tI*xF*Uva(kE
zj;(*1oG{F4ud5Qm*77Q2l3!WNVK5t>l?o(wh19AcKBNkP=S{0$PjnDgr;v@`t1J?x
zjEHRrE*XA1D{@7{dP+XAb!2}<$&8~wWlpi3WZ-nmm)>hyhq(h&NNGqKhXihkRMrzG
z-t~te3%`7VHJtFNL3B|OKUQ6xtdtWF(2F21FUi){*6?8(GBUJ*RqkajIXP34Fq7QN
z69e<=(yUga{Op&n22-i0-d(MtaBlH}WzH>Fz!#6*&VBKk*&v+~FE?nFX6@DxC6xw&
zc8N%y+7s3|#Vz;Bgb>5ed-)rw<g$5fO|*3LQw!C2%G_zE1iF}1rb@p&IXvDIw|R+A
zi%$NN+-;vUX*=j|WW=bds`|vJ#7vgq@0lc)B8~2Pp~XXot@DX$F`+Szf25y?#k#Z)
zjnmmPEUS534Vrs>Kh`+hKKa6T_aWc8`Diuw9lZBcX4WE{JO$8aV6kuYR8;VW%Pc6I
zoSYO6R@>$$>Nd`@Mk6a?pIwkgK#QIfRO)mml<idIo+Bm0IKFk<8rtOzR!jN8g)*lb
z<RUDSCObaraAc-(lzjC5vW?cg{sQ`kQIHE(63vB?%n!w};?@9Pz@P7FaN=J5Q$9CQ
z{}gxo$QLJAcp??3m7-SpzKB=|(Lb24kEwE;*UsmTI7|6Vut?N7qTf(Nk{%v;lGVq@
zN8xaDfLYp=4dTT8&-!&FxlIS{pN)x$!FHZP;*cZ+ju5;%{Y-=oQR#15-%(<&d~10g
zx-*UfPU@ZbaRQ@)#QEmO5@lA2eCWq;r{l?yk;Ih#zH9}d&)!L<>l76i-$SBgkH!Ft
zC4?|Z&dyE%;4?R372KYu9=qB~A=2EY*5lWJtkb4-#b2Ca|KotIZ8mRB6-r;fK1(+b
z%cx+aGC3~~T?CSw1{1WNiuYKq61_Ob!o8)*yB%s;%4pEQ+hs5AwKz=W;o&j*@+J-7
zlhz+x=w#)mPoMIiIdf0@DdbjoXQcL1xDSG<z#R*C_a!aiu$GXxILd&7wNBLq`u--n
zFvzJH{evN&7w~Jp?8YRhVh0^{_~?1=Kg(a2)Ruqew|(M4(OvQANAxcN;krSJS!<(F
z;!;xLyn6L&?}~)~wllN^uRN>vPbH)DE7@(FoK6dc89Lq-tcT=i)AvLLN?%?+ExOuC
zJe|^M_(k2XLz7Y?yFcsSW&F2aB$;n*Z9#X_$<7-7@<wl~)$yJ&{|RyjNtV$sU)tK*
z2;RJT11sIEMsq%Vs>xqX_~v40(bF}nlj!ubDc#{YdLrs~0%aq^&l8!Y=FdZ9mVPz7
zC090hBDcf`W9TW^?}5X0#T_1Od}2QDC%Le&;My=9cT%F~nTpog#TuqFlS$dx)FdP%
z(4V~WM3PZ^cQ;vXZZ1Te*PK+kP9y;RIg#K$6h@_{ygbI&uAPK)y;5uC=_!7r=Mmmx
z-Kh8PwI8IW<`xtX{`{2FIrj4A8sICtg~8(8m4@Gab+>;SY3U9IAyIqdA||=Hw}a&F
z_3vFx_oEXm1F@O_7d@_@Ywr?PMA6698kynym?`xquRzC)Efu3M`pmDzb0DCN<r)7V
z@;0u%UalwmJS*U1cMp$yOVUm|7oHmU-n~nvt@YzcV<Vf{Yv+3{*2<SJ)6&t?&#tTp
z15!$E|Lj<Oi6y)(ibfLv8zEo4v}YX~{9%bF=HTGK#l<;$2E)?;jBh69)~9&cC@X?1
zvLN%0AwnJaZOO325hG2K>BE*H&4oWXHlb*4;k6qhY=j{Tk&=UwY$_2p@;ZEP-00EK
zhF7kHO-z_?{+{M^dwVDaxgS3`I2aI+<R3fT(`*+g0kH-h?x~a-osyPjvI^W{7%sO;
zg!Sw12Y6u$VOsU_W$wVQ3}+fj%J9rgR@9T5pR%E9=Hq<NYG9JxwVm>e;3m=O=QN#s
zWK8Sw->5`Ue7}k9<Oa~KN#T?K{D@^~=p$~+!EBLg>TI<T9i54an}BMB4T;V(;{Ief
zv1cJ)`TF`k(M;sX%TG#A`S5+dKcB=h=+AYBYEwvw&}_{N=qa@WoW5*jg`k+^YXG8J
zAokXF-X60<0JkKc(sUiyEE&QmD6k<YFewQqgol*x?a-|9L`+Z9{WCITX#Edo_&*_q
zNQ+n8YvWR=j$`(~?f1Of6~>gIgpTu<BAn^y&}33jqlGLRYr7?{V|PA`kF%&nQntj3
zyuq!my3cR<Esa-GKogLc4{0glVw$iG^e843)_AYorV{|-)z;QRL%#hi(9X_I%x`nv
ztc>a+w4p!l-Y9-7$j`5Q_pU7F)F~TD7LWNpZW%c_3=a<>2@MC5eO_|1U?098;il{)
z_nFhagf1I;*xK^`suvF=!skX-@td({P$4FX*%ihOq`I5?UA)v4X*8;JHWHi9@C}*0
zTXm<H=%SZC((u9i7myVLaE`DKrADy@1qC;P^#HrRtf;I+CnfbHzkDgtk;s{Na1c;X
zSZFpe=!3=5j1QF<R|XyWef$3XoGS$x**#zz+}+(@Ch7xDW=O;dL1Dd7@a!4!#*f~%
zKN~;Sj`n{Opg4jLJdm_U9)Zuz%?<du^;>sJ@}oz%o7s1X4bG56-~XO+kzw_A+`BPB
zO=!XCA2-QX<D;)VY>hD=lpbuE?4W`8&6`u8^(;QM^gZl&fTt0s-6lNxeqR+c>%)Y}
z8#XjF48MP$4%ip?bj-PP)nkVpaVOh|xhEx*l$3V=q)%sbL!Rk{&5DSKxGa#0;JR0)
zXR{tZMnE=2Cvk2%zleyYsw)2dloaUeCu%1lsSptn`5tbK;Gx9C#7vs}&O!RP2N3LU
zBWa&!#IVFhk~%?tPq!aQmgn77K2}y3jkl;Q+v*Q=n!C|ia$ew1E<%}nynPBI&Lq;b
zN%#5lXHn~B6b~O?-jzb#qz`V2S7#R9Uz3|}k@MT+7q)D~y$Re`&F{wnj^ufu{j@cl
zjDBWz7O&d$tQ<AEnjU&XGJk*nrHMLv09BnxgnyL(-TdJLg;l^BQP9z!n{jNibkJ03
z0)qJQ@o_AKB=i)zUe0+&c7@)m59%`!65wpdz_689I3stOT#Jero584w;YJFsi*0ad
z$mYvy7hH67%W$cgdS|ZpI|oVI&IESA>B?|LzwM<l8F@N%GOu{LPFHWQI#5<}-b*<6
z>&)%7ogKGBYwTcIrHX?_V`F25bql_Eqc1u&H5G+)Do73y7pH69;`Bvs7639<gBOZ;
zy^}cEh%O(74&Kn#(t?+-5l_b+F@CbxFy?el03Ed5UESR++sj{&1A!ze*o;O8XjwG^
z1J6T46Zz_ur1_hGlc*;ni<vIa?^H`kk@hCl)cDK=1Aab%^!@`LUd!WG4Tb*p9~g0o
zM1^%yB#}u^2hc-Lmp6A={_Da*4APPQ{P<W+TbukpMf2P_24Hc%jz?ha0Xj8A0)&O*
zNpR=R9TZAlULG#X0AvGWRswEbPKxv_>gU2~XyS#QUS;A=eLw5G%q<rqs*yr{rz54}
zeuu{iJhVPV2DI3)yiv`~%9QMKC!jR+9(AyZTBAm=d5magX57!5U=CRHHh8$pCMG!O
z(q1B^sHCinIw$x@mXeqN)!f-h0+<4zPZG3b18DT25`qvhsYVvzO{o+Q#IxUVVK(w4
z2ER7{xzu;ku}^xQ#=YN*mi+>byv>_ghu1;$YKz7;ZDu$`gFQ6c@mI1B=Qxj>5qjdr
zPJ{cbYRZ^1j)MnwVRzRTU^4gFvs4!KUTA2p=HC<#IYPtT<{+78R^z(5zD@@}ExHC&
zT5`Lrmxmybc=`D^Im7F{>DZ%i^Vw|mwN!T^9cvxrn#bx7kCFzv<nkvf806-zq3A}t
z{3-yR4&qM*9|t~2OPluQjK<<iOV=1_W0%u`dMdYUa=jZ2l}2-#RzFi;Umq{51$Fv#
zVde^}tY<+8<i(vxV2P?^I9}t)7(a}ks1o`Mma_yuXsEPb<1f&52qdmqjC`cJK<)o0
z<XDpIHuP&O*LJ)QGXUa*12qzzpzLM5zP`RXm#Adh9D;{SNMM8;AFqsM62ZW8cP4Ub
zIymq{J`@3Bu|LoT5SUOxng!ROCzZgR>-@#_q|&Uq5rD_7p02PYFs|UEycYUsSe`!a
zQ*gPbx_~GD2W80oI`^Y5aR{N*`bwT1a3NXHQ@yzMlK9V`KYfc^y(=5PgO6Fu%F0T!
z;^vPLMsTFeqAlZ`Q&L7oj7k`BOJ}F1mk`GBGz$xh``345km&M&R&L%xZv!&W$VVF#
z8@j7v^87oY+FoNT#>OZYI%M%*P+Lj0`^$0OlsuX29Yjy5aXx^(b>=5Cx12#{?gKxl
z3Y-j)%S95|#FPf3UDbc@El%XkF8;u?OAX74fH?r}l|H}~-Oblb)KFGNd9O@F!@}5h
z-{-}}#I%Os5zNfb^UHCduki!CL&)Ou=h=}{j{+9mYhjQKUI-<bm4ur=LO>PcnZ7f`
zr!P|HJ^3oMKzZ)+hf;L_RrOC(cuDk){K<yN&0zJcyt1S;&uoQ0ZsZi8l;=trMRKCO
z{?T^L+m`q5$uB+>x|g4SDw6VKD6FHepP%hm)%D%=4<c*5($i1WqNwxo4dhR!YWLUO
zz6qgfcYJgx<358*;8aS0%&Sr|WINM*FDoVs#~}m#e`BM9p^;JR&=3P0DuCy}`yloE
z&=m=3^(v$B)&XH!Xz{xsrozoa5zt!m?I-iSi&uZ5=cWj>)3xDlFPvFMX}0r2@ZmU~
zEG#&lIB{Yo{bqIx<Zu>E4XBwm@R5(?eT1Rvhyh8A&?Upejq?fbt`r;2kCbx)`U+kC
zQZonGy4G(?ATta*XuG$y7`3<OM@U2z4LE;O*~8NwQ)L2`1{HGq`WWDJgtC~-i2uxS
z7~0XAdIdAh2)~D=c?f-kYHmnuECmYT|5sL4wDk1cCJF(a6_uBF0z_7C=W%@a_%XXi
zJWIrxE6=#Z#Por%61S_auHOBgW>w+8vvTA2&f1#IZ3uQcpn8(q*DKc4YMqRC<-#`{
z4Uk-#Hx22Bl$^X3h+~!-ULKyA;NzWO1WmVfbPz*sAVNW>fs47|Wd>wIO=Edp9=^A?
zcj==@T+fADx6V*=DBvKfRLe(zr{Ut}#xpW9LKeQ7T#A$oniZ3mH>iUCd}p%RdV74d
z|0Sr{?bjY~K#@^VQAvA{*mnWnQ9+}~z2yX*r{j=7ZCxE%iPhDsLzvEoT0FR|r7=t=
zAI7m{!O?59@|H-}WLEfKd1-M+S<W$&u5jr_o$Ym}&@lPSUfX*STtnyN4*U<fYa4K~
zCrz#tw(XqRpILAaDF`Ypd(*CH9Ni^;DO=%W`P(Bo=qw_DK%cAZ<Y)t4h)_gC1iBF}
zC=9c+vrs<K?5XqF=5?M#6imWcdwU*W*FNz{qw!GCrXf@xl&q}t{;xg!t*yBLh73AZ
zHv^wg8cU@Ltn>NL0|OC(*UCyt)cpM9mX_PIv<7~8!S*qEU(MAmYy(0lVx*?Jn!&&d
z)~#@WtoRa^C+1Bq2YGtc^#qIZ&Tc!|iprr^U2DJ1t)d35ove6$>(3{KI@K7Da_`U%
z-RS#Y1MHU?NiCZ2ND>do<;}#DmKL{&+2n#4rG~#QY}`)8t+c<oH6RKUl#IMQ0Sdu^
zAPEF)_G@-^bs;hJ<5P}^3~zrE0M-yJMg^b&qLH{x*5^WeDg4nJ{8(yQO%4PsF$y|E
zkR?=0hR7S-vvFEHu(1enwAx04W?<kgcmu2f0#tA~Qtqw5fB;dqDTVNe2(see%ggbJ
zihHC08C4g4k5&>;FbdI1rR{-A#iFMHiJJsui~XQOD}H+cuPxnh?IY{L^2$3EEwOIr
z365!VNc<)m&Q*Omp>ucc{4U|Ek3TcYxM@<U{QItli1^X?5f&0Vw`q3U(bDnXus_<K
zbo<(7oUNKv^C+Fnn=K$Rrv&SzH|V%^=`GREp7gMtDSsO1(pOhkXS@#WAA&SDySNzZ
z^5r$Rygcjv_OhCW2JznZa;)xSITgN?=mae+KE%Tl&;=|Syww3x$Uf$WZW^))RP5@!
zD^H^?n_8f%CrrIg#6G!C)J#bKoMrpTvsu!;ct8vQu&!Ou(h4ar7Xe0oz`ns;nCUeo
zVSm`ynVAqVyN)Mtu(U2*XaR&>=`p8K@bsz8*LTt8IZ0iiscC7_@@Cfp0u<!#ED#nO
z3UD!?^}n?GfQlk5u<+i!^tGWECC0I{Y}j1Ydu{7nFns{=`PKfg<y2EfQr5Apb2p<d
zafDnXkC=hF`lH2OoLgEt?evB2uXp|Z=a6IuxoP&>H^K;2Zs&9=jH9ey6UWJuC;yXp
zo2^ABUM1@A&Hdy+_Xi{NYBj~r^2Ztre=fP#tf}QR_vg*g>dV`UE+qTS-h5<@#ogBQ
z9o>sfu-6}~vPSV4vTWyGE^SR`Ij>qe#?OPPXNTl`I@=dt(CCG-t7}=sbrLv1dcy`K
z&;(;mOqicPe=c)JIGLyz04f|7L~qT=U_ph%$5R2lg74!ac9Dh_olXErxyWL*Vl7Mu
z)AsavfZ?yQ7LD&*u+r5Mo#Z=DKsfXkxCz7w9eXFH_&T?V`0q<`!PjM$zA60D#Cl0<
zByz-`(oEozkzoe7#w=>B{Bp64hFKKlSiSegk(rWs-b)S&NZyMVl)SvWGQmd*$SH^4
zm;ss4q|&zicc4H#-1%^sId!$^gjQgarN|4Z$!r}yC+~3;RIU3g{>0>DuAu<J_e~~-
zhT<ct)oI<VkbVb36wn1wun^+jQ`XJw)YboZ0rUdt(US@4jThe>tbgDa5NMVE-554H
zY61jG5tYIFp;ZskwsK=Vt{Nc<4+U90Y-v%&S2oWN;@qiSeZ4-U_6D)L<imVVU1Vc+
za(>g$NV6h>WKZkuPv@huve;0{0D}z-3^FEvw?<H`Zg0Q&IrnydrQy*zC!?1;V`F1B
z<F9V*LWlRe;m;@=RQ{A-nYov7e;UuF>o_<#&>Cc}SOM*A^uh!WI2q_`Fp7%Vr_7?q
zYvJOE+#`^RJQ5UK;ENE{kxmsWBY$UURBUPNC5h8WSvwTY)hYkqWSJ}xxbZTOnt<FO
z3_2^H;IPAybm2OPHv01VMpr0NxHVr=YNN8y0TacDZXP*&bn=+*@!^iXrDf2UK#++*
zZ&+E}{rzTN_{PVqyI;`ZP~u@%h~#`$0<6vjUF1u(UHd6vbMy2M+vKZGuNC?tPgf^E
z96|P*``RA63(6THi$d!M(o`&n2Q__9FEYWl1N`55cq;5m3N-f6`rG_A10W{%SXavh
zQ?JC>&qauW;zY?NLkscSyfYn~NJh_R4SEVLF_8%*nsHK)>tu>7JW9(>FnzcUpiCGE
zN&0?{^D{P!H1L;yCE^U=K+R)El0JNgKYb38kcO0j^hnuHNlB@+X^SKo=HjX6{P^*s
z@AlZu<j0TWAu%E3*74rRTUs`@xO)U7p6n7B72r+XK}exw7M*#v@9gZHF*#*9Hae;d
z5((%#XoKVNYA2vpqJT2@3Kn{%GXLvW^!E>GcF;=0rZXOGl?U4Z%9<Z4VS1+fI0ins
zr6cmnC3Lk9;1L<~3ii6|=?K&;@$C1FPUtIRSfK1n*CD+u2xb6`;}rn`0T72>8xn!#
z)X2Y8Nl6z?$d_>RObR0>#{zW|gzPrJ5$@Yd*S`JuVSBK<MoK|J&z@LhA<7ajb7$cM
zbk9jSIW#B`RA%Ble7I7zxs$cRn{=;Vzn+zm-qmBG6Sf+ObiBY5gX`iQm=P+xmra)@
z8<+%A$q7iP-xcZS_mx{I?e6c#CTL^v0m#gm^mCY(2`j5xoLm~zMRY)YiV{cBuW97$
zaVQ?t!59w@#_>T?QV4Wm(;iC3#>QYCnEMobBys=#eN*m}$ufYx5drP%++3y8h%kI6
zF_<hMp072IkDDUg)J#iQ{D4)7p}@P&&bIb;b$5_AKG=CswX%X-zB=hMJ^S*LVw=x&
zU=;7ght_S8RER?Z)jT!D0R>d*!x>D_?koww{i*dM(1lw+ek^04N2k+7=tL)E2U?fD
z3M+k&AUBL7k(BtEJV|o<tBW2LRUC9$Glj25%Uhs7vhR5$whMeJG*9<{GdKll4777u
zSy==UgTmL5z#a<_7%3$4eYw7ir8S~AKeF}rN1J9<SD%9i6?LCc=0DNcj4R@W#YPw{
zXjxEYdi$BMGSH@~J~bfI)x_GT3mx^~Qc~zG{eJNx@>$o%j~&o8SKR!>u{8bG%Inp>
z_Z-O=aM1LF0RW{82vA#Fo0fs0y}d7bg{x0i!;278<z9G^rcHeS@8dA=eSN`L6G0ns
zys>kBY%c~?@SD>vsbVmY5p#iN0&YOMbk1w08IOmTw*?FrAJRlRfxZ>>UNY%<EYCF8
z5wB)$p7mr9Te@h}+X9b>nweM4BWVVb8WuoBMWxTNniO~}F!F$=*_Bsaasj%Ed@+}C
z$rqMQEWrQVkoBWQLA~#Pc~cKGaWrtUovX8LHg0Z-a5+j&MK<lH23}Z9l`?cZMWv+-
z?D@SW^<HQEwiY6il4xI8G%&tB+7o0vYZ?wrc1IvWh9N)-|H^&RG=Wb@_$|G+h?i4~
zpO@LR!kp8_UK6y&PckZQBbr<unSfoT0Y*yf?)taPj4%}9n+RBKAwl4izJ*0RPzXXm
zaZpiF?f&_hLsY=)C=cx+XmP~IkCCEz>sB#zD>r64NL!{o@7!tM{57!lEkzF_DfysR
z1(RoFEk@P4y`7ue?{FDVvu{IH)%}4Ox_$z%twezY!VWF}y|6F>NfxTFU%yt|Sv2bJ
z4>yJ9%l2D1V}JcRx2Pzs++*tQ?rz`N2&Pq}g+*D5r&rjvlh2+SmUf*W<>cgqgoDz^
zUwgKH5jLBJB?dsGzbh(uf4OFLZLLMHym9@>*NpRzbE!#0oQYA`+FQ8~%_wxSZyBEI
zW7d@}wGNWAlaKQ7<axiSuV(>uJ2EF{`s~iHi8={NVzoJOCJ1ZyJ9nB&?odPvrFYw_
znwzsd(B?&COOdw+bga^@&Ymq?=)<>FRpQQ5jVBT244BW1$<C>%H>p<ukhw96m4AMp
zD#)E?_U=lZpuG2zGB|3qFJ7z+_~VIc{`goPIXOrwT%Bowk7CHWj}uY>v7b@IG7cfB
z>^c(=`nkfclbVuMiV`8<!0(0CJYhTA8j+;^6d0^q0dKcfSD~K2YraR|_D2Q^Cc-{>
z)-?_dFJI8sR)x+Qp@@TzHgX;*-1SV^xAI>;_wCy^gxiitO&zk}+=2Y;I~Ol~+0-;f
z(6~%te)PtodyCW{%X1)_BHU}r^5*)^Vx1m{;&qisH0$0F8~Ff2nnb(?0rV$J0HKj<
zXHNd?%T+_69_x<OB77*sPU-#RqSE_7t4}sa18fWjLJQ%ze*OB@)zgFUFHeRX>s-0f
zA-8{iL};L!AG7Xt+vB6)<{mBgxMHyn-(l(<EiEncPQ7{#M=&31kHn20Wg8oA6c<o_
zD1^m=CN_Bu*Q5P7ml(VNGq)B+<g_yfON>RJ5n_beg$j|L@+G+aqo)P5o(ku&D(^2`
zJs&=(0TGsI9@h_9dvkH*_v=InmiR}~ZX6teoBa<HMzt`>GbQD3RN-A~+h5%5AkoDe
zUta+V9_}Lw98JK>`|5RG($)PTsr@$D*qaUNELWO9f3J^d0i9I#colOzU9J&%QNgkk
zBRTZ5kdIoS<gdLKEVl;med)$CsCX1p0UMV<qTgJZWJO(uP=K~hxnw9NMrZKZ=;)~R
z?&>TglF|A*f}Wn96)s<-z<6K>GFsjonk;u@F*!Xw8aloQ4<4XU;C%tp3M}aCuU}_m
zxPg6V{LJO-1{qjT4nJAFXu#O6t(Z4SITW4zNsaqTbi&{dUw{AUSIacPy9B^L_wVu^
zTV4gI39cN-Uq0vd%*sddfd~Q+m84xRZ4T_<`o;$62eUta`gB$sgIj>I$uf8#F$qFu
z1|n#6XD3RBuVL{g6k{N6d`I@DDIgeu=<oWze#7ecczD8TAgilOtMO=W@sY$$YD9=c
z#79W^kL2zMeDjnAui4@k#iJU8O#w^8*W$wL)|%-c|B=L(_E+njWDIKQQGLZ~Rt^uc
zSUTA%WVNt_$OW`6U22EZ5*mCk#|g|)DDXF)w{No+q7U~L%WdrJ2xj&S7Xd(lhlKuQ
zW>K%l?b~g@$R&V4g)rWr$}+J>Ydr#Ve6TqHHd!+>Gp}EtiI5Zpzm*k-!yU1TxJXMV
z9jp>e1??}(AJus;XUrF2U9wGc@5NR#vFH$gvpUZ!moQc<<|(VFc$CfIv$uJRhldBL
zwg5pu%s`BNhzkRB8p)NBzCKj7X*_71UpGD}MWv-N;p5|j{FsoJhXIQm8pyGhmKIVf
zstCxDj6x=T1;hTon@GT0a4Ay?I#)dK@Buxl3RIrQY%3AK8Yx#AYU)T3%0Rqx@BF#?
z`J#wQvEkQ${i2Yz;LoPt`5T|Iso*LvUc9)$xj&zus9I`7P<vz<G*?+yR~Nq@06@>(
z+Z$oy0s;cyJd-$&zxq{XnJJ$xdz%lRfM5;E!5qjeyYRd2>)$RkZjU>G4YFx&&JDnn
z?^-((ghv#V%Bih6p~=}$cCe|fR?<YgwJZ2*uJP#w{sy$iR-nY=pvnMVw*x+q|Lj?H
zVA3j~o-6E3+>YOm3J<piBN+~Cs*f^?mdSw%0x<K%UQ$@!fA071#M)Y!IF9oy0G>2+
z6a&HPiE#{*3GM3v^NKR2ZewUgfI(9gJsA(+{(+&Dvg?=_VZ$m<=knmlh#j=eNW39W
z(<AX|V#{u0W23*?$#`&ZaAwHZN*%<Ss;Vl{;G+QOzDCE#rN*oZ2M{wIWc+vF`O!Bv
zCehT?M9f$a`tJ%Zr1X`VNh29G)9p=5&e7&+h0Q*-_$wB_w)_CR$FWLNpVEkHesJbW
z1*H9@sU}t=twYuaSvCR8Z`{1RM2o|vZE)%2oSfh*509k)76HFyeRS{#szN(_0ytmc
z56F--O_tNO&syE$S<aC|U?JxyG%QTabwWmq=YGeJc_hj}hX<t<=?9SJZT<TPG3QZH
zK#Wh|b87-!R^9>637O^{0LZn8+n;8k_a>#L*6{RPw>V^<LOjF>v~BSD{e!0Q`14qU
z5_nPaUit+0>(BehS>#7GnUm9epG*|s!-)F;0ESKIgz}GvI(P0|#p~Cu-qXM&)!dN8
z#2SEn2A`Sh^Lwi1Kw86I7?KEbwws%qZv<;5vYgXa765{l3FN1AxMP2gYBg-laEl_T
zK;60bi)AyOcks+5*|D0AkiOHy5Z1=H+%f_xF?f8SL{8WrHp!<$2O%IR2)!^lVv}s%
z`gPxbbN(J8JaVRFX2zDk*(1O5Oy}0bGeZH7Z(V30ZKYi|HaFh^J_zm2l%9UYsO#)3
z^k#oW9@PwyjTmJI2MytvatR7jfCcrF;=a()W}%V|7$--Y{G}tQSiL{|6tjItiV&@c
z<OHq?M3J1{a$0>yulFQ37Z2`N1sb80&L7{tSx+_IC8eZ%_iLb#wa_x~Te2oXfWbbI
z0OR;(+3B`4F%?TVqyj09MJPFrY(u`%h>Sy{xQl|D7XWcAIwK>)|G8VY^MmdAS)avY
zpDu+HBdqQ(>N&(w=a+}vZ>A*S#?!&Og@6RBpg+9z-pc}V)XD8*P&xo*L;)9?2b-X!
zrBz#Uoe4L7|M_zPHaRa+IEkuPuQK@Uti*#FADy5Mp#I*pDv16m+V-SgF&*9a!@!@=
zO>83pJ?&pe0(2Er?{}b^iGeQy@gP7LqRNhy9b7X$fV&Os8VEQAJ}SPqg%#n^!5Dpl
ziD`P%1UMH>&cIjQQo+H;9mi$QpSOZVf;H2SNyM`87c;OL&YlfcZ-Zn&vebxU?k^DY
zwhK&xK%H25>L8|cxJkdTxQIY#XposHBRkANVUqzwo^qJ!@3Qi+-=VR*6TF1vDSVey
zfx<b0E_G@|3^y+46G3^;QTF4|$%^tu>ZYY3cROUglA#f21383&fdORJYX#7SL(t4j
z)ZGRSOa(|EVoK%^uyTPTVWkMF;;mb^pzY8A!3fnnImr$gI5Y)~?mx5z(Pq2}oPJqC
zvN($GvhxnU%mQ_(^I9(NpQwb1I`lp1_yRlK0jvXfdDA<#wi7tck1sKVrKhJOIULOU
zku;oBOKZSVX1LIHbaY5v%cQNHnlOFkfCe9Z6lk7&JCnQj)~7k3W8<r_v$aK;AKM&F
z1K*E_0$RQIhV!k{df9}i#=V6S->o5IxBBCQKO2`^rFX_B;?Fq^TlZ$1oBvWHvb{1H
z%Vg=B>%sGV8E7Tj!6JQ8W=fMf&+q5s<#@zTL*x9|<sX=5?RJ;l?@P(-dp^Jt=<_J>
zhdh$;zyJ<Ox_=+BKL-F20)<x{jv|x_tR_$W1+W+(yrQp4DR8&GzB0f-Ls=s%JO%it
z4Z|9acsPNUg##1`E8Ri=7K`oe=U)a&BJt~ctK{G<9#so@Kqk-P+2t9LmJ8nLespkm
zcW)1PUd-v!Kc5A{3z8%as>a+7DC%<!=S4+VpSjTU=`=&B%N=qQyR%^EwKR%`*zldt
zjmkLR;WlE|R2jKg`a0Ir?%fP{XC{^gOAU%&bua&bx_0L5+4jNWE3@<SxT_^bv*2fX
z2TsCOIL>w;O4kG~0ci7B7(By&B8Kl2lUtmId5tT3XlUpv=;JehPwIa<|0vS-p^_3H
z=@%;^0^5jJrGBeDe$dci;f47pk+GpjUy1P(1>YSw8pFYGg<uXQVe@7<LSK4)Zi$OX
zNlAf8Aj-5<1Nc>Nlm3;x-~^pVF^+L<Zf|!0x~lZsazt!izo*#$Z53f#xIrANR!Mdx
zTfIXXl^BSd=ZA(Y@g8VIs1Vg$WyC_Uzqh-qERZUdaDU$ozzlT$Z%Q(trQ8KA1%2{l
zG!(s#<>Mx)oI0qn_$W#a1!l0F+)8GFs;mZmG&GomFz(R^1RdfUN8%B<Cdq3fYc&T4
z2cov`aG9@JDTxmNX;oBLt9yBQHKdz9d-jIT>{D4t$uolbKhzQCj>%Sq<LBV`m%Sd*
zr2DZ*IeNu1Dhd*pEgVD~j)NA|U}1}TE<QeDAid}Q2spk3JPYyw4HFX+KZtGM+?vO$
z>XI-8&y5O%nfKD@J&@;+^9r)uJE*(SFbmUh1KGcTfg)g{bD(h49$8C>!>kP}>lb@7
zeSNf)+Y}`Y4GqWulWWY7*%7~7jT<L&DlEZ;F;xTVG|U$O6S9)f2}4ViUUKa4kQvTa
zprBbOvaS^voPn(F%HlExeqU7c^wG97ST0og_K*4Ar3xA&<ylYf(|%dQjb0X{&p+?C
z*$k{F2?~~`9l{E+w<FG7_|)mMO7vhniU-bO`c)Hf4gmQGh$)+^u76mO&yf{N*U@UI
zIJrm6uESR$>(ByZBb{@qXiNBeVtQx$gO6I+ho3-x(>uRqm~tm4y68Gk^59{x@}hL?
z1Ya?7Oo3Pg&I&LnFz3*9huDLE@UzhNi-%23=MimAKVJjyuRk0XG5~t!U1;A|yHDrL
z{KLvJW`Q)j%pzt>2t6B8?GQi${u_j}yli9?dHut~$t`up;KM~L1l9skVgr00K-P9B
zIEYx<JT`XHujweE6WBH|%>MN$8(Rf30IY6#b}19l0jKvR&Em0`7U*2p0FZxe+L;Q|
z$ZvthG#;Uo&YU63lyYfR1HT_)T>uTQ>;3yypcJEl(31S}q=8lX8Dr{)@4E{n<rzAK
zasYw?HhQH;-vr9P*;^2VQHCv*^H*c<)NhJ{Z&sDJM#DJg9(Xb+8m4#*b6W`c3v3H9
zpyE-1I==f;>G+HsG5S#R`gLC5mhUMB?r|w7n0v8RiGnEusYyT+x^FF9MS<zs<s~=+
zFjYE?U$`0(!U)DXS|6S@tAIu6C|-i2ZVgwFVF!ss3by=()mxxv<~Uhv)Ya8(etmw*
z3?SWt^E5mk!rwyrSsONU!>d(f-+X?x1nkb{h!JYc3X?oB81SmHvIY4m5N84F-KRmB
zMZg4P#y$ulzWr&3(8F!!n(?5y?MW$xrv|?+Eiy{^#Qn(b1f|U#5*Lhv0V8g^Io}UM
zKdq1?*Fem+gXR+<1Li6b1p*j`sqIHK8-+@TWKiA_Du9-r{`SBj;tBzd3b{L*iCT}H
zSFXR~N0c{N5=y^2x{y_gER&#kd{3*Mbqlh%R3&5(O}In_LR2y-lu877!fnq$HvlFB
z4<4pNSK~7C7C?pw2m=^SfMR1~O9Gk?v@L|=!<KeKi2z4LClVajKQJzRsi6T12mof_
z-ewEu@vbJs7&lM&VjPeU-6Wg=AqYbHR~7*<5Au<hRVQMC>4;~I0-jd+sR1_^7Yc<~
zZ6(WENX)IKn;j&j{Nun!4wssR%nsELDAaVFXMFn*d;Y$^n|56N4giRYd3=4B`Fdw%
za{rm##7DsDZsQNLvf=>bcYL9XE*SfX5C_OC6Of%in}<Mrqm=Xet--rsQtQqQ&qM^&
zCEy%Brk4{d)pe=Jaq-^!1n=$OjM#bq{{2F>rS$C^qAW>%M>k>mj8ML_Wk6UNJa<qm
zO95_BQ&U&GezT=60*i?-07pNAj|$iPDbUpkpn0^y;Rk@O@&68Ju_yK<!3+>^Y*deR
zNl-{v4sJTYLuocZ`b7y<smM4E7Ikbr&c6HOq{4dFsrl)*N|G$`Gxk_}YDx_P^7o%w
zN4Q^1#>LgsL*fVF`*3AW>*;5{D|65bk^%kHdMZ)DQ?Ii!{8z5qo!l636`Pxz3l`?e
z%No66O|2InEzY9-IcDa?M%UvYKI?GxJbvj3_oNLhIXXS3J(g*%v}URmRuohzNPaYb
z+c<E9{M!EwtXing)LjzDYGC-omcq;<qE876Q~Pf(semU8k~!=P8Bzqn4gx*|TK1G4
z98@~EtIJoe5R@8GKz>KeomyP6&a??9C5Zp^bt8QJAQ-a1>ICI&mL}r{xG0{07}Q#J
z;(2kg3iO096bb5ysgkg;uEXJ@N6a=`B_$>LmoN9pWV|>#n%cl}y*lkl(wW)oO#E^Q
z>L!i8Vo<We5fKrln85Dc0%!7X+x^sF7I^%Z^Ehmcy?8p;y%Pq}l#>C_!o6ND*Gm&g
zo&9AaLg8;!UQ*Jzk-tY;0{@{h5wgSW(n6s?oZwGNZ+S8yF3;)F&R1!E8K2^{HyPu8
zUMG)ZDkyXu7rb0nKt~^ccxQ2Vw(e(k7-UbVd5E2sUEbRo_yDA$8I_rL>+@h;;s8<r
zK7(!`2K*nVKvPD}7rb9cj-^^3-mGR}Veyg*(r?{EIW*d0<n<@_$wH`n)`%g`ca<yX
zU`_L2cP<eiA}<UJBBl{=yh71OH){V-QczG>Ee}4tI2t+vxEnd{z`_a?1OaJQ2^A1D
zEsA1JCZ;I3S8mbEI;RAnz5|TEGG`cHdKR=9ux_Y2PRG=!@xONz;oi~aV-2U#f;xS)
zp7z!npz`ICp}t>UknzCc+yZUvT}<Y<B%lag_B~8bwFFqw`#n~=#)i-@6$9B7@z}!$
zb0oi=<2gReSzTw?IFfZSG=8z)NkO`K+oh64gFcChe#Rk;#z;p~lR5W~o|0(&r+t%#
z8JAuE!z2BXo&15?U;i#H2;hi{2>oAjax$vSN>T>8j)RSyV=NFOin}ueEoY*OWD^n+
zkiotBJB!3kmNe-)d~$TvV&J+$^rr5V(=hHE22NBB<VZWV!MP(fS4uA&azsEXBy)r>
z021UKqR)U$W+sBkQW@qb^q$P@JzxOdlE!{ddHCj`KX0n`mkEaVcnSJl!be0fB~nyY
zE*cxDY_<ak)Ee}81cXDk0>Ffpkuef9xmM7EycF^pY=wUl(G?SVQ+vC6dC(eI38b`q
zY#~edcy{DT1ADp-`AtpM+PYUu(`hB$;fLSlPAsw}@g~pNkeplkX3$k&AR|B8+Iyj=
zUfH0ypf|9LPXCK15I#VbmRHJ>q5xl;RN0e3B3qq};>-hk1{AeN%@43%@-=3jxo_TZ
zXg{rD7)WveJp?pe<R(2l#6lK-nl3Fb!=y4-n%UAXD-34uvCt8H9<<EB3XcLK3K%U8
z#_mud8u_!SM%Y22)Pv*EhK-Qt3kQKSpC<45kL_F{#7Z&}8BWl;@Bhef%DDFMT6}8!
zxl8H8CbBoKT)EXCWtwoVRQQ}@Os!k(jB62Tmw>$a@CS?F;K5ziPAZ07cacwvN-MHE
ziZtmQGQV#tc5>YNXwy}y7`60%0JnaUtN#8aupX&`b^e)7I;Nf+-EWa{frokKMXO>d
zH7~J*_pLix-mmE`VsY~iC0Tm!h)_S!;z4Y@NU7{Pl{nLte9HX$y{`Arry}$u2tg{~
zs?W-LF^4Z*e;d+v2iT<a0W8kcY&-@Kityu{Fj4{eB832E3RD5tT7P(W7FjgF5h2rM
z6@Fi%Pb2CRQq7^TxPpiAYLUOBXUQk-!gWea<G!lEED;e*WmHc`hI@||4&`p{aOb`x
z(R)X$zt0FbAGr6t;^JQ!)6tI~KVDm|nYINUSxZ;9gGT9qtB4mp84a=%_;hD?*XGm|
z6>I<F1&{$43@)QYJObj?1XWfekO+Wr2eO7sQIP}Gi$*&@@EN^)iOitY05L#_diU;~
zHE1Nv^4>zQAVkuqFix|%+fga++0R{K;2*rf-KX9oAs$_{oZiF7_Wqb?HoIwNrTHhr
zKyNoM9i`XnHp^1#Z38k#fkmt_kVD1Lu;L+3ctmxzZI5w6sxs?zzy{Sby_^I<X&^I`
zI$Xf|fb?_%JX4a$xH-~g0nY|8#jM~f3Jgm3^)8pokO~o|1Mq2VnJN}*)Wt^<4zg$)
zwESYwQDl~V(8YR5<gWC~SiYhkOT8(>#j~LN@xhYjJt=xq3)z)rIRd(9tQW82&u;+7
zvzEDtti{Z*BiwM4fNSQWID?d7C<I7~T%(fmv73OHV01(h$!0T%^)BE~)z#8wJIVAJ
zRw>m{u?gqOCuL;`ANG@JflGKKew`dh>ttxIApj0QRF(0Xr{&Yl@}1<3EIxE}a6o)0
zAoU?Fm(%NwPbn}R6?uu8ROx8f4%Hk@e|i<nbr=Z=dT`DFBf~7`#ZOGl#$OAj&9IMU
zlQyp(^^I+D^k2ml#XlJSJ{NFJjwk7fUh}bsj8-q-&tpNXy*dofdVL54IuWvWhLpp=
zIy7OQYf8iNe~i!GG$W0>(6};_T67lFOB?9(k;pK(as|>Eh<w<Qyc+^%&a~vJMIoS*
zIbK8SsR8J7A)itUGGTfP20UB8xV=5g^A?Ue3?*wxp(jsfveB9fF(OPDGI88<W8>-<
z%izO13NZZ#1?TBi3%m<h7d`GP+!GSIUSJA3<(E$#wO)flt*j+PHH;g(BePodS1dVJ
z+&xc7Q_iYu400f&Z0hQ{&|Dy+PhnwU^V(kYH(s7Z8g96`q?DA~-?!k~2tYHxEu&Rq
zFO_LiX_AzdcA5uB-WJ>07y<0?JBSO;z)iFwEG^>k1g#D6CP3ms6c>>AzAl!(jePr7
z2{DaEmubVvUB>#&dP}k_%&g4Bu2`<+EhbTkHpit@kz*g3+qh*adD8<fW&oQ7;IX$W
zuYrC8o=nm06}t)032<5nP{{DW*vr!(_aoCfwI{DpfFZx@yC)**s;-0Mb!}_JilUn@
zPu?!&roclHtAv>uYsAq4wZhnPXLAz^1o~Pt0nH4ErShN|G?%zRZ%vx^JV^b`HvTsQ
zqOSzMBMdPN7SYblC$BM(U#BX1xFFU%)60!8Jk^4Xu|yC{ryQM}kQq^U?u`}Wdn<K6
zIZiM#=IP}qg8lFw<c@iGT>Clal_g_f!o<bJMei&OmPCdfyFf^YX7(DFy@yH8yo3+-
zQnQ=i15qK&Z|xlng+p1BOL3xg&0{aHP4937F?93RbQaNYfAdH>dDMFYcnE7`papo`
zs~61-KD;y$0!{>$3e0~tut129H3|%<gM7}e%KT{`xKp*=>dW?mUo$W<)m%M`sq%#T
zPjeE<eO;{W#ei|N5ih3fws^+X+Q1n)iGf4Pnio!a7;ot=E_nI0nC;uVM`{Q!<0uUZ
zbhhtTGK8fE{~#Pvaiem}c(*ryenhGX)l0x>(oA4$LF5VsK_lhD0gIQ3h5`|Wj<*XP
z!JE$hvMxUX%;bQtVGh@`Eh^6ylwcGsJNx9rAzW9?<1?}$O|pv~xl51c1+atUlG-=1
zzB>#SVa`c(;*_F(RwwB=lG2_;|8p~^q0L9=UIc-GA`BscSiThhe4}TR@0bq$ZFJ|(
zIdD!P<|P|MH38BHjJDu^!5YI+sA|IFM(L!B6M;<Jc}4Gh1hB?}<z_;R0U*mD<_lz`
z9Wd)F?Kfl29mp6Bd`UpdmoE+KN?+{7Z2$tJLcm(`S55>=PakMhC@4WL{W=(YxjHWb
z)HkD>&DNhOT-C7H;cqqaWKFXUMo)|!n}hC>>4?Ub0eOt|icrmM2HJCV>E$N`xFd|B
z+#NcL4~`7rT5kCLrVR)VcQ96fv_eQsZ2P(3!Y+&lkiu|SKLf-A7n8-4b3HjqoG`!E
zax*6|65$sRO%I+^*t|9b)I>Nq0Wdt{lZx14z`g<|-~89F8Ns!Ss7L^zK&FJRSBXKN
z(wG6)(b-K4Jn(bug0d&(G%SQ_1^|or+z}pT7w!_Ns+$W#Kr@>+gH$gDvJWJXY7<3$
z%^%J`scEW}-Y-%l8wXP7t~3&)L#@->+J79r&%jG6TEhLZcX1Pfxa<8bNfwZoiJ=C9
zk1B7`q!ffaP+XMxQs6Y;2kvi2<mBYco5BhiS9as&!XhCvHZ3Ha2sJAcbRfgR%8HC=
zL~;0$f(5;@Uht0P-R}%+=lxC~#J8cq(50Qgh2lcdyX*D{M+c@=gg5w7#GC9waVJ*e
zZXkD+fUbwKmi>3_#r4^zWG5`b<d_-kwIp}lWx!RAU7m`M3<KvgSV_K5M%46y?L<;a
z3L_=O2rxEp5kkWv4XO#AvT)iQXndU@OVY%)<LCF2MyU4N=Hxth5Duth(N!X|7oq>5
zhW3NTj_~?-rXU+8ATkStWE-MCIS%rJ!_DfNt)z&^@Aq#ICKtF)WTI#WBDZPL8MAjO
z<Ti5x(=xWYjFN$>-Q8yu4cTBt844|lr!=OoPH>MXp3Wjlj+vSSqM|F~>9c2FA=%Z!
z_f&`?B5GY-bQzBhY_@Ke%stp?SZIstOT~s5aGPJ2fCTiTat{wo%?;~IO<%r${fb}w
z>bu@&s7c_mIt8<OFK>QwP5$(t+y0T59Wh71stVYMI8nfQgx%fvf#e&QLW;m*@KA^l
ztf%<LKK2`3|MBUR_q0ol(Hqw_ldpQ*Cs4%*4q6wr^@qBkP<nSlW~5-9;b^pj3kYV?
z5PQzY9Hm5r$r2MoZ*Onsl~j{~B5oCY$bbkhaQ7%KetulyF2qQQ{T4y%Td=}GW4XrO
zBn?BgczDBiJTD8US=HfR00svJs2m&}>mA4Xxz!?Z&o%8gR44P?-z@R3mxyo;2aA6D
z{IEeH9cP->Zrd_@aJDZqV`h;#PMG67z$kpwU`vSA_aU5Ey;_SrB`j$X`$b({4REv|
zRBFm;J!BjOG=z6>cKd)Nc3LTT$pG^i;12JE19RETi~<F-7Mh@xrJ2EWRmpX@T@Yb=
zGR`e#jDU&WAG&K;d^F6<f)I>^2tXnL22dUlEsfM+V%>F1M-`mCEU6S<52Bhc4n()9
zU<7aH9xy}1=H4I6_nC3;zk@525iVDL`ie}hjttUu(rvB3(fbI@UkI%O)Q2r-rO4zK
zXk^Kt?-gGuj42!W3Y7@y@d0?CLJA9!G798EB4c0{)HZ#ymwR{!uoVkB*e$U%1Dg>X
zH_(4#;2y22&%&dF_!QpIuusKVtMd;8%oC*TcX{qA)b{Hu6tSs$Z8oQUXeV&<$FUNA
z);W31S3mQyJxI&rj0%S>w<iJ#MAWW>(!ACk5B5pz_f%7PgBAE(LI1L?_F+1@xm5`)
z8#q-tV1Ytek%CKE!bxlS=|1np)ls`d3kjH_g@^l!kRu5HNkqZujy$^koX(9?pVx!#
zjoE4d;6KPxh3%}5A^65@z1pFuPA2g5l2%Gup2$D<fe3#Hldde#w0$!Ty!C8g_Z3fn
zF{wnS3GfDp?Tit2c6LPdg{e3S2T0k2e$Fc|)L=RQY)SA{23~D}fZD<4k^J!CaAYqe
zH?0#Qm$=xjy3WsL^4F91-zm;*Rt}S15)wDm%5>MVB&P4lbAsnVpZl;C83srTk}WC(
zGF~n)@}*AaE&^RI<uV=)eJ&G_Wh0YOu<e*Cbt!2s!;&F@L%<fuyVmLfx<x&K@!C>H
zxK=Wkcrjb$Q8zQA6qRT3CE;vkDOD;W;sZq<p|k~e6HDQG#PHq^l*bY?w4f*y6n|(;
zqgt$KCl;jP|A4lJu)?eTnw+!HjsedH2b#eGzP+TVcl6t}DhLi_?A#ad&1wg$TQhur
z5b81*-#`$agYTjV*M8axCxK9!7vp3+WD>(?M-agFw`-^YlMmLItDcwY#Jt;Thl)B`
z=p}RV99X9pSFT*4hZ&~3Lrx$mLg8%3p-&N3<Adu0Nm4L6lnYZTDENXBKnL+K{Dw?x
zFKWOJ3_E)&BNO7t8HB;!c7#Y_V1N-9=m3DsM<GtFCUF_Ct<zt>Lo}LxNH9o#I&Mjd
z{%<z|i;0(T*&5fvDZW|@Z4Gv~7(NGDGk9pQ*1#Y559AyJui969<urE_Or#;hpD4uX
z0g;MkVj=;o$sGs}fvtk-WLk<esB4H80CE|yp{EgL174d;WAun5{hNa)ui*zYb*8i7
zeri-Pi=ML|8b|>SMf_82hI_y>39Q&iqaxHR^Lnr1Ws1WUpqZFu+_|y?)-M2qcUb>P
z7nqR+et39N6nyy$n84bh6T?Hr$HzCaz?k_1P!VzdCv=%$>a_g~HEHyCG>Zi<^$!=-
z@(`Q9B?I}22plem33v|#CW2u^b^T{=CP}I`e4P_1_a9h{<qJ`DI*|}^j#MqK$w}G;
zwdN`D#DS&ra#;54G27PM1FI)Psi#W57Qj!=J~#Y~^SUP6Q-MPo#J?}2k2ijE>A+M3
zfCIz>4WMdq<t2Qx(C+T)M|>D@h6eibXxJkl3z2%>JTP!Y>kSyh0OypwP?;V^zPiHe
zmqBeq14VavX=(fZ@3%C-t8)JVRYCed-jQnQ8}+KsLTJD;k-c%YN`anIe=1Jyl|(y<
z{*VcIbt{=zP<P%UY5(OfIL<c*Cxu-<s;)0OIaRp!ddk${WXlEzrEGsGy>6RSKgIL*
zbuOCo$mx{%<H56gb#&>3J3Y8C)%<tqOXlX|o_Hgf^uy=oU+mQ)9SibuX@gGZ_-N-X
z(*6oskq{HFQ@5J!ep1U%pWYT?ycqSV#^lAmF7EP<!#5`puEKRmr1LWTY~%Wo&(T<W
zk=(s(iIgXa>&*ZuzrA*UYIsF6AO^y)7JlJr>yUnH+6Oeulzj6<fyWgn;{3|*8&u7I
zsBZST7)YA8D$qKJpK{u$Tcjo{J^Y_LD*g0>I2_I|csI!b3<HrF4WZaJ4a!!jcP}VP
zpOlUpE4D`7{`tXr@JlaP707^yZ^3{E5uYe62Zu1vF_<V?bB+&gBI7>Y$Dr#(Gg$>7
zg-LZCzM$x$jC^d}2T(a~;m;O%s=Hiafbuibb`wokE!%V<s_l;%3N_iZ*Q~aDx;9zH
zQIkYL)JZ`uiio^hZ_J%`=TiXjy%yU9Cx7X(j+~n`nkKlKq=TVpj@AvDL^TKH#TN$p
zyti8ME|N_&+~t;gA@Nm*BT3ow`hG^deCg5u{g|NKBkprJ_=?E4S^)^tCqd7p5!M-w
z<wntAF;AJCHN$clU8xmoI~&R&JWD%<@>G@n4^`hCk7e8ce<^!qCnF;_-Jw*Z6v9;$
z3R#t{BvFLyku4FGqLP)eM@7*PQYfK}qKu40iHx#;??XM$^Zi|~{<s_W<+{%EJdWe@
zd4Jv`U6H*`TW{@&Kw91ig;Gle5Xt|%%6P9IW@v!C9MmzQG&K*6P~s4I`9_GHi^-s%
z?eppH|I}nKPp}D99QCgqC_QL6cx)o*wYb{Z+||d97^O**O&qVd_@dg)_9zKn;If?s
zw-uNHw$1pg_&5BW?n*Of1el`hcDV_sZRxQq(ojubDI{~6Hc^4pke_%Xsg+;ukiD6+
zc;ofMHarx}1cTWGx1?Zo6W0x#;owfaJTWB*d;(H!T6r??xN=KhNh7__wJp=u<-vRF
zh>8^k>o@GB8lS|}irBvqH8v1=Qj4LRfc4!<C@(N2XN=!VSNhoNj@)az&7Ct=+YHwo
zl~ycQLEQ=+jKp=aIb@b%e*sjXbHvVW8RaPo6+&ES`?O)Hz@l+Fh$8|2s+w$EC8Fq9
zfMgQ8y25}(e~?A#yu7^Du3sO95sJCFFz8Dbev?KrW)DZNB)2K%Gn|*)+$dqEwNj*G
z`MUVM05v$>9r5%MfTN_zzV94`YWxCAEQ(JSE_hGAs-PVLoz<y1B0}jRy`Sv*HVgGk
zpUdbB$;()6{SpxvkkzPpK<n8xFAt}X`{CZ?P>d~{Xh(B6;^GTyX?nmV!y%N$UUy`R
z^>C<yT*CG4)T!WmNsrK-Z+U6IC1vOJ$T#`Fu!}XXj~~3JXMQ<~zg%MJM&ipxqdM0>
zo&Vcq<H|bw;&#5$Mu?JQQ2s|oP~-L-9qi+Z$*nt3lfN~TfAI*mBgZ#ml}R;ON-TN#
zLdw1jk@*_U`47FfiIsnqk<qo&5>?w@PQ^c4x{}dAVAjUl+^l%DePv|sg_aeJ?n=Ta
z`;{??CQ3{S>Ha`I>HPTwM`CDFlKBY}M<b&3!a^8rKT!~gt+$-T{laeJ{(sG~>vmrL
zv9l&)jwUiQgU5b@_?;n{TU@J@9_^06hZr%8Z%v8v6FKsgy;psDqn3oh#HcO@fZ0E<
z_EjjcMgWnZTiV(6)X!zKTaEGz)2pD@uU{X2`lrUG$ulo9`VL|*8Fmt^4s3~Y`5IJ;
z**E1LpgHHlUe)BNT;O=1MJ|MTdMdjnO9>6d`#1Y%uBTMRZt-7g<@?f|MLoFRTon$b
z<)Y5A^Zf4<aN-ld<cp)=$&>&+``n_GGI(pOdB#4kewCG#mDm*<n}8#S;h-31BZQL&
zYakSOX<eaZn(1jhoBlL)&C6u?SlGXu3hlfVe8Kiuh}Q&<Ido^^9wR~iYU$2$9&6sj
z3!gO<-*sz$KvVJGZ=ab|h!%+ON7tKqwwxd;CGb|{yeI65m&3?NiF-JFHmcudXfK`g
zK<5baQBo%%ka(1hy%5yeH?X-Og397Ha*lh~BfS>WpzJWNhf1DN`P3sV>o1FF>@i<$
zxXEXNiGO_=dz=xY^GMLsq}}4Or#Ab`D${Nc6)^wk6S0H}kf-~5pl<%2RU90F#V6W6
zV`LEa;DI8zE~8*quUWH3!>#y{G3qbSgE9N=0W%TUbeP$CLID~@>rTix1j3-UNeqo1
zF_@13lHPBe%>6N}rtG+T|5o$IB9zAM&ivna94ewCay3)5O`dxQ(3tpapjGnVHUCp5
z!19GDr#yJnn@qJ}@j{sEC{~H$QG3IakgFEn#O;TSbjYTKA#sk+$z+I9(d#l4&kS8k
zH^}0F#=9CMLF-M4w8d-E7F(;+rFJy?kGeARnupe8AHTew(IqOV+Ge0jDs6jX?LR^^
zZ>(eY<_>?c_`63tC~w&Y_13Xt$IgqWuI_z#fq`NMFAD-tU@TvRLV*Y@FhqhDiR0tM
z)cg<U(Q;Vk)*QLJ#ZiN*m_{~AH$L;=nxWcTr=QPARoN6_l^-eIDkPB>LzTQ#@u8lH
zD>7j&n^|_R_Fe$NVz((|JJ#R){#)RvDp7F&HgWt^#tX4$H9#O53hZZzjQ-=ulpla1
z7?Uo_{_^>f<k3Ae)J{*f$M!I4^FbAYmKlLUgjsCG!u%{IV`nik00N+q(?aH|_$zX1
z<7s>UdjEZ^uc|96EY=KoEX*?06K}tZy5uQnHpBPd;WH&e<~(_aAJ|}7qN~L}Ae>IH
zlcUjXip$98xYcjpWZx+2_T78<%S%vC?yCusOL%uTwpUG-13ivk&gQS)xO8S-0dM;F
zyW6Rl<`Rby4DyLywgToQ#Oc({4aIzrS?An4nAFfdJ=>|b{oM2uQ}MXbN=fpL4!_6&
z;XmiObU;AI7YVA<3Je@dzrMZEN`?xlslE$1(mrfS|G-$<EE%99=CHDEh$Z%b7d-`z
z57w)d=@fx@IzSafc*O+5{=3sE@m%-iuLXU#dw7_IV*2QeEb&q#V|1MTLDxC;HIK?E
z^w>=AkzXaa%S6oK7nb{){z_H-4|+0X%yIQU-_nYsv!`@>HtjRZ-TLoCoG*raQK23k
z<d$+WSz%m6#2)aTAwx%k`QhBQf$hd92(IXR2rHIn<@l{NoW)6~Y+Q2f8P*%kuPbBL
ze`MTY7)qA)owl@n1y~|Ax7}TSoCR7-c~(>@jJ;nLKkd(}nD@H*BN$eZk<ZxJm?t<k
z1PBLc&jc((ffV#}Ju3pB%jk*U&K4PQ=s@9NTILgXV~t#x&6RDOKCe3tb1W#9Q&9z|
z|AZ+qxMm>(DBG9pf1kpqN?$)txBsw=otgaVkXoCNsOA-H)qFRkRXnQZP=NsPdRYB4
zQYlRiCPb#Mx4L*K#_-)%7mpKvFRSh*OKu}x4oA0a7oA%HgazHa6;J{)Pu!Qh@EvE|
zAvVls9}082^QgpB6q?xU-6`08++5|c@b<`jw>GSBGS{x`dZnxNwCTw1<GU2*lur8}
zRhZjB*Dt8MHt%hT{X&H1CAo>SG?_B_FXmJq2j8Zj*BzI>U)(5oUiQlVW_|a`^PzNn
zHy%G`&}25~PI0~;*~|0y`h7~k(FzkrGTf0pTp|kZ7h+=(7#K*|G4`CBOv<3ef*mgU
z&9oTxb#da)Wtop@vy12aysUe}b8=!L<lYOKm5TWs7n3dFBl<Of>q@duDbMoN{(DUQ
z_gOP>9eXMxqneY#uVE)^Uo^)Y$v)fnPaqxt@ah=N^2Nd1vzD)=D@Rpj9OE3tmsKkF
zS`{VeH|eqNk1j`5`RCPKhh>leiVHvk%1#{}R>HUjDH#Z%vGZo}p&Ip%;$tWFz_og1
zlYM6R6Lr}#F8xPVfvdj$JiT+7o5HrLPc~BcqwwbX97P$vC|#N0r3H_7S3HKt6rxV+
zk&mV1yMhx+hGxqs@1^CjbMCUP-_>V$vqHIZ3^iymivpK2t6WueF7;vUK~$OS(F89b
zue6v7GMEXND;i{`QrlXpl?6f-Rsx!hS+1VeReJba?2%DAidIDr?SCFz{S)_UX~a17
zefl%LhIG<Yadrzyl4n0HOx=x6yo^0~&xQG`{<{-2_y6AR!MSuy<eFJntOojBc%+s~
zY?>HkIXf4|I^w5Kg)LVP`rt@PK~Sd06E{e7Ic{O8<x|7LAdexN_+4whdvxyZlbKDe
zOtp&F>oW`T)bdIWnLaJK&1uv!`T6ItfS%Es!SeFzw`nh19uK*$&1=v+=qE<+wwg`#
zP9xteevJ5#Cpt_$6O(Yzxy7ad!J&ayzYD~!U67rT2?R1N9i0$xF)kLg4R6E_Y2EO+
zX?PNibxPwLwq=(39~}=08s~sNhlKqMfs|Q0b8TMbflHQf6)tPCdVX(33`fGLor^`M
zG@8>VCg{1jM3NX753O&rk}nx_QD{~caG!Cwu|fOg*^kkKBJ%Y9`^02_ejr=-7#gZn
zoEoP0?c-wSfRn0W5?zZ;)Zr8+rBGg2VF!$e;BYlLfY~{*B^~*MrWXcoR7lz?tExi#
z81RoUXJBOEzSnmPRA5jqr5$&E0aF5NLF@xnVt4#4rbCgdlj>vAa^6d<5K(;mt6y+X
zh;OY#fLWK+*qdy3MpiyoGoP*7XtxLWuC^BIRR8IhZc#ZEAhX*(@Ivg|zD31*)faM5
zrB>J!b)zU0s`(T_lk~E(t;jXX4lp-)>349szT(G!xNGfBh+Dtm<sPqpPG=O8(`57H
zgf=z7XMcjM<hSyP&F4<I?!pv$>6P~A86KuMQM;&BbnQBqTX;!%ZTF@oTW`b0%BI|%
zfd|)=v2%V3L(S~w@?&f^c2ufaA=2xO#L~|MX-#?kM%edh?dxB4YvsxCt4bVuuI<6k
zEG~O>k-o2PNVJ&p1l=N&_ijPXPo>z2O=b8p`=%N*&B${3^AWZ`U-}41W|b~!_KJIl
z$Mri5=``bW8dWymj^1sdwsFpP$MyM}ala(320D@-Xf0HK^UZ9T8l9Z7pYj<IRw>*y
zuHfMF`{V9~QS14`GOC+5FOPB9{@Zh#;;ga;ch$!6`to9}h0u>mTeoHR^!9gt)0RB8
z{6AcPU#zP?KKRg8zBiBh^U{2(!CIG1{X(rGdrFlSZB~!$9|}HpC%QDRliFN@*8TIc
zu!wxMzq^Y^B*1cUHJ(kpN~LEsZ~=}O-B3r<km|Qix_ETTP18XAM(o_jxmWma?kmZc
z4l|xQ<<b>!rq%1W_x3eKr%c4A4^B&Uf9XA|w%q1$|G03PP^z<DrCye%)b{;udmnmp
z_-))SG8tWFOrFO-ue2PDL#PsZnTO?1uWRTwXgNroW)0L?H<El*<kSl4F1L!_3+Go}
zOg_^mZ<AIP+Ap%zjq8c+nvnh<9ruLiy+galzA~x*{B_GjG&Mjc$9HjCarIUMf115n
zLbl$|t5o*6*Q-35x?R;_{>}V8)y)GHx&+sHVIA^~CDd(g*T9y6DBQ%RL1KjG@Y2Ma
zbXc^fT}Q)mo+xGuXvY>j30!K^;5x08^+YeDAoxKN%eVeEOg(C5T84|?dwe}M!OYL>
z!C(n!oV=8T;Mof!D-6uUrlGNXY!zEm=S;q+Dd6E4K1YoutgKeD_Ljl(GHhzW5r_2c
zrkU*B``=%B!X7Ec-80a=W(hv!)DKTH$b#CPz`;zl1-ZG;(F}wh;ciNbJGXwP><Nsf
zmQd5+h?5#pSDoQ_Ww?bW-NKSQ=A2DVQB%4*Q(R5>UweOhV-b$t;BO0vp|0ds(fbBu
z42hY|M(|l-^=IidPY4p$)~V0zsBVuij|P$ussCwp@S(2~f8OxY-dhnr<d-e%Gq;jc
zwDx#&$WY<Rx&Jm0iqB|Lc@JS+PEID^w}c){q({JqcD0G8J?b9jVP(ij9~7X<s-N0a
zX+IEL8>u7!zbb#VH9cQ=n4e@ZJ-XMHR#v3*--@A(ot-hS5s2Ic@DgpIh9IV}7@Y^A
zA&VqGn3jTDzWU$LubHiFq%U>SSr?ZY(75)uQrON@Z>{&gQR1(s|8K`N7~s6qK9o>Q
zFkZ5lR)S;<Am}oJ_RmW`n%cxHxUTWYrk!GOQRpw1&25no!nUvX{P#Z^>y0Av|1K5k
zV(_t%+(yE8B{OM4@&vAocn4z1LVi?caE^lfXU|=H6w%f-tPY_`-MeAaqqDB+*&Ewk
zZK8i4e!=d*MrJTI$k=16*H4)K{Zr=ggW%@HDL-e_n+Ml%eqQ|#4#RtivV218lo>8n
zpWo-OpZ54;!nzR$wBT6M{m4i##qi(IKFpB$S<Ngk{lqb(h-+qgfsB<z)A)=F+H5pi
zf5q-O!syHtY;fk7hZR(BL|O!w=6e7Ax=zY<(%ErI2b?t%7weB~fu;Sj1^=5U0wuZk
z?p=3Ey1XQGTw5DF;o+vaIVIt`bwja9htK?)-o1J6s}nt6dVK|_+15%Il=c27=ZMdw
z%3hAGDt*Du;JPo>(sx4Gj}K+}y=vC1ko`)cX|3+1JLyB!C1tv7iWp!0sqV=geoe%o
zs2gDY%>+6ce(CD!7MjX{Q;<C9O?A84y!V*ri6A|Lt5MoUs|JTST{V+k;T8dIe`R&v
z^zwC-_a$FwRSr$R6F$12c<)2VwV~{eH(gJE@u>Xvtup^|T}QIn;go8?%Z1IRWxh#;
zbLN`NV++~AvX?E{Rx1AL@b0>K-*Rk3O11Nhg*}zBaDPH=r+CHn2EVi8Qm6GTO_tjp
zZFdU5S1Gw$woFJ?XP5dHm%qC{rMkMBP((ltg6z=alOKT(R<0bL-Yy-tcPxD6O0_%}
zu8WJ~ia9BU=SqFRe8mG&iq*G?SP_ntFZx3?0;?qQ3jmb_7v<UeBHvb)xMG5^>@i=D
zj^oo<U?Z9(2TiHDhd+E-9-O%O_7$DyY4x<ixr>#)uW#-!HC^ZrGJBn<WM=BGZ^6E>
z)lb6z{iP;JFWmvN=2gEc_O-K}@+%XjyLxx@&Z-@7=qcV*7n4$w#r)`Q0M%gdi2ddI
z@^_d_b`pXw)FWU~d~z$tgv+!gkmx@mB2GpxSC{VX?xrv>gbSWDSr;F>xAX{ioWi&(
z^VuY`==R%tFSqhO)8-Z}<7(2)x}_E?{ZXts`oexHmC{)7Ztb`FgBN~=iGQn`p{Y&%
zrIE+u8oq*)h6wkd;2>Ihf+1rnyXiyAtJCmRfaN%o1DPv=bsoB0U2wej!h>>oA<p6r
z29Z2sC*yt>PjKMEoZhKrtR~qMk-w8_yi@FexR~dTslHr|T~+<k8tP=Lai9wC0}F<z
zMTw*@S=G`_cxnAP*gkGKj~Od7&;VL4o>sGvH}6HsCZ`{zUvrtCXUFQ3Q!1H?%~r>z
zR@lp#EtVd`91eW=Zd5`j;+E)X_;AZc+0~JbQKR;i(ejQJPs@?s)4J@0&SE0&;4b3#
zh*R*)ZG{tEGrNVe9k%L;-Zf@UWDZuBwDOoDTW9yPjO2~-j@Ljt5Pk!PlF<tRbIZY*
z7ey(y>rX|IwuRIa`rCQGeiGsAobPt{fi0oQNI>@(O^x-26Cb!E=Y8B>cB?eixBoir
z#k=Y5_+j~Eu#XLHtJ~Dj3cdamV&6rJlS%2;UoQS&=acQ2y>5g>{Z=K5M}0rwM3~a#
zyv=3lvK+2l`f2?(KmILS)}(NL;B@o#m}P}K*SmzP4}oPM+@Z|F9X4^I_36s-bE6|e
zld@NIsWxkmp3!1WTw6x3AMOsfZlch!DH6r3?D+Av@=F1rkO=)rBly5i&Q};l%J8jm
zi1wkFT1kEHOj~Ln9DE?jB<wW89w6=-MMXuiuL_iPyT8`5v@Jb@j%*AexaQ3=legE%
zrP6PTWNE&;72mDqMWU>v)<2t?6e1k2?7Ks<7~ro$)cAjo<jm+jyH;Y~O`9sHy|E%`
zdI#lSp`($pTAtBB!Eb&$_fD&Rj$Zf0Zh!)g#2623c_IY*ixFs1$cN6|i9y609+AKg
zHo)zbMDJyAy1($FP>N;PPiN&|hd=U4#|{Xihz1xFIL6dZHAe!5yGMoxge*vmk%0G;
zS3CcPMJb~0gI?9yo!7PU0mgyEDj)9uv2(CkehgP9jQ;p8!p|-gWqd_#?H(<EP>U>l
z$97py!#D}uDFnfM7f;D9l+OF{9Elh0Z4{J~Uv+)*rdpLx&nFY92U`CA`#8U~oEaTG
zePYUoo;}*B$JfWxQ(>_yi$=q{#MIwnA7`_J&%>_W{^fmxS6(%&4{2|r8M#^C@Z&oF
z2b<|%6F<&(YKW)A{;u5Ty+6uVviRm9C*dpQ_XGr_ieu*loOl<;ItBGkJQp&_zR#BM
zz*G2!(D%^36T<?9-N6dZhEcTzStkl-db4%bX&BT{zrCL6c_3!}=8!~~PKbe>RGC0r
zy;t{HhC;Xs!!gZ@ob4*UTbqYYTo74;orM?Km&OzyxX@Jro&uaTIyuREaS^jtF?dU;
zsHmv_R79Q)5J?y%u+&1BsthGmPk;YE5&=*$5>nEgJ9lE|fM=?sCaW!xv@9QI`S#5y
z;_J$PBvci1I2@vvm+cB?sx5chC?cTgGQBNvx^Zh(b-!gzaQda)rVCdJ&Q#W?Z}y(t
z3$r`Uw`Z^IPmNu(b;n={eVy#_hSu~!p;qDi5e%D-cPNV9y{`Irhfru=!6&z6oaRau
zRl8mXy3NPPU*DSFE-kOco^(FxU{`u}5f7e@c@+u6$o$<$4iIPt-lxkc#8?peMstsH
zFsRU~-9M@zeperw|FtqQ&xz4GdR=5tXlX_Kuh)eVYx#;DflHHAV&aYYaoAEv7XTo3
z6bTJ5y*ATY0E0meO5l*t$Zccgo?|vd{0<oy@5OTaCUKCp#6Gk>uj*h+{iJ03xhg%h
zFIn|cXg;S{<<EwsI_j^%{DjfbwPMrdHy;Mpyg$jW>US%4?%vBcshY`J9~Hk_h)v&l
zw^`4>mh)~WyZX<sQ}4IExBmTJ!i}AU=VJN5r2yN(gxaN+@bIq-iW~iGc2ufAH~T%_
z_N&CIy@nb>A5l7bytYonap0jtL8IuuM>d_YTMW#8!S$`yBK&zQlO9MG%{D~j2u%JR
zJiQ_dqgF-MOo6&htk;Z;m}=V~LEvZoZwrnM23w5zX+0(Je~)mTIXWL$_rgYKLwl&o
z*+r=`zPNfYdCvi$Xs#gRDgdzmr(OPBjd4Gfvb4`&r-{jm->1c#bbg}pd&JbLwE)rU
z_W+E(pt=x&`Ica((Kt+;n0zE?mKw-nlrm&@GrA?xU_bRV-sn)Xf_#aMp!i)ohYW6d
zx(h3T3J(y%x#vLQv6E#N;M8eRa4MbrNc=Ek^yt{5j5qqV=G1pS$b863vo9v^eHgyY
zLL2U(6`Y5lVI*Fqa4R4lZN7efV-T`%6LPM~;+!jSw^|~*!rJ#d>F2~9cvpQ@-Qv^5
zAmT$;WFCCDh>i2Zx=SBu6AdNH=;K<l3JNsg)j}{;hSi=#sZ0rkM;j5TljjM}bTA!c
zJrA)j#7;qipvX4Wg9?ha7nFfLWQrbIZNmxR*aOhZ(gAXJFD%MYSqLoOej!w+&eAF3
zBw;|pY6@Bu3I$X)_{qHhlt+a46bcB?j87~sOiZ|8_!N<=jqC6L&<X>Lu)+LcAsCI>
zl^Zp6qYv!=yuDI%m{Q+~%}qxfpDlb0u-)0n&(0KiGFV4a12(-FLAdr8gK<v0w1^KJ
z=uPAppuo8@D)Sn-$U2~qg0;C;TAKT!<-vosM{6T<&l|eYEWLYL$I+!nkPAZ-o;Hh0
z04#Eex**RSjDNY>hZQ=n-x!9_P>Avtq!}nrIUt@PW=Ys?Gz<+dci$#=(jy;kMUcrp
z8u^(6Wd)Q+$7c<w-~aPe9f`m3{bL-D64Ty9;q`3;-Bz8u6J&+BxD?tpH|z+8EKU@C
z6NS*S$T<M)y%H!N=pbjy7w3Q2Xp>FxQwCvO!Ytq!{wn|RnTbJtKtE!D!U3p*p&R-v
zjA}!APVzE-_2JF(J!0i-Oi>PZUF1BtU-R8V#bFGx5$YXII5Cy}Y$pO;@O*iAc~kG-
z=OIqq?@8e|IPx8|J4h3VIDU!512Tmiod1d56xRIRU?OTBIB*GH+$rimkATDashg+=
z7(;bJE3e4_sak0sS+foX6bumlmkqd=nz372TB6jY$s<ekP%GR>Kn(EjA0~h`293ls
z<L~$(4LTKQH@QqM{yB#6?`wui+kM0~dWtKRhNS0co(Dqu{bscd@glutaFY+`ym$v6
zJR&f>mbN6mg&=kRewVyJOBrm<6NEyTA*8uyRFOvd9-{7u)lFtjPSrVHX{#cb^;ayT
zJVO`ENb&rdan%_@$T)?i{iOZI^<YIZSx+=vPXJ1(I4!cL5~cV-SpE}wCTH<ARIzJ7
z@F$6GP^JWdV`YtO3CKZdM?1J~fqI3WE9g2`vFyTt<3BJQA{Hii@QJBOsNp&?)g}Z-
zq8tOY=kaOY=TJbB2?5~;6EP3`1n3^XW*o1yDLZ(?6Jjbc55^C+g@GH2yIF-<Hh7D}
zYghJKUfZw>TyBVm0wn3Dni}XSmzroVHNF3wOjcU!G5F2F8-~!CC=@ip_z|`7-*L<Y
zLV*m|FRGCQR}kv{9yp{!e=P=gX9%VV_Dm!Qr~CIl9)Gjlvp?QM4(cn)GkgdaG-c#Q
z3q<dBc3t|MMPSEa9-Mq<_k=33KwBE(fFcc(g43Vo2@{xu!?`25{0~(a3K%kZ$;{2A
zLfQ@*lpeB=z+xnd15a$6KZa(1TiRV>>YWWO89dgBg{HKivJdNeUuKKLk|)ZU??Jed
zv0@saapgM(;KPl=(TUMfOM;#u#29se4R#5L)`)ZmUxW-2+bn$8pe@}%hFf5-2vLa5
z761l>9*JzgZ9Sj7FuCH^_xAQ&KS9AE{lf=7)d5AlhAF;HR|cD&^o>SoEB^2h>f%xG
zn_vf}_(>DlA+KKLL<<zNChDB~%zdh=w#h>4srl$;P(ARDo3Wkp!^0d<$aiKbE#nth
ziXcPMQ^>U=DQ}}4yh)41Ge+za9{4k%nlD5pRZ<Vm8efU;wt4Z@7}i_OI~2;84Vx`@
zXZ<6BxtYs?#x@CwnNJ`d9GV=zbn#;B<;!lLoE8t-Oc2W)z+Eug?XW530+S2QbqNPn
z&#lDKk~#nOs~d441q%v}k*~k78-%|=4}d+!;KATl*44^c2WR=r>MB=Q`oI12wP6~<
zRMA_gcue`Obi>YN|1RN(yPDTSAn?pH>dOR#?wKt^d=19y;^7XY0rww0^ZIxOtlUg@
zBiLh+c)O>*+jJz~#kavMV(dZrJR5uzW1pS*-=4LlUYk4OY|=~X9~o>l**i#5m+4{!
zc4DTsJNrd$s0J)tnf3DaCMNX66YS~$dDe1lk*^6O9m2`R$dn(-cvSL`j)ec6XzxCG
za(%LfmT=5}H~iLt@9&w&QDrDzRekOfWJTI;NR@=nh87cGY*9cMJ`=BJaK{_V0|M|T
z=+bqPRfI7tt0vtFsMJesr4w0`9ngO#6Ha@Avd<P`1dFGr2da>lgs2129I+7t)kNLc
z_{v~5)#;oEX^Zq9Ij`R?(y=1RvGr+a25+MvYu?9*2*+sf2v`{fg6B&YciT<7c%E1W
zjBFk17dfm{bQ7ciKt!N-O9_$*Y!C(qNn9myjRyq|fSJfrL5NLgwX!#(1DW)=`4Inh
z!1ywUOIVphKW95Z-ue4;)CLZy&xm=u{N-{Hy_XeBcfXiL@Rv|U&%MO@_6J3KE|k%m
zODl?tAMP9GIlrKmu5V%ayhqgClkg#IFDRd>cjllF>FI+`Z<dA~<E$j2g%tIZLoKlD
zeGXKC^p&u_k0uvku&MW2Pj>)5(lew!V83_zcwS59Pwiyx*%QZ<ux+HB`*NG>#sbEw
z#lpgPLC(0mp8y(?57a0WBugtVJM~iZVY?5L{jTN`h`HeQ8Ure;;0QN(lBgWuxW<a`
zF><y|OgOf>vXC9TJqj*2U@m-e_JLv#Q}*ue?#8gZi08eVNzsOfv+C7a<?{!sXZ2E?
zI+I>gTlXQA0xU@|SReMNzzKqmBCnue?p2S^EO{xbWft$xS$r8ZQ6W`-A#hvEN2Y+7
zP$;m2?zmDM5z_k?I#IfBsm<QwHs(+14vba0jM|4V#R;($ifGrC5)9ctXI!<wgq<*2
zDa51+@r;_jzK<iyaE+)QM=+LST(ROMXbnWYNCXM*?;qzTd4t4YG1+fw73wI@Pe<D#
z9amjNDk^yt*@qqX2)wx>^!TSc8rS<}r?+-^F*@#Q-a^tAh~Fn%4v1d}{w-((=ioOc
z{VF2gv}hS@E+~cYGC`eZVGjI;YFrJAG;B&Q02LFH6ZmG#WPj(D+V`9sl$hFPipB8@
zag@+ps(l_{MlL0+j=-I&8;-*8ozFD?u|>+08X7ToM!YqyWA&Y6|8vNb>4g#jVuw&z
zq7f4!*vW~DiS^ViVCj<m0SP%`g$lUFP2UkXL&)SrSx!6yF=E5Bt`2u~<b{#~^)U<t
zwD#`hGLO%!ci#EFjy!%Op97<Cd>HghjF@P4_xIDmtYoW_5)J|N^(%jA!l2Iu<~7z0
zaliLt#LaPSv8$oQUK{8mCSV%2w&{&?D2c46(-ywvMds^=-@V9e_PU1iqU8besJowI
zu3P5hX&iJZZJc}i%4Xc1jZOXT*QrES$-bj2;bOpCt0=%)(@+~?QLysjDY*pa5u1$y
z?lv*9^yy4rHt?pU>Pty<*~^bT^cEDl9D{NxP||h{^Zi6ZWx6}Ftj{P(W1zwYi-_^-
zQ8|9N-)@kB>3ceS2S6Sa-cpV#4MTLEi(Rc4H4-0n6j944f1YdSPk!)8=q$-4okg=a
zD}7abR~+CqP83-!Q{TQh5`qIUVFGS#h+X;`)hT9qrSa6j^;4gT!7Wi7pd=?QQ6Mfr
zrl+Q>%Z7ypRq!Y{+r-*ZUCG$c&<YPGifQ$quU@?xgHnh26>=S)p^9B2zX%f&zAFU3
zxO3B+nK@LQGz+^>G!x@1iLOI^M+SOFJA2<3dzHr$D$G9NY2NLM-?^^@D9sC%ocai=
z>|OP!%`Y3oCXcC2_qY6>?4xs>J$|*&*P}p7?1S6v{gz_S#o~Uwwa@1g?&Re^V~DZ{
zGHZ0nNGn^tzhRMnuUc1tyOZ+5?+3hkjE9uf54Sz*9TL{7{mwbGIvQ+QC;^LR-*507
zO+D`H?#lP}rM-M=?k%oP|Fbqf+2{?gnD72V@erH#4(7b7eSfg8_g~2_d#aZao51zw
zW`m;bkm``7+3J8|g19q*j{_Cp7=-v$SS-LVs)_vtUR&&!3J<4e=-SAV0atsMN$Gmj
zzr<4(FZG|nK|#b#1B`2PEq}xUKpuOhK7NhUKd{E=cyLP7)HEU17hy((4YkCPg$)(x
zD$)`i*W50!R}IdFTmMYA4ecwR{=CU8nz7>6D%+604gNiTZWl(*F|D}3b+LSK`WMHl
z4z}JqQKgr<qf6oJrV{t?)U|}1Da+a3?Ti=BqlbIYQ2&nZhu<{4Qh4Mlne)T55}gN&
z*_WuaE2P4$W`8w^$vi)-_4HJ;_WPe-hZ)&MIWKm7U9|i8(EI8)UWX8`=de;_@A@$v
z9?L$xE!5Ml?!@nP{Rt^^I**lxEzEZO27<9F@v6&@%NG_!d^}e38BIT>nY+1)|BM>X
zLD~?PDmJ=ZE)3_*INF})EsK|rx_eq$eowuu^Z|`~o34J`q|nByv&+m?vwkNnr`}FJ
zwN~rE>G_I=>D|Gb);gPxG`PFO|2TNl$EtHxQuz4lkDcZFVrL%Me%?5gel)0=&j9jn
z9G2%0e@l#*|HOzI-3FDk!MtA*whP3#6_tvWt!*&WP`&W9)Bx8GxlD|4T86>LBOTb`
zck$=NMhFNbVi;;1#-&)|%)Po0^})l=YhVg2SfW%SbFZbA36{`6>R!OYlI~GqaAw(`
zM8rSu)UqCQR@llDl^bnV0CI#YjP=CR+QH;W0bk0#IWx{77ds~20HsTQU~NS9%Kv^0
zD%J2ivWimm))H~7mS3p)Sr8J`9JuwrA3b&~9M#w+TVij7)Cy(>28Oc+oJJ%m2nXGv
z3BkdFv8QoOE^-6c7M4EB<RPfpn{Mk0BHqk2b_S!EjWCL$MZQF_3F-{`6|2v~(T7ML
zDS^<6q`>5InbewWi=|ux{bI%&3I}`rOnmRrwGQa!Y-Gc8t$ZOajuRcS)v;qzZam~f
za?pnvY+v7z;xSz8UN~lf0!?CxQ_|CyySlp4aB;Pe{2uZFS$}@Z7<;PvV@5byaFwyG
z0RDIH)d*$BzO`p5%X%URy#XqGD4|5uhNO9zA%SH!cqG0;MCCb|7UHajaiywj`Krx!
zR;cmi6cl#hd)&YOUL|!Vy1^N5NM1qUFXAMMyLy#DN=ho>+BIfjQVYo6%FAPe!8l=S
z7+~dzz!-<~)|p-YGhd}hlL3PyHPG5qb8-&bk;G0gDG)f|K@x8$vEW_Z)-8R@*>dVn
z=mV~W0tewNqszkeTmJZGoiw2^3JD9l{2DQ56_95{u2r@lNt=Lj8JsCbB&6X(q~b#q
znX>NRzh7-g)$1oMj8gJMi=|JaDZTI1!3T2?X%8Q?FyyCW*bHyQ)a&nGrvHZv5cZCf
zDV(ixCMA>c{Zrw=R<`#KIS1b{@GI;cH2m}U8Ajbe5;qV7SOEpWrUWPOD0+pSALWZL
zVCDhUNjO|Gs$gV;yHrU^Sh!MI!z@u<lk}k8)~H;FN;Nb_&2Ts(lzrz0`*wM&zM8j-
z4Yf3LbAhT0l-2@<H#d3Z1;|dv$=RZUf{_aU5VVg_FNNXJP6cX9?2HQiX3vmQ2@alA
za8rmZ4=Sl{nD~<PhsLHzvPDjO?^sEGCIoxZ?%%MLj^X1PZQV`n_xU|41~jbcsh<yY
z9415WN$U-HKkW8RqFomBbvquqtU3Ad;n?f%2(|>O3YwE6+ih{K)dixtIR*xM%W=iY
zOWt{Y#+~4TKOT?AIDEsi6pmi_b$vbjhNvwMOG>stNX7vjC{SmV7)B8_RLb7#!6?b)
zwrqh_ld|a5R|+b|XfvUjLK*iGMNaZId_sg568C;XLmu=bDS3J3S-nsMk3zm93McK7
zl9C$%5)s%XdeMRAI+tVlSG2Sg({pjQusd?J8ij^tSUxyOQ#>(H@b~8V3=R&G$i<^C
z9G#r3ot)SZN)}Olbmgj_qoWiGPz3~*29as}bDNE6>FL&Xc9L#FY4rWLY(T~<A@dA}
z@gSpc_+aRb9bQ~0@baC@;RJ;ZijDaye_l?$1>MOm`YnzW$^CpvWJ!}dcp8t(L)R}c
z?GF<ZS6!1op0;ISv`2%o6L43KvS8Er$44G3usjeQq(#9JSx!Mf6Z4Ib!RkD?tpPxD
zums>rG6l{@M-w?MdT(9$^eUV);b`4eRJ3^5!tI->)tF6+i>>wMKR;7HZ(v}69#1&o
z;o=IJL_-Esi52d&h$gJ!<3}N<_WN122~-E7iB3rN;b2P3%w(i&WyspMSxzqS?Af#J
zo?x*yJve0vTZ|Mq5BzNnX|=Q@Shp|^3zP~as0ho+QFL@DmX_V4)~N^n{%Bc6yVv|=
zwOzS#<>(8|Xn~B9Q#eXsjE@ue@PV;x9hXQhk$Pp;<Ny4)WMs7W^4>oi(NixJV8o3b
zV_jlG1!6KK7FtLGHQd>kq;6+tmuL$`6dk%ul9&ubXMq+}<Se*Taj~(K0|(x1>L9Qz
z>0D9w!+^x2Wrtjvk|?Zb@vvmKZH7cCqS}U_$z*OqMCoWYvp#7$qfYn@`w%@ny_6C+
zcQ0a|hr}Y1<$y}%XrdD&eD+OgbR@L~?q1nG&eX-GeDm%0$Bz>#f+L4Pa3#I9dG9bw
zNZ7J2f0l_Mx``1s8B<X=IKgIr;rj7u*YoOo0^hG+o8w@nQ{}-^ob7b}{CWA4&ZefO
zaAbRYjoWAiC*_sh?Z){<yNeu?yTHFI9u~w#mXMG@1ZyF9cCcEJ;UJ1%S6|@_YmWIB
zn-GkmTwID?xD&A^;(&-}A=D*@5?zsN!G^C5lLwRd9g=Bf>Xw!wWL`*eEV;OXkmOc@
zx(iuX&|u&fN>=^tBL`i_58~uv1;{{ZCpfvSeP;*(n-owdVjs57ONZKU*K6DksHd8C
z)ORii)8FR{PTba^QSzYg^m{PTpczBSJ4E9%#dsZxg`wgDVhqmnpI1ds$zFaHclG*#
z`lZIPPKqmPbiwnK{HErWw-)YHbfn5QwnR<VlVEJ15inUw$8ibOuuF0C*!XxAYzDF0
zO>JVU1myD>Jel9{^#3rCLI1L*h7s{<FVMu3FNGRJ5XNdW6lhjc>xKr0hi!rQnPSQX
z2O8kPsqH1$6LCqc;C*gnd_mwg-Vgh-y3-uqlu%0W8&R%El15JRcIhxEd-3kx#PGi!
z9^I<1jsvnZDvmtW5Ym>QsL-ysaXtNJzi^{Js9jpN_0glDkTG{7@&KD?IPB8`-wuRe
z{{uE6;jErEQ{UAkg07c{Lk-y~0mSscx{mm>r9j>5te!ng9=vElMHSaasi|Uci8u%U
zy;iGr;yp%>IpkaXm6QT6tN@{4@GH$S+4TR2`3zT1&(0$C+-Aw_04_0vsnXfmDTff+
zt-|{A1K+-};;1I4Fyc3M9uHnk>GU3!6-(DwCff<7s5yRqK>mr25q{;E2ZqM%G3@oE
zv-V55!vJyF*TrK;7_a{s#G=%oC-nF%!z%|D?rr*|GW!5u`PtjITxj~A;Yk_8!<vGY
z3!;L&?1$9JeO}|=cYR5Lu?QiK&B22do=W^M?`KCm2#)`+M{T(f1BG8%(T~c?mbAjS
zK5{ra$T^SM!!ij8F8Ts8=vjKDDP_Rr5dif=aEB7dV8#5Tar*xK``BHga1?fvqZpSB
zMHE9rK-WNM1m*G@RmwxC5E94WhJFLHj?g#TtSf1e*kX-a0k>U#Y)+n&qAdwzui=X9
z3+>W~7+Pn6^gSxIxbpK4kDd<maE0~<g@ugA6ZdCtTiWSbjAakqr%=d20I=UU8luzL
z(1vYC>v(bj31D~n4GV!n%nl=}=RSF-EqZ$=zkzopoG3B}YVou|cR8>AMSkPP3V>h4
zB%GMH;hT~Cn)suV-ZzQ37NJ<s7prTEW$XW`8}quqEN+3dZhq5KLXg+d$sJfa+zby#
zLwiANl8M2~$H!jWK&oKS!+?Hsyz+}<OExcYJOXBgYRT$nU#(F>Lgs>;y!>;lIcI2y
z(Bywk^5{H^@(ywxJdok2L{};m;KmX2A%Nkw@ZN50Yb*HN*R1#Hq04fpagV+L1q(DD
zPC)Q@ZkJKO6O*Ox5U;>NO0HcHhD<o%Lt>UVlf~yvW!AeUs}YSZS+R2P^d{!?#C?<4
zvXdPL4Z4hLe=<~0kGKjSZ=@v1ALk-#7f6J(bk301PO5v&JFlNz+a5v~zr%Zikyu>7
zf`bJ+7Igb8il;iCqXspaRaJhPwvCB`{9+jRYY>V5%8YfzaBwr?F5H2$1xFknc&nv?
z=6?->;R$#}#&2IBW+WIx;M7;g<jROmIl~~UT~mS!NkO+>yYxpZ$K%AIKpEj+M3u@O
z2Ufzt!I84%<cSkeSWkq-2iJ_H+}z^gmMNAc=dIYGhKB0_D5Pp?YIZ(2`Ki0F&(M8E
zkiZU{)9`-CijJzb!Erzg&F$ekQ_rx5i~s+8GO4eFgOGF=?R{gUrxyt26Fb}&_~q^O
z;%{8ae0uNx{g8}|qwzORz|NF76(HW~z&~Hsc-Nteh83v#OVMoi_y1c&^f_3gc!Ee>
z!Ntu@{3`*}rrL5U@Z(RhwCLfA%YY>cpJG(_L#dF(07G7l9R!2BUfy7;6)d9vz61&p
zSxdZDfNT&U1@YQRU4yr3YHG6YD3K*L{gaa<SN*b2fO^_hxD+tstr2wxb6!?fGH)cd
z<M_cOnZ^3}@!GZk2WNa*IH3^X2Od+NTmEQ9B=_HBUA}y|+R&Zbw<{YO_7~2eJ_5xd
z6oC$+lkeWWgTm_+qwyjRZwxC!pFDXYIyrnUA#<zQ;xOMnky-8785a+3=iawYA>n2m
znf^OYnnzRLi+n*Rxj||5g96Q-z8}ZvvI9io6n}jF-Se)px!hGGEV`Z9;QG1^(!ESq
zUN6vl&y)o<H@w!WTABQqMOfLo)8y@@)7{%`H&0C7!e7+XRrJxFNGXuknSZ(e^z_q1
zbU>3sa$h?>R<*t<I@wQm`iSzrm$`e^gk&Wrhie}H;8}evUUbrRyicpP=6Sb=3ioY)
ze;NakV@w9uBSJR%{&qG!bjSwB6$zVabcqnSjMD@>1(lsS#(89y<QEiVAvcIPaDV_r
zJcZCG5@kvSjvD~oVKFi6<Y0k{<};SY<GZl}JO~yH3kxH8&B(Kt)|v+&f{gV@p7pH-
zbo}HxK7(&g><oIQDoi@F$`ZYGA0cQHMM#e0lizh^aQ?EY|5|n{%ZvNITcO4enoo3D
z<;TXwOg|h-xf&iNIvMcEjidF9@!;1^-PoCO#~p!uvUxp2twsqmsw3g%f`=ytzsP9K
zo2$-+N!nFD-mTaF_3N;8<5~~-dyeb0<_9~NMuk-Pbn8RCEFXz@(VdEM35_p}6rC)f
zN_0=Ao-A~Hi>}*bF|ucNNPDebpSYN`z_#e#LQ#7aC6CiwlA3BXILVq&?a&`2{U~;;
zpq2(?q^PS0i55p{@q-M5IWDIXy?*c!2ZwX$`40{98#a80v=Wso@%Pn|PzSz2tQ5N8
zJnoEu4a^ps5Q{(#SU?w?O6&Eb5pRpll7aNHgN<sjGkB4v1p`>xtOD$!G|TBzoDN&(
z$9Nqo$CNUtxviz8n%MN?jAAynM1>dI?jBhb$$6-xSi^F^^~p@%nfI3VUq#L4M`OqB
z`>i&Z&KIVCY0OS?p4`8BTX5_5BS9^y{=ek~5*gopOR2jw<dA-%;Z5NQ+1DF72M6Kb
zenR#AdzSI1Q%U(-=qFruo=7YeopjG=9~~TtxYEjp4-n@t?pJ)VZ`ZWe>$0ej*K2PL
zNUAgaBXj=S=-0=hsYn#XOP{x~rDYdd6fv$y3I)|ftd_(>|9Pdq8s$i8+M%G!nW^1D
zZ<daqX3>6LyI%WdJ2S6Qsnz|GJDL8WJZm>ipjTXaH=`%$k^q6SWYeb<6>TK8j>5uh
zDQ=;UfWH_mkIPK`&u$_h;g`iZ(|#ea<afZY^E_%H1JNM|51x&P5V#H2unn=FJKY!2
zs2?G4i$n~8C#U{0#X%nR8Vp&URaesy@My(~aMVvEA)H|5a35Jlskrh`iVHm&kp47>
zqDAL~7qgN(wBU&EJsBz{?j0sMb7gRYVDy2a*_5?$Rj(WAl@6KRhWgiGAGRHms)`ZJ
za>$(3Uy3<6leh;G>Wia6T{G3;E-bdeC7_|8N9Pn0%56UbMjs`^WK|5pAR;`eK_mh>
zA|kuGOvJy%v+Gu$JdD>JbTp_Fj<;v{GIT3_e5{>5?j;DDA<Mt<dr2DCp+kpEVKLF@
zGL(zvOKjstP7M09L?;1BH)Zp74G*V*^<Y(fj+b~e!jZMnB{OF1=g**0ujyY|iV6y$
z$cx#btbAfCZ`|n8wx8y}>HRsVkuxHqF2iu0zF}HmNne0=)U~1ORk~9G6ZhHZbO3!9
z8|~S%WKe(tg+(}H-vRJGfMr9;W;GeVH$$}eYd`RHQsJ7GI3_mgJLKU8V!VfkfYCq|
zLDa2p3dpnuv$g+w;-t}|VEkMc+*?zly*uE;;HFoRTUt*@T5S?fD;ox2;yKz)17AMi
zXU!!md3nWv@j*$9+>O@38)%7xZ{A!jBO`P3ktL>)IIR#Uwj#@L=Jh<W1W!m{L>oUi
zG^FBMj#ztm;Av=UYxm~v$Cdn89#CFs8^8A-Oy;eutO9ZTxB4J*Qv*V}B9vxoLjh<8
z`&BfDkyd*TV>Rez;Rh!?F|p2KulYMKI{Pgq3Q2-Hs|%!z^oi9hz%nXeztjw)0F~7Z
z7&*cJ6Jr)Bt$ADr)Ws|S)QB(}btSA)4~(42Fo?u<sv&vF8rGBwH)6E5+N3Tg%m)Ai
zmvdI^6W!7)dAEO6&6Q(-6(TbCZTI~xhp5QqwzjtSPj#){y?Zx+uk#MGsJ{q*gkL})
z4Z8^T>9bi`|B`5-O(uz?9yc*ri4Bd+xG<JNs*2U|<B{OC^<u-%_FD-m7~JSll2^QY
z$1N!-2^0N5V2$vD6D&TDkq`E^V601=8w?s68YHY||IHo$-m?S%f-0~vQ-96-jt-Ji
z47+kNR>N6jq+5-SObw$9n952@?slKU|J`(J4|huu{AOOF3b(i&W{~w9mLZJ~U2MVF
zPPl$O9W^a#D(_MeS_Q2JpEoMSXeCV#A5KU8VBhsrnU#&LXJ8-<EB(s0tGCCqH@Qy!
zt2;nH`mGJlkc@D;@?DtzLP4Op6#g58XQQ&+Zf^2u<e6~2AgF!z{5dTOyZD99He;M`
z2ou9k<LVp6zFk~GL;Su#{`nlO*s13;e*vMWfH=m$HWgnJyr_~Ama#B*?vKo_$<@ry
zdD+HQilTgb8eg8=dQCunr4kpXeiA7laTYJ<<KrWlm;kw4evex3@cqp}0Xt(E5_DjY
zW`!F_45;C)P7)=NLnw}d5V1OjN&DeYfU2B0uM@=woFhoSAn~6DjloD{3a~8T4`Lrp
z%!dhS4mZsj+c0T(i8&duiKfKfxWNh(1iX;Sjt)VTxGzznlRUM)FJEl&ohW+?OtCab
zwi{pp)Xt#@Iwi;eD$i$7(Doo27YAb+$dg372%hG(El$jo1l0hQD>4U{2?`35+&qj2
zD}k(?$H<;UE)4=|p^%ub-*}^*zP=FPZEE}XuRu6cV0n2!^!&$6|IEQ{EOS4+K9o2e
zCTZ<}Y=~JXiJ>90MbJ?T@874@EA`6wYBui?c5dLsxW#_h77%xMY@j^={zk{ggF{2<
zz}k^EA`WAK`ytn<M#M@8d7QD<h;1|4RKO<=-@a%noPs-f&8t^cxB|`~?hp~Uvd`3W
zfU!kTI|8O4bv!Ellt%qX2jaj@UWPx+efYy3edS#25)z_!P3Kcbrh4|h)jK=`33nJt
zZ-B4$WiNdl9?oe$h8F6t)li+;*WNE*7;x=L5&{Mm=q(O7y+XN7E)-$4QAmbH3La;2
zi@~9{a&*jXe?Wc@+7Vzyg18phYQDa)jT<F`fx;%3yvIT7jqb??T^S`1WHAo%1;9lJ
z!ze2qi;NNN18(J&_2%c`+D0|%MQYT6`2cJ^_#<XGk8FBnyIWJ!w!bbO3FCO;4P;HW
zqHrXhdc=AO7mhS&z{S$=+pRuRxekdhVStntu(*KVebb)T$+XTI+~)J;jWD52ffcnJ
zzJwHws<%Ee7ogVx1<?~>3dBYbBe<g1UJ8jOfp)G&?@Ci^hikF8#?Zli#@!>vb4q8v
zG433nX1Kh5V`6-KD0*XZYEvl4HYApWC`j|KiLkaD!brXfV}BB@g%qJGq|d;9ncybC
zLP*j9uz@{;2~>p`?ej~dX8k8;=6n97-Z2k0l~DV^*I6icmwRbVmcZ2n#?vHPtKo5Y
zCw9ChJhGcaelE|@?)Jls5$M*@E9$fqw1{wtFqk}6WO^Pb0E{8uhQh5N+;pz|LNU7I
z6%>$f4n&g*5bN9XdvO~91ZTXxL($CAQmjBh!1V=B<;>BiY?MIKpQDLF!GU3bK?^ZD
zMVDiwfSh*t3!X(D)TUp7m$W3{dXVX6?(lw{_nXz*PSe6g58y!dK=%PK+w40ZZ$mYa
zf98hdICG<8zJrpYyqB^UpY91{nm+L%;JlIRXcnrcRN#5>+1l<`ae2sX$H4*~1zA?u
zE$v)m1^x_zfX{U9t>5!J>3$AJ)bqg>o3%26AJ88}WWI?yf1Z{CI?bbx-(k&w)QO;b
z_cl|-<Jzb0h&;#hb9L1ZY+C5Y>dxbcVWVJ1W@6`C6DxLf;Uf}!0d{6h`QV@il`#$*
z5O!Q3C|OV8jlh^Pu@C}ZV(9fryKiVBV)R0Fac((<B&yXl%Wd4KhWQ^P6f`^C8y?(w
zCBAUIxY#{o*q|z-)^!R(bxI(N<g%S`vFzLOh2fw9g!7eSj=hC)My*+k+eyO-@1Hxa
zwG#g^bTM^F0hBLR8ZJLNmBC`S1#D1w?_L$jbH*V-kY^xkvaYAh(!zqY&$n)b+K5zy
zj~;C&pB!^5OwP!TjdKui#!FbFM4DK5?E<7uUO}5T-`~AbT9Mygup&Dy_w#aGQFS=4
zQ7Go-pH4r3zF+0n?>i=I_efma)Z3r5e_i|)pP0oF={T)AvuQ&^L)w)4tG<$>e;cq6
z^B&u*ns-V#3{|krHJ7N?UyhHb+xeu4X=mOZjfoxCb{tE7QxVV*8NeM?%@Ij;V4FKP
z&~k5TdGnX-+pfJw2ds=<s|LCjijlAcr1UufBO(!KWDmrRk^|VJb4{KJ-S2O>kBq|K
zvg*y7eS;6F3Sn%ON)l+B0XLd_dOQOSB&pt2W(L>z{`%xWAUIgOhi&&5A^K*pE&KYL
z$hU9qE<QdafnZs0dJOYWPNmxF{iAFivjK*m0Fxa$F!NwnR#wJ}PXE|@stfA?wM1sg
zd8}QWwvF%JRiRZ%F15twbj#p)Vg{phIxu$1-mCxmcKW@(>EAtv`B09Z@%Y57+AXH-
z+aEC}-^)8xyo_R2?r)y=0r?3C@L@{SnR-p{Yt*=O=M8ujFs>2fiyA{<ih8QvPDB|N
z=OH!|Ege}A`;p3ot2+vjgY~mOnrs(r&t^}3SCOp`9B}86kK48|W~udXrMo}Jha;B5
zw&-dRQ{_GK6f-M!oN2)%YZ`y3aQQE&k$IZTzDW~}6P6g0;^g6kB^$QnHQTnW0*aF@
zO_fI_?V|dXb3|E-vrjwAuyyum-44RpCj27ULDV@)1EDvS3B#HNc3B9gh)o>;gGTWY
zs(@Srd~dm;kI6M+%pQ7<i39h%T`~{gC3_xy7RhBHhqlA4vfn~vwGC>&Y}u%GWOPtS
zqK+PhVs?k&FbMFHLc!RYLYV{A`slUKKny`sYKc<OaR`+6dV|e02hSFnn1d5zX*WyG
zD(oGgtmhtb#g4PiTiGic$Rnk0y>9Mtwm-Q=-6+c)3K;PGXgW9yMJa1#Z$>TQjKX<S
zeZ`&o#l`Sd48h!=lyu8(t>I$FRB#RUf(YzW)&Sj+#Igw?UpW8HBa4TkmzTvxL8<Tz
z&<vr$0^8NftRasaRbijS<rTuR%Bi*-abd*k5zd<=e+@^I%);;96{zURB?o614w+}+
zGE2x%DJi1S4cmVwg|7Y5>Y<=6!3CcwTvDwBib4wV^neLfqrvX}{5c5hJyf2(Xg<js
z?n_pQCYe~c2G5DbE2gU{I0lK+2!;nyxV^nVJI(@1D;@x0|6eR(wo|80`C@$$XF=$s
zrhar?z(~G@HwFP{yHUCk-4o$h;iTS$$@Z3$A0vp@F>+3E)rn^b;tL8P8{y~Yw?e;4
z02)Hx>IzsSfgdm6c$*5G01L|oUeWI{bR|JXIKe8>&B6q3CI0a;kY9*}5U_PBrisP=
zBAE+KecQH!Lrj)4iIN~uni%(!NNwavoz2U0%CI<K8FJfI<5C+pg<$)n;Dp_?C=|R1
zTCBubEuT$80k^Ue;494XOvzLYbHNI@kCHhY#w4Q-Q{E#O80R^Qpv(}^Qa5~5sAj4O
zq!6&64RRXDC8a!rJsY7v5JN0<T$aTrJqb324=2`(qFMQ#)BIkc_Q^}TXZh;-YwBKS
z63BXJ!OM-)<86Kgy3h0O*Gcyy4~`uj92r}TBsSzcOc43mN{lf$F?}Y9Lhy28qoDDr
zz%-Ty{-3NQdI!{9;*d&;O$y?PLIHs9`u;RfJTCL2Pm;2*&rIUWcZ(_~x&pcm)B^}b
zTULe9AW152c^Dn7n>Y{wFDU{PAg#h`GV9JI*h7m__E6OfhgwKy4eGm@xjFjHg7yvI
z7NLSgF@_cW0@6b(phjf+JpZffC3c66*o=*?U>Q|_EfYVBGJ7d=2vaFUTccee(+t!K
z+0swJQ>b;_g1R4!bMVI_9bz1gAGb!!3{k9kqNPpd>&I{1@|QK*pQP}X-=t)aKu2G0
z9P`AU9sQUb2CUI1;I!F;4Uxo}5;JW|hBhbmV_~w!ApDkIc)`utvKe#mjertJHxl?(
z@e!b9tP>5iDaHG}uI`06<q2z7o7A(8j>GAG{inC#<R#M${ey8A*7w?-{<`7qQQljG
zeq&gV2@9e66B}#nR-vXc#-DzD`xh)3^5{ddfeAM|TF<&C4tnBoJI05SK(4bv<(#d$
z;n00XJZ0-Vx)#@GFkW77$q$wUi7SE<87(Cy)eiMr@sY6Hm~ha-?nn)#aY{i!!I^xr
z-6B*4<yba-)Y-E%WO1Q5A!0UQ__%(J@1%WpJ;HF5>=bB-i1{;&&klS)&~-H<Bip3O
z)mVnl7N3I%BwzvycU>0faZmr)p8550_G>db`DDRsk8e_*VYetma|cm>jzH?QR!ShL
z=1Bw}3Ruu>(v?Wl`^pbi2$RBsA(PL_uM;|3ODg7+NFgj}+z4QWH{P30@uK$C<kLm|
zyd{!ObcAN_jFB^9s|HGE`B$wXeohLPGkt+D3dkOJu(2_)ZpE{6b$fn#dN2y16NXhE
zK9p9@N672rlLL^);i-_ImGoyBqvQhGTW-jPZ@~O(jkp72a{<F2ZA12&kz3{(RD-RQ
zl<UP(-*mS#-uMg7XHMDvMe2uc4c;GqxFE+$X)K<m)aYzyzV^pN%;IKbBokQ$=ut7d
z(*qJU$@d>ttdi9JxM0D>fc_L$*A@32Kd=4E1j=S6WzEFI1j$PI3?os@8_CQQgB^Of
z#^TGWX=*YM_bi-cAu-y9$Nt0Pa*ughBIez}rV+D)CGVP{Z~3afqrDlCB{CB#u6=jq
z?k+zcvTtc_L^V2&W`U%zfe%F7Ua^`mUTD#N%s?2@ORp9ZME(rS8tIP(C=JMsLoS=q
z@~jAfFo+=>-xt#0%;p`zDzYFJ1OPo0V$TnqgEJ6msd3OBb&bsw7-Lpn`YhR#%si!y
zEn7H{p4LLczHiF@-3NA?NV#%uzuD8Q`8S`?0Vh!#;&w(2O5hTt9rA7w#+-xj#W7_>
ze|Iko?5tBL^-1l9^OD5Ue;M%sR|XQL4Jf%nut&4~gRy2H$z8)MyZ`g%bmxD?c$5!&
zqW09mZ7{3`K?6dc7?>}7My0BTFh$rZ#i6l+=!|r?h||TE^mfo!pVI{&ay5dyu3fsi
zVdjP360PF;-Vg2VHYgbfo3k!C=wM2Wwb1zXZ6zu~Ivq{J#p#)uG+Z!D3$u&<KS*)G
zFD4>$%YSfLhrE#2+kZ<hti7x!zF9S%O+C$f0QWor$z%ks0wP29H<XANU9Hm5)wMoy
zgchdEXc`vAuvxsw%(~gA*qryGuVJd=Lcr;5@@Mx*aOEC;kpJ5B{~(T+&t>)iJr4{I
z*Gw$JwkZS|4&XNKwY+ef0FRFZZ^P`IpjrUIiO=Y#M^a0{R%$~ld3Z!+WLPnB3TnLo
z)aTYyv-O&4YG=Wt;Br#l_<2OK`(vbbP;<cL%gjzrvNXeY)HF>r3um%wk4Hq8ZLpem
zdrqxpGEx0UnC{R8eM19U*^P@sdL6uVybq_^ba&`z@!=gU0wc7KjmcRg{F+==y3G2^
z57C*<HEKhzSa=!BPqQ@r5@9qf&$M__AZfR&SAFh>o_gd|y;)P|<M76~h|F`s_&O-J
zt5A!C#;D)o0{0#wTf|_brgJE4#q$YyS|H#@H}}x#T`R=Vza;<U1dvmB{GB8}KcnmZ
ztKd&|qx|8@_20LT9h)H|dRnz16ui9%bKuGK2R@28RR%O>Ovb9)w;#rAVNXG7lJ&=C
z4prIq3qg;sS>~_pw_f+wV3V_|WYLQg%XiiV8{WEOCq8mAXy>FzkBc|o$SxMwcN2XK
z7h<P>GufnvOz(=xA7Qs2$PE?T)8l+{M1PD%B6IGeuh0iS*0Pz(Ia?)FMM2G9j=?ec
zazCU$l%Ae@*vmTM9$I>OjiO~tvi~f%w^C(ugY?==#UD8<?v8w(n}0iC^~mm1Q&ZX<
za$N)fAV*Ne(Q3%znGGX(iiD_w!!3Wy0%H2awr(|5cnTUl3FZX>E;Kxx0iAZw_wUg-
ze{F%xlUOf`6_BrN=^cyHUqs%2_>lFc2viTmF?@!`ib4YUtQ&4B+XJ#^+hLQ|eiWSo
z(eOZ4Rrm2g3w#ugy}3bF7jS0EWsB4kz)#P7Zx_Q8h3J9`+%cE_I$jL#$&3LlJO$HR
zTC_-{Z@|PgdtAkTj(<4=`*W1d;O3-cWQ0Q5KuoUlnqKNsn`1E%m)pF#iuhGy)U7u3
zz`ki2@*v1jN1JsD99ppS(oj1GH9sgRc>y%5vCRF1+R)*H2MPRV|KV<i=%lQSj1?eR
zt?y+=1_t|U)@cPVj!k0%K|koSutuo-5x2*?HCD%uta!D1N?kMUi`*8ckZDUl+DENh
zf|^Bol-NWHOjii{Sa-H{x?Q#IXs@-G-MsSF)1!Av-_K1m*dKQi)AH*$duh0OcBo2f
z<5rJF#>;EZ1}Cy89h~3;(!$vyhoLR@UWkS6>jXO^ngz%Ta4!4Ac}&$n&_lqbcGqS!
zK6>bbDNt)k!2W?4Mc;R5VMgr{iisK-r@svUqrAkXwi~}MI{8V7FctuHBGb=wQq?N$
z&<o%}Qu}T^!{x+%W}^^j0ev3WyCepK)I$_!Oj!s607Vrg5cxuX>_8F{6Nx7q=66*9
zUJ=#E0F%0{+qYlMQp6|4jE;sR-jg<fG%65Qkqlv+P!1DC+?)zLOXz%@9V$}e$K3h~
z_PK4W!i$Sl(@~MBYuDCsZ^czOR{c-)!x>T5y_Zw+4w`H|G{s3S4%1FR@YqGcOAuJC
zJk#~Ihlm2P^P$E!d@uyQ3h_lo{zd%hs}dyZqX;blB@km8;^l$F^a^0K#SOrtNRSmg
z1R06BDfn)oWZsJ=0oZNkfTp(L><%a9u#k|<7coNHf-rT6pr&o=1<9Qfh@K%k=_PW|
zF`F`0n6`VJaF$pAp|CtUwHZ7UKz@WILuzW0@E!Oz#~-+m_aCkXuS3Uu>1(Ur@3goa
zuFaH+Iaivuj`)?7az`GegkRKSul(^x;^H*$J!Xa5NQ*Dw2w<#Y_$${}$sav7`jadH
zfeq@o3?Nk*&mTnfVp;z7S852Y^-1@g61Py^5MnGQF5wQ};G0fZAaIih`EDaN05~+-
zzO(`-Q%dU2%6vuW_|RpvBqwj-*|vz;*@`9P))4%+kdcz$6xk*<gDK^Yx)(Tb?OSYj
zD@8X|d^93FX9>9t8#bWiYKh-VZUAaV!U;rAo$~@mHl~??!zU}3;J!6$(l^=S<fo$m
zH!mOciX6HNM?Yd(3wH6^wQGqA2ASPwL2x4WhJXh&&Yal-o~`9bOK|*HSMPTBA<R|k
zrZuUC(SNuV(Ss)h(+-cDP$--$_V?_exq65k>}f;qau04M6GYr6a3o+ys!C+s08)tR
z3Y&TeXnas(qx)sn)_%1zKLge8mfIN_lbgmOhaLxP*zAp~gez=~ltTU;o`nwBr-&de
zLoj49DH!}`n~kbvlH%F>eB=KS&(8R~y6Z*st<5#kulbai<IY9z5qz(3a)scflE!yi
zc$C;VTbBDKBf$F0D&Rnbo`4>nP|*l0l41qKk3><RYjs8?ef5yV<%qg3pp(MUmdyTe
z=YWEuoZ!Bucq#@rbq!!9d@q8f1=1`>rAIaVLfTNwEqQZQ0(j3wWL`Pv>$Ylt-&eAn
zW(?BLtq{!?`;@CJ#y941K4P#xV?aWDuJhEO)d+LA-Fotzr5&mP?vm{LlL@I^bO^I(
zG1`kH;XvqAar^=NJlNBsZBxXLArC0=`H}CblSIRc&&J8c1xS54oM<hIoj@KZi3uE#
zHe(Kt75c06i$ilJCq)!Z1%eiGTR*ZH;Ur~dVF7vXAFX*HUVzvb4X$UdXNH0bT^<<o
zg{T?T{69pZOiMeoR^b=T&6;<qIT;rQUdV>VXc*F?6ENg`ds?L1?V~@BbX2TtfQ*|E
z%DD_Bw?RsxPT1g}{liC(s+ya5C{{;~gaCT(Mh{D%UyM1DeOAY5(Sim*QZ0z5HlA3b
zrL?u}z2?b3u}FFr1j?qRiK`7IBM2T~8P}~d<_e%mlAtVfUSQ^iI1rrH6cRA}^D(4g
zn|R*D{j>RhguMqe)_)&Ae36JEB$bt&goY7XR*EKBNfD8qnIt1363R$~%<NgRXCV?L
zD-;bBg@zrT*GIqm|D5MM&pFR^?)%*LZ^U(dug`eDU+ayTFNCJT91+6d!I1e#K54fs
zNN?uAI>L!SkK!PAdHjR;VVlHKGu)pW4|wdw!=R!-`%328{>0PkG-o1Zxn;urKb#bb
z+uLB~KtbwMoCO*B7A0$0JjQpkMToqvP3A5=?Y3+Nu*}Cc`Un_h>)RNLIwu>>agphf
znqOmv?1)MRNJF!7r6RgAGO1lg#$%|WB`z3(0?}(yjJ<Ei3r|kFum!06b=7HuB^hd0
zLLUOQJ|q1B{1-q8-fi2C*f8yn-Ln*@OPC1=j`GRqg86ykJ;WQ=SlAmUO*i|yEk;Fi
zZdL;il?GayhuaUI#fJodMW88MB-$nrGvTOGg*XiOp{3Yr&Ku4cia-zAroBGs_HA;B
z!@>cv`JGE~D$e)IMK%YVPI>tOofCbG-(g9iCS#e$0fz;?xvFIh#b*bsDD7SIr5`hY
zzLxwKO_Cne3Ze&PwHD`w02Ly!<(!_FsK(q<7U&;vF>Txw@u4Z9+?|RHCk3-Gb>va1
z=VQ#I0f0iZgUdz@sMc3$mgQIiJSx$>rY^|3*TY~@lo_oBDt}GRek8&Y2ov@*u(T^E
z2>!T%`y|3G4%b@b;o~Y=^AOa2hP%y2GhIwF2lq7CrcQH5W=gTD9(P!HI1P?WzzH=_
zar)3Qk|=r@aaU7jU~X%VGf>eg>T~<MTRprkPpaVN^(S|sLJ|#+dgy2Jrk7=U<|)J&
zY4=`&*uV09KU)M_ZD+DL>5T>b(b#52V{bf$aGfu#amGc#@^#Lk{Wp{}41NY{<%$_r
za^dc_T;RmORsvl@j+6fLD1j^pgADv&0$4F(UNA2K^fv)af$U)x7qcuMa0?J1-Ajh&
zo|vvXbp0)}#gM%RLjgmZ{bUXO7q$$;!Q(1_+JnA4J{itdTVKqGZ|2j!W6Rp(w3?DD
zl2nrSeq0R8qoT*yX#}Yy%GjEl`$KyO3;Gc+Exd)(WI8X_f(K%?0Znt03$t?O(`s@$
za0mIHQ;Or~tZOpHabMH+;v70Pj{My}C{~|4MMO0Qa%Yl6H11WYtYgv0Wj|-$m9^(;
zHxoFT<&K;qt13>0KCoLbs}=nsl!JNJ?{*d1fQ=wZRKUkHhjnf(iRSu97*^gUW@w0N
zXHW~<{hhJ52Ec=Y@e}JN73Y}iPgu&i9#sQBi|mb;Pr<DdpK9a(r(E#w@2`pYj?(Uh
zsyn@DOx(SUQ~-&HFSOS5dj9-MvtE<%d3S9CgWKRGNpWrS007GIxV`kK@2El-)-IWz
zYak~$WIrU|nJ@)t#z={V-Lp=fgGHMq|L_`R?sb4{wQp%1+3?<mmYQOlYM^lSTxj+3
zQzjqPa?9OSUyY-D@hh)YZ=-*>6#@z{PaWsBK;MC{GP~gSdUp2Oh6Xwm!WzAu<*trs
z?vjApYwL_3%vaQFAg%du>%#^d1Qg1nM~_IIk0(Z<T&kF4&y~m3s(7EXC(-6@4mz!>
zPa#-YB2I$*tS4<ht}A?y3w#SNjjcbIcGIt8%_vgC@R^gS1$9U66VWojCs1&$eHT$+
z6JQZ|#$sE;LhrlB7s<G<C)V#+>+9?D^76dENE1*N%Tjm5-ux3TJyJD7UxOe2vu*e#
zOe;W4^Dg!8FSG$NiCI&AK0eQmbfq)sE78R_-BnK^3>5<4VGv4vx(|`l=;h@e+#tIo
zn7!g|=(-SVS6ZIo)cB}`Ahbk>DJK|q+3lLrC7-VBm4d`pqObu}-ZC8)3$~F}-Y@!4
z$Zw;JvvW%)9^B|!s#2SIwM8I4TPVHfY6uVC{{5Aj4Unqxf<cES`Z?|~tqo!eXK>X)
zK65uL>~~uOl^3MG<*vW;EIUMT!jQfO`X)ptso*>$wIwJ5BC#0B-`k20{t;YXsrt_d
z-E)eV3y7O-LTT9T*B5$1lYj`%v;nfkz4-U{3BA*&E3}3(p3|P^c1FqL(C`Xpt1LZC
z$xDZyvljO-QN^nWycZ_7Jk^Dln+3!!ul;h+_0EPt#dI+$Qbi>$k&BGPk)iEf?u-FA
z))D50h!U|uvD-R?KOJYd+B>XOZo(+Jlj={3P{x5Sm9U9}(1V@_+EMI2F9U)sW3m`u
zrU}XzJ7?!Qu+2DznygP?LM-}<Z4pbZnDt2nClJkdKxWB7yJn*ozHfEI!ANlOxIqc7
zW!W1O6N4epx6r#)I*nyK_e=-0O0By&raHDg(;)Qjh_7ntBWosB>67n%9gVp?yME=t
z5_2<ipAVOx@v;1NVGZy<{2sK&%d=Ml{WHh_EutsGu0HftM5UvpIWJAeW7FSE#Z~Qn
z4oMMjo>-hB>I*R8<lDg~0BeLzx)E0$#1v2TacrYUC8R@;qt^50<~6920(!3C$C7ss
ze9Q7$4$##<AanbMjB?La*g;pI-^E4K&70){xb8-<f=WdAnftzwKNR1+Mk~}5ih2yU
zxv#^Qs8xaDHcZi*^pZl}+nMLff%j%?Cv6m3ZL}xLUj~_KG}NqUWCTTpzv!Sv*pB^`
zm!;?-vLfscU{qr~ZX&lY<OU=|IK(W&Djt1(>FMzoryrYr1WBpO=?U{0qLM)2Rxq>@
z*BluWhW)40D`Qkmx#8zm+udEP{!2$1R--5^L}=*qSv!Pwe%nn&k*T+dPfFQl2a983
z(@XJYCB5%FU$!~CZjYBE!ohobXTIuTuXxj=fJ5&O#uskNMA}rKT);QKtPj=N5^&Hf
zO#6ogD-7D7amO6OLH~DQLAo;4#5AM~JSlQ;UW47q@ISj1%{{T7wAZg&w+=&07yIUS
zLzI;Ws2xDHM(+@w4DJmrON-9v`s1uDcfD{=7Ec%dDLSGrW6vhbs4-cxB+D8#n+`o(
zqNT%@9~{5f-kE1y@~ym#CBt&J6y9>oX0-TQlZYMwG7ok2rknyz4A8(=Nhbd#{RSw)
z55OA(6xNSDz=Q(N#o0qi?g1mLxB^jRe}q<#V{_QQnDfG)O)X*A`NJN?i(00`G8*nW
zz0o&#czDQAGF<z44uegi2P9yL!cGC%9^3O2=j}%vO&PhkWHT4u$BSYV3^Nv;HYcVg
z?8k*fESQBhUAcybsyr5&+~0E`P!@ZMj+;AHv{gTX>{-^6#+_$h1pYG0>ZLixm~{hX
zvG|zj-=Hei;7>|_HP|A}r>--zbcn}_&nG0gZvRubncgNq3maKQD-cszlVHk2DTqUJ
z#nGcjo6b(<!uS9MAS!|y&~mE*k)gE7yeC9Vi)hKH(8@irea`{d6*kp6K=q6e=*zdl
zxj>?S@z*uDRq%0Daa)p_8<7X3Y|NfQBM9SP5Gw3=@O2QD-+1ugOqAQ&vn38I!6+EI
zPe%0oBv6Hx=s~Os+%Wio+Akx(sY2@p&H1=#RfI~!BQ4Q<T-#E%eH(G!gTf%185~*d
zQ||IEpmAwoHN!y|QPLGD{QFD)?-@K`b9gYoki1SjxEd(O0RE1pvvh1Pp88D~ntOi2
z5<q$wKuuH>l%;+-w%<mM;%f3Cqgh48r$=}a{Y1BHp+mc<HQ)5=mHW_6tY@hHg;4Qn
zrMvyMa4-NJjJBR>$qIAe(9IQ(yzK2Q>&a;`#ED8~ychn2*%~2KymE4<r7ti3bwy($
zAg#R+`7sr_C#XxizG|j#&zRaG$e+<*+;)jV(T~2&E=#X5nZMH_@J0KPEBq%8skB+h
z?sg%9IPIPJhi^{}{WEwJ9yK=RC+RUT?S<^0TLo;c0iYIEoH|DSf)9vzX9)Ql`b<c8
zSADql`xuI;c=*tW=n^;xnY#<l5C_+0tWDwohob0iczAN@E=92Z2`7y}E5GjStdBJG
zpvM9rlZ;4}21JYy=NWR4CyAXdpaHH^2ND8EnIA5u&oGB5qje{;0_Z|JTwj)15)Lfk
zB!z;i|Nm2ZmicqfF|A5+Sb-w;!<0sx{IqD+35$S=!E%1TdoXH+c_McJ^=M>n4#krz
zb$H87Y@3h3_eh69!d=k$kt&yD&jGx`y|jW75#75KcX0_(5Ht`8efre4ElV|+iylN8
z8IpqTt`_=0z!a~+8ClNqtkb^9DrEN~VJ9RIKmWVp+kV3V%#A&5`58Eg0?d(3`|6>5
zM9RfRX-l?8$763HH^E%%NZ4onr`GfVDw?({?16NVaoiN6pP%Xge0r-o8@7W^ISGR~
z2(^e!Rp9Ti*O>vp5-J+(^`MN_ML<Yk%M}G#wXIq*YH5)auMo#}fmeDri@veV6Z7x9
z@W)yv33YmB{$TzGK;}LxxVuj*==+UeL`{Prv&Jum)~FUW6A5Rcc!h+ly*fGKi{>&C
zR_vo|h&u4-1DRveLr{XDhk>Ozy*c6v?7;*e!fM%x&@jkHJ7kb(gaw~AHZN)aX(``j
z*j^)IeslGK;pX~X&w{J-=tTlo-4`|+)U*vSvn>fvi7<^eGd0_K)?j?xY{<LklXRJB
zsD5;msqI#Y-LK9%ZqQ<0dHTXdzjq#<8y%HT)_6Vp<yv*FvPrTvYW(Ki&jYh9zxTE#
zg5K+{bl*e#6VM>Ckzy1#H&Q`@A-g>DfDfs8%1Biq6SQDw`Gm_Y_g7a|hviO;<t@h&
z=`M5Gj)CW-7t)=*mKYrh4Jtr6VsGp&y0jVJDpg-8JIN!%aio1U?p!X({?3+k9UGQc
zgD3$)j|tF4XF`1a{60ZtR}Y<NK8)X#YEa)9?l%1|0ihY~ikcvyhM=e6=h2C170Os^
zSthrN!Y%Jq4U4)c*Yxk-WuvY@P{<tPJm9VHs3kT{jycC+NDTrs@a^0L6^7H`%;rFt
zp0wJt1J-w>&$^Heb+w|hc76$EzyzXn2ogsAhAX@#jAP?m%klqruer;7No4gno^bNT
z$n@2A70<}WHj+=KiVNTTsc9E~y~*mL+$#PIcOgIijYron?ZmNvNzdzV4D}b&BfnQG
zl$idEF2E{Gz*_4ZGK$&~cZHgdk1rxlU==kjBZt;lHzp<#77WtzdwX(pGJk><s4?1M
z1a0^rj~f&a*{>8S%m-QUozPJTkU}A9Ap9E`e%8|SM?@0PFOiCY|L>T&6)+NltDzQu
z9@5axAzr&oSD^PZd*Zr^F7-RS{QJJu`O~m}xs%)fv;f1p=kFO~slvCA)wS8qvr9(9
z#Ox`hA*ah0)fj>r^`Dq)e2mm|9sj{k1hn8|$`_`3iy>}~u=+4s4*8y9>RUb;nWy@O
zM{JmBsK9AGb8OwG`!SzDwLM&aVd&v7aAY1>qr}hOI3xS?M(QKE+e5dGTws<~<r`c3
z>1vJX1@^Ob37NeU=D$;ugR!xaCE3pF{&G3#&)~$wMI&@zKuKlt4rXBY`lOucX*AA{
zhasZ&SJbg%I3MBO9WQ?J2wN}5u<G2$D-bD{q);Sx?IKBE9t(5Ya~JiWvi1Cs<P&{R
z*j46oiCF0(w;R&hS(&j>o}jd^)dP@)5%WZl6!Bq|V8(dUy^+lC9Dfjn_}R5+`yRd6
z(PZnYY_N9>U76sp<Xy(dEF^yPe5y%5D|2iTN1DX49*Q&XN!c{jv?PA!Y$w;B$?W$-
z1N|`m1dt13d*zIAA&2p_{5X+RVCNH^2JGhpfG)zE)*jGcgE-YN9x*;pxRiXBRb(dI
z0-`pbGQI7HnR9h!W|3xb(4Z<wy8~BGES@v>wTIRBes8lEKd>qCQk<&eu#RQ}f1+sH
zrojAduG|cN<m{f<`_yULuNCG`oKpyI%p17*Ij=DCk>c*s$1WvNR>!oVTcqGg{)!M7
zYj-Mh-wKSRt*n!!?o6Ob<AuRX$r)Pf4?NeDdHlAqF}=?*a^E4-R`FFzB|`^I-Qu+`
z-#%>VQbs%8`KR)&G<iS;qic(u8E+-73R3lffVpfa_j`{_)KSE?PnQb^G^wm~<L&w?
z>U^8QwolA2@RM-uUMC92amT#o)!WO-1hY|x8fp7DJY5JiNILglj2W-ST}P(vCP2AN
z7%g&C;oe)3Ru;hlPX#;zRtSDY4PM#SP=)L=^@|)bZ2!!pvH7PgU$>-kwyl<U)|=ql
z4$R=zpLztidEUf^=F-af$|1W^wB$#{3Bkmx+Yx0*%*)7A#igP5?LI_KdYqyv+JoR&
zG}y?Ygx@2=wlr^)jAaP;CgaauAP9Mdg>_&aMBU53o1Hf|pLKtKN>ek-jXgpRRR9IE
zqE}H~C)Q;@JQRX~h$zn@Ij!>b>=23w?hdh2wa5;mz@S>=@EN)1K*>&}<oc#bM7W>8
ztn=2+cUMfCavo%)wKyB^ilX$DF4|t)u@N1Q8U#xC(>n7#m*I85sfxGIbhZ}~-ppTJ
zO2=9wmv$SBtcuW|`a$)5#(9fZsRa{*w_StqQmlXCl`coLF@89*k?db=@OIIob%hs7
z43a89bkO<eOK~nfpSuq+TG)yG0bnc(YaJR2)g&TKkl#kcR!DrgurEh5oy3BT{u*Rl
z5Aqt98&2=x;D=-BHiS$-^COh`5RE3zyF_&>u`bewVjS(Mf;n=p<8Z@22GftHF2KvH
z5IU!#HH}%x1H)Rgagu+g-#0hcC2K5sDQ$iA{Yy+U=lx&5j6~R2D{{?-l7=S^o)k@<
zeQrAR>e};i8Rdc#*UX05+h^u8KInO$ZS3be5y?mwvAUENJzl$>^@s9J;4W`qwD@Q-
zWc?6obfCf|;Wk`QER*5SM~~i~%!m4%1`3_dN4@X;6a7IP5b-EM33ptZS+boPagrzU
zAViiS%~P-BHw3>VKw|V&0~8UYsUUF#P)Sor1_v2|`RK&MV1#5w^rqTCyG~}-VXPRX
z&+-a&HHiK5G7qwMAAq6vwzs!8L&?j^rx5FDmG*7buo*?QO&LJvMgrcl`yazOiXc+O
zh=-~2U20Lp(+}VSnHu>#XHfXIy?u-90u@R})zgKA1u6<uuVP)i24$CRe47(Rl(Kak
zT{SbNM=Xa145t^fGI|d%JC;E;2z5nAx2q|d|2&*Sl%lZ*Xd!_=tk%zd+H<TWI`O2)
z1hR}jrl-c$V!g!UT7tt>OE0v9&)DY<kse}4o<5+c+huHQOoldV>($2Rkl|P`T6p-L
zQ0$1nhd&QnT&JgXdm)lcjUi!&axo&=1K<6}5&fk{bKoSkIJPhfKM(fn>^c=XcPTg~
z$bP)#`c0ka>B4ElZ-Ks&e$}nL2Sexf`H9sY3Oje>grvoXUgL8QG?K<MQ&hQG!@BTg
zuMBG1ban;ZGgVt%lhzX|ZC61+CcOQ&<6;XJnSTsT;re?PYzJBNvvm@0w7-4@)Ef0j
z9T5hiBnNy<yck5)il<>Dqia(#0(RR7%?XJGA&N*O&!E9tO}KKX`San31E3tS=OS)g
zB&_Z^hj=;C#sc|58Z-h_3q4L0S#>cHnt{kU(N2=220Is*2FTu7Hg8@F&*t;GIx0%L
zr8KFeabc4l80*akn;DG(e5-*_zB)|0p_p{};EM3D2`XqM2H<VMPb{T9r=6Jb4Jcht
z(WMY9ZeGxOXw+38OvWMO()HkXQD%$tFI~h%)PP+jtTLEd@(qx(6UBFyNrs~h{9arH
z0YDQ${IiKrA3&p#H+FYM*PAyUe?@fHJ1~hA?dfOV4d;-WEVZP{n{j)vbtpdM#4|k|
zEe6tjPfXxS@;W0)i?+dyjH<*BQ^n&;3Dy!o{|x>U{ue$(j~C3L1yDkMS@NW^QgYNQ
zO`R%Wz@hEBB5V+`M`kHzofHv#P$9C?C!MV-<?oH5mTXy{TmF{527JoD**(lZXhyH~
z5MfiFq4TZYtq?seouU-1Q*+LIlgO(B{A><#wbbq`Uj&Z%=tx9%e4QWZ_#615!&E8T
zskDPj%kG4DVtpbyZ+t~0_ZN+Xs56&kGWr%DochLpXMU2f3JL{IGYE{=lQcyV=rCB{
zZ$ba0x>|t?`j+h#e^?2l4bX<-g;AWiop4j#VHhT%y|~wlTGTZ&JCIGgmN;d?Ig4<G
zOI`ZFMHC+xa?CK$>0nIiROngO1Pyt*orW$AkqCX&iwpBVF-es~Fq524z|{fZqQlRK
ziep_s&bzlW=WHOL5JD><ktLRuhU&n=pLKC=V{)WZ1jKwDcC>XT_Fe}TM*xSx%8M@-
zs|u|0<h$r8dnemZflaB!%F3SSz6_ZF7#Q!evNC4_=bRsqW00kbFRdP_$9aes22d@4
z!LN9yX?NW4V2BVFY&h>Neh|AaX!$y}slAi=%KS0#;Hh_9)*`6$x5XCr-C5RB-qdtx
z<*yjnb<}JHSZ&YqyExB3y`OMLI(F?EW@(!164~eT?<Y8~x{;kYHs#8j*n@JC4g}%e
zzZ1_i4<)GDf&(UL=yqt|)3iiKcXflb5#4=UIPAT%DsEU2Y^Iwyafw|YM^$ZBO_e2}
z2RM4Xpx0JLAu20d+*W^6<Q8cNKe!Ha3vStxg!mtjxDg%$^&ruz7N$5+DF)-l0?bm{
z_u*F2Jh~9<OGHFl+h!_;lATH>pY-(^;7<#Pi8Yt1q}3U3-F+D{W0VGFDXphsM`?;X
ztYZ*1OJagilL+kCF%f)3TX(o84!SXDA=dwwrq%hby2w+UGJ!C_&$uE<83S;Ao7B#L
z<W9b}4S}+bcLeLveQI#4H&qtSUcO`1zE71^ntQg?`tQA8>$q(CChizESDfHYlwbda
zW2-X_e})j**l>Z7kVW;Ui44+=i6R($`9wADNYa~lu>RUSh8;gtXcW_aHOT|`cCk6W
z7y%Rl8Dp7J_tKuq%E=MSjpw{;hGQ(n3u57jQJu*i#8mm;&YEjY)y(cb=(PI{OXZb>
zB_C_1&@TiA%Mf^Xu<LH-%Ny!ut1sPT*KFX+*sY@Im)Lgg%eR7$>GTSY7cK6^V2F>U
zB`!y_Yu8ay6HN*P2N5lKWSDfjWeY!<@{>Vtqf}BQ*|C<CY8`izW6nILTVYVkw{Di6
zFL9S;#JV2yvQMu&Y-B^Hw)Ooe%+vjKs#C@JmAQ<<MKQnVuFV=@_K+IFr?CRVMak4E
z?4NDiyRS?b{im_y*hK!%h(bEX8}>O;yc8l=1-xbMIDqwMzi)}P?br5C3QgO;a{Q4}
zN%+LGQ$=gfZL<|q26H^Va|TX~S2;K$9=G51Ph&!e8WkC{fGg|G<BKPV|Hi?!u9pVp
z2G(SO(gScNcqt`#VXQ>EhJQ4d2HQ@=en^aNW$~$@1lD;2sW_~OXwD~X3%mhoi?})Q
zGYu+TIfMi0c`QBCv?DGxA8T*l_x0_Hcf+{?nKK+MA7uS?c(~C7wF_>1rTV?(302fV
zi#+W#K7~6(4MK&voItV=vP6x_B}|@4Bj{USi6j~lG3Paj!fMPk1a27FcueQC&jeOm
zX#4Icu#T+KJC+~ybyj6QnpNf>^(TX#1VnR0^wDx(bL0Ejb~UDHhB1!Av$*?bSieeP
z;{_Sx@xje^PCKoaad2~D<Z519z${IKpmq)pwGcui2a|vv>1v!;fXs2;;93X&g8Wq>
zWCTv%v~rO)f;CA5;szIMfHJT-Mu;FoS}pvk<eXoSU!-z89@vQV+Mom8?|YQkRM_qf
zsd3*-(<tlU>htIH-I{l|H^FMaNJ3n_{W&?lh;;z8cFOYaRZ(%iExC#_9{ZCxgmBfY
zpb+~VnVE7_m}IxVhLJ>8j(nEMpll^^XPQJ>zS`d`G$pD98j4zh$RM9OJIk`kS2up{
zvD{aq&-~6}{23-dD;uW!JwtHX<I&N9vA$7J;qGe>u<vG2P!J7$Kw7?<y1EkRh(A{=
z1Sf%-mmJ^$`SY-w*V+qZ60Mt?8!~t}Fj<$x9+P~)KcihIM$GZBJRckle=K)c>!t0z
z3j?a>8MJnt!Goot<J3(TV7W^}g2xetOsW2N2H#R|NI?upLqn=W5{rg_P*@|h8i@Vb
zqK05cDunE+Az(Uj4o)aqb2uxAwNT_1)HlTZgxbCV=Lufg0S_v6^GtcNUS^<^?J33+
z_7~g*@aD_ldv~=<ip;M8@X3RH3p$HBfOuX|-W<_1Onr+%k!pzO@K38e?~_M$6*1QN
z&Z9a2vsX|^auZ5BwU}Fk2IXKCpRN<K`~{{=6ekmTg|6bf!T1`2a&kHu;|C*M9nRNh
z0B4ecn-!^TFTQ<5E|vLvCwtD+IUFM(&inSRPAXwj%(<(1h>t&GCuzmtxF&<>z|v7s
z&=DenhO{Z_F~QNd(0YGD(ee!Uw$YA_qTG+6Z3nv3Wxs3r4P+d%W|mPvWkSF$kYJe8
zTHLo<W1sWRdD)!7EzcUCZX#Ec@^O24res?`krxA&slH+2k;;M@OpohsUpp%kV?8=H
z<N;8GT*MpBU7>=?p?tI(*kK|nN6hymUUVShBO*d1sF36hclN)$unT<|0o8EzsArFt
zTn4F4%o~vY`96HO8CB$six1t|VQ51)klxXwLRoI{Hy>EcG5P-EhX%Y1hyL-)X*vFy
zbn)|&$TAib7B0Z5Ik7&Hec<Y%kiWM-CDJbaGqbe0#O1hEP63!dRJ$1d5cg~f&J3gb
z;%0y2=2l-e_)EKKrS3?Ut*R<*&bf+Ps8|3<_c?zjlhweW8B|<dj}7V^K5-8u7Wpd#
z?hp_kTi70gHv)|TawSgN*ddZq2H^H5JTM&=XDv$?$Cnn@qTe9{Cs1!v^+ju3L05V-
zZto_`LJr)+H;amjo~!*+B%tByR#=ZjwbZxTvWw^`er~>`GO6IuqNZnby-0a=_lqCN
z!OHOFy}5i{b?`QN^W(|m5jfa<Fo3~shx2L_6pG9j?z{|lj^o?DeJ$)V#h*W4t>6K=
zl=Dm_;(@F^KN($$FXkiG46#CBk^;dMyDXT1>%eEe5!hJv+tPL^ybV5#p(6VnDGww(
zn?fO93w8{#QIOdOu(phryONx11V)5ldwoY2HZvG$*tAfJn;STOa<FQ9a+J84;P786
z3XW2jm&q*BW)2$HnJO#z6rTHLB^};gEanVU;S~m&4s>J}hq|%jAGc7g7_ihWdU+Oy
z1dI@yh#u;lb&gXr7@Z(ejKD_`+!_`Z=7R;pCn@>(Y2rPB$d@}Y_yg}{0AB#eanb4G
zH|Ea)vf85;ciXS)T(8prxBH#AIFGKKm>wwWGRAdVwWah?)4T3&&<fSBess3IN7st&
z0D(cgb8}M2{#(F#<R2a*Gavr@kR#~eR#MO^m(*?_(^V48wwhIYy(utWqon%m*)q2c
zEYd;}mwC*8ykntExQ*z14H9DA!k_Vw47or1-tXn^6UxFb-fT6Z^nEXMjsyKbBIrPq
zvdUO@pR{Np_!oXZWCfRe8>)cWKx#(<kg<<dv3Fm?Gfw<56gc?SB5MW|&H$EG=WmF$
zK8}yC&DIG-O4fYKIA(I=QUzc~MS-Q(50}1jwhpmil9VLOeIIIlxHhm(bF6RVcEy5p
zKCl^RE3z6e*Aw6-fs}{=90M7MKLI)*WGmVas?VmFv;rUon}L)tGCdrrY*+{fk&crU
z#LGX>dAkFCyzd_>dasH>SO*JI-nl)vdsk3UMX80}C8wajWr)ED#mSR2b3h!`xPAOZ
z?}ocmQ<zb(*|R3rZ)vaCm$Sh@RY!wk@x{#c9#@wzc~*@I$J3r`PW$R+y;5{>v$*%L
zNKcG-4)8HTNCyfv;vf$R&bC7mFfm=U{f-0x3uXnpKbek4W_d&MK&S_Ta$txAnKMA(
zdIE9cHl(w&${mA4@fu`gCo?;yBhIG2{fLDGTdM-BRand<>mOzdsOM-&imi?gJ8|Me
zI`a{qh2%0&#p|1lYt2uc_QaOM$9{%y<|mq)YOG*lLIu@9>~&<M8O4iOY_JbSBqZw3
zT&nQJGbIUfxB}4E-c(I0$?fvt$c;RAm7X*kg#CcfZ}0DB$r-5n5;`p36qp_%m<(!l
zJkvY!qT~*PGJ9E}cc;<XyV2PD=$)hz$NNC@IJj|t)lAdJyxQx3xb-+GM_pq2xbIk%
zGez+#r+O`#3;#TN8EXfW!)JI4pW~ep%lW9uW{(*@^ep7Dk+{tDEwOO+LkC1#^0)KA
z>2sK=hRehRRmF~|qfmpAfhI&pLXJeV=*k#GLR4s2eI}Tf($SLn_8ZKv$H|W(`e<C_
zD@qK4;kn}lJWQljP}i7pA0VO|l#kf;gaku4p<Z@=2OL|0r29nIL-8W1mDm-yWmJL8
zqE;ta_Bge9(Tm~8qN5N!Et$6i$R%Up>dDW$p_71kkP{JT6fYu6gz}QO(#f0}gszcj
zUe?s*Xu-suY9Up~9Pb<r|M)YW&!66%MLjBrJVhb5loeQxK1z?m<ry?8!X8A0oa{aH
z(_Fz*>eReCHyve%MVla{8Y+RT29Ko$7i<JhKn+i1gl)SZSp@ls=}+L~$&_fgj=kTQ
zKN+jqvWj|fVcb)h?I5`max11dfONpf7iMYtclf|k002a?|Irwbi&Wlu<ie|Rm@?6K
z5<&N1%c;zcD;QaZ%RTF*l+HzKXLufn{kgM0Fwy+|<#_9;H^eB1#;k`+>@6UDfH=Bk
z!qOh7P>CHL_4s7>fu-3KfVs$}K;+zL^aai<{VNc}my2QQGrkq=$P7vgD2R}&R{pWY
zc}*rv_N0<Twz4>`O~i(v!Bf$Aj@9hk@EhH_`TK@eB##mL7fw=2Ur%K2Q7EaQOoc9#
z_Ipp^YT$TcUEk9f0_&(1nF+6RG0v41)cbmp?E_o~FetRnr2S&1NodPmOF{2?5=)^2
zF$CSckxR3mYUno8P>_2=GI^kjK**d;%JBSLXNcd<j_t4jK_hMF`x*L9baoLQo0o-P
z8lS8Gm3%B686+qa6)5gURb<n4<{F0J(r3jS`z)uw`6dm-dx5IL7>k#<K$*DZwa%Lv
zY`S!5Tz#o3gUpNr#`p<!a05xG#-!oXoB{|AblIoNc4;9@T{ukJ%G&u|x0XSaLi*qS
zqa!#zC`D!*I;y!{<U6oMO9qUkbNtBiv$oVPUN*8xUlD%HQs+l)uGNAXFXNP)H<QG@
zxvFGmv{o$Z*Zab}iz%Zu7j~VF<jHO(x(Lhx!`2|Y;8UIhIAVPBi7A#0rh?z)BP?Fa
zBbiWlm`g=-G>*YY0q=Z_tbJlz!-<Fc87c78Wp{Q*58>9(V6#W(&JYsy$_!0;PQ8`p
zGvQU#ME(xmK~PADPekN!C>ML6?=kh|H%<S~Ho&N#pTJxe*7eL>;$Z8*aC~?=`ZBF8
z$6wlh@)q07qI!`g8l9Ts35-st1vwgmB8%i2A>F3*b9VvlgQ}{k$!;7w+oY<R%_IpE
zj_wU;4i~ZN;%b_JE|lb*WZvyd6d~{%>2i=;^uXcqMrV0$a-D%o!Kkq;eG{wbpddz+
zoJ6k;ZFFki=sXThGF^&^X>w`)#z-gh;r1B(^#tA<bS^67xw5nCoHxM#*%n);zaC2h
zM>uK(B7g_vcqcmg78WZRPb2HOxED3{;#j0+aGiDSm~TbG$-{52Jkb+Ft&58zw0+r#
znZdAp5?43=ao-WG&hlxI_j|asq{(&vjCxgDx+sxDmHRmlwRzMjdsLtp)f#Pi;b5<L
zZ*2{Aw;o4*$RE(j&q&NP`r)z`9vUx#i^3s8!p{j@f&n7{sOYeT(J*w{Z+8E~LM}I)
zfkdo`<*S1D4WcI{(JyGc)9!vJ*#-D1Bs_|Q?Emiu6Fcn<K*j{%Gft$8sqSbQbmgR@
z!+}Gfd2dXs!ioKNzJJLtGjrP-TIKBa6u(gvc~8dzO3>&aXs58v9>K^&wLA42%ZXdk
ztU2w^PC7a|i0^rDJp6Fqy?I=P+3RNS)#^^w=&H@@WX_i7rrcOAdPg*$6+ci=(F)o0
ziq@c6p4O*?jY@(gk04^sYHTWdzjSe)Y4r>W?|VGaMpNBiG88#3kg>hxR*CHPmp(>R
zJW{2KR_WJT<+cVS>YZ;?z>N=v#~69g%Ul1a=fXIOQu`l0(l8k^5!e{BVqkeUl6gRK
zfk0mnSsl#6;F+%w0=cyif14DMm@5Mb&W3o+_pUnjrz;8;XvYN+H-2;N=K3SsD+qZn
z04R_W-(i`Lo(lKPm&*DNi_VdzB@2a_-}OvqO^l5Z5>2kEk035rP{3f4sXq7jxJaOo
zcIwIvi;jb~1cmx|jtPyT0?NCw(OGClgS)%C86;PM9<%%T=^$KIY)H6313|(B$1+t-
zRTQgN`zQ^>>wIvt6W_>LMCTxPsfZcZzs5rSLp^#zf<UeNL_b?_5+M_CGe_&s3r<!5
zcQ?lGM-oAJvWJGoa_SfA$HV&l@86#dog)~8m{@r7<zs4%`{?jB<Xk>ySKv%=ldxnB
zxzAKKx?JkL!2%Y02F!AOD+(+mQc|~dnlwCmZ)Y>aueK2SiU0dw*L^YX$HvWm^=oTo
zHSFeOs(p4g`oNNLERWUG@7K#7n`G~;I=gK{n!x+>JI~yi?sfhfuJ2)3!g|)C;)>xu
zU{Wa!P&M6#C{tH<KL`f^G_U5|a8_NL`<6<kxDj_ZRDt>zH;SJt@0LG~bM#6Mpnmlb
zU64roL7-}La00?dWt9*gPk~J1rcP$X=hj?pZceBR;Q$^TU#aNz&RSx(Do$rv9W%s=
zk{+~V8PNdGLB`quO~n{d3Wb!1jZ^oN=Byy=Dm<rrZgGqG|Fi&*4k2j*$=UFFp&8UJ
zY54r<Qyxe&4x4`id*4(5<OQlz2eATzy2RAnaY~~csD6`w**5w85!6OZoA-;ICVv|x
z%aNyBZr2!9!@Cv#=CI`GE2AA+v%=@xBf{OS9wzg49S#U%mbST`!%}?Cj^1KcP4>>*
zYM#~C=HecEa|kG>Qw=W`uPUrRmpFGd=72NTl%>T7-9IU+T>I@;GFhHc8;Rt&E$bjm
z&rueg+h1UH<i^rFuDwr&nA?I=wKtl~IOgWoD@jBWF^Xn`$f2Y+S|xWiV)Lcd*bd(C
zTf8JZNs>Zv854BLInH;a2=?2JLG@+98tGbjMY$EzY>LboeQC=FUt;t#T0TS{tX!Jw
zs=}Qvp|*eu;4=v-1)Cu_>GGq;mt<REzlFB<BT*bh@=f4D>YAQ7>Pbl^IIl7bpdF)?
zMAd#o-#Ha?{ejS?lTBI^&TH<G#W)>*V(o>8Ck~-OXffka(R*y3F3=$t6vpxSgFDNs
zGjr;iOD~j7c9{KIGedhA`Imo9dUuR(_Ib1GWOUSQ@4D8HKD$-<UZk4o3h$`#e*bIq
zMbXCBE#LhXeP~zshz9rC4_xu_Fifes2<aTjNrSFtV0=6T4m&br0furf^%T(}wrwmF
zU_7580-EL;9vcfrsD3#HCPECF?sI`Oiz-^ZqP7=cT-UxA?r1e-<Tz~iLd9dq`>>^G
ziFl&Bo09O0gz$HVf#lV{;2mH|qDu@Kn~=^-309tLM^qpSCublY4ROi=D1b4`7i2nA
z$k{K0SR*#=y|M#EFW)L^q8K7w0v6VDu#RgKnd7xY$svz2?=!BsuWwV=F7r3m|H?V4
z@(0IMoF1wlyP6Cj$?IhxYsZ$j=%KHiG902MQb&%mcCZUwm?-E*UpA7D!e%pJR0y$$
zye=@L0)m1x1jv2BIj#2S))Rek0F4Chf<V8b&xFtgD7eYvLOH@OC$~x2Z%qvz(N5^?
zA#PO*kN(#JQsZ=2nNClWTKkki`g=GBsdQj8tTKW(B3cwV$T0Dcp^Tt8Xd&HU0-2=g
zJH$(I`rge#))i4bmt*cAy3^-S0R|9I0<9A%Rq>LxGcXSK|GHnUS4MgKj>gxoU57?Q
z;Cpy+fnHwE=q{}M(#9BsMUt<os+PDLple~mjSp~~j)J-x5o8tW^X&h^#uf<O@TQ-4
z74=25SQ<g4V(dl__pWHTAcxW{QMDNd0A@;~Q6J@QAaSsi_^#3H2Iz*N#j@PDnY<kq
z%&OT*bo)UcA|70X28Dj@Y@Gr#4Unn#Z`S8CH<-$KQDpVl$P-YN06hs*sDadO7{G!l
zZ#8q>*d8R}3CEJtSl=en`-il*YmuDUj2HO1PIM@Jankr^W@ZXbzA}?^l8+TE4EVbE
zuQUtNPglE{)O9(rJ=w?VDSM6W+QZxdX_{M&`GOk@kdnGvUf{)dyKy8dF8YyP8h&2A
zVmm&aEWPW<DVN=|2Qk)-rMJIAf(CAzibBH0QKRu({mQr11QoT05IPnl@qu^SrnZ2i
zVa(@Pd-Z?6J6hAsXKMexyz!<GS$p@jd)q4^wF#4F*N{uc?_UHbX6mH0T81tljR|8Y
z$rB`Ize=Rdqet8f^Cg(n)x>y$qZ>(Zjq;ViFi3<E^ZNBd=Pj|QGj5CAg2V?<VQ72n
zt5<`!_Ng$gA?vFHpxB)&7TkN=Jyu<|S{?PGN;QB3^Y7SKIavgKi8+vWz*MrdXk*K0
z<VaAaF-!2@cU)^E<r7lgX^b3JX}^J|T@I%K32&iLa5=)ck?HguyrD+qo8?uU#vY(a
z$IEYdco)MdyQCfK)!uF4W!QypRYUs8o6ElX-)`(Z%YOXPIBw9p9nSe1nY~UX33$Kb
z666m+y2xj=F3-^y{i6o})06avM9afX`>80vDhCSaRTys@XbtdHePHGtxYo&Xfz9A9
zXbnny>>I|l9PgQpFNexAut=ZBTLbjC1JbVlgZO({SxFnAdL=!!l9CdBMY~}5Mh3F<
zs8wi+pWql}c~7@t&xspuPR}=QTB=Luq#H>18HpdSvT8!6_Xt24^m?6VF*O869bgpq
zk{6*%cn#1&%@SXE$8NyJJ2q@Trvl9&dOp(q03k$IqL|x-=rJZ&SJ%$d$IxuErV3da
zggP#Itz?Is%`4k`w?hC&MxNOld}ZPWy@!{FK>4SsX^61FdUO5p<Hv%NHrCca@CCt+
zsA%=)0TTT(H1rAjn(HV)ApxOKNKPgWyW1qy7kxb8s38dQ@q`SD6bVusj2I{LYV$RJ
zOF+J=fKMq3ca9`KxA6Zuj$eEJRboo8;xVnK@jgih>E(3iaBn3Fzd);s$WXvp;Ab${
z8N=PMhcLQ=^Fx$5FKOp5n3^U4Zp2T3*#02YwgeCY+FcEcIuH;a=!3EECtfSBB>|S`
z6Uy7)1aSEN@jfE5lMU3ml4M~nU<Cg(swQUj2}4&}kFb6(dUh-rmWZS`<~x9TfB-ij
zM-^ULbRbH2$!roRs6L@E0m9N<?kb7%6|<jxsL<RZjPejV7Qma@2qfS|HHx$>)-78s
zx&EkfhwtJ9g`*$&6G_Jz8JtS=9+CZCrztUF>MyFNS@|;zVY?tQIz8PiSUB+n>VuBl
zmf?kpy>;I-*BaUt`h}J&HOf2sV-I-lr^F*luQMNff9M54%PeYxS`5^zo#jyp3U9Q2
z9#pG!MUDnvB}c;11RA<8Q9D*Ec&YiU3JU}==s%W*a6~k^m}~pLj}|BkrbJVO?aI`<
zi})(qzr0Y9`@(-fSFyeOezUN3)ePg*xkake2Q4%k5>$mxy*^*Hcb!>{_mHqO|53EW
z{QRYvfDRK-xob9%nZ>Z<*6(?bht{6lsIsYCit5E%)AODnE!Xd*W{6TB(Q99Yqo3;z
z9zb4of3j+F<4WpCb|HRxZj7IVbJm61@%Io>p$`CGF643ha`#ZY=FbYed3;+t|D`rZ
zmv_c--_^M9jbcPL^yxb8zj0t0rK#U}GYQvF9|+r**v$nTdw`5Ag=i^LYAW}P9&+m)
zI9D&fLsdVL?lRYlDGBGLIAAssb`GB}@L0H-g;9a{B%nPEgZ7Oh%$np?kQjWh4AN8O
zTE9UY-~4;kWOO#yUTAtvIT!W7FdIlX%K@e4<90hlK3e1>vc`YJ7lv9ZpB?{P>+9>q
zkhX**5hPizp!g@Qa&eiKf2Fp(vH;cNq}lVLYj9w?-SBl=w)o#z<&@x4<3t`o^1Hwt
z`H`$P$TkilR+pqoVHA;iOx_3yyvF1sqL-EK;k_Lg_zaS_mFT%J;VQpAF4FQQXbG7f
z3E7yZD{khFc|C|_zpNhAo}YI8Hz_Hu`@gUR!$~}IL<<e?+a};LP;602MK4wK1LG#M
z3`pRn;L(3Zx58uL@)Q;g_lOurRym<fFT~U%<kc$>*5QpxjKt*;-U?2@3CK}RAir1z
zuMfg>$4iQinLI}OzZXE-;p4|8T&@8uBN`<#8dGCeMh+@!b8c^=9i#TVlN$ibfmp<`
zmIk)&Bl<<sqY*1DATErlJ~cvoMA(uI%iR7AEcw9)sT7K$#9h|Zbi)r<J~I-)lZdQw
zpyy#6GkN&Hl<E;Ce+wCLP%x@<KR#}Q=lej$M#<jZzT`8w7Fjr2>5=Fb5c+1peEjW-
zDfL4$@+*~@i^T-BUmf$IKPddTeJ3&xR6r8IyCv&5v<7Mj5<E{b`qz8h`)}h0l4MkE
zexYeDly$}3SF7K(D38ca)<YiS=eacBk0jVyoZh^Vy4N^=PEY6lF~Fo(6d5GQ8C!Zf
zOB*gD*qr#0Y^I*_>(XG3?T=EV3qhK?hDdTiUrlovhUL<~k5ab{{J<x`hifkB<O=(@
zMzuM6x!Jrabh?G)WXouV+Xi+eGVz?amI0j+8Ih&)v&FJzW2)fTM!%lpe8=1=W|9Y_
zmdDmia*~@m;&783Zas&fP+7GFl7<870<AX$(OJz%tK9+gnPepaV4+YTf*}I?SG)V4
zw7ia%a}*|;5)$Eo8=$+;YQt?+0oZAY$A#S96tA>25rSGlt=`hoGHAJP;m=#5)%s6Z
zs#9{=H16uiR=iVwq?Y&nJg_*kf*41@Fuktw^duM+4qM7Z^@_D9*_Hu#DC~VmupiKC
za`9q!XSM!(xb&9}N_fPRknQGAVrI}kU|5u|vw^nZx(ruqCQa=@j#r19-_{sWMQ>GM
z`ydvI`&Oe6ny+g(2TAO8eaPmu=mr7#9_Ajsv@=Tliv6$sFH|hpj}_=qVvhq+p;H^d
z^$WIoC3$W|jpB_epk~<;87hz0Xf~wlV(ZdEF^4$VLiNwkb^rf!7UV9$e;^&g7eaGy
zyJ7J2=PFtHg$$P}XN7q@X3nF_F+pZI4w7f6!*)PHZ<P%klg!}6idQf)GAgvx1)zC`
z)6>&biUWgAQTh&`BO+0Hkhc=&17Q7?P%aTrpG*j6@ToazySWTbLV{wU6C|8Ga#_hG
ze&WCa8{Yjh@HQU@2eq@iZulqPy(nd(YfG;#?xnL;`4HXuX0x#~YATFx0ms44Lu9w>
zVH448m_2hL(SP6XhT57I)zG=<#U+F&kxH1H;~BTesn;uL+8@2IyH9S7gGRJzz(^9z
z4oHh5fTH&Pi@*Y2w6xTdaec_|I~@|jA<aM*1MFlwFxb_>jU8SzGMx?<1EL{>3wkFY
zf06-Bux8}k5G5-ZI|s&E)Fw$#s6*ITGTzgX^8e`s>9>cQb(m211kOj@-8onWhl5!K
z-v)Gl=)B7NT&J%dOS1)4?N42F>Cu_R9<D!-VTd@is`Fb85LOQSMY;(Ae;|rNNY@?A
z21H(wVz~M2dLXEDJXaJ7=7fZVa0Y}v@la8@kxWJh2xULSzXxrr+K8?yGtCQ{P2L6q
zDZ!e#L?=WdjkE?G!AE+b{)b4hpwGTgKZp#d202Vpq+d#Z!|0)NWApi-wJ1bQwxYPO
z;UmtpV5H=I0*e6Vkm8kFImdw9ZUAu;?p--HM1_aC0_r%??>~qWgD`JMYfH2HP;qWu
zhP$q9TwVI@rB6TO6Y?~=k8ls?6zp=?Av)x|DpBfHhNk=0b7TFpIY(wXZQZ%~c^=Z|
zvsZa+_jD<rZ`>$bYH9bW>f+A1*jyv2`xJ_h{E8kXcjbay!_@8BF5{Wx^u|kT30(>B
zFBQxuRZ><KO7DlVncTlMHEm*paHf%r{>a|xObK|FjDW#H2^gB}fv@=|KAZYwfqY_2
z8Dtv`eEK0op!T<h$OZNEr1ov78?ho#4rjH^{Kr|qnHBLtWP5MeI6%a~;D+bV!{^;`
z_#@C*!@$6L-&F1D?M7HhdXLQ*i3N^`^q`)qMmc-l#Kf_`(5e?pgNlN!kU4Q5lL}zA
z_E?zI{W^J-X#GsIU5}+r;odSRb!G5Pah+Kh(Q#th$m^V3=HTp7f4gqHdi(1$p>wSV
zcGADy;7)&Osnn`+siS!DN8b55J(}6wnS~s?#j3Zi9$2$lcrUB>?!`R~b9z^XH|bT`
zaYS2L<u1N|;(FD%!p#unBY7px-DFh47$+ZOMI%l^&oFTgUu|f|hJzZv<}_cOE9aD<
z+g4Cm=!<Adq>cFeGZFPWU$e0Ic6RgjdB;fDB!YXp!~3@p;BAIy6U6a=Fi55eNpT{q
z4@EC>Lq&g2k>s)yVleu4#E&C-+NLFAfA(#^#Sewv>-MZBE2Y9i%rs6DKYYc9b%}aX
z;sUqwks<(Br2>j<0|RN-!vd28Cl6}S1|_ZIiajfozJ1-nV%c}kUTM3J*GxAudYwt9
z&tYIG=*`j9VEgzz-tR$}a&|`*rv}^1^a`?%@Or;)oQIi3BZ84jmRUZQ3CMfdupZ)H
zyL);NcDgQNzXu}bDaE<(UgB;fIwCMbH;CsCKEZmT8i0A6BlE|=0Hxxb;YT^3Svo(;
zO`FU_HzRAN)&$isbYKV3JB6#uq>XZ(@#R=M=R`+|DX!#nEwv6<ceg!ytLb`VX8iYJ
zTt4o#Oi?*tXU%{>CRqRdqmLL&AZmwz8+ubzHrJ6FMPh5W?p95GOA-NJn@e5!p5HLG
z?InyO6bb?KF@4uKcHgRaRm=JL`4tr86rLWN!H(T_+u5#I`qbBRNABSmUgynA@}-ef
zvUR!}sMFohP;Zpn?<eIB%kK{sXBpsP3ah)ZV(WwRyLA69E(3pLl~}GSeh%%Lc9d#T
z4SGUI8c!YtA%KY6eLQa8cLI|lL>&&>Ix^CL==1mP<@*0+CZ;QNbudPiSQ0^H5^ND*
zl*IH*fA&6TZb(`!iO3@QBUll@0N0d;3>X6GUQI%sNhN?(cmOF>xLTZiK?M{(0|77r
zvWzH{h>8HMP(F%91Pqhp7(C?Z$*U*dn<xFa)ts0KUG%4rKkoi*m+#55b`<c3Z4xqD
z-N)u)7)nOXl2;nvN8;C_Y$ymeZcSl)AyFye9DX$y;tS|l55N;)M?^#5kDh~Eoq$l9
zU{%B%f@+S6g2RK95d=;VgTxsZG2tif#Ks1ZN*l|qb6C9BCAXmBansaL+1YnuwFm*J
ztljV4k$zU^cWxml&p;xwCTckHyNC>zk1wIN8~wmucw@-G0kp$GsJ3T=TgxD`{yS>8
zM1*++-X>{{q*8@Y(eCSOKJbH8lhG%CXub&lI<R2{6?NOMFe9n{rMPgdrS%V*yCX7{
zjdPgfXj{kt9r8-7_mNx`;7>7V-XBc8aEzuD;DWjju76vba4?}nLPT`$;Fj_{&;7qN
z$!$??XrP~?3fPwIGVc&$iosq4{qNs-QQNAau1;It@e=F+L{>oJyFt_|W_C~)7a9)y
zAP^BrgvNS8c#O-na}bmQBP;*ivIhmfD)<eY@Y~+IOn8vHCFgS7mE_qlyP|1d11(^#
zmTLKpw_{7>*`mbl?gqYoF;hZ9M9?W${kggiA{ZcU!k{!lL)D8C`4Z;>Q9D@vvSoXe
zsNlKose>EZU3Ko%nx9isenmyo`x-2yIEXopiM7MpCPE#Z*bC7hXD6o{Nl7LV^$gx#
z@$vEVO)70~<tJKmc5a(R@p$(g^~DnZmSCBa{xSLg)J2n#@2vgmjBD`uh0-rUthb;b
z&C+2AP*Wo;a#<lWz)Gbtu!e;{@ln?E3-v8~xfzzMnJ(KuxgPpT@!#7YhFF8|7C-7O
zA+-f$;bGXLZfK8+ljoH!<#$gbEgsvSAL>_H<%?qV9p!<{yu7cal}6yd7vzh5=Y7^y
zo-JC;&iKEG8fX@G%B<o{!!|jObbSJa*<sAd=udPO?C@r@3JX~3Neup_THQS1u<AxE
z)T3Pdmhq5`wwUtxy)(Wi&#=HK1U%AgHJ2&%fAKTu($k=Gqhuou0%*Dd(x%=NTIrIG
z7p91^_I!U1F4X^DzMmN_D6-4$;c-NR@XlrM5LZ<tXR0I<aE3+NizRD9<lZG_hKXyD
z^n~S?hCuy3C|wDJi5uwa`?B3$XGZss)L}^YLuEH<JWJdip^)uBOG5=gF~R9pu#bG;
z`L;6BeyE=6isj#rX1NjatnasMkvnkU(0?ay(B^S3DFw)giH0Go>HvDMC;A`!1v4Z#
zOSnsEm{f8HhIv=TY*kX3&W-A!OBh>QOTC&Yd6C*DT?;Vc|6VT7H$mu$tK$ut3N#x^
zxm^Mn%ZY*@6j-Q_(1!Jzhj<fJ;|qtM9o~La(YBG1tK#XpH<cA&(OonwpAPud&}?4s
z$wd_5cqKl%>7v5H3!t2^gAE~@Z)&EifL<YZi+}W9jXNLt`ag7EKEGa~e=p<qW0Y&L
z-YLNcBSfUn+_|6n>{!w!ZYi%SM!XUHtPC7X2Gzjp0V{7;TfleCCwu=u650$ZfPChl
z9`V-jE354C>=mI&4o+UUrEJmrF=y&}P~Y(py+0dhAFnLKLhK2o?ptqQ|B#%BIVe(q
z=MvaJvmY2ak^->|k(ta0!D}--9k+30GYN;$|LZ<Lw&PAj5L>uBG{PETUA58-QNFtG
zxUz0GMf=Ek%Fu0gEVv$fn?NWS74>O&_$G>1!T<pCefbicZfLR${-FCa+Vd+U0i+T6
z51=drbmBjdU3>FVC{Sr_vB&h)OXGt<CmU_v^dt#bueT)YK%o6iRY*I_R~g#Cm2SVB
zUdTkB-@@><X^8jV!UD)j;7+7W=RGd_ZB3J+Nw4wcl|<s6t?k2~^gUQFc(2ho7Yk)3
zP~7U7i;g>9n*05D_)5_Yur|3L@I2^EeDdzjV#H8#u!2^u&|nOOG9&M+u=Vj>S&WD7
zSedl)^#k#|1;^7G(*O51c}`p+$NT^{q{N(}c=2|LK$x#_>Xc>MPak^Le3q4ahc2v6
zs;m7!?~{e{fsZdy__O7QgW=Imb_Mn~Gh<WHn5WT$(!Kex<yx@SY@_BOivnwv>;Ih?
zc$^1*O3<@NZ!&!uNL!++9&^J#IvUdVHA9+QN8dBV=IfEu;oqzM-#b!@lUNOK?oOck
zjGV7JgwoOXjf$R{KbrFWpyln8ekLXOZCc|O{sK0GgJdy~*H>FWCyBxd?a4Qjf;fRU
zCEVExrc346ZQfP>Q|{F2Q4C$HB|aK=+NTRhtFs4gO8Vf*kuD=6tDy($5f&)~LV)v=
z{|yjr|1$Bp1FHYxUbiIO3CA5#L4=GEZaSWzx@UZBMtpD!S5=(9nWdp@ckyBF%r|*@
z%_2L-lxCFq?wl5zc>FP>1bH@!ICy?GOSTS>g!X~!bR>@h3hD$<@^<$2*Hx3QtEYGy
zoSUrbI%XzG+dRYI5HJ;2s3(S+=(>LPz2yC5acF2ywBX@w-F4UV2g_Y2>@)j{fu(n4
zzrTvSoL~6d)hrkmYb|&uZyoQAJIvG<c)uS>Jgl9?AR?k;k)5q$zoD?$TR*b&Nhf=@
zmSXlTkL)HU>htuO1<I?=M=%RsFuXXs@UgXBO6<H^+|i*0d88}1_POrw`f_1c*zDeE
z_u;SA>9P7QU#)_^JNSs11-{Ae`w;S~W_^YOx2A<gNvODlugb4SVw7SZWDT05+t4p@
zNYO#y7S2epYsv@@(LR4Z0UBVfvuCrhD-ECp$?cMR=sL6s_0tjdz@4r6$c@snv&-Wv
zq@qx~=03$b;D1!x_gpPL9?qj(_*i|gqmJPZ{Z9GNGma%OqL;I@B1Yqn7K~(DEWLJ<
zytG4}oieZ&9zYA|_BcRH3d!}OK)$Nh;+TM{#Kzp6LZQU(H)mFYY~d=9KQ<PY^M@B;
zTdKgg3QNS@@=YQiZxT|t(?GZV+P8~}wqa~yV%`pudr{jYBs!s2%mVqUJQ9hCx_No8
z@AOcNXf6e6`|Jshc<5r}^F2Uc$Ka9Y((ffU*3M6%56kwe2L}H6_I-B;e=<v4*HO)1
zX*&vkjO31}r*<w5=Uuk_kgK~`sdAo=iqUXM>ekDq$&r}tY#bMw*VKm=FvRKNF;FPg
zavv`Lef>Wzz#HBhf{ovtn!Y({_ERJxUw$54;gxFfqJr*hsKN2bnByjkm8D-o-wuYp
z?U-oYW)}9vMmoK;DP4N0@ooBNZ3nfbs!IFT?GiqtE6C5L#H?mi5<wi>bf04g>_Lp_
z2Vl=U7Avof%G9sj6V^K?{8v;)rt{7mVMCqz-KW!0e&C-$!QzY&&fA2A>+j5={5|gT
zqg7)tt^HG$oaJWL%(!e#(_K*<$xIX}Au7d14$oda=kGl&&=L$~znLuhd;EvIbj$DP
zmxFhnX(wA<`fD%zw`u!_4)>zebgN;VAk8Ju1;;V@_mbZ%>z>%y9A0>5C}!HfZTU^x
z9miM$sWWsZerQ@~Cmg@WorK(pSK*Bp?nXA=H!WCxYLvc+cV8)#*Cz{3*hUD}-4No4
zp`o<2e0(xmX0lCNv+e`NLy?#H74ox=Y*7!%Nc)w1t>SNoaQFSeQ&jPlPd6JT?b{-=
z_4n()s_)8Lxl0TrBKBNv94z4EUb*~7tAD)VrFkv3Amx^KO7BF}F?OG~l<IeN2a{j4
z?m4;TWw3?rhpS1SrPNY_k4DJv(!Q`%|5ar94^nJH_TiWAiI{TW-r5{0=*)Cvhb$u{
z=0V3x-KUo$M;|OaG#~!yY{}iO_;FD~wZU?i`uC^|yRmy0?miT1`HDajVOyxEEc+`f
zn$1jd0$LgyHLhMQJwHiBiF;+c+pJ@K(g=029r<8;<Weuc_~sxYnifW#C4M`21!a<T
zQ}^+F7WXH^&xN!1NEv=RyQQ($SX%zxNA>Tc+8_Ib8nf8XJu{Uq+9}UU8DM?f`{RE*
z83wj*|Bsz~p{REfDF7De-X`Vyq0^G+=e-nll#%Q2nx?=EXZS1*s;Y9VqfiF!LNh*S
zC+h^kGoGH;gLuVN@$Y`_a(mgjO=0K90M$cOUccKOcPMag-=FM#YhBN;1&5R;WzAm?
z`-v7_((-joUF_wuwsL~vRp=?4@b6I`j&x=}z@(IH>`4REOu82?Btwo4Vnq;Bzg%5A
zmVT@bJ6-xchsJfD<8epzmbbe>yY*woQd-w@afzYBj);iZqzMGm5Py7kvW(YiWgljR
z$8UwoBjs-jy%6+~*hK5K`ohjL?Ntlr;w{lhD^f3KoH8%xr{4E?tBz%VvcvFYHQ|-6
zjA0#pEng2)yl5xz2rSV<5MxeIFg3-i7x2Qcv&O3wNK1vnG~P~A88EXycQHYPQ2OK~
z(5=bFEe|(_)0*oy3rlWSrxd5KSAJB##gp?zMquw#xfy{}4@;)G`y~wPDbd>^U)GIY
z#`>wINUeI`4WopWdj&d4_sI@j+7zt$NcE#vyi0FP?~?aRpe~==`g~M$_WquY|1H+|
zcL|DL9|2iR0I+5ORWP18RzLH(Auz@NR;APt{NQ$kFn$>y&&Ac$Azz6O?ks+RyxoA;
z&KinOv4O<@{$+fr?B>j+321b+pxjP?|0n|qpx@By=2!%yF|>lT6N3bH$qKzL4qiXr
z<-KpIEx~)=FwYu_ZQ7HJ!%V5~ziwnqtj#~fq?`U`=fAhG7trwT-)LJ$@!u|Zsxy!s
zC2m-0fH|3n{*4pdDkY@_QOKQ;kdPX+>-B}>Uv$=M$7#7b2fWds41~^Ciku0Ys!SaZ
z&X$sI64D*(PGD`&U$3)0{N>;dd8XxG;&o_ZmprAV6>@KBAcMX)ZypJaIiAYn`h~?A
zG6w*d>PL=9%JRLQQO{(EICD+GEbJh+vfD<8t|jCSjPuGFT>i0VHa5{)erv|TElC1r
zSqFn*TDHC;ID3W_LfJLPA?LztU=yBC6ijO?X=yEibC4US9jtaC5E{e*Xo0tm-A7M#
zBJY4_nPB=^yWO)_W;eyqQY0!)Rvn(SmOgpLQP*PO?!nTm-ebQef9;m<Jh;{6yJosc
zh2YlLyA&$#<<M?HEG1qOT=?Qn9z5v#=x9_TB%~P#x|D8!#{J;Fxvxq3#dZnTzj@UU
zI>}jb+`w+5O{MbkTgUn4P{S=pr^G-0GF<1fYB0H1zHQ|;m+w010;fXRjJ|C*Jd#y|
zIlPqD8xc<HG*C?i0d3QrIpm<c!Qv~#2pp{7Gtff#eZjT5HD2+nA0`cUK+EqAj)7tA
zT3s!zaD470@Qj&Ox;d3MFUn~apNN>+wQ}As(^TQf8+%H84e-WyJ!f-#Pmh1o`{`l&
zuy5qv230NDskHX|<9HkSjje6(2UPGjirX>dE)ChZX#mDM4<T3|O1N!>m-;{CStFK~
zd`d71B!=*v_0#C=Yj`YzWIaPl@~5b3;h64O=ZNo@%b;DA#w?p-QM>BCnkytV?Jv;|
zWn0%d$|XQmy)b*mr#<!6;p{zs3M&^peEF$+GJG%Hf0N%5b}&?p>Ym51$Imx$)8frA
z5;t!~TH0#}KrOj9Bb?q^=7D0mVCvg+IQm5;B~Kx|O~1;c0y6Jhm|~jD9&g3KQsJ=1
zJiIX}VU8QN8PHjQQ+kA`a0@54>Kzi1N5rqyY8kkyb1Ow%r>>TKvNoV2Kap$U<WvQh
zMA;krmIt{hf#uCCzjm$scLOm#x0cx^E`AJPV76RmiQ~nT<m8R_M;Z(ug3|?APYz&H
zGc!wI%1up8nmRi7;JS;5i?cOXrBFC<@l$MJj!Qq$dw1jqhs5JXYbP#>xm2T#>fz9G
zaUGv9gN@NX_^huCkrLMqGZ43|odUk@QWQ%M+P{}0Z}?kAW=<P|KEYrK+jQqCV#0}P
z12rZ7%|x&~rRb0&vsdcRr+TAOblk^pGR5`SoaD#PyHWt%4iX14T7Et>`+U~EZao8&
zZS;nK+4*@ZWUiDB$GNpb?*2Bo@#RaaFIsq9d(D`!(2GlIZ-tQxR9jB;cChA=l;2*>
z40np;YL{<SbBd{{seZn(FD6uo+<cY*Vk->~M!E=7en#Qp*TxL)>sv<~toJX}H9*63
z6;}~iK<}(&gaid&ybLVTqN0q2T$IE++<kBISy9hb>l7}Dn}UrapKa_CH*3qUP3x{y
z5_aAC+$~G&5hY0g$L9cRdQOgxnb{*ei<G7&br?W#e56A*<Xm5>>^NZ!rOY4>Q#fK5
z`j-|SBBUS-sgoH%>@l%mmpl)Jvf(x@AEgMkf#~+il@y~J>TV^w)^0y(`}4YfgFvs}
zrgdL0Hdk)Zw(b{b6f$&N`|lgM@p7;YI<!Klu90S|9txpG_`Q3N@V7W)a?U$_YdZ+>
zNCza?5m8a;<h)1JNz;oLC$C*Ar(MgF1+*^7gqmXae^vJ$em(F1|NqNAM)qnUC!?WA
z8Og|qLeh?snFa|(q;Zbo7#Wc$4Ha$aD3VfE9g>O$37M^vSw>Xf>m%p=yL|qIUzhjg
zeU@J1`FuR@<955<Z}(^ReLK2L3puQ2viXft)Ba(RMl)CVcKWdB+2G-64*jD3{4^l`
zLQt99;SI9RE0vYbcppq$HFHw8mV4cMe5l)K&xe;-Yg6ezWPEb8zM~LlKX@>3z<@O_
z@q4P@yx}y$hK{Pmbnq^V2zd(N<iy1;8qryjK1#+{QfZJDWzenXwma_<Pdr{bf%rT$
zcY}t?Xq~0<vpvSv?=7eZy}Y#azCoJfdh6X6GBf60p1$quqR~M=?@b?6RhoA}l~0h~
z2s5?r;$uci&fdCpYvQ{n<)=j$hc7hJ)%`=HCEi&^zPrSy=j`2mw`hkdtE!rMd0kUz
zlJC8jyU;7)y{mTICKu0`hgR=D?(MNH|6QVazR#o{U$nNJ=q1tYvcgXN_PldrN2@Nq
zYM5A_apHMgbgb#TF1P)IZ=UZFSbID(bDVO1pP<sodh@hPW2K|p+A_^e&;2wxTRp7O
zsOiAMK-)*@Guy(;XNFgs(a(GqNhj)e&F*lKG8$|DaoB{@<U{5rN|SA~?XET|OLAH$
z@~>gqAvJtP<%NW6j2CJ$ku|5jkF>_EqGnKAjZ`m3$M`c%*cX|<AEp1l5_i&CRf)?q
zb9LLM{W6i4!!oQ(2W?%eG5Ot!qzTiX)SoGLe)y?vdGeLyUE`!vEfTZhjd$l>i`BdK
zLrYnuLB{Ke=FKm6lM>_2ot@9mS8bxkmC;+b?kv-2uU)1bCv3RNT~5VbjCF1LfJw&O
zd2!ukW$TYlRO=vFJVa`CEDUBd?D?SK!#6ExfvB*@i4WQ4#Aq+{<f*9LZ|CO97-#dx
zmR;L*Wz|As7+)tI!#!C?Sy|+ihNgMNnk6UXO>}Q}|8DuSWYWn|I+7|^g9CSaCD<67
z#omcKT=TSA>Hs(1rzamv1RhVGWMvhVV~;nKfh<!1u65PC{4i>MpFu;0IN^FGikIGU
zQzEdF=17jkg_0nqH@`Cln}X!)$A|x3KeX!O-zOh?_mu<(&vLp@+>cjx=>W5NmTXgl
zL?a-YM~O)5|F)_Ng{<I$wUE7{v6#n5v;~KW$uqGVzW!W@voZQC|3lp+A4L6~?x*bG
z*L%pnM@0uRl;kdczHVn8WSr%NS?K#uC{lkoVIQ(5=~GE1vi(@^wpIgu{iQ@}jfoQj
zLqZZz6+}gKkj&cM*T*Avn|ZvO(fp}4hFzL{MMWZhE(e8&3nNt(44X(QW{gO#O}PL)
zU^gjbkgRM1+&87}ohMJO-XFU6rUL&<D0Q4QTjDb&Qunu<ah)x9Yh9o<ykgy65|YcU
za7>CJWuWvD#dOpw_s-tlZ4ZY%zqB!Au&QMQb~!4eM$Ka$(EoHVNtJ^LBP!&!n%~+#
zPkLZoog<-;BHT-*XsyN_yNlAtd(7MpoQwLy9#Ul4W#>?4vp7E3;EWO9E?9|pAbuhq
zyG%WMcB{3<DXC?iw?^>`W5<s#I@z34I3?6@<I@Z??x+d$bzpF?qpRZS>zG5wfjEpM
z<ub=wPEDPLF+Zy)4JL&w#^cmYCNrlnFeoVB`LYvc_IpZ_6BA9%%mTp}tcQm2K=L!|
z^C}Ic=scHGS^idPDO1+}zUgI@KJ0uMWV|a$RG-{ASn9@THEr!><<?vM33I8NU+HbE
zOkWBfT8@w_iEG??sVecF-9<$uWo62wH*ek~{_K)2&0A<X7zUZx9eZ3eRJ_;4<J?nR
znnYYwR#KYUaFXz8PB8F4{qO3N11q)628Ta<|GBiNckR??xAvYv#Fd7=W!?Jq3j>&t
z5eIM7=+QE#do>LjIy9E?LI)HR6)9`>Qfv?G+jpg_s^(^tHh){HHa8`E(>Q}~wBXq`
zUUWD&Rm>%ePmn!sWVIQ`+C~RFIe4V>rD&|gmB%CMn%DPy+xW=Hmx-^2c{VPc?9{ho
zSUpT*8CrqYTdurjIW^9L+)<9x6Lw5U5LGcTS}wh5=B{rSkfNN0AF@(Cp|_~3#J)EN
z!~X9r2l~XOe>#3zEhe)dSG7-P-b&NL``gXorc0Ne5#OCHzJM?5LMJn7bdhK-W$;_o
zE{f*68zv=?DG8E){cQ0+rGFq-Z1E(`JK&g9SDqZ+{G}m((Hsf8ZRx|;uB+Y@C|c>2
zQd@b8s{X_TlWp5BG2azOSdo_RmNV)Bh<FobXRB7N(wndQJ315o_RObGpHgw#O0vFU
zjDPdrFe43fx=CDyQ$1R4-kc&Lu?So8UD99Qjo<H<ICj3>F{{uUFa7DJ`7SELo|WWB
zaYfTt)p70m^<X0B?r%;gHP~9Tn^s7Ln1kZmf<ghOyaX5xb4Hbt*jgRSNN>s*Wl7b+
zdwSh+zCXUaDV$bB(E(B+K0e-<lAd^D24F^urOUWV351Q5VGrxmE#IHK4p_Kxnd{Ny
zAqyk<7?Pk0s60(vB-kaLM@dvyOO?J8>$oQT0UQv==;i4SD5gq$>T**mqyGHW5-!H~
zZ{JvY5`x^lD$)If%K(JDqA(gXY}kseewK!Yhxk%WNSH(oDKm9Tw5W4twthUk6MBY7
z8;$PfLO-p_Ll57SwEk=eh(7sY;-*|Fartj#eW`HL_WGHQ24n5T$F8>rS)j0vy`Rpn
zh$q^|SK_sWDZz18e@CBF8rrw-h`l#gxhQLDac?qEU}-@w^4NxNJNU5n_~U!03CdEl
zFQ*`4Y+?La25x)N=wsw2O(uDG9vr!1IA0WT76<k3KN7a0>HTwCVN))y_&4_x>wP`G
zU$f6v{`YK;Pu)${u3R4VN?f&q2aH}G$F4gl|15;B!_y4682Rl_-O4#qy^NPFi)R4R
zo>4C9hEHF<hy<hQQZ<XqHdMaP;^T)87fO52Z2cVG78@J8fPh0+5O`;;<7-!m&$&Ol
zo(VxVTMp){h3tk`U(D=xIVPdw^y4WT4Gi|vB1hS_ehSjw^7^m4`T0%MsnVxq(#Jou
zOSsVzRiftGqD7VIvu8`W&2}XsT$G-_F+dM`xg}TSY*5&AsFm>H>k)!la8|e?{3o%d
zCC$G0Mvc9TNoi?mMV7~9UOO-}6r)TRkpY#I_U)M{aUVewa%*);jmTMUZfS7(OE1dh
z0H>(xbF0@DJi4^;lv|gtrOL5c%etfm<-gI;?!gr+QWZh`PQWOG1q+JU!_HGsrOAd+
zCSzmptbmAB($IKNv%qC5MvTojhqp%H<nM$TiuR6A(acUJQys^lrW8pI8mV|f2Z_IL
z*E6!>@<LkqAL^hv<-`?nZ)EnckqAw>Yk^idFh-8m&zD28kKA58LAz^kKX4lq6GoU}
z6%q4AoeNRVW9Y`3fFoX?L!J+n?<Xl*@aX7Z@lujLuD~Sq2h^__Sy`O|DmZ+CC=BU+
z`pcx?-S%VcE?R*q6eW8#sXS$f5S01>xBB3rE)x5>_b6BS$>}5>oMy%WN_dF;r6=kU
zw|cN=eWEkz`%_bs)(C~7Ov-T)%PKP560FsN7@>|9>3hxfP7HxKz-zA0^n?rGq#7~o
z|NFU)5=~7KUvRIz4;2oaN)SDTPp%6OZT;TiUPlzOV3>DShi*5V_Kt`7->rSDXXBMY
zCC{*MJ?PuI6w4AUb%NFH1L6}Bv^G7{5&2S0&5)xgKrI1?_ZL{3{@BECy7?E|lGI-4
z*FDxQ1w$06=xgJ{B?%WUd`Miw(HZ&s@3Kn$iZZ9l^%j{8LyuUiipeu(tPrQp-ahtv
zt<S>*k|)?v8e^Ibk5cBKs4{YO6}>Om@tZfdpX5ub&VTFKJ3{}T8CMpY9WpSh$QM8u
zO`k6gXtMGhK2U#64JOYh#!;#mELmbfDcQ@g%V`_T{m);0r_J|sKn=x$$73S!3MgS&
zrh5tti6hYoi?&6LKp@$(HT_?0cgGk-rO`UYRhqnmPqk&L{ZSzxme+AN;^4=~f5tvT
z3D)W8u?_`G5b_^CeL4|qu{bpKdU+l0FJ$OnbIa=mQqmf8Q8t(mo)I0NkB?Ut6^UA3
z@izI%Ga?<=2l6xye3>WPM_jcv-|}r+BU%0aIdKhgVE4~M(tXco)G33E4Y|!$+dFse
z7#JFw6Fez%HVVwLi-xScMDi)V&w|xoM(`3vCu#bXIKc&>(=OXRz6oyTzz5g~Gc=H#
z4)m7Nd8ErQU?>i)t;6m8gQ(18tQiAfX_!TII{BMf<G`awmtK56S`tyaib8fyT78AV
zy89ii#?5dXQd!V`gMH>RsZ)wSof}j8b$eu;jYwzF!@r^&UHA}{P6PpAahQ&kfJWFw
zAZ6G~b0ZgS+0|`HL{_S#QB;2-j#5wP&F>`h6mx~ZW5=S9WF}dr%%-{`S?)i0P*Yn*
zQnmgh#iUUu@8abFB@;AP44zmn+B1p&_40BHB7vY+6qOk0rt^P5G2xgi)^VGenTb2S
zbiyB_3u|2+sqF+x#$d4q|3&Z&84Z`7*!RA%#trQdwq2&&MZq62Jy@IU-ICBsiQ@ni
zM0gez{a@0)M~+hrlkb%j6>lAB7AceIjSZG!R%m)-(VQN=drwMiWkN8Pl6|Ogc0|Oi
zJ~|(K81?8C7ye#5S|R9(<%)rPZu_}|56ojQ8Sr~uK?+CDv-NxRPQ)SN4l$6sFkn?%
zLWjlTlBBJ?-jQ-oKY%%K`lRzzLDHvtjX!Z`tubl06v2T=S)vDKem()eXea7dnMoT?
z3jL+)BFP6Rh%*35$}lMxMu|s}p>LfGuhpwxI+SnNq?)y}Nh7f}!_BzuuY^Ib2RJSe
zJXVg$(``@Mk4&S;&!Em3Xpo!G2G2(yT8ETx(xzvLghmxaN;R>wj=8?P)tyhtb(xDd
zyb*8+Z&aM(Gx{t6S!SZ$K~X4)P|ee4u;}&3%@X?_P-WgN^|n$;-g)xdjl!G74V8Hu
zy8T%khe`$O$N<e%H#&=|oT<40@rH<5>X`J<1od@oYItHnU<D965j12|N{98Ys2&+6
z!$B)-^Su{0_m<L`>n-Xc3G=s57cWgtLZW^<FRxsC5}LE=jDTuye(B3&RIp*}EKv<5
znkU?|+nV%>dqYN<ICfA`<}NrWQRc9&RUxF{9ExHq+-KHwV8(gI*S_s9TU3b?Mh!Jj
z!q!T5xgayp?DpcLT<^UACNU^_%^<Zxi#Pk&C8eiZf+dPK6`G%{Pj^*B@Bt+_c&MwS
zVDQA>2+#nF0<2?}qwY~kRPj=u(GG&{WT`ecew2Piws$tX#6}n8i0}LT{mnSBsf0^_
zxH>9<Wm!AZQ2S%@I|V1nMcDD_<#qP4^WzB?Kf`j@YB<oQ1Y?Kr_r;5Y08<w94$3!M
z|5z_)LAO4`N<>FOqSiMi`J=o!OW(1Z=oWr^^-}IA=f)M0t^ob`NEJ-;SbbINEU_N?
zUvuRiA*zjvzt<gLmNOD_BbKro!1TH0ioc;P9%6da`EtVFe;YIDb+@q4V#SIu28Y=o
zp(>hP;$|rmmg-;5mn_{<GwoECy6We%RW4I{CHVLJ$_SkkZyFE{;w`=!VNf%rLnkmF
zbJ~3~$`;X4BTtl(kGr!~JCuG?q-upGyx-|8dc{dq-4MtSKRKORpc>6-ffYrmkJEXv
zLmrE4zy0~_r02_hV;Wx{TN3u=5>IdqmNzAk$-Ax><W1Z%_084q4pzPXttE6xTgVMj
zCEO5tWY{`K9lH);@uAM{nb-zV1)ctKiN*-~f<@#?APtAtYZouk*WU+3?416h>zul2
zp@wD_7P7{&Ierta*#DwSt2WPQuKArDALe%O&e{h+1M2BZw`55BA{l%Du;*AX;@^ic
zg<V41TvZ1jyLM8c%TUcfTyB}h#l~8qz02*K(?vyHJ?N>kg^|$_s*sYK+|N`-`t^;$
zM<|miqQ)G!z1l@%4b}Bn4Gm*@+R2Hnw2Ed{Ry&HTck=BWJ-Zf-C)kQB<<m`Orq{0B
z<%r$_f`WobM<u{0O#l>S4e;yb46N?Zm-WqgO!M6EJlhs8fViz8ET-$QPhuS#Za3KZ
zS*8^*M45OR5hb8WUV@3`Q7kg9ckdA4u%@5ef(Oz;o*iU5Wqb2}vtFqvD;s-xwOY+2
zJIQM9zO#09_DIcO4oFMGChv#XBPU~g3o?;2k7b5}m3)`=b%m&wxYRU!NAWR==Tli-
zxpIYw?K9fHtAe6Fu6zf%@Wm6V{@^cRNCKkPh_5$n3@`&!{Q4HAZ%twNL>a)KOGV8;
zIbEi1`}ZUR8A;4Xj8#`(LI=HNiX8B>%Rb8_bMZy8+gjdhp_zAUs4(NH{x4B!I;c+Q
zp+8gYhcWq0UEK?M>RXm*<PL78?638o{LF}tDc(rUCgUIH%b;@$NkoBor!nmlL+%rg
z7e-n;8Qsz(OhrX?YBDx5iXe<S&?W_VS<iL_$ELo0(Pk7Ot#N8m-Gj634;+=nsRFm^
zT#l!p@XO01{2e<)bbnFWS5cqip{%B6x6xTMgYj>=Um1Dvp*~C@n$yWX<e+G>MsnHN
z4z5}M(*j&hvbkz*VZGUp2Pp!9(h{1?a&(`K2)^x^P@)+M(vAD0SxHs3M0QiA!qum0
z2X2ofNMe3+v2a9{LZW%S!{JR7d(U@$_D6vzPT;U%&;GeXQL%F0(ggwJMN=cAm9D+q
zBY0Zcy9xnf_~f0M;hEulwc(O>fv&#H_?14w@^(53P~~ys%qS4GaQ2{%EC`?Z+$S~X
zd-vYG%|-on<VXZ{lL>{hX15OFrjg|ZOR~bd#jW7f#Q8vHYi3$YhV3JdD+&w;j*J+p
zrQJhk5#CEqH^e3JPfSfUWufkFzO_p(n>regK6NgqAQrV8$wOu2$k`CA4!5Fn{1B8>
zIs^^vTy-V)kbDBxxdJC*zR--=B33R!$3XzVV}*?RtAl&56{B<m;2vb<r*64bGxWo(
z`G@3|QlB0?B*MP?aUUPaJn%~D`}J}c0f6+40V4_~&}>eXJJhwXTS++=mR4j(1vN{@
zDWU&kNU-Ecz?nhtqs=7iR9Mxip@vbfI?RfCvE{b~0etpzoU5ocIi18qN)t0tTR??X
zo6o1e?7<%Tg#ylJD%HxVwvAz>x-)x+1;BuENJr{}4nN@O1xq`)zgNi=QWxtWBC(qJ
zXLU*}2NwWgB7MN*m+%n6^hQg|j95c5y)8{{@Ok?e4`55W>pjR;M{%D~#?{b=YW?>K
zgar)4a)#b?xlKgdu`~n#e|PWSKX0Yi_};tiDyvUN>9bb(#KT%s3LF1}Qhh9pjU$=q
z(Ze%~e|x;t(ah1&vEaRu<h1&M+mXzNg?;=&$Ot@se5``PA8+P&&{X+3Z?~cY8(?;_
zhYY2_+7wH~3VmLHJk#1`|9qqV3!Kz3q)&k2iABh7!kd<X%(7K55l>rxxv9n?<`FEN
za<P6eV`1M3*1oJ;>F5raqAVU3CeO?P|A6VxEJShH9?Eqtion}PR(|?UQ^d*UbY}qZ
zGXX-R<$FrhWTa*<2tZe$i3`B_Kd(p*>wNl)CC6u|FdgXL_3}1P6aNbRFO91>Bmyvf
z%AO5DV=3x$K0bZ@I<w&sZFS~^s19lPbS-CwB}frOiN|><o#6N>gM8lJf3qbHtcyAp
zROlqbdT|VQaG3M$F2<2QMX2nlH~BZH%adoA)gLN<Gii!kZ-FT=E0Z9y(&L70b}L1z
z<+6OeWABc7^<cP|DI~+fxc&R6n1(>dmx|QbnCCoz0HrI!81-odpe|YYoIvZD{L=_2
z(3{`I=d<4KzUjYFW|1Ay&!=ygB+I%Jew>viPqxVZc3Zs>4}L*p(Mi>f4=E6c3@pC;
z1zsnX)=qis*dpp%K^XY?`WDG`uDbNCzKeA<@SrF1i1knJi9jL=<|)(=i~{|!-uyqh
zsqkxp_GhLeDoT<g&F|8~Eav>1En8OMkoTzJi=JT@N%diA-<)}Ml~*p3FL2r2;3{f%
ziKNnS#}2Uw^(?<n{1J>LBGYw&!z7D?OC1AC9j7;JWb8DBi294W$XA^sS=?W0cGbw1
zRO!v_y{Nc`5}zLzo-|ZkShZW+icg=m=o?dBQ6cNYX+9CeFOQNL_$A41EEP4g3-o*^
zz$=PBN+J+3RE<RXQr6r{yq4s?Ga->G<|u`vqCPY_z*4XPIt{x@hUC#7iN~ADITdt^
z%@$6g=7Aph6_7=Y5^XQ9xtgm{^6NF<ji$D_T)E`{$Hm@tp!;8$nL2}hl?<X?hp|0)
zsIw&KAW@?E>-%K-7mlKUkVJ~$zKAX<#5v<m&0m68Xs>$HW{BwOa}^zFkuYf^9PMVP
zSQ&K~9mwiUYLbEzk2jdJj|S)-XfB3pOqnPi6GhN6wTdMn{lXngR*P@!B9;U*=@CR*
z$#j}JB!{eg_kvjt4sp~db#j-NCDlE>dYLLDXDBsr28DK^`wu1x2g=JQ6%`fH733Tf
zFDj}U2Hn)u)rT317Y;rz#?0u1A7J6?C|_7uSibS{({o~e?d{3itBYy&5?O8WXPcXg
zE)Tt^t&*%R>XonXNa9md|4Fc5wmKzj<T)QagfR1{tqrLgitc?>9-iJg%>8BeUm}+r
zD|F*4IGApSp1YOXG&2el2!bH%oUTikyc7Lf6KD`+@)DdhV>>v!I1hG<=V_!6JVZy&
z4G^LUDsPv;+pgL`jaOXryvpL^B<T4BhP&vUfzXanR}}Z_pg9M&NPNK$MkJkLHVZ^>
z&jd?uSMQ6d?s5)V33iaKysHZEX;rjKH<Da&z^H|yvBMR3(nisH^7IXblL>Awj)R@f
z<6avq!X>JZx^Ev}HNQNdE@(Lap%Z5FU?eq72XCQlim}O+melS3Ul$_c0`<9wVU7(6
z@5mguoh<@{4CV9NR4>-As-}B-=H2G`=Jo3Z0YpD7axcjG!>YBCsug4e0YXaA;0uHZ
zTtz@RrwZO`uVJ-DBHzEB;9)n_MKZ)UTTXmyiF+&uSYMpmy2`-=2K=s1atDAp9Y|vX
zOatx_BUa6`0x{D{N;2d+Z5)Tvm{pTllPGYHN;94UAkY+wz(au<2d*n~e4x*?SEVwQ
zH~#3<EtUNci-R@e9vg-{2WA&DLQh>EE*71yVqyAo@gZbD5<E09B;~lOsH!faFrM-1
z5xAm3^XA7BtE`!S66Edo=4QwMNpK*!r+n%qG*~=&3j(gSd=ZR%fYlSpeOam5>3+C2
zGyT#Wgm<q1U}WTDbmeAxS>vE~bi<_Oi+2|dc!%8wfFFZt5#SS0JWfE(%IfOLSCW|O
zy8n&AR1_Ma+qNm_9HVIfQlg1kRy|40v5QacMWC!@w(2$+u-RT+&t7S?Rx5W1GV9{v
z0DZiib#Q-k9HFX&Rzgp2PndV>4xw`;Xnr0w)7vDg{VP`%QW!3+uiwAF2XV=O2P^(_
z8<2Ym@TcqdH{kDYXPdlrHQ0WCeEtRvY2)W_ws5DcsaE%tZ2FBhSdeaz*JhTMu}?N-
z+fXV@qS|neW<0DOPawBL;sf%>GhQ8<CNuMRG}It%Rk>3l-v@RXayXDs#Lb7WUqJo$
zkT`feFz_r@y{5}*q;4+;V+eSr-|u=u<)aRX;)hgP!W56}>z|98yhO-y@uqWy{ffVs
zd{hxH!Rbs;ca6?)w>####Mb-!4(W#@el|p|7NEWHRRspL>@fytMqoBdpV;&FlZLJ_
z9VLZt04Lxk{=wV(?M=tDc)q2W)TN@4y0h-`lTysmwjS~zbu|H+vf^+#UkWtIoRVK?
zo`{4FSQS84{pRtwqrcC5>0p0qI+P}*^~$+uFyb;YDvDHA*{B5xGFp_9dU`90rgD#H
z^h=S~?C|g?*S^L+Uj{<QYHBtb&l5GE&*(zMxcTWp9W(tM`WiS6&W{AdaR7d=xVk;L
z9Z-Y$rQR~WnC=bu3PXsx(YiXF4o=Kmb?V<zzIgIe*ZUJhX@uFHtb&x!))#~1D{tU%
zsFF|`Pj#Umz|)L&ON-(67RU@!FkO;rBG5IRat#HQ*7F~qQxKH-3ETkxSfQ}sR+llp
zymaC=W)}ZmmbA0>Y*5`=wz+45((tvev1AV@W;#Z)0nt##`HpFICzd-;t@hk}#abwz
z2pgq9>7vdT(UEFy$J6)Ndh62h33VrcbQ3<8`pqAePa|0&SR>t7Btkj(d|lhkb~fET
z6>&`#=8WQc@W$e>Tz!|ZP{0q}h^WEc?fTvrN5{qWU#1q6@Ya%3Cv-TByw>B3V+yhj
zCX#Ir>?O`<BV0<k%vJJ63p()Pq|49Z+1ospCJ?WZ)1r%De8myURm*AJw9xz7(v%WT
zWyOa`nsfS0j}}!gTE&bvCzgv&eAJjR;h(QndbhRg<Y6ks5>?O(sDd71U3!+|R`SEE
zFnp#h4!z<`3vd<m*8w;}FuKC5wH{oBVJ$GDi})kx2+19rjvno3x!0a9!s4g_>x$0o
zW>`g&P-vGHK9W=8HFtMk=On;Q#M@yk9#zzDe)`@k2E3P+%oh23+PNpF)=Y(ELlZ$N
z<E^C-jIp7oe&3j|BC<-I&1Bea)Saorynk+9w{~p}6D8&Z$inZQ<o2EXQOfsdk~^EI
zmjr0C>_`gBQVQ(z77VlX2T5GQ*j7L^fS<pRsbsfZYi*2~S<7=5q}cLya<=HK8oyVE
zKOE5x`jkAuXW`xkicBB7gBCkooS;=R-Ud54UYkU9scC;iv2>Z8X#lHrY}cQ7Hps$S
z0j!W3cG50OeQ`1c0YV`e{5C?-N_Ftl9o3;tGnOz+K(S-~xe48}Q|o3~z><QGM3{IQ
zBkf(^KHT2hjjIKRD{?8KGeeK|GS}y=zlDd3lMqv2$egr>R2is<hN4t^*zn=%$Kt+>
zZsOAkVw#iaO4YnaO2vWBA;7ib>;jPd<?lcDf}8?-(z4y_>l+I#SobXJDg}{+(*}3<
zboe;+@$Hh=h$~aE%xwK$=v@aWUBm$Q<I4!Eayq5l8e18hoV+P?jd?;->Ke~oy9_-l
z-)y;q<PRG4?V}|gQmvJ>heDH?ywi6J33_nYqi0!p?Yjea)@r}nEWVUPu<@9XI}%j`
zT0M&YEwo7I%=oN_5IVe>nGEebt{H07N%&qEnkSfzTb6o65gFf^vUJh`wMSj#L|o98
z28d?@82L=uMs%y0HJ?S$m3+uXA(Dy#vnY~!@0}xr0ME0t*`4TG(Q^}TRP>YYppz@=
z+SRxBFU!Ll@+{A)I1CRC{}^cSfLY%+uG&ty;z$xc-R^%knKMB}C1Hg$gpy*daA0+7
z<tPalwtDj>@urg4aKerw5)wKk0dKs^s8<8K-2Zk2->lo3t_nJL1r3g$qZ^e9PvTA}
zltD5w@odsG4G`FU!ubpXab6IDHeG$*V#saXSr->0n3eVF+4*W$7xkI(Eka~VO7R3*
z_6!VA3N71)WPkHxtV>kALbDi5e&qr*Qc%U?oT~eNn)rA*69woE3@RillveAIRC2sJ
zh%_dE>q2f*eap}KHu)df9`A*;djGzC#U@Lpn_YY!USRhS@ijv!ODDb)j4t4-0H~k0
zwSJ|EEN78XGQql5?E7?k2G0Ey|M2Fo*=<S0ks1+Mopz(Psb!<l144c@xzdp-0V1>D
z)7_vEm8Vb7D*Ime)hoQP{e<|e2RD46Rc^h;RWvG^3ZvUu-p)lwyee#8P(2p24P2-+
za9W5*ObMbl_0?AZf<>Trb@Z21HigTgE?h8zXECLAsM8!>7){g>Vh7er%eLb+9~Lap
z`8YSJFzC;q$I;)*{l?~8^=b>Taoqhj%y{YM6pBUJG-tl62+H{WLgr6$+{njCA5P?c
z#qI{>sZ+n7eg(Nwm(^&%6HSgQRGE)SX%_QA{J4zf^oWLu+Iu8CK2JnD-aTc^f3Ua2
z`5ch85H7PB3i@c=#m7R1>4?=bv=#?PmPexfrK<GMjYbQG8jLAGP}-PtDCy`_QrGfE
zL+qkDNu&xkOsYS%KCBq$nS<@*rX*+HG?We?qNP0hPj}WQfy(L>=$<YUtt{?AVgRaI
z&VzO!tEp_cmVfm{q4XHN=wt-G74a*k3iPa{&LTU${L!BAMAEFr2MhO<b_<?3%+}{e
zr0!pKh1^%j%u>L;|L-kjwOwNcBFWtsQ?o+e@ulJN^EyVJjCFOdPt7w7pJ9^?EZf9y
zhHm{A0#^K;1NMnccT>Qi2+_$a)~IW7$zTc`nZnvr^6Nc4qW-%cMV$%qr@y$*n&pE&
zp!gq$qDny+;;bxMSP<+5f{Yd#ENHFC2Ge;dS{EY<0;l6UB?ss4{zbq<bXNkx#H_sW
zu1a!=lf!~4V+@jf5vEKh_S{FUqBst_ZQTA)(cYCg8%6;dPPW)t*sE(*t3Ag6{8|is
z_-9P|@;x!lI|HNO`=!KF*^x5h#zpL-T^HQS3XcXY0aqEu9P6r}7Wd$RaRZ18PhxeG
zC)FzCGqE?J{2Lf6I+^*~uC|<Bvyf^cRNQcqz)}#&tSzSxP@SS30-R90Gqte#8m@3<
zPPG>4L6qqPpM}(5{^l3s;|*{iaxMvOy8f!CeKVia2)oe>sBvYQ`dosg<ysWQI;{yW
zT=t`F+N3dK=0jnY(1yt>hbh)#BW1yIb5kp;#}RJrr!S2WBTpJUkz9r5OOW}f#$gSw
z5?0r`>S%Tno$tCJn{!N|ap(n~Dx3zACKS&2kx2C;+#j~yd$UDgQ$3`l1QaVo&HzbH
zr0$rcV^r1x1*NAKBJl+a{*ew+(Se2`OP;H$63c;<#*a4=Oe-u|!O57god#qEv;bJ?
z1f#_$$@16^Hdc!<(dRC0I&|#kFoCdQEWRD&2FOHanI-f)Yk)7%qtV#>)PC*_!*<6p
ztt|7h>x{4KD8B{$2?FAFQ);sH2hUj#KT%zXQdEo$;3q8twUD-vQ`2%qm?ZSXeQu%I
z2r6|HNN5~0y&`7`=0H)%Nko{V=u&L?+OCwLk4k7tXAx;Vor&evXAs7`f7rIjTU>sa
zAQF!oOfwj~MS!TQC{+0$9a#9+!k5;Elf_sIFAq07EA;(zKvjSYx5w*(O66XC?4l{B
z9@Qylt2kML_k4GyI_%+$2`}d(xhR+4UU>#ro~}t?fM4yY69s472>(0Tew?_;Ifp42
zoA|bnhid|aph^jR3tncz6zSk&PRUv3)WXeqGh4-C;#y*U`S7-H^QDIASW=)KYkZ|D
zI;bnw=Y!v0Wz};NV%r41eFm$4Nmns|feTH}Es8UF1=lDfa>OJzF~A3pB04o5sxe^S
z^vJPx8_@Y87czsoV93#!Vy9HlH)4ruW@hHw8@0b$wf~9dMfw(ED1=tQ+HUb;lz>%;
z&t@JLcY;NTf8eK(2pcZ=lD2PoGsUE!1G_SSIb<d~pB8FDPkc7TCw013?tHnP?`A=1
zf{=6Pv#)RZcz@f?QE?R$g{*ALyckux13QsM5EiBn%~@6uOo=P%4M8-Spm9w_=SkyF
zd$7sN&CM-z>!XHkw3hiVT*krx^qJmxMBuujawRt|rA-68QLW9+UO4%gdlZ?giPWvk
zVHYSePW+LA4qTC$US^39JmziTMeD5vTB2$bkSkP>AXMOT&@sH((WZ01>6;8wADjxI
z?Gns~Mo8PVy|ol^MnI5Q&W)}8*=GLXTgUr-q-GC&OQ<uW(-5dbRFtLzz^WkDWi#*7
zBWuB-7o40B_13Sw7>!s(r;0?*h<dCDN8sZpPg>R&7*`o%IL9iEe4ug79iP^cC+j#y
zWu$1vdfHT`O<N8#mcdh(%KlgSPW=ErMJ%XEnboP<!`?rmU;>)xc<j*%1Vm#Mt^^f;
zFUonXZS)&Lo%PJ?$3~I80-}M$99irj)J}rD7KEbRE(%xJJ7tn5D#Qd_Wiq(+LVd8p
zX!ixUVZke7B5Ac%Rax0UK$--6A^poAcezHqKQYN9K^L>iU804hS>|$9d)o3mp!*)p
z_(%A5XU*w?WER>KfMNpq`EyF*@`x|X5Kof0ep#HLrWxz$QNz5N9j0T`EpiYWjBM(+
zH;v)S(pMOhu%d?#4`JhhAm&lC+;{5Q%%)c~<ASr}^~z4I+BKu@^S5@{1aMRxajclj
z?U-8-Hhnp4qNw{ug|<EKG0ZmS`-?$idZnQA5VdMsa+XJ*{qaBBB(2Zp{V}ia?9Q6+
zB7C}T={}($Vd{W$J_k)#jDKG>%5SFT#6eQG7uH!y`ww{_C%wjO^r!wWJv-m(p0;4r
z+zGt~?e0IcpRY~duV1@etgc>V@yKkm$+KkXn5lijrx$K5^R(6u|GX=VD8T+_<5$(q
zCCASvCJqcs-DPcNHi*8X2ydi!BAvNal<Ssnt~GJ84SzJIIIXL<ho>;S0KB?IBNw()
z_18pwGz0R#vETH1Vd3FBx^;bpy)^iTOQK<qXH-M`lBWW>1qTO<?N<s@e@j&y1!Ai8
z^5MgW#SkJ2Q)4HA@!_(tzryS1$N6IUC}6Aa>@$5^WAl5g$Z*vkI2@ncU|Hf+Ckf*y
zmo}F!dHeF^NJaI!x;o)5Us7J)$=<4a_f$m(T7Pz|sAQ}^j@b_?kk!yJOWelM_%B+K
zIUC8r!E@(Ev%S+k(el8)eb*5nAL2l{EqD$^GdN(kWJI9B1g5woIYcY%=p{Ba8rdZY
zN8W6?@*zXG>T&u7{`Fx~@9H5-^ImQWUE7zv9A#x?GNY6l<CH>Ao!SQ?cmuy{#3_14
zMtz;1UxWYbf95Icd$3u;RHaD!5E1b@|I8oRgA*02z-XEj{HNIDcB!#&vMsj^H0Z)2
zD;^|=%x)QFRoOV;vWRnRogfH*>)CTaU=OjDk9;Exf0E92`_G=Bc5I-5e^zXXs;X)c
z)z1)}k#&o1+`Kt250%yUNt5*E%(*$hHmWr<K3`_W{p;6tU3YzbXH%K&ZR@n5{u>Qg
z#HCAvDQGUFruyO@)0${~rN2aT*vsBhQiJ4wU@Cx5+Uy6qwq(&8v=6v{c4ep?lk4^s
z0(1KW;}16Ltnn)90f`SGUaSZj$`VNYWW(x~4YM6LPd%IJNHPneEw>OW_Ns;ke{#p1
zIMW$lucaj^n<kfeR=ocHIU@fj@<$&`xWGmIzB2jLMIzSHN|t=x)TMgJQ7_BY$9qMx
z<CD^G$n>mpn-<qnf{GQ1MBH4INk{P=;_99G`|tkvUErjX%YRStZu44nN}~%v@xaon
zGueP~NN(z%P+lu3qQBAm2qRgs$BK0s{Z727*H-g%Y^Wd5Q_{n4c52;{aG9U{=3H)R
zU46aZH3*jr5Vc93V799Pwc(QdLPDN={#O6X-J~uxI3(mK9@nsuV+ZySPX`5P(alK1
zbIj1}oBQCMV`*@(jHD>o*D3U0dzDLWna!<xAXREs)#M4?pz8)lM`i6v8X7y#m0l97
z6BaI<<Ki;@?fuZ6==)BdI_1zPb_C${sjC(wmWtfuQ7d}%=r3TN=}l=LR^%tlvni<8
z51$`r+GF3`j{D}0Ox2n^dCpnmeHEz<3m*+*nW6urreTpu+3Qn7+Z>up{LX(yUfva3
zr#ag@cHE0BdZGSYT~j+d@r}gt5cK)i@y3)?4=Fy7(^*0K78s3Rfn8~1l2Su`{Q>Y8
z$t<i^NnJ;Sg1X7d$|5z%DJdC<*zFs_nh!`06-XwHw<>A-#5RGS#@!@r9vC`po2oGC
zW=T|?)~~DnWU!6nhhz?5T3XtX{5Kf?^=~R|?ye)b;bCoRdQ07mkA(hmy%DCL{90su
zd=E)gP0gcsWqs$i4Z_q-;zK+AVoz2DpxWF#R(u^XI6Gy>Cq=^0T$R-zRK5T;-Sb$X
z#gnmzC6eSn?ROo+#4~^#YA*Hj5yac6F=x*@+}c+ApbJsrHgr1M;{Sp_VU_v_T;zRT
zB>VL$<WZK0FtqA^Y@DgGH8xqSpyp+}N^q2y$G9a(-Bx&0`97ljt$F&D>@&Ao;zPv`
zDR%X0(Qt`%x{L0>7^AUb@mJQ54TooZcEU4D7?@*8&fa@saSFNjMfGXhx$=ic>DBsC
z49&5zk%xj)cPzhrE;ZxarQn_2cOc%Xx%0PhIrOo!@x*Fl)t!uG_V!~JEnYm&EyGPd
zXhzzXrhOX^sA+v8{Jv;t7^m2+6uFn5=vq&kw0t!-HI4o}<ySxn+cS4$Kl5ps2Cc0>
zHen$<fTC&+DMFaE@pvh)9gUE-XV|c#hIGyyBsAwkHqJ4&u-Hqzuxy~yh7G+qsv^Mr
zUM&3eE3-T-rf{NfunI@)`%Th6iVUQBhqMe}{TWl`7ytgN>Sz%(dBZqfx&OY-%uI;&
zc2x6z9uxLl8?2EDSU+pz(Us^{*<(d@>W|_=ftt+T#QPmh>A3x+CuW(9`O3E6^J(wl
zr;E4SZ`?SJ&bu@5lwX$EUuo+up;wFCftPN-lCrY>_s0Ypta5g4z18Y^rtl|y*$sLt
zA2@QcVU1R1$WxU!KS$QNTk9@RKi#}gEa;dWl5x(J*Hs-iF74WRx5W-kdnz33Vjh>+
z`Zyx?Mo6Mwi_hG?6Q4Sb8#S3tnfu`^qBx>Nl9<i4#!lIpnQ}hPZr@8&8XS_ldUHwN
zJ<mRj@{z%Nw)Tp#yk@Kh3mfw#?dCm4)d=Ha>pNJ|O)S1qru@O5!zOFh#*I=0;gcs%
zMl2libd#-)afdvE!0{6TQioS|u(q~Q2+Hq}`g7)6#4)>(&lL}rovXeZu(orTE-w(=
zu`=m4gHlUvZ4blay@Yy}`H1+Im0EdZ7qKCkb1@QkGC=POMAK1Y#9nG(>y`BE4E4-_
zb9Et!$W3vPkU7#BcG+ta6caOs6y);nqY0{(7bL=hm4P#2a+20F9u>0qoSi$hiT}QJ
z#%+F13O9SQ00G%?3V<P>OMz7hVxN(-vl=BQHl8#+ofT#%EB8QibBnovrA9_R^nz?6
zTcv(84+)G1`s{vU?G;H!JaKR$SgF>PD2ZJbCvn%MOgM7<xG#BFb@`=Uq}{HPlOZ7@
zVIZ&jGdWpZL{$PK(APg}HHuq|_s#v2caYO@@5|L_-yXgsytk8+Q|R+Pp4ArY3HAE+
zqTAZFYpbxvD;vA_-uHPWDJMrn4>4>>ef?@*TuhwO(~SF<v>$(s{!+G#K|Yf>Q{oTI
z+bkQKP{UUbV#=<~_;v5y+_WSGSzRmY63(6LDkUYgD?YIO;nTKWxpD+`^-&y*$!ag@
zy3*b12Xi!xOiXrXWN2$@X$f=QNZG`VDO$&>F7MxaL}v8Nzq$pUy>R*Re7YF1Y|82B
z@B3@!Eqy+q_|nYaotsh_Dr-xgJMudElnGS&a?`f;0#e+ak}?4W!agLQtKJ-6hDj%6
ztUv!G454S`&-sVNNFs?sZ1_DQ35d0GV4+<p0>-3H)7IYqHe<%CP47L8!f#eq%28(D
z20uxm^2R?4w>yR3fA0j9R9+!iqN9O<en@H*K7P_k_FLhMj%N?M5rj=Oyqv^dX^AWb
zrYQPyz}o(j`t7$_#>V|;v^1KC@I#3*7xa@ejbW}*^DYp))wtb?%*ZVPWFKwN_wJi5
zs<=AD%GfzOA403Q2FXL@M^9wD-&h1fUfRoq-1U3+dSMtgS68<Kl-Y>j`IsK3?`%>=
z<x2hc5hCGo$cSAhz6-9IFYp2|OeXM|<i2|3q0vjT_NVF8%uGs2>4%)^Cf>JV5u9XP
z$mifwr>KCJm_9QJfu~;s>LoT}!hZREA!D!wNJ<3Az~KdV@7`eO;v=3OtC?c8cXrl8
z!Y20JGP||d^&JH&n<lT{y49JH7wy9dyb~^9Z{NO+O-w9_O~YTL4`-F)A#3#?ZA&cm
z^z>c|iPh)nl4bp)xzKq8k~2nQi?G;cbCzY)`SeTIBvmli*WugddHmW5((mXymIK`1
z*E2HcIuC@)t6V8}505#ckEKcfs44y(fbu3EfZYcTVONP(Xgpl70t^j(+B0+-|KR@Z
z!<pP?o0E`0bVXY~hpn?{nZkyy|A*(9(=Hb94byR#sH>~5dak<>Giz$`u9O7+(OaFg
zw6tR5<Nr)ZP*Lo^FV|CaA(TjsfE)##t#<5~$|3)dE4%5hck>;$vi6&X;6Ok~^MiF>
zELW0{u|ItHQ1oa_iWc{y`mmDQ&KC#${a@hK%#k@$&{yj$q}X=<Hw`ERW})mpva-a@
zy1AsHe9RZCQ^CQj=|0&`UAmyJGs0sh+dBW;tgNi?C;#O<?E%kHqb>Xc0|yb>Di-(d
zI8MbK(8ATAcgI<10{RXd*xAG5$B-G@<>ch}&&mUPY<O}>itB`N^4gG;J(+_wbf~u2
zYRc)nbV{!CvSA;-HZ^UWjFHfZ6DM4@ZuNM%{x0sHuo?~L!|F>DhEOj(5pT47&h&*_
z`W9{U0tf5~X?BwU{bc%Zs;$p?{iD|Qh4e{x{}>r7zT(XrH$*puC6xjWe26Ti1U`wC
zWc~DC1=Np3EVBH0(pWUr!!wi})l}Nh&eMxZN)-32tE&2%D*MI6RBx62pIvNmel#bI
zwb!ETpjH(oiSI@kG@jO+H0dR~=driWW2N(Ww>r}~efGtgtp4*zuhr=4rnpo|fNpl9
ziS?F*Wu0q0R<XRi$Rdw(d4rvy1kf8RcCi6w0VN2nfuE&K-q|Hdo}Q~qUu~T3wXr<j
z+2f~DFt8&5{a)xai;{9EUGYr)gsPamVa}XRbO96Wi$;Wo^-ydt`$mr3IcPDCte~pc
zo836BWYSkg#<51BUsiuIe{Vdx2|QHDtEg8hrycoc78?nr=bD%d*tKidN;|tfw%#pf
zuNGF#D?4sJc3}FJH$CWgSFT%EYL}?i9-6PYeGm7$bdA1DoGPMkM|-Qs>8o$|z-(s~
zz>~c8u>U?sWE&KXrm#IG=gfmaIJ0+-BK?Eds|qO5iKSn2`pn;is0XE!mX40j2XJ!E
z`-7r6BkNL6j!@hUNjFy1U&2Wc7g5bd7wu3gS6@HBRRl$00%~O?jm7E&unIAdK(*Cg
z68!o<x4xr)yA#2Op;p~Xwp{=0+lkaCDQmhhQZihuO2;lfHaU4dk0;ZTB6CDWj!eh?
zn@Ip#(OH`}e_Z&z@&CMFzuG<BXKTdn+0(W1>h?Y~Pl;Mjb%7H7&_T&bRysQNz&DWg
zlqXp%THsrEk(ipB&sGk+QBhHG<mP{ds^aoQ_Rj6Buyl+IODVW+@I^hxPoB)C&M|<!
zE7z=9Q*-A3yl-(&{_y|#=YC7a+z3AOKg+)MzlQ!c!1w>J|Kgvv4zY1FKi(W?<tU(9
O{dtCS&(B`@=l=t(_yQ#W


From 3d537dc0db6f5d1f97f3f54023fca71e04beedf4 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Fri, 18 Jan 2019 17:41:14 +0900
Subject: [PATCH 03/41] =?UTF-8?q?=ED=95=99=EC=8A=B5=20=EC=8A=A4=ED=81=AC?=
 =?UTF-8?q?=EB=A6=BD=ED=8A=B8=EB=A5=BC=20=EC=B6=94=EA=B0=80=ED=95=98?=
 =?UTF-8?q?=EA=B8=B0=20=EC=9C=84=ED=95=B4=20rsc/lib=20=EC=95=84=EB=9E=98?=
 =?UTF-8?q?=EC=9D=98=20python=20=EB=AA=A8=EB=93=88=EB=93=A4=EC=9D=84=20src?=
 =?UTF-8?q?/main/python/khaiii=20=EC=95=84=EB=9E=98=EB=A1=9C=20=EC=A0=95?=
 =?UTF-8?q?=EB=A6=AC.=20type=20hint=20=EA=B4=80=EB=A0=A8=20=EC=BD=94?=
 =?UTF-8?q?=EB=93=9C=20=EB=A6=AC=ED=8C=A9=ED=86=A0=EB=A7=81=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 rsc/Makefile                                  |  10 +-
 rsc/bin/compile_errpatch.py                   |  41 +--
 rsc/bin/compile_model.py                      |  11 +-
 rsc/bin/compile_preanal.py                    |  41 +--
 rsc/bin/compile_restore.py                    |  26 +-
 src/main/python/khaiii/__init__.py            |   0
 src/main/python/khaiii/khaiii.py              |  12 +-
 src/main/python/khaiii/munjong/__init__.py    |   0
 .../python/khaiii/munjong}/sejong_corpus.py   | 272 +++++++++---------
 src/main/python/khaiii/resource/__init__.py   |   0
 .../python/khaiii/resource}/char_align.py     | 140 ++++-----
 .../main/python/khaiii/resource}/jaso.py      |  10 +-
 .../main/python/khaiii/resource}/morphs.py    |  26 +-
 .../main/python/khaiii/resource}/resource.py  |  24 +-
 .../main/python/khaiii/resource}/trie.py      | 149 +++++-----
 .../python/khaiii/resource}/vocabulary.py     |  14 +-
 16 files changed, 404 insertions(+), 372 deletions(-)
 create mode 100644 src/main/python/khaiii/__init__.py
 create mode 100644 src/main/python/khaiii/munjong/__init__.py
 rename {rsc/lib => src/main/python/khaiii/munjong}/sejong_corpus.py (76%)
 create mode 100644 src/main/python/khaiii/resource/__init__.py
 rename {rsc/lib => src/main/python/khaiii/resource}/char_align.py (90%)
 rename {rsc/lib => src/main/python/khaiii/resource}/jaso.py (92%)
 rename {rsc/lib => src/main/python/khaiii/resource}/morphs.py (87%)
 rename {rsc/lib => src/main/python/khaiii/resource}/resource.py (72%)
 rename {rsc/lib => src/main/python/khaiii/resource}/trie.py (90%)
 rename {rsc/lib => src/main/python/khaiii/resource}/vocabulary.py (89%)

diff --git a/rsc/Makefile b/rsc/Makefile
index a99039b..36d1a3a 100644
--- a/rsc/Makefile
+++ b/rsc/Makefile
@@ -1,6 +1,6 @@
 HOME_DIR = .
 BIN_DIR = $(HOME_DIR)/bin
-LIB_DIR = $(HOME_DIR)/lib
+SRC_PYTHON = $(HOME_DIR)/../src/main/python
 RSC_SRC = $(HOME_DIR)/src
 PREFIX = /usr/local
 RSC_DIR = $(PREFIX)/share/khaiii
@@ -35,22 +35,22 @@ all: $(MODEL) $(PREANAL) $(RESTORE) $(ERRPATCH)
 $(wordlist 2,100,$(MODEL)): $(firstword $(MODEL))
 $(firstword $(MODEL)): $(RSC_SRC)/$(MODEL_SIZE).config.json $(RSC_SRC)/$(MODEL_SIZE).model.pickle
 	mkdir -p $(RSC_DIR)
-	PYTHONPATH=$(LIB_DIR) python3 $(BIN_DIR)/compile_model.py --model-size $(MODEL_SIZE) --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
+	PYTHONPATH=$(SRC_PYTHON) python3 $(BIN_DIR)/compile_model.py --model-size $(MODEL_SIZE) --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
 
 $(wordlist 2,100,$(PREANAL)): $(firstword $(PREANAL))
 $(firstword $(PREANAL)): $(RSC_SRC)/preanal.auto $(RSC_SRC)/preanal.manual
 	mkdir -p $(RSC_DIR)
-	PYTHONPATH=$(LIB_DIR) python3 $(BIN_DIR)/compile_preanal.py --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
+	PYTHONPATH=$(SRC_PYTHON) python3 $(BIN_DIR)/compile_preanal.py --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
 
 $(wordlist 2,100,$(RESTORE)): $(firstword $(RESTORE))
 $(firstword $(RESTORE)): $(RSC_SRC)/restore.dic $(RSC_SRC)/vocab.out $(RSC_SRC)/vocab.out.more
 	mkdir -p $(RSC_DIR)
-	PYTHONPATH=$(LIB_DIR) python3 $(BIN_DIR)/compile_restore.py --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
+	PYTHONPATH=$(SRC_PYTHON) python3 $(BIN_DIR)/compile_restore.py --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
 
 $(wordlist 2,100,$(ERRPATCH)): $(firstword $(ERRPATCH))
 $(firstword $(ERRPATCH)): $(RSC_SRC)/$(MODEL_SIZE).errpatch.auto $(RSC_SRC)/$(MODEL_SIZE).errpatch.manual
 	mkdir -p $(RSC_DIR)
-	PYTHONPATH=$(LIB_DIR) python3 $(BIN_DIR)/compile_errpatch.py --model-size $(MODEL_SIZE) --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
+	PYTHONPATH=$(SRC_PYTHON) python3 $(BIN_DIR)/compile_errpatch.py --model-size $(MODEL_SIZE) --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
 
 clean:
 	rm -rf $(RSC_DIR)
diff --git a/rsc/bin/compile_errpatch.py b/rsc/bin/compile_errpatch.py
index 34e4486..caf7d8a 100755
--- a/rsc/bin/compile_errpatch.py
+++ b/rsc/bin/compile_errpatch.py
@@ -5,14 +5,14 @@
 """
 오분석 패치를 빌드하는 스크립트
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import argparse
+from argparse import ArgumentParser, Namespace
 from collections import defaultdict
 import glob
 import itertools
@@ -20,14 +20,16 @@
 import os
 import struct
 import sys
+from typing import Dict, List, Tuple
+
+from khaiii.munjong import sejong_corpus
+from khaiii.resource.char_align import Aligner, AlignError
+from khaiii.resource.morphs import Morph, ParseError
+from khaiii.resource.morphs import WORD_DELIM_STR, SENT_DELIM_STR, WORD_DELIM_NUM, SENT_DELIM_NUM
+from khaiii.resource.trie import Trie
 
-from char_align import Aligner, AlignError
 from compile_preanal import align_to_tag, print_errors
 from compile_restore import load_restore_dic, load_vocab_out
-from morphs import Morph, ParseError
-from morphs import WORD_DELIM_STR, SENT_DELIM_STR, WORD_DELIM_NUM, SENT_DELIM_NUM
-import sejong_corpus
-from trie import Trie
 
 
 #########
@@ -37,7 +39,7 @@ class Entry:
     """
     error patch entry
     """
-    def __init__(self, file_path, line_num, line):
+    def __init__(self, file_path: str, line_num: int, line: str):
         """
         Args:
             file_path:  파일 경로
@@ -65,7 +67,7 @@ def __str__(self):
             return '{}: "{}"'.format(file_num, line)
         return '{}\t{}\t{}'.format(self.raw, Morph.to_str(self.left), Morph.to_str(self.right))
 
-    def key_str(self):
+    def key_str(self) -> str:
         """
         패치의 중복 검사를 하기 위해 원문과 left를 이용하여 키를 생성
         Returns:
@@ -102,7 +104,7 @@ def _parse(self):
 #############
 # functions #
 #############
-def _split_list(lst, delim):
+def _split_list(lst: List[str], delim: str) -> List[List[str]]:
     """
     리스트를 delimiter로 split하는 함수
 
@@ -125,7 +127,8 @@ def _split_list(lst, delim):
     return sublists
 
 
-def align_patch(rsc_src, raw, morph_str):
+def align_patch(rsc_src: Tuple[Aligner, Dict, Dict[str, int]], raw: str, morph_str: str) \
+        -> List[int]:
     """
     패치의 원문과 분석 결과를 음절단위 매핑(정렬)을 수행한다.
     Args:
@@ -171,7 +174,7 @@ def align_patch(rsc_src, raw, morph_str):
     return tag_nums
 
 
-def mix_char_tag(chars, tags):
+def mix_char_tag(chars: str, tags: List[int]) -> List[int]:
     """
     음절과 출력 태그를 비트 연산으로 합쳐서 하나의 (32비트) 숫자로 표현한다.
     Args:
@@ -195,11 +198,11 @@ def mix_char_tag(chars, tags):
     return char_nums
 
 
-def _load_entries(args):
+def _load_entries(args: Namespace) -> List[Entry]:
     """
     패치 엔트리를 파일로부터 로드한다.
     Args:
-        args:  arguments
+        args:  program arguments
     Returns:
         엔트리 리스트
     """
@@ -221,7 +224,7 @@ def _load_entries(args):
     return good_entries
 
 
-def _check_dup(entries):
+def _check_dup(entries: List[Entry]):
     """
     중복된 엔트리가 없는 지 확인한다.
     Args:
@@ -239,7 +242,7 @@ def _check_dup(entries):
     print_errors(bad_entries)
 
 
-def _set_align(rsc_src, entries):    # pylint: disable=invalid-name
+def _set_align(rsc_src: Tuple[Aligner, dict, Dict[str, int]], entries: List[Entry]):
     """
     음절과 형태소 분석 결과를 정렬한다.
     Args:
@@ -265,7 +268,7 @@ def _set_align(rsc_src, entries):    # pylint: disable=invalid-name
     print_errors(bad_entries)
 
 
-def _save_trie(rsc_dir, entries):
+def _save_trie(rsc_dir: str, entries: List[Entry]):
     """
     트라이를 저장한다.
     Args:
@@ -309,7 +312,7 @@ def _save_trie(rsc_dir, entries):
                  (sum([len(r) for r in rights])+1) * struct.Struct('h').size)
 
 
-def run(args):
+def run(args: Namespace):
     """
     run function which is the start point of program
     Args:
@@ -338,7 +341,7 @@ def main():
     """
     main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='기분석 사전을 빌드하는 스크립트')
+    parser = ArgumentParser(description='기분석 사전을 빌드하는 스크립트')
     parser.add_argument('--model-size', help='model size <default: base>',
                         metavar='SIZE', default='base')
     parser.add_argument('--rsc-src', help='source directory (text) <default: ./src>',
diff --git a/rsc/bin/compile_model.py b/rsc/bin/compile_model.py
index fa7b122..1c10d1f 100755
--- a/rsc/bin/compile_model.py
+++ b/rsc/bin/compile_model.py
@@ -5,15 +5,14 @@
 """
 compile trained model for C/C++ decoder
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import argparse
-from argparse import Namespace
+from argparse import ArgumentParser, Namespace
 import json
 import logging
 import os
@@ -21,7 +20,7 @@
 import pickle
 from typing import Tuple
 
-from resource import Resource    # pylint: disable=wrong-import-order
+from khaiii.resource.resource import Resource
 
 
 #############
@@ -40,7 +39,7 @@ def load_cfg_rsc(rsc_src: str, model_size: str) -> Tuple[Namespace, Resource]:
     file_path = '{}/{}.config.json'.format(rsc_src, model_size)
     cfg_dic = json.load(open(file_path, 'r', encoding='UTF-8'))
     logging.info('config: %s', json.dumps(cfg_dic, indent=2))
-    cfg = argparse.Namespace()
+    cfg = Namespace()
     for key, val in cfg_dic.items():
         setattr(cfg, key, val)
     cwd = os.path.realpath(os.getcwd())
@@ -159,7 +158,7 @@ def main():
     """
     main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='part-of-speech tagger')
+    parser = ArgumentParser(description='part-of-speech tagger')
     parser.add_argument('--model-size', help='model size <default: base>',
                         metavar='SIZE', default='base')
     parser.add_argument('--rsc-src', help='source directory (model) <default: ./src>',
diff --git a/rsc/bin/compile_preanal.py b/rsc/bin/compile_preanal.py
index 556f5ea..6874d2c 100755
--- a/rsc/bin/compile_preanal.py
+++ b/rsc/bin/compile_preanal.py
@@ -5,36 +5,38 @@
 """
 기분석 사전을 빌드하는 스크립트
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import argparse
+from argparse import ArgumentParser, Namespace
 from collections import defaultdict
 import glob
 import logging
 import os
 import struct
 import sys
+from typing import Dict, List, Tuple
+
+from khaiii.munjong import sejong_corpus
+from khaiii.resource.char_align import Aligner, AlignError, MrpChr
+from khaiii.resource.morphs import Morph, ParseError
+from khaiii.resource.trie import Trie
 
 from compile_restore import load_restore_dic, load_vocab_out, append_new_entries
-from char_align import Aligner, AlignError, MrpChr
-from morphs import Morph, ParseError
-import sejong_corpus
-from trie import Trie
 
 
 #########
 # types #
 #########
-class Entry(object):
+class Entry:
     """
     pre-analyzed dictionary entry
     """
-    def __init__(self, file_path, line_num, line):
+    def __init__(self, file_path: str, line_num: int, line: str):
         """
         Args:
             file_path:  파일 경로
@@ -59,7 +61,7 @@ def __str__(self):
         line = '# {}'.format(self.line) if self.is_sharp else self.line
         if self.err_msg:
             return '{}{}: "{}"'.format(file_num, self.err_msg, line)
-        elif self.is_sharp:
+        if self.is_sharp:
             return '{}: "{}"'.format(file_num, line)
         return '{}{}\t{}'.format(self.word, '*' if self.is_pfx else '', Morph.to_str(self.morphs))
 
@@ -94,7 +96,7 @@ def _parse(self):
 #############
 # functions #
 #############
-def print_errors(entries):
+def print_errors(entries: List[Entry]):
     """
     에러가 발생한 엔트리를 출력하고 프로그램을 종료한다.
     Args:
@@ -107,7 +109,7 @@ def print_errors(entries):
         sys.exit(1)
 
 
-def _load_entries(args):
+def _load_entries(args: Namespace) -> List[Entry]:
     """
     사전 엔트리를 파일로부터 로드한다.
     Args:
@@ -133,7 +135,7 @@ def _load_entries(args):
     return good_entries
 
 
-def _check_dup(entries):
+def _check_dup(entries: List[Entry]):
     """
     중복된 엔트리가 없는 지 확인한다.
     Args:
@@ -151,7 +153,7 @@ def _check_dup(entries):
     print_errors(bad_entries)
 
 
-def _set_align(aligner, Word, entries):    # pylint: disable=invalid-name
+def _set_align(aligner: Aligner, Word: type, entries: List[Entry]):    # pylint: disable=invalid-name
     """
     음절과 형태소 분석 결과를 정렬한다.
     Args:
@@ -173,7 +175,9 @@ def _set_align(aligner, Word, entries):    # pylint: disable=invalid-name
     print_errors(bad_entries)
 
 
-def align_to_tag(raw_word, alignment, restore, vocab):
+def align_to_tag(raw_word: str, alignment: List[List[str]], restore: Tuple[dict, dict],
+                 vocab: Tuple[Dict[str, int], Dict[str, int]]) \
+        -> Tuple[List[str], List[int]]:
     """
     어절의 원문과 정렬 정보를 활용해 음절과 매핑된 태그를 생성한다.
     Args:
@@ -224,7 +228,8 @@ def align_to_tag(raw_word, alignment, restore, vocab):
     return tag_outs, tag_nums
 
 
-def _set_tag_out(restore_dic, restore_new, vocab_out, vocab_new, entries):
+def _set_tag_out(restore_dic: dict, restore_new: dict, vocab_out: Dict[str, int],
+                 vocab_new: Dict[str, int], entries: List[Entry]):
     """
     음절 정렬로부터 출력 태그를 결정하고 출력 태그의 번호를 매핑한다.
     Args:
@@ -240,7 +245,7 @@ def _set_tag_out(restore_dic, restore_new, vocab_out, vocab_new, entries):
                                                       (vocab_out, vocab_new))
 
 
-def _save_trie(rsc_dir, entries):
+def _save_trie(rsc_dir: str, entries: List[Entry]):
     """
     트라이를 저장한다.
     Args:
@@ -270,7 +275,7 @@ def _save_trie(rsc_dir, entries):
                  (sum([len(e.tag_nums) for e in entries])+1) * struct.Struct('H').size)
 
 
-def run(args):
+def run(args: Namespace):
     """
     run function which is the start point of program
     Args:
@@ -301,7 +306,7 @@ def main():
     """
     main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='기분석 사전을 빌드하는 스크립트')
+    parser = ArgumentParser(description='기분석 사전을 빌드하는 스크립트')
     parser.add_argument('--rsc-src', help='source directory (text) <default: ./src>',
                         metavar='DIR', default='./src')
     parser.add_argument('--rsc-dir', help='target directory (binary) <default: ./share/khaiii>',
diff --git a/rsc/bin/compile_restore.py b/rsc/bin/compile_restore.py
index f602096..132df3a 100755
--- a/rsc/bin/compile_restore.py
+++ b/rsc/bin/compile_restore.py
@@ -5,23 +5,22 @@
 """
 원형복원 사전을 빌드하는 스크립트
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-from __future__ import print_function
-
-import argparse
+from argparse import ArgumentParser, Namespace
 from collections import defaultdict
 import logging
 import os
 import struct
 import sys
+from typing import Dict, Tuple
 
-from morphs import TAG_SET
+from khaiii.resource.morphs import TAG_SET
 
 
 #############
@@ -33,7 +32,7 @@
 #############
 # functions #
 #############
-def load_restore_dic(file_path):
+def load_restore_dic(file_path: str) -> Dict[Tuple[str, str], Dict[int, str]]:
     """
     원형복원 사전을 로드한다.
     Args:
@@ -61,7 +60,7 @@ def load_restore_dic(file_path):
     return restore_dic
 
 
-def load_vocab_out(rsc_src):
+def load_vocab_out(rsc_src: str) -> Dict[str, int]:
     """
     출력 태그 vocabulary를 로드한다.
     Args:
@@ -80,7 +79,7 @@ def load_vocab_out(rsc_src):
     return {tag: idx for idx, tag in enumerate(vocab_out + vocab_out_more, start=1)}
 
 
-def append_new_entries(rsc_src, restore_new, vocab_new):
+def append_new_entries(rsc_src: str, restore_new: dict, vocab_new: Dict[str, int]):
     """
     기분석 사전 빌드 중에 새로 추가가 필요한 사전 엔트리를 해당 사전에 추가한다.
     Args:
@@ -103,7 +102,7 @@ def append_new_entries(rsc_src, restore_new, vocab_new):
                 print(tag, file=fout)
 
 
-def _make_bin(restore_dic, vocab_out, vocab_new):
+def _make_bin(restore_dic: dict, vocab_out: Dict[str, int], vocab_new: Dict[str, int]) -> dict:
     """
     두 텍스트 사전을 읽어들여 바이너리 형태의 key-value 사전을 만든다.
     Args:
@@ -140,7 +139,7 @@ def _make_bin(restore_dic, vocab_out, vocab_new):
     return bin_dic
 
 
-def _save_restore_dic(rsc_dir, bin_dic):
+def _save_restore_dic(rsc_dir: str, bin_dic: dict):
     """
     원형복원 바이너리 사전을 저장한다.
     Args:
@@ -158,14 +157,13 @@ def _save_restore_dic(rsc_dir, bin_dic):
     logging.info('restore.val: %d', 4 * sum([len(vals) for vals in bin_dic.values()]))
 
 
-def _save_restore_one(rsc_dir, vocab_out, vocab_new):
+def _save_restore_one(rsc_dir: str, vocab_out: Dict[str, int], vocab_new: Dict[str, int]):
     """
     출력 태그 번호 별 원형복원을 하지 않는 비복원 사전을 저장한다.
     Args:
         rsc_dir:  resource directory
         vocab_out:  출력 태그 사전
         vocab_new:  출력 태그 사전에 추가할 새로운 태그
-    :return:
     """
     idx_tags = sorted([(idx, tag) for tag, idx
                        in list(vocab_out.items()) + list(vocab_new.items())])
@@ -182,7 +180,7 @@ def _save_restore_one(rsc_dir, vocab_out, vocab_new):
     logging.info('restore.one: %d', 1 + len(idx_tags))
 
 
-def run(args):
+def run(args: Namespace):
     """
     run function which is the start point of program
     Args:
@@ -210,7 +208,7 @@ def main():
     """
     main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='기분석 사전을 빌드하는 스크립트')
+    parser = ArgumentParser(description='기분석 사전을 빌드하는 스크립트')
     parser.add_argument('--rsc-src', help='source directory (text) <default: ./src>',
                         metavar='DIR', default='./src')
     parser.add_argument('--rsc-dir', help='target directory (binary) <default: ./share/khaiii>',
diff --git a/src/main/python/khaiii/__init__.py b/src/main/python/khaiii/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/src/main/python/khaiii/khaiii.py b/src/main/python/khaiii/khaiii.py
index 5f174e2..4bdb193 100644
--- a/src/main/python/khaiii/khaiii.py
+++ b/src/main/python/khaiii/khaiii.py
@@ -12,8 +12,7 @@
 ###########
 # imports #
 ###########
-import argparse
-from argparse import Namespace
+from argparse import ArgumentParser, Namespace
 import ctypes
 from ctypes.util import find_library
 import logging
@@ -26,11 +25,10 @@
 #########
 # types #
 #########
-class _khaiii_morph_t(ctypes.Structure):    # pylint: disable=invalid-name
+class _khaiii_morph_t(ctypes.Structure):    # pylint: disable=invalid-name,too-few-public-methods
     """
     khaiii_morph_t structure
     """
-    pass
 
 
 _khaiii_morph_t._fields_ = [    # pylint: disable=protected-access
@@ -43,11 +41,10 @@ class _khaiii_morph_t(ctypes.Structure):    # pylint: disable=invalid-name
 ]
 
 
-class _khaiii_word_t(ctypes.Structure):    # pylint: disable=invalid-name
+class _khaiii_word_t(ctypes.Structure):    # pylint: disable=invalid-name,too-few-public-methods
     """
     khaiii_word_t structure
     """
-    pass
 
 
 _khaiii_word_t._fields_ = [    # pylint: disable=protected-access
@@ -63,7 +60,6 @@ class KhaiiiExcept(Exception):
     """
     khaiii API를 위한 표준 예외 클래스
     """
-    pass
 
 
 class KhaiiiMorph:
@@ -384,7 +380,7 @@ def main():
     """
     main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='khaiii API module test program')
+    parser = ArgumentParser(description='khaiii API module test program')
     parser.add_argument('--lib-path', help='library path', metavar='FILE', default='')
     parser.add_argument('--rsc-dir', help='resource directory', metavar='DIR', default='')
     parser.add_argument('--opt-str', help='option string (JSON format)', metavar='JSON', default='')
diff --git a/src/main/python/khaiii/munjong/__init__.py b/src/main/python/khaiii/munjong/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/rsc/lib/sejong_corpus.py b/src/main/python/khaiii/munjong/sejong_corpus.py
similarity index 76%
rename from rsc/lib/sejong_corpus.py
rename to src/main/python/khaiii/munjong/sejong_corpus.py
index c8b9681..fb19fe3 100755
--- a/rsc/lib/sejong_corpus.py
+++ b/src/main/python/khaiii/munjong/sejong_corpus.py
@@ -5,19 +5,19 @@
 """
 Sejong corpus parser
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import argparse
-import codecs
+from argparse import ArgumentParser
 import logging
 import os
 import re
 import sys
+from typing import Iterator, TextIO
 from unicodedata import normalize as norm
 
 
@@ -45,7 +45,7 @@
 _OPEN_TAGS_IN_WRITTEN_SENT = {'<date>', }    # open tags in sentence at written corpus
 
 # close tags in sentence at written corpus
-_CLOSE_TAGS_IN_WRITTEN_SENT = set([tag[0] + '/' + tag[1:] for tag in _OPEN_TAGS_IN_WRITTEN_SENT])
+_CLOSE_TAGS_IN_WRITTEN_SENT = {tag[0] + '/' + tag[1:] for tag in _OPEN_TAGS_IN_WRITTEN_SENT}
 
 # open and close tags in sentence at written corpus
 _TAGS_IN_WRITTEN_SENT = _OPEN_TAGS_IN_WRITTEN_SENT | _CLOSE_TAGS_IN_WRITTEN_SENT
@@ -58,84 +58,67 @@ class ParseError(Exception):
     """
     error occurred while parsing corpus
     """
-    pass
 
 
-class Sentence(object):
+class Morph:    # pylint: disable=too-few-public-methods
     """
-    sentence
+    morpheme
     """
-    def __init__(self):
-        self.words = []    # word list
-        self._wids = set()    # word ID set
+    def __init__(self, lex: str = '', tag: str = ''):
+        if ' ' in lex:
+            raise ParseError('space in raw morph: %s' % lex)
+        self.lex = lex    # lexical form
+        self.tag = tag    # part-of-speech tag
 
     def __str__(self):
-        words_str = '\n'.join([str(_) for _ in self.words])
-        return '# %s\n%s' % (self.raw_str(), words_str)
-
-    def raw_str(self):
-        """
-        raw sentence (words)
-        :return:  raw sentence
-        """
-        return ' '.join([_.raw for _ in self.words])
-
-    def morph_str(self):
-        """
-        make morpheme string
-        :return:  morpheme string
-        """
-        return ' + '.join([_.morph_str() for _ in self.words])
+        return '{}/{}'.format(self.lex, self.tag)
 
-    @classmethod
-    def is_opening(cls, line):
-        """
-        whether sentence opening or not
-        :param  line:  line
-        :return:  whether opening or not
-        """
-        return line in SENT_OPEN_TAGS
+    def __eq__(self, other: 'Morph'):
+        return self.lex == other.lex and self.tag == other.tag
 
-    @classmethod
-    def is_closing(cls, line):
-        """
-        whether sentence closing or not
-        :param  line:  line
-        :return:  whether closing or not
-        """
-        return line in SENT_CLOSE_TAGS
+    def __hash__(self):
+        return hash(str(self))
 
     @classmethod
-    def is_tag_in_sent(cls, line):
-        """
-        whether tag in sentence or not
-        :param  line:  line
-        :return:  whether tag or not
+    def parse(cls, token_str: str, file_name: str = '', line_num: int = 0):
         """
-        return line in _TAGS_IN_WRITTEN_SENT
-
-    def is_good_tags(self):
-        """
-        whether all tags in sentence are good(correct) or not
-        :return:  whether all tags are good
+        parse token string
+        Args:
+            token_str:  morpheme/tag string
+            file_name:  file name
+            line_num:  line number
+        Returns:
+            Morph object
         """
-        return not [morph.tag for word in self.words for morph in word.morphs
-                    if morph.tag not in TAG_SET]
+        def _raise(err_msg: str, file_name: str = None, line_num: int = 0):
+            """
+            에러 메세지를 이용해 파싱 에러를 발생시킨다.
+            Arguments:
+                err_msg:  에러 메세지
+                file_name:  파일명
+                line_num:  라인 번호
+            """
+            file_pfx = '{}({}) '.format(file_name, line_num) if file_name and line_num else ''
+            raise ParseError('{}{}'.format(file_pfx, err_msg))
 
-    def append(self, word):
-        """
-        append word
-        :param word:  Word object
-        """
-        if word.wid and word.wid in self._wids:
-            raise ParseError('duplicated word ID: %s' % word)
-        self.words.append(word)
-        self._wids.add(word.wid)
+        morph = Morph()
+        morph.lex, morph.tag = token_str.rsplit('/', 1)
+        if not morph.lex:
+            _raise('no text in morpheme: {}'.format(token_str), file_name, line_num)
+        if ' ' in morph.lex:
+            _raise('space in raw morph: {}'.format(token_str), file_name, line_num)
+        if morph.tag not in TAG_SET:
+            _raise('invalid tag: {} in {}'.format(morph.tag, token_str), file_name, line_num)
+        for char in morph.lex:
+            if 0x1100 <= ord(char) < 0x1200:
+                _raise('Hangul Jamo character: "{}" in {}'.format(char, token_str), file_name,
+                       line_num)
+        return morph
 
 
-class Word(object):
+class Word:
     """
-    word(EoJeol)
+    word(어절)
     """
     def __init__(self):
         self.wid = ''    # word ID
@@ -143,10 +126,9 @@ def __init__(self):
         self.morphs = []    # morpheme list
 
     def __str__(self):
-        return '%s\t%s\t%s' % (self.wid, self.raw,
-                               ' + '.join([str(morph) for morph in self.morphs]))
+        return '{}\t{}\t{}'.format(self.wid, self.raw, ' + '.join([str(m) for m in self.morphs]))
 
-    def __eq__(self, other):
+    def __eq__(self, other: 'Word'):
         if self.raw != other.raw:
             return False
         if len(self.morphs) != len(other.morphs):
@@ -156,20 +138,24 @@ def __eq__(self, other):
                 return False
         return True
 
-    def morph_str(self):
+    def morph_str(self) -> str:
         """
         make morpheme string
-        :return:  morpheme string
+        Returns:
+            morpheme string
         """
         return ' + '.join([str(_) for _ in self.morphs])
 
     @classmethod
-    def parse(cls, line, file_name, line_num):
+    def parse(cls, line: str, file_name: str, line_num: int) -> 'Word':
         """
         parse word(EoJeol) with single line
-        :param  line:  line
-        :param  file_name:  file name
-        :param  line_num:  line number
+        Args:
+            line:  line
+            file_name:  file name
+            line_num:  line number
+        Returns:
+            Word object
         """
         cols = line.split('\t')
         if len(cols) != 3:
@@ -192,68 +178,98 @@ def parse(cls, line, file_name, line_num):
         return word
 
 
-class Morph(object):    # pylint: disable=too-few-public-methods
+class Sentence:
     """
-    morpheme
+    sentence
     """
-    def __init__(self, lex='', tag=''):
-        if ' ' in lex:
-            raise ParseError('space in raw morph: %s' % lex)
-        self.lex = lex    # lexical form
-        self.tag = tag    # part-of-speech tag
+    def __init__(self):
+        self.words = []    # word list
+        self._wids = set()    # word ID set
 
     def __str__(self):
-        return '%s/%s' % (self.lex, self.tag)
+        words_str = '\n'.join([str(_) for _ in self.words])
+        return '# %s\n%s' % (self.raw_str(), words_str)
 
-    def __eq__(self, other):
-        return self.lex == other.lex and self.tag == other.tag
+    def raw_str(self) -> str:
+        """
+        raw sentence (words)
+        Returns:
+            raw sentence
+        """
+        return ' '.join([_.raw for _ in self.words])
 
-    def __hash__(self):
-        return hash(str(self))
+    def morph_str(self) -> str:
+        """
+        make morpheme string
+        Returns:
+            morpheme string
+        """
+        return ' + '.join([_.morph_str() for _ in self.words])
 
     @classmethod
-    def parse(cls, token_str, file_name='', line_num=0):
+    def is_opening(cls, line: str) -> bool:
         """
-        parse token string
-        :param  token_str:  morpheme/tag string
-        :param  file_name:  file name
-        :param  line_num:  line number
-        :return:  Morph object
+        whether sentence opening or not
+        Args:
+            line:  line
+        Returns:
+            whether opening or not
         """
-        def _raise(err_msg, file_name=None, line_num=0):
-            """
-            에러 메세지를 이용해 파싱 에러를 발생시킨다.
-            Arguments:
-                err_msg:  에러 메세지
-                file_name:  파일명
-                line_num:  라인 번호
-            """
-            file_pfx = '{}({}) '.format(file_name, line_num) if file_name and line_num else ''
-            raise ParseError('{}{}'.format(file_pfx, err_msg))
+        return line in SENT_OPEN_TAGS
 
-        morph = Morph()
-        morph.lex, morph.tag = token_str.rsplit('/', 1)
-        if not morph.lex:
-            _raise('no text in morpheme: {}'.format(token_str), file_name, line_num)
-        if ' ' in morph.lex:
-            _raise('space in raw morph: {}'.format(token_str), file_name, line_num)
-        if morph.tag not in TAG_SET:
-            _raise('invalid tag: {} in {}'.format(morph.tag, token_str), file_name, line_num)
-        for char in morph.lex:
-            if 0x1100 <= ord(char) < 0x1200:
-                _raise('Hangul Jamo character: "{}" in {}'.format(char, token_str), file_name,
-                       line_num)
-        return morph
+    @classmethod
+    def is_closing(cls, line: str) -> bool:
+        """
+        whether sentence closing or not
+        Args:
+            line:  line
+        Returns:
+            whether closing or not
+        """
+        return line in SENT_CLOSE_TAGS
+
+    @classmethod
+    def is_tag_in_sent(cls, line: str) -> bool:
+        """
+        whether tag in sentence or not
+        Args:
+            line:  line
+        Returns:
+            whether tag or not
+        """
+        return line in _TAGS_IN_WRITTEN_SENT
+
+    def is_good_tags(self) -> bool:
+        """
+        whether all tags in sentence are good(correct) or not
+        Returns:
+            whether all tags are good
+        """
+        return not [morph.tag for word in self.words for morph in word.morphs
+                    if morph.tag not in TAG_SET]
+
+    def append(self, word: Word):
+        """
+        append word
+        Args:
+            word:  Word object
+        """
+        if word.wid and word.wid in self._wids:
+            raise ParseError('duplicated word ID: %s' % word)
+        self.words.append(word)
+        self._wids.add(word.wid)
 
 
 #############
 # functions #
 #############
-def sents(fin):
+def sents(fin: TextIO) -> Iterator[Sentence]:
     """
     load from file and return sentences (generator)
-    :param  fin:  input file object
-    :yeild:  Sentence object
+    Args:
+        fin:  input file object
+    Yields:
+        Sentence object
     """
     file_name = os.path.basename(fin.name)
     par_errs = []
@@ -292,14 +308,10 @@ def sents(fin):
 
 def run():
     """
-    load Sejong corpus and print
+    run function which is the start point of program
     """
-    try:
-        for sent in sents(sys.stdin):
-            print(sent)
-    except ParseError as par_err:
-        logging.error(par_err)
-        sys.exit(1)
+    for sent in sents(sys.stdin):
+        print(sent)
 
 
 ########
@@ -307,18 +319,18 @@ def run():
 ########
 def main():
     """
-    load Sejong corpus and print
+    main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='load Sejong corpus and print')
+    parser = ArgumentParser(description='load Sejong corpus and print')
     parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
     parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
     parser.add_argument('--debug', help='enable debug', action='store_true')
     args = parser.parse_args()
 
     if args.input:
-        sys.stdin = codecs.open(args.input, 'r', encoding='UTF-8')
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
     if args.output:
-        sys.stdout = codecs.open(args.output, 'w', encoding='UTF-8')
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
     if args.debug:
         logging.basicConfig(level=logging.DEBUG)
     else:
diff --git a/src/main/python/khaiii/resource/__init__.py b/src/main/python/khaiii/resource/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/rsc/lib/char_align.py b/src/main/python/khaiii/resource/char_align.py
similarity index 90%
rename from rsc/lib/char_align.py
rename to src/main/python/khaiii/resource/char_align.py
index 65af965..1682606 100644
--- a/rsc/lib/char_align.py
+++ b/src/main/python/khaiii/resource/char_align.py
@@ -4,29 +4,71 @@
 """
 형태소 분석 결과와 원문의 음절을 정렬하는 모듈
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import codecs
 from collections import Counter, defaultdict
 import logging
 import os
+from typing import List, Tuple
 
-import jaso
+from khaiii.munjong.sejong_corpus import Word
+from khaiii.resource import jaso
+from khaiii.resource.morphs import Morph
 
 
 #########
 # types #
 #########
-class Aligner(object):
+class MrpChr:    # pylint: disable=too-few-public-methods
+    """
+    음절과 태그 pair
+    """
+    def __init__(self, char: str, tag: str):
+        """
+        Args:
+            char:  음절
+            tag:  태그
+        """
+        self.char = char
+        self.tag = tag
+
+    def __str__(self):
+        return '{}/{}'.format(self.char, self.tag)
+
+    def __hash__(self):
+        return hash(str(self))
+
+    def __eq__(self, other: 'MrpChr'):
+        """
+        Args:
+            other:  다른 객체
+        Returns:
+            같을 경우 True
+        """
+        return self.char == other.char and self.tag == other.tag
+
+    @classmethod
+    def to_str(cls, mrp_chrs: List['MrpChr']):
+        """
+        MrpChr 객체 리스트를 문자열로 변환하는 메소드
+        Args:
+            mrp_chrs:  MrpChr 객체 리스트
+        Returns:
+            변환된 문자열
+        """
+        return ' '.join([str(m) for m in mrp_chrs])
+
+
+class Aligner:
     """
     음절과 형태소 분석 결과의 정렬을 수행하는 클래스
     """
-    def __init__(self, rsc_src):
+    def __init__(self, rsc_src: str):
         """
         리소스를 오픈하고 초기화한다.
         Args:
@@ -37,7 +79,7 @@ def __init__(self, rsc_src):
         self.middle_unmapped = defaultdict(Counter)
         self._open(rsc_src)
 
-    def align(self, word):
+    def align(self, word: Word) -> List[List[MrpChr]]:
         """
         어절의 원문과 분석 결과를 음절 단위로 정렬(매핑)한다.
         Args:
@@ -78,7 +120,7 @@ def print_middle_cnt(self):
         logging.info('total number of unmapped pairs: %d',
                      sum([sum(cnt.values()) for cnt in self.middle_unmapped.values()]))
 
-    def _open(self, rsc_dir):
+    def _open(self, rsc_dir: str):
         """
         initialize resources
         Args:
@@ -86,7 +128,7 @@ def _open(self, rsc_dir):
         """
         file_path = '{}/char_align.map'.format(rsc_dir)
         file_name = os.path.basename(file_path)
-        for line_num, line in enumerate(codecs.open(file_path, 'r', encoding='UTF-8'), start=1):
+        for line_num, line in enumerate(open(file_path, 'r', encoding='UTF-8'), start=1):
             line = line.rstrip('\r\n')
             if not line or line[0] == '#':
                 continue
@@ -105,7 +147,7 @@ def _open(self, rsc_dir):
             self.align_map[key] = map_nums
 
     @classmethod
-    def _get_morph_raw(cls, word):
+    def _get_morph_raw(cls, word: Word) -> str:
         """
         get raw string from morphemes
         Args:
@@ -116,7 +158,7 @@ def _get_morph_raw(cls, word):
         return ''.join([m.lex for m in word.morphs])
 
     @classmethod
-    def _norm(cls, text):
+    def _norm(cls, text: str) -> str:
         """
         unicode normalization of text
         Args:
@@ -127,7 +169,7 @@ def _norm(cls, text):
         return jaso.decompose(text)
 
     @classmethod
-    def _align_phoneme(cls, raw_word, mrp_chrs):
+    def _align_phoneme(cls, raw_word: str, mrp_chrs: List[MrpChr]) -> List[List[MrpChr]]:
         """
         align word with morpheme which is same phoneme
         Args:
@@ -163,7 +205,7 @@ def _align_phoneme(cls, raw_word, mrp_chrs):
         return maps
 
     @classmethod
-    def _align_forward(cls, raw_word, mrp_chrs):
+    def _align_forward(cls, raw_word: str, mrp_chrs: List[MrpChr]) -> Tuple[int, int]:
         """
         align from front of word
         Args:
@@ -192,7 +234,7 @@ def _align_forward(cls, raw_word, mrp_chrs):
         return word_idx, mrp_chrs_idx
 
     @classmethod
-    def _align_backward(cls, raw_word, mrp_chrs):
+    def _align_backward(cls, raw_word: str, mrp_chrs: List[MrpChr]) -> Tuple[int, int]:
         """
         align from back of word
         Args:
@@ -226,7 +268,7 @@ def _align_backward(cls, raw_word, mrp_chrs):
         return word_idx+1, mrp_chrs_idx+1
 
     @classmethod
-    def _is_verb_ending(cls, verb, ending):
+    def _is_verb_ending(cls, verb: Morph, ending: Morph) -> bool:
         """
         whether is verb + ending pattern or not
         Args:
@@ -241,7 +283,7 @@ def _is_verb_ending(cls, verb, ending):
             ending_tag in {'EC', 'EP', 'EF', 'ETN', 'ETM'}
 
     @classmethod
-    def _are_first_last_phoneme_same(cls, raw_word, mrp_chrs):
+    def _are_first_last_phoneme_same(cls, raw_word: str, mrp_chrs: List[MrpChr]) -> bool:
         """
         whether are same the first phoneme and last phoneme
         Args:
@@ -255,7 +297,7 @@ def _are_first_last_phoneme_same(cls, raw_word, mrp_chrs):
         return word_norm[0] == morph_norm[0] and word_norm[-1] == morph_norm[-1]
 
     @classmethod
-    def _is_ah_ending_verb(cls, mrp_chr):
+    def _is_ah_ending_verb(cls, mrp_chr: MrpChr) -> bool:
         """
         whether 'ㅏ' ending verb or not
         Args:
@@ -269,7 +311,7 @@ def _is_ah_ending_verb(cls, mrp_chr):
         return len(norm_char) == 2 and norm_char[1] == 'ᅡ'    # code is 4449, not 12623
 
     @classmethod
-    def _is_eo_ending_verb(cls, mrp_chr):
+    def _is_eo_ending_verb(cls, mrp_chr: MrpChr) -> bool:
         """
         whether 'ㅓ', 'ㅐ' ending verb or not
         Args:
@@ -284,7 +326,8 @@ def _is_eo_ending_verb(cls, mrp_chr):
                 norm_char[1] in ['ᅥ', 'ᅧ', 'ᅢ'])    # code is 4453, 4455, 4450
 
     @classmethod
-    def _align_middle_zero2one(cls, pfx_word, pfx_map, mdl_mrp_chrs, sfx_word, sfx_map):
+    def _align_middle_zero2one(cls, pfx_word: Word, pfx_map: List[MrpChr],
+                               mdl_mrp_chrs: List[MrpChr], sfx_word: Word, sfx_map: List[MrpChr]):
         """
         align middle chunks after forward/backward aligning which has no middle raw character,
             but has a remaining middle single morpheme character
@@ -322,7 +365,7 @@ def _attach_to_pfx():
             raise RuntimeError('nowhere attach to')
 
     @classmethod
-    def _is_share_phoneme(cls, mdl_word, mdl_mrp_chrs):
+    def _is_share_phoneme(cls, mdl_word: str, mdl_mrp_chrs: List[MrpChr]) -> bool:
         """
         whether middle word characters and morpheme characters share same phoneme
         Args:
@@ -342,7 +385,7 @@ def _is_share_phoneme(cls, mdl_word, mdl_mrp_chrs):
                 return False
         return True
 
-    def _align_middle_by_dic(self, mdl_word, mdl_mrp_chrs):
+    def _align_middle_by_dic(self, mdl_word: str, mdl_mrp_chrs: List[MrpChr]) -> List[List[MrpChr]]:
         """
         align middle chunks after forward/backward aligning with mapping dictionary
         Args:
@@ -364,7 +407,8 @@ def _align_middle_by_dic(self, mdl_word, mdl_mrp_chrs):
                 idx += int(map_num)
         return maps
 
-    def _align_middle(self, mdl_word, mdl_mrp_chrs, raw_word, mrp_chrs):
+    def _align_middle(self, mdl_word: str, mdl_mrp_chrs: List[MrpChr], raw_word: str,
+                      mrp_chrs: List[MrpChr]) -> List[List[MrpChr]]:
         """
         align middle chunks after forward/backward aligning
         Args:
@@ -417,7 +461,9 @@ def _align_middle(self, mdl_word, mdl_mrp_chrs, raw_word, mrp_chrs):
         return maps
 
     @classmethod
-    def _align_middle_preproc(cls, pfx_mrp_chrs, pfx_map, mdl_mrp_chrs, sfx_mrp_chrs, sfx_map):
+    def _align_middle_preproc(cls, pfx_mrp_chrs: List[MrpChr], pfx_map: List[List[MrpChr]],
+                              mdl_mrp_chrs: List[MrpChr], sfx_mrp_chrs: List[MrpChr],
+                              sfx_map: List[List[MrpChr]]):
         """
         pre-processing middle part after forward/backward mapping before applying rules
         Args:
@@ -454,7 +500,7 @@ def _align_middle_preproc(cls, pfx_mrp_chrs, pfx_map, mdl_mrp_chrs, sfx_mrp_chrs
                 sfx_map[0].insert(0, mdl_mrp_chrs[-1])
                 del mdl_mrp_chrs[-1]
 
-    def _get_pfx_mdl_sfx(self, raw_word, mrp_chrs):
+    def _get_pfx_mdl_sfx(self, raw_word: str, mrp_chrs: List[MrpChr]) -> Tuple:
         """
         get prefix, middle, suffix after forward/backward align
         Args:
@@ -481,7 +527,7 @@ def _get_pfx_mdl_sfx(self, raw_word, mrp_chrs):
                (pfx_mrp_chrs, mdl_mrp_chrs, sfx_mrp_chrs), \
                (pfx_map, sfx_map)
 
-    def _align_forward_backward(self, raw_word, mrp_chrs):
+    def _align_forward_backward(self, raw_word: str, mrp_chrs: List[MrpChr]) -> List[str]:
         """
         align word with morpheme which is same phoneme
         Args:
@@ -497,7 +543,7 @@ def _align_forward_backward(self, raw_word, mrp_chrs):
 
         if not mdl_word and not mdl_mrp_chrs:
             return pfx_map + sfx_map
-        elif not mdl_word:
+        if not mdl_word:
             if len(mdl_mrp_chrs) == 1:
                 self._align_middle_zero2one(pfx_word, pfx_map, mdl_mrp_chrs, sfx_word, sfx_map)
                 return pfx_map + sfx_map
@@ -529,7 +575,7 @@ class AlignError(Exception):
     """
     음절 정렬 과정에서 나타나는 예외
     """
-    def __init__(self, pfx):
+    def __init__(self, pfx: str):
         """
         Args:
             pfx:  예외 출력 시 보여줄 prefix (카테고리)
@@ -541,50 +587,10 @@ def __init__(self, pfx):
     def __str__(self):
         return '\n'.join(['%s %s' % (self._pfx, _) for _ in self._msgs] + ['', ])
 
-    def add_msg(self, msg):
+    def add_msg(self, msg: str):
         """
         메세제를 추가한다.
         Args:
             msg:  에러 메세지
         """
         self._msgs.append(msg)
-
-
-class MrpChr(object):    # pylint: disable=too-few-public-methods
-    """
-    음절과 태그 pair
-    """
-    def __init__(self, char, tag):
-        """
-        Args:
-            char:  음절
-            tag:  태그
-        """
-        self.char = char
-        self.tag = tag
-
-    def __str__(self):
-        return '%s/%s' % (self.char, self.tag)
-
-    def __hash__(self):
-        return hash(str(self))
-
-    def __eq__(self, other):
-        """
-        Args:
-            other:  다른 객체
-        Returns:
-            같을 경우 True
-        """
-        return self.char == other.char and self.tag == other.tag
-
-    @classmethod
-    def to_str(cls, mrp_chrs):
-        """
-        MrpChr 객체 리스트를 문자열로 변환하는 메소드
-        Args:
-            mrp_chrs:  MrpChr 객체 리스트
-        Returns:
-            변환된 문자열
-        """
-        return ' '.join([str(_) for _ in mrp_chrs])
diff --git a/rsc/lib/jaso.py b/src/main/python/khaiii/resource/jaso.py
similarity index 92%
rename from rsc/lib/jaso.py
rename to src/main/python/khaiii/resource/jaso.py
index a93a64a..1f5e184 100644
--- a/rsc/lib/jaso.py
+++ b/src/main/python/khaiii/resource/jaso.py
@@ -4,10 +4,16 @@
 """
 한글 자소 관련 유틸리티 모듈
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
+###########
+# imports #
+###########
+from typing import Tuple
+
+
 #############
 # constants #
 #############
@@ -31,7 +37,7 @@
 #############
 # functions #
 #############
-def _decomp_char(char):
+def _decomp_char(char: str) -> Tuple[str, str, str]:
     """
     한글 음절 하나를 자소로 분해한다.
     Args:
diff --git a/rsc/lib/morphs.py b/src/main/python/khaiii/resource/morphs.py
similarity index 87%
rename from rsc/lib/morphs.py
rename to src/main/python/khaiii/resource/morphs.py
index 69184f1..be6d232 100644
--- a/rsc/lib/morphs.py
+++ b/src/main/python/khaiii/resource/morphs.py
@@ -2,12 +2,19 @@
 
 
 """
-형태소 분석 결과를 기술한 문자열을 파싱하는 모듈
+형태소 분석 결과를 기술한 문자열을 파싱하는 모듈.
+TODO(jamie): sejong_corpus 모듈의 Morph 클래스와 중복되므로 정리 필요
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
+###########
+# imports #
+###########
+from typing import List
+
+
 #############
 # constants #
 #############
@@ -36,14 +43,13 @@ class ParseError(Exception):
     """
     형태소 분석 결과 문자열을 파싱하면서 발생하는 오류
     """
-    pass
 
 
-class Morph(object):
+class Morph:
     """
     형태소
     """
-    def __init__(self, lex, tag):
+    def __init__(self, lex: str, tag: str):
         """
         Arguments:
             lex:  형태소(어휘)
@@ -57,7 +63,7 @@ def __str__(self):
             return self.lex
         return '{}/{}'.format(self.lex, self.tag)
 
-    def is_word_delim(self):
+    def is_word_delim(self) -> bool:
         """
         어절의 경계를 나타태는 지 여부
         Returns:
@@ -65,7 +71,7 @@ def is_word_delim(self):
         """
         return not self.tag and self.lex == WORD_DELIM_STR
 
-    def is_sent_delim(self):
+    def is_sent_delim(self) -> bool:
         """
         문장의 경계를 나타태는 지 여부
         Returns:
@@ -74,7 +80,7 @@ def is_sent_delim(self):
         return not self.tag and self.lex == SENT_DELIM_STR
 
     @classmethod
-    def to_str(cls, morphs):
+    def to_str(cls, morphs: List['Morph']) -> str:
         """
         Morph 객체 리스트를 문자열로 변환한다.
         Arguments:
@@ -85,7 +91,7 @@ def to_str(cls, morphs):
         return ' + '.join([str(m) for m in morphs])
 
     @classmethod
-    def parse(cls, morphs_str):
+    def parse(cls, morphs_str: str) -> List['Morph']:
         """
         형태소 분석 결과 형태의 문자열을 파싱하여 Morph 객체 리스트를 반환하는 파싱 함수
         Arguments:
@@ -98,7 +104,7 @@ def parse(cls, morphs_str):
         return [cls._parse_one(m) for m in morphs_str.split(' + ')]
 
     @classmethod
-    def _parse_one(cls, morph_str):
+    def _parse_one(cls, morph_str: str) -> 'Morph':
         """
         하나의 형태소 객체를 기술한 문자열을 파싱한다.
         Arguments:
diff --git a/rsc/lib/resource.py b/src/main/python/khaiii/resource/resource.py
similarity index 72%
rename from rsc/lib/resource.py
rename to src/main/python/khaiii/resource/resource.py
index a2ebc9e..3514e81 100644
--- a/rsc/lib/resource.py
+++ b/src/main/python/khaiii/resource/resource.py
@@ -4,18 +4,19 @@
 """
 resources for training and tagging
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import codecs
+from argparse import Namespace
 import logging
 import os
+from typing import Dict
 
-from vocabulary import Vocabulary
+from khaiii.resource.vocabulary import Vocabulary
 
 
 #############
@@ -33,13 +34,14 @@
 #########
 # types #
 #########
-class Resource(object):
+class Resource:
     """
     resources
     """
-    def __init__(self, cfg):
+    def __init__(self, cfg: Namespace):
         """
-        :param  cfg:  config
+        Args:
+            cfg:  config
         """
         vocab_in_path = '{}/vocab.in'.format(cfg.rsc_src)
         self.vocab_in = Vocabulary(vocab_in_path, cfg.cutoff, SPECIAL_CHARS)
@@ -49,14 +51,16 @@ def __init__(self, cfg):
         self.restore_dic = self._load_restore_dic(restore_dic_path)
 
     @classmethod
-    def _load_restore_dic(cls, path):
+    def _load_restore_dic(cls, path: str) -> Dict[str, str]:
         """
         load character to output tag mapping
-        :param  path:  file path
-        :return:  dictionary
+        Args:
+            path:  file path
+        Returns:
+            dictionary
         """
         dic = {}
-        for line in codecs.open(path, 'r', encoding='UTF-8'):
+        for line in open(path, 'r', encoding='UTF-8'):
             line = line.rstrip('\r\n')
             if not line:
                 continue
diff --git a/rsc/lib/trie.py b/src/main/python/khaiii/resource/trie.py
similarity index 90%
rename from rsc/lib/trie.py
rename to src/main/python/khaiii/resource/trie.py
index b380ee9..f843285 100755
--- a/rsc/lib/trie.py
+++ b/src/main/python/khaiii/resource/trie.py
@@ -5,18 +5,18 @@
 """
 TRIE 모듈
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import argparse
-import codecs
+from argparse import ArgumentParser, Namespace
 import logging
 import struct
 import sys
+from typing import List
 
 
 #############
@@ -28,14 +28,73 @@
 #########
 # types #
 #########
-class Trie(object):
+class Node:
+    """
+    TRIE 노드
+    """
+    def __init__(self, char: str = ''):
+        """
+        Args:
+            char:  (유니코드) 입력 문자
+        """
+        self.depth = -1    # 노드를 배열로 펼치기 위해 필요한 현재 노드의 루트로부터의 깊이
+        self.char = char    # 입력 문자
+        self.value = 0    # 단말 노드일 경우 그 값 (단말 노드가 아닌 경우 0)
+        self.children = {}    # 자식 노드들
+        self.child_start = -1    # 현재 노드로부터 자식 노드들의 시작점 사이에 있는 노드의 갯수
+
+    def __str__(self):
+        node_str = self.char if self.char else 'ROOT'
+        if self.value > 0:
+            node_str = '{%s: %d}' % (node_str, self.value)
+        pfx = ''
+        if self.depth > 0:
+            pfx = '·' * self.depth
+        child_start_str = ''
+        if self.child_start > 0:
+            child_start_str = ', {}'.format(self.child_start)
+        return '{}({}, [{}]{})'.format(pfx, node_str, ', '.join(self.children.keys()),
+                                       child_start_str)
+
+    def insert(self, key: str, val: int):
+        """
+        문자열 키와 그 값을 현재 노드로부터 내려가며 적절한 위치에 삽입한다.
+        Args:
+            key:  키 문자열
+            val:  값 (양수)
+        """
+        if val <= 0:
+            raise ValueError('value must be greater than zero')
+        if not key:
+            self.value = val
+        elif key[0] in self.children:
+            self.children[key[0]].insert(key[1:], val)
+        else:
+            new_node = Node(key[0])
+            self.children[key[0]] = new_node
+            new_node.insert(key[1:], val)
+        logging.debug('INSERT {%s: %d} INTO %s', key, val, self)
+
+    def pack(self) -> bytes:
+        """
+        구조체로 packing한다.
+        Returns:
+            packing된 구조체
+        """
+        char = 0 if not self.char else self.char
+        if isinstance(char, str):
+            char = ord(char)
+        return _NODE_STRUCT.pack(char, self.value, self.child_start, len(self.children))
+
+
+class Trie:
     """
     TRIE 인터페이스
     """
     def __init__(self):
         self.root = Node()
 
-    def insert(self, key, val):
+    def insert(self, key: str, val: int):
         """
         하나의 문자열 키와 그 값을 삽입한다.
         Args:
@@ -44,7 +103,7 @@ def insert(self, key, val):
         """
         self.root.insert(key, val)
 
-    def update(self, keys, vals=None):
+    def update(self, keys: List[str], vals: List[int] = None):
         """
         문자열 키와 그 값의 리스트를 차례로 삽입한다.
         값이 없을 경우 키 목록의 인덱스(1부터 시작)를 값으로 설정한다.
@@ -59,7 +118,7 @@ def update(self, keys, vals=None):
         for key, val in zip(keys, vals):
             self.insert(key, val)
 
-    def save(self, file_path):
+    def save(self, file_path: str):
         """
         TRIE 자료구조를 파일로 저장한다.
         Args:
@@ -74,7 +133,7 @@ def save(self, file_path):
         logging.info('total nodes: %d', len(nodes))
         logging.info('expected size: %d', len(nodes) * _NODE_STRUCT.size)
 
-    def find(self, key):
+    def find(self, key: str) -> int:
         """
         키를 이용하여 값을 얻는다.
         Args:
@@ -86,11 +145,10 @@ def find(self, key):
         for char in key:
             if char not in node.children:
                 return 0
-            else:
-                node = node.children[char]
+            node = node.children[char]
         return node.value
 
-    def _breadth_first_traverse(self):
+    def _breadth_first_traverse(self) -> List[Node]:
         """
         너비우선으로 루트로부터 전체 노드를 탐색한다.
         Returns:
@@ -113,7 +171,7 @@ def _breadth_first_traverse(self):
         return nodes
 
     @classmethod
-    def _set_child_start(cls, nodes):
+    def _set_child_start(cls, nodes: List[Node]):
         """
         child_start 필드를 세팅한다.
         Args:
@@ -135,69 +193,10 @@ def _set_child_start(cls, nodes):
                                                                                 else -1
 
 
-class Node(object):
-    """
-    TRIE 노드
-    """
-    def __init__(self, char=''):
-        """
-        Args:
-            char:  (유니코드) 입력 문자
-        """
-        self.depth = -1    # 노드를 배열로 펼치기 위해 필요한 현재 노드의 루트로부터의 깊이
-        self.char = char    # 입력 문자
-        self.value = 0    # 단말 노드일 경우 그 값 (단말 노드가 아닌 경우 0)
-        self.children = {}    # 자식 노드들
-        self.child_start = -1    # 현재 노드로부터 자식 노드들의 시작점 사이에 있는 노드의 갯수
-
-    def __str__(self):
-        node_str = self.char if self.char else 'ROOT'
-        if self.value > 0:
-            node_str = '{%s: %d}' % (node_str, self.value)
-        pfx = ''
-        if self.depth > 0:
-            pfx = '·' * self.depth
-        child_start_str = ''
-        if self.child_start > 0:
-            child_start_str = ', {}'.format(self.child_start)
-        return '{}({}, [{}]{})'.format(pfx, node_str, ', '.join(self.children.keys()),
-                                       child_start_str)
-
-    def insert(self, key, val):
-        """
-        문자열 키와 그 값을 현재 노드로부터 내려가며 적절한 위치에 삽입한다.
-        Args:
-            key:  키 문자열
-            val:  값 (양수)
-        """
-        if val <= 0:
-            raise ValueError('value must be greater than zero')
-        if not key:
-            self.value = val
-        elif key[0] in self.children:
-            self.children[key[0]].insert(key[1:], val)
-        else:
-            new_node = Node(key[0])
-            self.children[key[0]] = new_node
-            new_node.insert(key[1:], val)
-        logging.debug('INSERT {%s: %d} INTO %s', key, val, self)
-
-    def pack(self):
-        """
-        구조체로 packing한다.
-        Returns:
-            packing된 구조체
-        """
-        char = 0 if not self.char else self.char
-        if isinstance(char, str):
-            char = ord(char)
-        return _NODE_STRUCT.pack(char, self.value, self.child_start, len(self.children))
-
-
 #############
 # functions #
 #############
-def run(args):
+def run(args: Namespace):
     """
     run function which is the start point of program
     Args:
@@ -233,7 +232,7 @@ def main():
     """
     main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='트라이를 빌드합니다.')
+    parser = ArgumentParser(description='트라이를 빌드합니다.')
     parser.add_argument('-o', '--output', help='output file', metavar='FILE', required=True)
     parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
     parser.add_argument('--val', dest='has_val', help='탭으로 구분된 마지막 컬럼이 값일 경우',
@@ -242,9 +241,7 @@ def main():
     args = parser.parse_args()
 
     if args.input:
-        sys.stdin = codecs.open(args.input, 'r', encoding='UTF-8')
-    else:
-        sys.stdin = codecs.getreader('UTF-8')(sys.stdin.detach())
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
     if args.debug:
         logging.basicConfig(level=logging.DEBUG)
     else:
diff --git a/rsc/lib/vocabulary.py b/src/main/python/khaiii/resource/vocabulary.py
similarity index 89%
rename from rsc/lib/vocabulary.py
rename to src/main/python/khaiii/resource/vocabulary.py
index f07fb1c..6e9ba87 100644
--- a/rsc/lib/vocabulary.py
+++ b/src/main/python/khaiii/resource/vocabulary.py
@@ -4,17 +4,17 @@
 """
 vocabulary library
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import codecs
 import copy
 import logging
 import os
+from typing import List
 
 
 #########
@@ -24,7 +24,7 @@ class Vocabulary:
     """
     vocabulary class
     """
-    def __init__(self, path, cutoff=1, special=None, padding=''):
+    def __init__(self, path: str, cutoff: int = 1, special: List[str] = None, padding: str = ''):
         """
         Args:
             path:  file path
@@ -60,7 +60,7 @@ def __getitem__(self, key):
         except KeyError:
             return 0    # unknown word number
 
-    def __len__(self):
+    def __len__(self) -> int:
         return len(self.dic)
 
     '''
@@ -78,7 +78,7 @@ def get_embedding(self, dim, padding_idx=None):
         return nn.Embedding(len(self), dim)
     '''    # pylint: disable=pointless-string-statement
 
-    def padding_idx(self):
+    def padding_idx(self) -> int:
         """
         맨 마지막에 추가한 패딩의 인덱스를 리턴한다.
         Returns:
@@ -88,7 +88,7 @@ def padding_idx(self):
             raise RuntimeError('vocabulary has no padding')
         return self.dic[self.padding]
 
-    def _load(self, path, cutoff=1):
+    def _load(self, path: str, cutoff: int = 1):
         """
         load vocabulary from file
         Args:
@@ -97,7 +97,7 @@ def _load(self, path, cutoff=1):
         """
         append_num = 0
         cutoff_num = 0
-        for line in codecs.open(path, 'r', encoding='UTF-8'):
+        for line in open(path, 'r', encoding='UTF-8'):
             line = line.rstrip('\r\n')
             if not line:
                 continue

From 4c040535c7d3ac9caad31bb31f1cb1900983a8f4 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 21 Jan 2019 16:59:24 +0900
Subject: [PATCH 04/41] =?UTF-8?q?=EC=84=B8=EC=A2=85=20=EC=BD=94=ED=8D=BC?=
 =?UTF-8?q?=EC=8A=A4=EB=A1=9C=EB=B6=80=ED=84=B0=20=EC=9D=8C=EC=A0=88?=
 =?UTF-8?q?=EB=8B=A8=EC=9C=84=20=EC=A0=95=EB=A0=AC=EC=9D=84=20=EC=88=98?=
 =?UTF-8?q?=ED=96=89=ED=95=98=EB=8A=94=20=EC=8A=A4=ED=81=AC=EB=A6=BD?=
 =?UTF-8?q?=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 train/bin/map_char_to_tag.py | 184 +++++++++++++++++++++++++++++++++++
 1 file changed, 184 insertions(+)
 create mode 100755 train/bin/map_char_to_tag.py

diff --git a/train/bin/map_char_to_tag.py b/train/bin/map_char_to_tag.py
new file mode 100755
index 0000000..b762ab3
--- /dev/null
+++ b/train/bin/map_char_to_tag.py
@@ -0,0 +1,184 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+어절 내에서 원문과 형태소 사이에 정렬을 수행하고 음절기반 학습 코퍼스를 생성
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+from collections import defaultdict
+import logging
+import os
+import random
+import sys
+from typing import Iterator, List
+
+from khaiii.munjong.sejong_corpus import Sentence, sents, Word
+from khaiii.resource.char_align import Aligner, AlignError, MrpChr
+
+
+#############
+# variables #
+#############
+_MAP_DIC = defaultdict(list)    # char/tag => morpheme result mapping dictionary
+_MAP_CASE = defaultdict(list)    # char/tag => case(word) dictionary
+
+
+#############
+# functions #
+#############
+def _print_restore_dic(args: Namespace):
+    """
+    원형복원 사전을 출력한다.
+    Args:
+        args:  program arguments
+    """
+    if not args.restore_dic:
+        return
+    with open(args.restore_dic, 'w', encoding='UTF-8') as fout:
+        with open('{}.case'.format(args.restore_dic), 'w', encoding='UTF-8') as fcase:
+            for (char, tag), vals in _MAP_DIC.items():
+                for idx, val in enumerate(vals):
+                    print('{}/{}:{}\t{}'.format(char, tag, idx, MrpChr.to_str(val)), file=fout)
+                    word = _MAP_CASE[char, tag][idx]
+                    print('{}/{}:{}\t{}'.format(char, tag, idx, word), file=fcase)
+
+
+def sent_iter(args: Namespace) -> Iterator[Sentence]:
+    """
+    sentence generator
+    Args:
+        args:  program arguments
+    Yields:
+        sentence
+    """
+    for name in sorted(os.listdir(args.corpus_dir)):
+        if not name.endswith('.txt'):
+            continue
+        logging.info(name)
+        path = '{}/{}'.format(args.corpus_dir, name)
+        for sent in sents(open(path, 'r', encoding='UTF-8')):
+            yield sent
+
+
+def _maps_to_tag(word: Word, maps: List[List[MrpChr]]) -> List[str]:
+    """
+    매핑 정보를 이용해 어절 내 각 음절 별로 출력 태그를 생성한다.
+    Args:
+        word:  Word object
+        maps:  음절 별 매핑 정보
+    Returns:
+        출력 태그 리스트
+    """
+    tags = []
+    for char, align in zip(word.raw, maps):
+        if len(align) == 1 and align[0].char == char:
+            tags.append(align[0].tag)
+            continue
+        tag = ':'.join([_.tag for _ in align])
+        if (char, tag) in _MAP_DIC:
+            vals = _MAP_DIC[char, tag]
+            try:
+                idx = vals.index(align)
+            except ValueError:
+                idx = len(vals)
+                vals.append(align)
+                _MAP_CASE[char, tag].append(word)
+            tag = '{}:{}'.format(tag, idx)
+        else:
+            _MAP_DIC[char, tag].append(align)
+            _MAP_CASE[char, tag].append(word)
+            tag = '{}:0'.format(tag)
+        tags.append(tag)
+    return tags
+
+
+def _print_sent(sent: Sentence, word_per_maps: List[List[List[MrpChr]]]):
+    """
+    각 어절 별 매핑 정보 리스트를 이용해 한 문장을 출력한다.
+    Args:
+        sent:  Sentence object
+        word_per_maps:  각 어절별로 음절 매핑 정보를 포함하는 리스트
+    """
+    lines = []
+    has_error = False
+    for word, maps in zip(sent.words, word_per_maps):
+        if maps:
+            if len(word.raw) == len(maps):
+                lines.append('{}\t{}'.format(word.raw, ' '.join(_maps_to_tag(word, maps))))
+            else:
+                raise RuntimeError('length of maps is different from length of word')
+        else:
+            has_error = True
+            logging.debug(word)
+    if not has_error:
+        print('\n'.join(lines))
+        print()
+
+
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    aligner = Aligner(args.rsc_src)
+    funmap = open(args.unmapped, 'w', encoding='UTF-8') if args.unmapped else None
+
+    for sent in sent_iter(args):
+        if 0.0 < args.sample < 1.0 and random.random() >= args.sample:
+            continue
+        word_per_maps = []
+        for word in sent.words:
+            try:
+                maps = aligner.align(word)
+            except AlignError as algn_err:
+                if funmap:
+                    algn_err.add_msg(str(word))
+                    print(algn_err, file=funmap)
+                maps = []
+            word_per_maps.append(maps)
+        _print_sent(sent, word_per_maps)
+
+    _print_restore_dic(args)
+    aligner.print_middle_cnt()
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='어절 내에서 원문과 형태소 사이에 정렬을 수행하고'
+                                        ' 음절기반 학습 코퍼스를 생성')
+    parser.add_argument('-c', '--corpus-dir', help='corpus dir', metavar='DIR', required=True)
+    parser.add_argument('--rsc-src', help='train resource dir <default: ../rsc/src>', metavar='DIR',
+                        default='../rsc/src')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--restore-dic', help='restore dic output file', metavar='FILE')
+    parser.add_argument('--unmapped', help='unmapped log file', metavar='FILE')
+    parser.add_argument('--sample', help='sampling ratio', metavar='REAL', type=float, default=1.0)
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From fac247dccf92876e1e7ea054edf835bce0078348 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 21 Jan 2019 17:25:43 +0900
Subject: [PATCH 05/41] =?UTF-8?q?=ED=95=99=EC=8A=B5=20=EC=BD=94=ED=8D=BC?=
 =?UTF-8?q?=EC=8A=A4=EB=A5=BC=20dev/test/train=EC=9C=BC=EB=A1=9C=20?=
 =?UTF-8?q?=EB=B6=84=ED=95=A0=ED=95=98=EB=8A=94=20=EC=8A=A4=ED=81=AC?=
 =?UTF-8?q?=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 train/bin/split_corpus.py | 106 ++++++++++++++++++++++++++++++++++++++
 1 file changed, 106 insertions(+)
 create mode 100755 train/bin/split_corpus.py

diff --git a/train/bin/split_corpus.py b/train/bin/split_corpus.py
new file mode 100755
index 0000000..4dda3b8
--- /dev/null
+++ b/train/bin/split_corpus.py
@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+코퍼스를 train/dev/test로 분할한다.
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+import logging
+import random
+import sys
+from typing import Iterator, List, TextIO
+
+
+#############
+# functions #
+#############
+def _sents(fin: TextIO) -> Iterator[List[str]]:
+    """
+    read from file and yield a sentence (generator)
+    Args:
+        fin:  input file
+    Yields:
+        sentence (list of lines)
+    """
+    sent = []
+    for line in fin:
+        line = line.rstrip('\r\n')
+        if not line:
+            if sent:
+                yield sent
+                sent = []
+            continue
+        sent.append(line)
+    if sent:
+        yield sent
+
+
+def _write_to_file(path: str, sents: List[List[str]]):
+    """
+    파일에 쓴다.
+    Args:
+        path:  path
+        sents:  sentences
+    """
+    with open(path, 'w', encoding='UTF-8') as fout:
+        for sent in sents:
+            print('\n'.join(sent), file=fout)
+            print(file=fout)
+
+
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    sents = []
+    for num, sent in enumerate(_sents(sys.stdin), start=1):
+        if num % 100000 == 0:
+            logging.info('%d00k-th sent..', num // 100000)
+        sents.append(sent)
+    random.shuffle(sents)
+    _write_to_file('{}.dev'.format(args.out_pfx), sents[:args.dev])
+    _write_to_file('{}.test'.format(args.out_pfx), sents[args.dev:args.dev+args.test])
+    _write_to_file('{}.train'.format(args.out_pfx), sents[args.dev+args.test:])
+    logging.info('dev / test / train: %d / %d / %d', args.dev, args.test,
+                 len(sents[args.dev+args.test:]))
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='코퍼스를 train/dev/test로 분할한다.')
+    parser.add_argument('-o', '--out-pfx', help='output file prefix', metavar='NAME', required=True)
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--dev', help='number of sentence in dev set', metavar='NUM', type=int,
+                        default=5000)
+    parser.add_argument('--test', help='number of sentence in test set', metavar='NUM', type=int,
+                        default=5000)
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From 420cbc73cde5136827460539fe1d179b02db5217 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 21 Jan 2019 18:02:41 +0900
Subject: [PATCH 06/41] =?UTF-8?q?=EC=9E=85,=20=EC=B6=9C=EB=A0=A5=20vocab?=
 =?UTF-8?q?=EC=9D=84=20=EC=83=9D=EC=84=B1=ED=95=98=EB=8A=94=20=EC=8A=A4?=
 =?UTF-8?q?=ED=81=AC=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 train/bin/make_vocab.py      | 96 ++++++++++++++++++++++++++++++++++++
 train/bin/map_char_to_tag.py |  4 +-
 2 files changed, 98 insertions(+), 2 deletions(-)
 create mode 100755 train/bin/make_vocab.py

diff --git a/train/bin/make_vocab.py b/train/bin/make_vocab.py
new file mode 100755
index 0000000..c5b0fd4
--- /dev/null
+++ b/train/bin/make_vocab.py
@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+입력(음절) 및 출력(태그) vocabulary를 생성한다.
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+from collections import Counter
+import logging
+import os
+import sys
+from typing import TextIO
+
+from khaiii.resource.morphs import TAGS
+
+
+#############
+# functions #
+#############
+def _print(cnt: Counter, fout: TextIO, is_with_freq: bool = True):
+    """
+    vocabulary 사전을 출력한다.
+    Args:
+        cnt:  Counter object
+        fout:  출력 파일
+        is_with_freq:  빈도를 함께 출력할 지 여부
+    """
+    for char, freq in sorted(cnt.items(), key=lambda x: x[0]):
+        if is_with_freq and freq < 2:
+            continue
+        if is_with_freq:
+            print('{}\t{}'.format(char, freq), file=fout)
+        else:
+            print(char, file=fout)
+
+
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    in_cnt = Counter()
+    out_cnt = Counter()
+    for line_num, line in enumerate(sys.stdin, start=1):
+        if line_num % 1000000 == 0:
+            logging.info('%dm-th line', line_num // 1000000)
+        line = line.rstrip('\r\n')
+        if not line:
+            continue
+        raw, tagged = line.split('\t')
+        in_cnt.update(list(raw))
+        out_cnt.update([tag for tag in tagged.split() if tag[2:] not in TAGS])
+    os.makedirs(args.rsc_src, exist_ok=True)
+    with open('{}/vocab.in'.format(args.rsc_src), 'w', encoding='UTF-8') as fout:
+        _print(in_cnt, fout)
+    with open('{}/vocab.out'.format(args.rsc_src), 'w', encoding='UTF-8') as fout:
+        print('\n'.join(['B-{}'.format(tag) for tag in TAGS]), file=fout)
+        print('\n'.join(['I-{}'.format(tag) for tag in TAGS]), file=fout)
+        _print(out_cnt, fout, is_with_freq=False)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='입력(음절) 및 출력(태그) vocabulary를 생성한다.')
+    parser.add_argument('--rsc-src', help='resource source dir <default: ../rsc/src>',
+                        metavar='DIR', default='../rsc/src')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/train/bin/map_char_to_tag.py b/train/bin/map_char_to_tag.py
index b762ab3..d868eb5 100755
--- a/train/bin/map_char_to_tag.py
+++ b/train/bin/map_char_to_tag.py
@@ -161,8 +161,8 @@ def main():
     parser = ArgumentParser(description='어절 내에서 원문과 형태소 사이에 정렬을 수행하고'
                                         ' 음절기반 학습 코퍼스를 생성')
     parser.add_argument('-c', '--corpus-dir', help='corpus dir', metavar='DIR', required=True)
-    parser.add_argument('--rsc-src', help='train resource dir <default: ../rsc/src>', metavar='DIR',
-                        default='../rsc/src')
+    parser.add_argument('--rsc-src', help='resource source dir <default: ../rsc/src>',
+                        metavar='DIR', default='../rsc/src')
     parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
     parser.add_argument('--restore-dic', help='restore dic output file', metavar='FILE')
     parser.add_argument('--unmapped', help='unmapped log file', metavar='FILE')

From 8eed3eb214586be3373c54658f64e2ec7844b704 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Tue, 22 Jan 2019 13:49:21 +0900
Subject: [PATCH 07/41] =?UTF-8?q?=EB=AA=A8=EB=8D=B8=20=ED=95=99=EC=8A=B5?=
 =?UTF-8?q?=EC=9D=84=20=EC=9C=84=ED=95=9C=20=EB=AA=A8=EB=93=88=EA=B3=BC=20?=
 =?UTF-8?q?=EC=8A=A4=ED=81=AC=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/main/python/khaiii/resource/vocabulary.py |  17 +-
 src/main/python/khaiii/train/dataset.py       | 196 +++++++++
 src/main/python/khaiii/train/embedder.py      | 102 +++++
 src/main/python/khaiii/train/evaluator.py     | 116 ++++++
 src/main/python/khaiii/train/models.py        | 115 ++++++
 src/main/python/khaiii/train/sentence.py      | 278 +++++++++++++
 src/main/python/khaiii/train/tagger.py        |  71 ++++
 src/main/python/khaiii/train/trainer.py       | 385 ++++++++++++++++++
 train/bin/train.py                            |  87 ++++
 9 files changed, 1351 insertions(+), 16 deletions(-)
 create mode 100644 src/main/python/khaiii/train/dataset.py
 create mode 100644 src/main/python/khaiii/train/embedder.py
 create mode 100644 src/main/python/khaiii/train/evaluator.py
 create mode 100644 src/main/python/khaiii/train/models.py
 create mode 100644 src/main/python/khaiii/train/sentence.py
 create mode 100644 src/main/python/khaiii/train/tagger.py
 create mode 100644 src/main/python/khaiii/train/trainer.py
 create mode 100755 train/bin/train.py

diff --git a/src/main/python/khaiii/resource/vocabulary.py b/src/main/python/khaiii/resource/vocabulary.py
index 6e9ba87..eb7a67c 100644
--- a/src/main/python/khaiii/resource/vocabulary.py
+++ b/src/main/python/khaiii/resource/vocabulary.py
@@ -60,24 +60,9 @@ def __getitem__(self, key):
         except KeyError:
             return 0    # unknown word number
 
-    def __len__(self) -> int:
+    def __len__(self):
         return len(self.dic)
 
-    '''
-    # 리소스 빌드 시 pytorch 의존성 제거를 위해 임시로 메서드를 제거합니다.
-    # 추후 학습 코드를 추가할 때 이 부분을 리팩토링 합니다.
-    def get_embedding(self, dim, padding_idx=None):
-        """
-        embedding을 리턴합니다.
-        Args:
-            dim:  embedding dimension
-            padding_idx:  padding index
-        """
-        if padding_idx:
-            return nn.Embedding(len(self), dim, padding_idx=padding_idx)
-        return nn.Embedding(len(self), dim)
-    '''    # pylint: disable=pointless-string-statement
-
     def padding_idx(self) -> int:
         """
         맨 마지막에 추가한 패딩의 인덱스를 리턴한다.
diff --git a/src/main/python/khaiii/train/dataset.py b/src/main/python/khaiii/train/dataset.py
new file mode 100644
index 0000000..7aee785
--- /dev/null
+++ b/src/main/python/khaiii/train/dataset.py
@@ -0,0 +1,196 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+corpus for training
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import Namespace
+import logging
+import os
+import random
+from typing import List, TextIO, Tuple
+
+from torch import LongTensor, Tensor    # pylint: disable=no-member, no-name-in-module
+from tqdm import tqdm
+
+from khaiii.resource.resource import PAD_CHR, Resource
+from khaiii.train.sentence import PosSentence, PosWord
+
+
+#########
+# types #
+#########
+class PosSentTensor(PosSentence):
+    """
+    tensor transformable sentence
+    """
+    def __init__(self, raw_sent: str = ''):
+        super().__init__(raw_sent)
+        if raw_sent:
+            self.init_pos_tags()
+
+    def __len__(self):
+        # RNN에서 길이가 긴 것으로 정렬하기위한 문장의 길이 = 음절 갯수 + 문장 경계 + 어절 경계
+        if self.words:
+            return sum([len(w) for w in self.words]) + len(self.words) + 1
+        if self.pos_tagged_words:
+            return sum([len(w.raw) for w in self.pos_tagged_words]) + len(self.pos_tagged_words) + 1
+        return 0
+
+    def make_labels(self, with_spc: bool) -> List[str]:
+        """
+        각 음절별로 출력 레이블(태그)를 생성한다.
+        Args:
+            with_spc:  공백(어절 경계) 포함 여부
+        Returns:
+            레이블 리스트
+        """
+        if not with_spc:
+            # 문장 경계, 어절 경계 등 가상 음절을 제외하고 순수한 음절들의 레이블
+            return [tag for pos_word in self.pos_tagged_words for tag in pos_word.tags]
+        labels = [PAD_CHR, ]    # 문장 시작
+        for pos_word in self.pos_tagged_words:
+            if len(labels) > 1:
+                labels.append(PAD_CHR)    # 어절 경계
+            labels.extend(pos_word.tags)
+        labels.append(PAD_CHR)    # 문장 종료
+        return labels
+
+    def make_contexts(self, window: int, spc_dropout: float) -> List[str]:
+        """
+        각 음절 별로 좌/우 window 크기 만큼 context를 만든다.
+        Args:
+            window:  left/right window size
+            spc_dropout:  space(word delimiter) dropout rate
+        Returns:
+            contexts
+        """
+        contexts = []
+        for wrd_idx, word in enumerate(self.words):
+            for chr_idx, char in enumerate(word):
+                left_context = list(reversed(word[:chr_idx]))
+                if random.random() >= spc_dropout:
+                    left_context.append('<w>')
+                for left_word in reversed(self.words[:wrd_idx]):
+                    left_context.extend(reversed(left_word))
+                    if len(left_context) >= window:
+                        break
+                if len(left_context) < window:
+                    left_context.extend(['<s>', ] * (window - len(left_context)))
+                left_context = list(reversed(left_context[:window]))
+                assert len(left_context) == window
+
+                right_context = list(word[chr_idx+1:])
+                if random.random() >= spc_dropout:
+                    right_context.append('</w>')
+                for right_word in self.words[wrd_idx+1:]:
+                    right_context.extend(list(right_word))
+                    if len(right_context) >= window:
+                        break
+                if len(right_context) < window:
+                    right_context.extend(['</s>', ] * (window - len(right_context)))
+                right_context = right_context[:window]
+                assert len(right_context) == window
+                contexts.append(left_context + [char, ] + right_context)
+        return contexts
+
+    def to_tensor(self, cfg: Namespace, rsc: Resource, is_train: bool) -> Tuple[Tensor, Tensor]:
+        """
+        문장 내에 포함된 전체 음절들과 태그를 모델의 forward 메소드에 넣을 수 있는 텐서로 변환한다.
+        Args:
+            cfg:  config
+            rsc:  Resource object
+            is_train:  whether is train or not
+        Returns:
+            labels tensor
+            contexts tensor
+        """
+        # 차원: [문장내 음절 갯수, ]
+        label_nums = [rsc.vocab_out[l] for l in self.make_labels(False)]
+        labels_tensor = LongTensor(label_nums)
+        # 차원: [문장내 음절 갯수 x context 크기]
+        spc_dropout = cfg.spc_dropout if is_train else 0.0
+        context_nums = [[rsc.vocab_in[c] for c in context] \
+                        for context in self.make_contexts(cfg.window, spc_dropout)]
+        contexts_tensor = LongTensor(context_nums)
+        return labels_tensor, contexts_tensor
+
+    def make_chars(self) -> List[str]:
+        """
+        문장 내 포함된 음절들을 만든다. 문장 경계 및 어절 경계를 포함한다.
+        Returns:
+            음절의 리스트
+        """
+        chars = ['<s>', ]    # 문장 시작
+        for word in self.words:
+            if len(chars) > 1:
+                chars.append('<w>')    # 어절 경계
+            chars.extend(word)
+        chars.append('</s>')    # 문장 종료
+        return chars
+
+
+class PosDataset:
+    """
+    part-of-speech tag dataset
+    """
+    def __init__(self, cfg: Namespace, restore_dic: dict, fin: TextIO):
+        """
+        Args:
+            cfg:  config
+            restore_dic:  restore dictionary
+            fin:  input file
+        """
+        self.cfg = cfg
+        self.fin = fin
+        self.sents = []
+        self.sent_idx = -1
+        self._load(restore_dic)
+
+    def __str__(self):
+        return '<PosDataset: file: {}, sents: {}, sent_idx: {}>'.format(
+            os.path.basename(self.fin.name), len(self.sents), self.sent_idx)
+
+    def _load(self, restore_dic: dict):
+        """
+        load data file
+        Args:
+            restore_dic:  restore dictionary
+        """
+        sent = PosSentTensor()
+        lines = self.fin.readlines()
+        for line in tqdm(lines, os.path.basename(self.fin.name), len(lines), mininterval=1,
+                         ncols=100):
+            line = line.rstrip('\r\n')
+            if not line:
+                if sent and sent.pos_tagged_words:
+                    sent.set_raw_by_words()
+                    self.sents.append(sent)
+                sent = PosSentTensor()
+                continue
+            raw, tags = line.split('\t')
+            pos_word = PosWord(raw)
+            pos_word.set_pos_result(tags.split(), restore_dic)
+            sent.pos_tagged_words.append(pos_word)
+        logging.info('%s: %d sentences', os.path.basename(self.fin.name), len(self.sents))
+
+    def __iter__(self):
+        self.sent_idx = -1
+        random.shuffle(self.sents)
+        return self
+
+    def __next__(self):
+        self.sent_idx += 1
+        if self.sent_idx >= len(self.sents):
+            raise StopIteration()
+        return self.sents[self.sent_idx]
+
+    def __len__(self):
+        return len(self.sents)
diff --git a/src/main/python/khaiii/train/embedder.py b/src/main/python/khaiii/train/embedder.py
new file mode 100644
index 0000000..61acf2d
--- /dev/null
+++ b/src/main/python/khaiii/train/embedder.py
@@ -0,0 +1,102 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+making embedding models
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import Namespace
+import math
+
+import torch
+from torch import nn, Tensor
+
+from khaiii.resource.resource import Resource
+
+
+class Embedder(nn.Module):
+    """
+    embedder class
+    """
+    def __init__(self, cfg: Namespace, rsc: Resource):
+        """
+        Args:
+            cfg:  config
+            rsc:  Resource object
+        """
+        super().__init__()
+        self.cfg = cfg
+        self.rsc = rsc
+        self.embedding = nn.Embedding(len(rsc.vocab_in), cfg.embed_dim)
+
+    def forward(self, inputs):    # pylint: disable=arguments-differ
+        """
+        임베딩을 생성하는 메소드
+        Args:
+            inputs:  contexts of batch size
+        Returns:
+            embedding
+        """
+        embeds = self.embedding(inputs)
+        embeds += positional_encoding(self.cfg.context_len, self.cfg.context_len,
+                                      self.cfg.embed_dim, 1)
+        return embeds
+
+
+#############
+# functions #
+#############
+def memoize(func):
+    """
+    memoize decorator
+    """
+    class Memodict(dict):
+        """
+        Memoization decorator for a function taking one or more arguments.
+        """
+        def __getitem__(self, *key):
+            return dict.__getitem__(self, key)
+
+        def __missing__(self, key):
+            ret = self[key] = func(*key)
+            return ret
+
+    return Memodict().__getitem__
+
+
+@memoize
+def positional_encoding(sent_len: int, max_dim: int, embed_dim: int, method: int = 1) -> Tensor:
+    """
+    positional encoding Tensor 출력.
+    embeds [batch_size, context_len, embed_dim]에 Broadcasting 으로 더해짐
+    Args:
+        sent_len:  actual sentence length
+        max_dim:  maximum dimension
+        embed_dim:  embedding dimension
+        method:  method number (1. end-to-end memory networks or 2. attention is all you need)
+    Returns:
+        pe [context_len, embed_dim]
+    """
+    pe_tensor = torch.zeros([max_dim, embed_dim])    # pylint: disable=no-member
+    for pos in range(1, sent_len + 1):
+        for i in range(1, embed_dim+1):
+            if method == 1:
+                # end-to-end memory networks
+                pe_tensor[pos-1, i-1] = 1 - pos / sent_len - ((i / embed_dim) *
+                                                              (1 - 2 * pos / sent_len))
+            elif method == 2:
+                # attention is all you need
+                if i % 2 == 0:
+                    pe_tensor[pos-1, i-1] = math.sin(pos / 10000 ** (2*i / embed_dim))
+                else:
+                    pe_tensor[pos-1, i-1] = math.cos(pos / 10000 ** (2*i / embed_dim))
+    if torch.cuda.is_available():
+        pe_tensor = pe_tensor.cuda()
+    pe_tensor.detach()
+    return pe_tensor
diff --git a/src/main/python/khaiii/train/evaluator.py b/src/main/python/khaiii/train/evaluator.py
new file mode 100644
index 0000000..33f5fe3
--- /dev/null
+++ b/src/main/python/khaiii/train/evaluator.py
@@ -0,0 +1,116 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+evaluation related module
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from collections import Counter
+import logging
+from typing import List, TextIO, Tuple
+
+from khaiii.train.sentence import PosMorph, PosSentence, PosWord
+
+
+#########
+# types #
+#########
+class Evaluator:
+    """
+    evauator
+    """
+    def __init__(self):
+        self.cnt = Counter()
+
+    def evaluate(self) -> Tuple[float, float, float]:
+        """
+        char/word accuracy, f-score(recall/precision)를 측정한다.
+        Returns:
+            character accuracy
+            word accuracy
+            f-score
+        """
+        char_acc = self.cnt['match_chars'] / self.cnt['total_chars']
+        word_acc = self.cnt['match_words'] / self.cnt['total_words']
+        recall = self.cnt['match_morphs'] / self.cnt['total_gold_morphs']
+        precision = self.cnt['match_morphs'] / self.cnt['total_pred_morphs']
+        f_score = 2.0 * recall * precision / (recall + precision)
+        self.cnt.clear()
+        return char_acc, word_acc, f_score
+
+    def count(self, correct_sent: PosSentence, predict_sent: PosSentence):
+        """
+        정답 문장과 비교하여 맞춘 갯수를 샌다.
+        Args:
+            correct_sent:  정답 문장
+            predict_sent:  예측한 문장
+        """
+        assert len(correct_sent.words) == len(predict_sent.words)
+        for gold, pred in zip(correct_sent.pos_tagged_words, predict_sent.pos_tagged_words):
+            self.cnt['total_chars'] += len(gold.res_tags)
+            self.cnt['match_chars'] += len([1 for x, y in zip(gold.res_tags, pred.res_tags)
+                                            if x == y])
+            self._count_word(gold, pred)
+
+    def _count_word(self, gold: PosWord, pred: PosWord):
+        """
+        count with gold standard and predicted (will update counter)
+        Args:
+            gold:  gold standard word
+            pred:  predicted word
+        """
+        self.cnt['total_words'] += 1
+        gold_morphs = gold.pos_tagged_morphs
+        pred_morphs = pred.pos_tagged_morphs
+        if gold == pred:
+            self.cnt['match_words'] += 1
+            num_match = len(gold_morphs)
+            self.cnt['total_gold_morphs'] += num_match
+            self.cnt['total_pred_morphs'] += num_match
+            self.cnt['match_morphs'] += num_match
+            return
+        logging.debug('gold: %s', ' '.join([str(_) for _ in gold_morphs]))
+        logging.debug('pred: %s', ' '.join([str(_) for _ in pred_morphs]))
+        self.cnt['total_gold_morphs'] += len(gold_morphs)
+        self.cnt['total_pred_morphs'] += len(pred_morphs)
+        gold_set = self.morphs_to_set(gold_morphs)
+        pred_set = self.morphs_to_set(pred_morphs)
+        self.cnt['match_morphs'] += len(gold_set & pred_set)
+
+    @classmethod
+    def morphs_to_set(cls, morphs: List[PosMorph]) -> set:
+        """
+        make set from morpheme list
+        Args:
+            morphs:  morpheme list
+        Returns:
+            morphemes set
+        """
+        morph_cnt = Counter([(morph.morph, morph.pos_tag) for morph in morphs])
+        morph_set = set()
+        for (lex, tag), freq in morph_cnt.items():
+            if freq == 1:
+                morph_set.add((lex, tag))
+            else:
+                morph_set.update([(lex, tag, _) for _ in range(1, freq+1)])
+        return morph_set
+
+    def report(self, fout: TextIO):
+        """
+        report recall/precision to file
+        :param  fout:  output file
+        """
+        print('word accuracy: %d / %d = %.4f' % (self.cnt['match_words'], self.cnt['total_words'],
+                                                 self.cnt['match_words'] / self.cnt['total_words']),
+              file=fout)
+        recall = self.cnt['match_morphs'] / self.cnt['total_gold_morphs']
+        precision = self.cnt['match_morphs'] / self.cnt['total_pred_morphs']
+        f_score = 2.0 * recall * precision / (recall + precision)
+        print('f-score / (recall, precision): %.4f / (%.4f, %.4f)' % (f_score, recall, precision),
+              file=fout)
diff --git a/src/main/python/khaiii/train/models.py b/src/main/python/khaiii/train/models.py
new file mode 100644
index 0000000..c1b3222
--- /dev/null
+++ b/src/main/python/khaiii/train/models.py
@@ -0,0 +1,115 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+Pytorch models
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import Namespace
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from khaiii.resource.resource import Resource
+from khaiii.train.embedder import Embedder
+
+
+#########
+# types #
+#########
+class PosModel(nn.Module):
+    """
+    part-of-speech tagger pytorch model
+    """
+    def __init__(self, cfg: Namespace, rsc: Resource):
+        """
+        Args:
+            cfg (Namespace):  config
+            rsc (Resource):  Resource object
+        """
+        super().__init__()
+        self.cfg = cfg
+        self.rsc = rsc
+        self.embedder = Embedder(cfg, rsc)
+
+    def forward(self, *inputs):
+        raise NotImplementedError
+
+    def save(self, path: str):
+        """
+        모델을 저장하는 메소드
+        Args:
+            path (str):  경로
+        """
+        torch.save(self.state_dict(), path)
+
+    def load(self, path: str):
+        """
+        저장된 모델을 로드하는 메소드
+        Args:
+            path (str):  경로
+        """
+        if torch.cuda.is_available():
+            state_dict = torch.load(path)
+        else:
+            state_dict = torch.load(path, map_location=lambda storage, loc: storage)
+        self.load_state_dict(state_dict)
+        if torch.cuda.is_available():
+            self.cuda()
+
+
+class CnnModel(PosModel):
+    """
+    convolutional neural network based part-of-speech tagger
+    """
+    def __init__(self, cfg: Namespace, rsc: Resource):
+        """
+        Args:
+            cfg (Namespace):  config
+            rsc (Resource):  Resource object
+        """
+        super().__init__(cfg, rsc)
+
+        ngram = min(5, cfg.window * 2 + 1)
+        self.convs = nn.ModuleList([nn.Conv1d(cfg.embed_dim, cfg.embed_dim, kernel_size)
+                                    for kernel_size in range(2, ngram+1)])
+
+        # conv => hidden
+        setattr(cfg, 'hidden_dim', (cfg.embed_dim * len(self.convs) + len(rsc.vocab_out)) // 2)
+        self.conv2hidden = nn.Linear(cfg.embed_dim * len(self.convs), cfg.hidden_dim)
+
+        # hidden => tag
+        self.hidden2tag = nn.Linear(cfg.hidden_dim, len(rsc.vocab_out))
+
+    def forward(self, contexts):    # pylint: disable=arguments-differ
+        """
+        forward path
+        Args:
+            contexts:  batch size list of character and context
+        Returns:
+            output score
+        """
+        embeds = self.embedder(contexts)
+        embeds_t = embeds.transpose(1, 2)
+
+        pool_outs = []
+        for conv in self.convs:
+            conv_out = F.relu(conv(embeds_t))
+            pool_outs.append(F.max_pool1d(conv_out, conv_out.size(2)))
+
+        # conv => hidden
+        features = torch.cat([p.view(contexts.size(0), -1) for p in pool_outs], dim=1)    # pylint: disable=no-member
+        features_drop = F.dropout(features)
+        hidden_out = F.relu(self.conv2hidden(features_drop))
+
+        # hidden => tag
+        hidden_out_drop = F.dropout(hidden_out)
+        tag_out = self.hidden2tag(hidden_out_drop)
+        return tag_out
diff --git a/src/main/python/khaiii/train/sentence.py b/src/main/python/khaiii/train/sentence.py
new file mode 100644
index 0000000..c36113c
--- /dev/null
+++ b/src/main/python/khaiii/train/sentence.py
@@ -0,0 +1,278 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+sentence, word, morph, ...
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+import logging
+import re
+
+from typing import List, Tuple
+
+
+#########
+# types #
+#########
+class Sentence:
+    """
+    raw sentence
+    """
+    def __init__(self, raw: str):
+        """
+        :param  raw:  raw sentence
+        """
+        self.words = raw.split()
+
+
+class PosMorph:
+    """
+    morpheme
+    """
+    def __init__(self, morph: str, pos_tag: str = 'O', beg: int = -1, end: int = -1):
+        self.morph = morph
+        self.pos_tag = pos_tag
+        self.beg = beg
+        self.end = end
+
+    def __str__(self):
+        return '{}/{}'.format(self.morph, self.pos_tag)
+
+    def __len__(self):
+        return self.end - self.beg
+
+
+class PosWord:
+    """
+    part-of-speech tagged word
+    """
+    def __init__(self, raw: str):
+        """
+        Args:
+            raw:  raw word
+        """
+        self.raw = raw
+        self.tags = []    # output tags for each character
+        self.res_chrs = raw    # 원형 복원된 형태소들의 음절들을 합친 것
+        self.res_tags = []    # 원형 복원된 형태소 음절들의 음절별 태그 (IOB)
+        self.pos_tagged_morphs = [PosMorph(x, 'O', idx, idx+1) for idx, x in enumerate(raw)]
+
+    def __str__(self):
+        return '{}\t{}'.format(self.raw, ' '.join([str(x) for x in self.pos_tagged_morphs]))
+
+    def for_pretrain(self) -> str:
+        """
+        pre-training을 위한 출력
+        Returns:
+            pre-training을 위한 문장
+        """
+        morph_strs = []
+        morph = ''
+        prev_tag = ''
+        for char, iob_tag in zip(self.raw, self.tags):
+            try:
+                iob_tag, _ = iob_tag.split(':', 1)
+            except ValueError:
+                pass
+            try:
+                iob, tag = iob_tag.split('-')
+            except ValueError as val_err:
+                logging.error('raw: %s', self.raw)
+                logging.error('tags: %s', self.tags)
+                logging.error('iob_tag: %s', iob_tag)
+                raise val_err
+            if iob == 'B':
+                if morph:
+                    morph_strs.append('%s/%s' % (re.sub(r'\d', '0', morph), prev_tag))
+                morph = char
+                prev_tag = tag
+            elif iob == 'I':
+                if prev_tag == tag:
+                    morph += char
+                else:
+                    if morph:
+                        morph_strs.append('%s/%s' % (re.sub(r'\d', '0', morph), prev_tag))
+                    morph = char
+                    prev_tag = tag
+        if morph:
+            morph_strs.append('%s/%s' % (re.sub(r'\d', '0', morph), prev_tag))
+        return ' '.join(morph_strs)
+
+    def __eq__(self, other: 'PosWord'):
+        """
+        어절의 형태소 분석 결과가 일치할 경우 같다고 간주한다. (평가 프로그램에서 어절 단위 일치 여부 판단 시 사용)
+        Args:
+            other:  other object
+        """
+        return self.res_chrs == other.res_chrs and self.res_tags == other.res_tags
+
+    def set_pos_result(self, tags: List[str], restore_dic: dict = None):
+        """
+        외부에서 생성된 PosWord객체의 정보를 현재 인스턴스에 설정합니다.
+        Args:
+            tags:  파일로 부터 읽은 형태소 태그(음절단위)
+            restore_dic:  원형 복원 사전
+        """
+        if not restore_dic:
+            tags = [x.split(':', 1)[0] for x in tags]
+        self.tags = tags
+        assert len(self.raw) == len(self.tags)    # 음절수와 태그수는 동일해야 한다.
+        self.pos_tagged_morphs = self._make_pos_morphs(restore_dic)
+
+    def _make_pos_morphs(self, restore_dic: dict = None):
+        """
+        형태소 태그리스트를 대상으로 B/I 로 병합되는 위치를 구합니다.
+        Args:
+            restore_dic:  원형 복원 사전
+        """
+        if not self.tags:
+            return []
+
+        self._restore(restore_dic)
+
+        pos_morphs = []
+        for beg, (lex, iob_tag) in enumerate(zip(self.res_chrs, self.res_tags)):
+            try:
+                iob, pos_tag = iob_tag.rsplit('-', 1)
+            except ValueError as val_err:
+                logging.error('invalid char/tag: %s/%s in [%s] %s', lex, iob_tag, self.res_chrs,
+                              self.res_tags)
+                raise val_err
+            if iob == 'B' or not pos_morphs or pos_morphs[-1].pos_tag != pos_tag:
+                pos_morphs.append(PosMorph(lex, pos_tag, beg, beg+1))
+            elif iob == 'I':
+                if pos_morphs[-1].pos_tag == pos_tag:
+                    pos_morphs[-1].morph += lex
+                    pos_morphs[-1].end += len(lex)
+                else:
+                    logging.debug('tag is different between B and I: %s vs %s',
+                                  pos_morphs[-1].pos_tag, pos_tag)
+                    pos_morphs.append(PosMorph(lex, pos_tag, beg, beg+1))
+            else:
+                raise ValueError('invalid IOB tag: {}/{} in [{}] {}'.format \
+                                     (lex, iob_tag, self.res_chrs, self.res_tags))
+        return pos_morphs
+
+    def _restore(self, restore_dic: dict):
+        """
+        원형 복원 사전을 이용하여 형태소의 원형을 복원한다.
+        Args:
+            restore_dic:  원형 복원 사전
+        """
+        if not restore_dic:
+            self.res_chrs = self.raw
+            self.res_tags = self.tags
+            return
+
+        res_chrs = []
+        self.res_tags = []
+        for char, tag in zip(self.raw, self.tags):
+            if ':' in tag:
+                key = '{}/{}'.format(char, tag)
+                if key in restore_dic:
+                    for char_tag in restore_dic[key].split():
+                        res_chr, res_tag = char_tag.rsplit('/', 1)
+                        res_chrs.append(res_chr)
+                        self.res_tags.append(res_tag)
+                    continue
+                else:
+                    logging.debug('mapping not found: %s/%s', char, tag)
+                    tag, _ = tag.split(':', 1)
+            res_chrs.append(char)
+            self.res_tags.append(tag)
+        self.res_chrs = ''.join(res_chrs)
+
+
+class PosSentence(Sentence):
+    """
+    part-of-speech tagged sentence
+    """
+    def __init__(self, raw: str):
+        """
+        Args:
+            raw:  raw sentence
+        """
+        super().__init__(raw)
+        self.pos_tagged_words = [] # list of PosWord
+
+    def __str__(self):
+        return '\n'.join([str(pos_word) for pos_word in self.pos_tagged_words])
+
+    def get_beg_end_list(self) -> Tuple[List[int], List[int]]:
+        """
+        모든 형태소의 시작위치를 담는 리스트와, 끝 위치를 담는 리스트를 구합니다.
+        Returns:
+            list of begin positions
+            list of end positions
+        """
+        begs = []
+        ends = []
+        for word in self.pos_tagged_words:
+            for morph in word.pos_tagged_morphs:
+                begs.append(morph.beg)
+                ends.append(morph.end)
+        return begs, ends
+
+    def set_raw_by_words(self):
+        """
+        Sentence 객체의 'words' 멤버를 PosWords 객체의 raw를 이용하여 채운다.
+        """
+        self.words = [pos_word.raw for pos_word in self.pos_tagged_words]
+
+    def init_pos_tags(self):
+        """
+        PosWord 객체를 생성하고 태그를 'O'로 세팅한다.
+        """
+        if self.pos_tagged_words:
+            raise RuntimeError('PoS tagged words are already initialized')
+        for word in self.words:
+            self.pos_tagged_words.append(PosWord(word))
+
+    def set_pos_result(self, tags: List[str], restore_dic: dict = None):
+        """
+        문장 전체에 대한 형태소 태그 출력 레이블 정보를 세팅하고 형태소를 복원한다.
+        Args:
+            tags:  문장 전체 태그 출력 레이블 정보
+            restore_dic:  원형 복원 사전
+        """
+        total_char_num = 0
+        for pos_word in self.pos_tagged_words:
+            pos_word.set_pos_result(tags[total_char_num:total_char_num + len(pos_word.raw)],
+                                    restore_dic)
+            total_char_num += len(pos_word.raw)
+        assert total_char_num == len(tags)
+
+    def get_sequence(self, morph: bool = True, tag: bool = True, simple: bool = False) -> List[str]:
+        """
+        태그를 포함한 형태소 문자열을 생성하여 리턴합니다.
+        Args:
+            morph:  형태소 출력
+            tag:  태그 표시 방법
+            simple:  tag를 1byte만 출력
+        Returns:
+            문자열의 리스트
+        """
+        sequence = []
+        for word in self.pos_tagged_words:
+            for pos_morph in word.pos_tagged_morphs:
+                morphs = []
+                if morph:
+                    morphs.append(pos_morph.morph)
+                if tag:
+                    morphs.append(pos_morph.pos_tag if not simple else pos_morph.pos_tag[0])
+                sequence.append('/'.join(morphs))
+        return sequence
+
+    def get_all_morphs(self) -> List[str]:
+        """
+        문장을 구성하는 모든 PosMorph의 리스트를 리턴합니다.
+        Returns:
+            모든 형태소 리스트
+        """
+        return [morph for word in self.pos_tagged_words for morph in word.pos_tagged_morphs]
diff --git a/src/main/python/khaiii/train/tagger.py b/src/main/python/khaiii/train/tagger.py
new file mode 100644
index 0000000..06fa0e2
--- /dev/null
+++ b/src/main/python/khaiii/train/tagger.py
@@ -0,0 +1,71 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+part-of-speech tagger
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import Namespace
+import json
+import logging
+import re
+
+import torch
+import torch.nn.functional as F
+
+from khaiii.resource.resource import Resource
+from khaiii.train.dataset import PosSentTensor
+from khaiii.train.models import CnnModel
+
+
+#########
+# types #
+#########
+class PosTagger:
+    """
+    part-of-speech tagger
+    """
+    def __init__(self, rsc_src: str):
+        """
+        Args:
+            rsc_src:  resource dir
+        """
+        cfg_dict = json.load(open('{}/config.json'.format(rsc_src), 'r', encoding='UTF-8'))
+        self.cfg = Namespace()
+        for key, val in cfg_dict.items():
+            setattr(self.cfg, key, val)
+        self.cfg.rsc_src = rsc_src
+        self.rsc = Resource(self.cfg)
+        self.model = CnnModel(self.cfg, self.rsc)
+        self.model.load('{}/model.state'.format(rsc_src))
+        self.model.eval()
+
+    def tag_raw(self, raw_sent: str, enable_restore: bool = True) -> PosSentTensor:
+        """
+        part-of-speech tagging at raw sentence
+        Args:
+            raw_sent:  raw input sentence
+        Returns:
+            PosSentTensor object
+        """
+        pos_sent = PosSentTensor(raw_sent)
+        _, contexts = pos_sent.to_tensor(self.cfg, self.rsc, False)
+        if torch.cuda.is_available():
+            contexts = contexts.cuda()
+        outputs = self.model(contexts)
+        _, predicts = F.softmax(outputs, dim=1).max(1)
+        tags = [self.rsc.vocab_out[t.item()] for t in predicts]
+        pos_sent.set_pos_result(tags, self.rsc.restore_dic if enable_restore else None)
+
+        if logging.getLogger().isEnabledFor(logging.DEBUG):
+            raw_nospc = re.sub(r'\s+', '', raw_sent)
+            for idx, (tag, pred) in enumerate(zip(tags, predicts)):
+                logging.debug('[%2d]%s: %5s(%d)', idx, raw_nospc[idx], tag, pred.data[0])
+
+        return pos_sent
diff --git a/src/main/python/khaiii/train/trainer.py b/src/main/python/khaiii/train/trainer.py
new file mode 100644
index 0000000..140d558
--- /dev/null
+++ b/src/main/python/khaiii/train/trainer.py
@@ -0,0 +1,385 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+training related library
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import Namespace
+import copy
+from datetime import datetime, timedelta
+import json
+import logging
+import os
+import pathlib
+import pprint
+import sys
+from typing import List, Tuple
+
+from tensorboardX import SummaryWriter
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from tqdm import tqdm
+
+from khaiii.train.dataset import PosDataset
+from khaiii.train.evaluator import Evaluator
+from khaiii.train.models import CnnModel
+from khaiii.resource.resource import Resource
+
+
+#############
+# functions #
+#############
+class Trainer:
+    """
+    trainer class
+    """
+    def __init__(self, cfg: Namespace):
+        """
+        Args:
+            cfg:  config
+        """
+        self.cfg = cfg
+        setattr(cfg, 'model_id', self.model_id(cfg))
+        setattr(cfg, 'out_dir', '{}/{}'.format(cfg.log_dir, cfg.model_id))
+        setattr(cfg, 'context_len', 2 * cfg.window + 1)
+        self.rsc = Resource(cfg)
+        self.model = CnnModel(cfg, self.rsc)
+        self.optimizer = torch.optim.Adam(self.model.parameters(), cfg.learning_rate)
+        self.criterion = nn.CrossEntropyLoss()
+        self.evaler = Evaluator()
+        self._load_dataset()
+        if 'epoch' not in cfg.__dict__:
+            setattr(cfg, 'epoch', 0)
+            setattr(cfg, 'iteration', 0)
+            setattr(cfg, 'iter_best', 0)
+        self.log_file = None    # tab separated log file
+        self.sum_wrt = None    # tensorboard summary writer
+        self.loss_trains = []
+        self.loss_devs = []
+        self.acc_chars = []
+        self.acc_words = []
+        self.f_scores = []
+        self.learning_rates = []
+        self.batch_sizes = []
+
+    @classmethod
+    def model_id(cls, cfg: Namespace) -> str:
+        """
+        get model ID
+        Args:
+            cfg:  config
+        Returns:
+            model ID
+        """
+        model_cfgs = [
+            os.path.basename(cfg.in_pfx),
+            'cut{}'.format(cfg.cutoff),
+            'win{}'.format(cfg.window),
+            'sdo{}'.format(cfg.spc_dropout),
+            'emb{}'.format(cfg.embed_dim),
+            'lr{}'.format(cfg.learning_rate),
+            'lrd{}'.format(cfg.lr_decay),
+            'bs{}'.format(cfg.batch_size),
+            'ci{}'.format(cfg.check_iter),
+            'bg{}'.format(cfg.batch_grow),
+        ]
+        return '.'.join(model_cfgs)
+
+    def _load_dataset(self):
+        """
+        load training dataset
+        """
+        dataset_dev_path = '{}.dev'.format(self.cfg.in_pfx)
+        self.dataset_dev = PosDataset(self.cfg, self.rsc.restore_dic,
+                                      open(dataset_dev_path, 'r', encoding='UTF-8'))
+        dataset_train_path = '{}.train'.format(self.cfg.in_pfx)
+        self.dataset_train = PosDataset(self.cfg, self.rsc.restore_dic,
+                                        open(dataset_train_path, 'r', encoding='UTF-8'))
+
+    @classmethod
+    def _dt_str(cls, dt_obj: datetime) -> str:
+        """
+        string formatting for datetime object
+        Args:
+            dt_obj:  datetime object
+        Returns:
+            string
+        """
+        return dt_obj.strftime('%m/%d %H:%M:%S')
+
+    @classmethod
+    def _elapsed(cls, td_obj: timedelta) -> str:
+        """
+        string formatting for timedelta object
+        Args:
+            td_obj:  timedelta object
+        Returns:
+            string
+        """
+        seconds = td_obj.seconds
+        if td_obj.days > 0:
+            seconds += td_obj.days * 24 * 3600
+        hours = seconds // 3600
+        seconds -= hours * 3600
+        minutes = seconds // 60
+        seconds -= minutes * 60
+        return '{}:{:02d}:{:02d}'.format(hours, minutes, seconds)
+
+    def _check_continue(self):
+        """
+        기존에 학습하다 중지한 경우 그 이후부터 계속해서 학습할 수 있도록 이전 상태를 복원한다.
+        """
+        out_path = pathlib.Path(self.cfg.out_dir)
+        if not out_path.is_dir():
+            return
+        logging.info('==== continue learning: %s ====', self.cfg.model_id)
+        cfg = json.load(open('{}/config.json'.format(self.cfg.out_dir), 'r', encoding='UTF-8'))
+        for key, val in cfg.items():
+            setattr(self.cfg, key, val)
+        self._revert(False)
+
+        f_score_best = 0.0
+        best_idx = -1
+        for idx, line in enumerate(open('{}/log.tsv'.format(self.cfg.out_dir))):
+            line = line.rstrip('\r\n')
+            if not line:
+                continue
+            (epoch, iteration, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate,
+             batch_size) = line.split('\t')
+            self.cfg.epoch = int(epoch) + 1
+            self.cfg.iteration = self.cfg.iter_best = int(iteration) * self.cfg.check_iter
+            self.loss_trains.append(float(loss_train))
+            self.loss_devs.append(float(loss_dev))
+            self.acc_chars.append(float(acc_char))
+            self.acc_words.append(float(acc_word))
+            self.f_scores.append(float(f_score))
+            self.learning_rates.append(float(learning_rate))
+            self.batch_sizes.append(int(batch_size))
+            if float(f_score) > f_score_best:
+                f_score_best = float(f_score)
+                best_idx = idx
+        logging.info('---- [%d|%d] loss(train/dev): %f / %f, acc(char/word): %f / %f, ' \
+                     'f-score: %f, lr: %f, bs: %d ----', self.cfg.epoch,
+                     self.cfg.iteration // self.cfg.check_iter, self.loss_trains[best_idx],
+                     self.loss_devs[best_idx], self.acc_chars[best_idx], self.acc_words[best_idx],
+                     self.f_scores[best_idx], self.learning_rates[-1], self.batch_sizes[-1])
+
+    def train(self):
+        """
+        train model with dataset
+        """
+        self._check_continue()
+        logging.info('config: %s', pprint.pformat(self.cfg.__dict__))
+        train_begin = datetime.now()
+        logging.info('{{{{ training begin: %s {{{{', self._dt_str(train_begin))
+        if torch.cuda.is_available():
+            self.model.cuda()
+        pathlib.Path(self.cfg.out_dir).mkdir(parents=True, exist_ok=True)
+        self.log_file = open('{}/log.tsv'.format(self.cfg.out_dir), 'at')
+        self.sum_wrt = SummaryWriter(self.cfg.out_dir)
+        patience = self.cfg.patience
+        for _ in range(10000):
+            has_best = self.train_epoch()
+            if has_best:
+                patience = self.cfg.patience
+            else:
+                if patience > 0:
+                    self._revert(True)
+                    patience -= 1
+                    logging.info('==== revert to iter: %d, f-score: %f, patience: %d ====',
+                                 self.cfg.iter_best, max(self.f_scores), patience)
+                else:
+                    break
+        train_end = datetime.now()
+        train_elapsed = self._elapsed(train_end - train_begin)
+        logging.info('}}}} training end: %s, elapsed: %s, epoch: %s, iter: %dk }}}}',
+                     self._dt_str(train_end), train_elapsed, self.cfg.epoch,
+                     self.cfg.iteration // 1000)
+
+    def _revert(self, is_decay_lr: bool):
+        """
+        이전 best 모델로 되돌린다.
+        Args:
+            is_decay_lr:  whether multiply decay factor or not
+        """
+        self.model.load('{}/model.state'.format(self.cfg.out_dir))
+        if is_decay_lr:
+            self.cfg.learning_rate *= self.cfg.lr_decay
+        self._load_optim('{}/optim.state'.format(self.cfg.out_dir), self.cfg.learning_rate)
+
+    def train_epoch(self) -> bool:
+        """
+        한 epoch을 학습한다. 배치 단위는 글자 단위
+        Returns:
+            best f-score를 기록한 iteration 여부
+        """
+        epoch_begin = datetime.now()
+        logging.info('{{{{ epoch: %d, begin: %s {{{{', self.cfg.epoch, self._dt_str(epoch_begin))
+        iter_lap1 = datetime.now()
+        has_best = False
+        batches = []
+        loss_trains = []
+        batch_size = self.cfg.batch_size
+        if self.cfg.batch_grow > 0:
+            batch_size = self.cfg.batch_size + self.cfg.iteration // self.cfg.batch_grow
+        for train_sent in tqdm(self.dataset_train, 'EPOCH [{}]'.format(self.cfg.epoch),
+                               len(self.dataset_train), mininterval=1, ncols=100):
+            train_labels, train_contexts = train_sent.to_tensor(self.cfg, self.rsc, True)
+            if torch.cuda.is_available():
+                train_labels = train_labels.cuda()
+                train_contexts = train_contexts.cuda()
+
+            self.optimizer.zero_grad()
+            self.model.train()
+            train_outputs = self.model(train_contexts)
+            batches.append((train_labels, train_outputs))
+            if sum([batch[0].size(0) for batch in batches]) < batch_size:
+                continue
+            batch_label = torch.cat([x[0] for x in batches], 0)    # pylint: disable=no-member
+            batch_output = torch.cat([x[1] for x in batches], 0)    # pylint: disable=no-member
+            batches = []
+
+            batch_output.requires_grad_()
+            loss_train = self.criterion(batch_output, batch_label)
+            loss_trains.append(loss_train.item())
+            loss_train.backward()
+            self.optimizer.step()
+            self.cfg.iteration += 1
+
+            if self.cfg.iteration % self.cfg.check_iter == 0:
+                avg_loss_dev, acc_char, acc_word, f_score = self.evaluate()
+                iter_lap2 = datetime.now()
+                iter_elapsed = self._elapsed(iter_lap2 - iter_lap1)
+                iter_lap1 = iter_lap2
+                has_best |= self._check_iter(iter_elapsed, loss_trains, avg_loss_dev, acc_char,
+                                             acc_word, f_score, batch_size)
+            if self.cfg.batch_grow > 0 and self.cfg.iteration % self.cfg.batch_grow == 0:
+                batch_size = self.cfg.batch_size + self.cfg.iteration // self.cfg.batch_grow
+
+        print(file=sys.stderr)
+        sys.stderr.flush()
+        epoch_end = datetime.now()
+        epoch_elapsed = self._elapsed(epoch_end - epoch_begin)
+        logging.info('}}}} epoch: %d, end: %s, elapsed: %s, %s best }}}}', self.cfg.epoch,
+                     self._dt_str(epoch_begin), epoch_elapsed, 'hit' if has_best else 'did not hit')
+        self.cfg.epoch += 1
+        return has_best
+
+    def _check_iter(self, elapsed: str, loss_trains: List[float], avg_loss_dev: float,
+                    acc_char: float, acc_word: float, f_score: float, batch_size: int) -> bool:
+        """
+        cfg.check_iter번의 iteration마다 수행하는 체크
+        Args:
+            elapsed:  경과 시간
+            loss_trains:   train 코퍼스에서 각 배치별 loss 리스트
+            avg_loss_dev:  dev 코퍼스 문장 별 평균 loss
+            acc_char:  음절 정확도
+            acc_word:  어절 정확도
+            f_score:  f-score
+            batch_size:  batch size
+        Returns:
+            현재 iteration이 best 성능을 나타냈는 지 여부
+        """
+        avg_loss_train = sum(loss_trains) / len(loss_trains)
+        loss_trains.clear()
+        self.loss_trains.append(avg_loss_train)
+        self.loss_devs.append(avg_loss_dev)
+        self.acc_chars.append(acc_char)
+        self.acc_words.append(acc_word)
+        self.f_scores.append(f_score)
+        self.learning_rates.append(self.cfg.learning_rate)
+        self.batch_sizes.append(batch_size)
+        check_num = self.cfg.iteration // self.cfg.check_iter
+        logging.info('---- [%d|%d] elapsed: %s, loss(train/dev): %f / %f, acc(char/word): %f / %f' \
+                     ', f-score: %f (max: %f), lr: %f, bs: %d ----', self.cfg.epoch, check_num,
+                     elapsed, avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score,
+                     max(self.f_scores), self.cfg.learning_rate, batch_size)
+        print('{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}'.format( \
+              self.cfg.epoch, check_num, avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score,
+              self.cfg.learning_rate, batch_size), file=self.log_file)
+        self.log_file.flush()
+        self.sum_wrt.add_scalar('loss-train', avg_loss_train, check_num)
+        self.sum_wrt.add_scalar('loss-dev', avg_loss_dev, check_num)
+        self.sum_wrt.add_scalar('acc-char', acc_char, check_num)
+        self.sum_wrt.add_scalar('acc-word', acc_word, check_num)
+        self.sum_wrt.add_scalar('f-score', f_score, check_num)
+        self.sum_wrt.add_scalar('learning-rate', self.cfg.learning_rate, check_num)
+        self.sum_wrt.add_scalar('batch-size', batch_size, check_num)
+        return self._check_best()
+
+    def _check_best(self) -> bool:
+        """
+        이번 iteration에 가장 좋은 성능을 냈는 지 확인하고 그럴 경우 현재 상태를 저장한다.
+        Returns:
+            best 여부
+        """
+        if len(self.f_scores) > 1 and max(self.f_scores[:-1]) >= self.f_scores[-1]:
+            return False
+        # this iteration hits new max value
+        logging.info('==== best model: %f ====', self.f_scores[-1])
+        self.cfg.iter_best = self.cfg.iteration
+        self.model.save('{}/model.state'.format(self.cfg.out_dir))
+        self._save_optim('{}/optim.state'.format(self.cfg.out_dir))
+        with open('{}/config.json'.format(self.cfg.out_dir), 'w', encoding='UTF-8') as fout:
+            json.dump(vars(self.cfg), fout, indent=2, sort_keys=True)
+        return True
+
+    def _save_optim(self, path: str):
+        """
+        save optimizer parameters
+        Args:
+            path:  path
+        """
+        torch.save(self.optimizer.state_dict(), path)
+
+    def _load_optim(self, path: str, learning_rate: float):
+        """
+        load optimizer parameters
+        Args:
+            path:  path
+            learning_rate:  learning rate
+        """
+        if torch.cuda.is_available():
+            state_dict = torch.load(path)
+        else:
+            state_dict = torch.load(path, map_location=lambda storage, loc: storage)
+        self.optimizer = torch.optim.Adam(self.model.parameters(), learning_rate)
+        self.optimizer.load_state_dict(state_dict)
+        self.optimizer.param_groups[0]['lr'] = learning_rate
+
+    def evaluate(self) -> Tuple[float, float, float, float]:
+        """
+        evaluate f-score
+        Returns:
+            average dev loss
+            character accuracy
+            word accuracy
+            f-score
+        """
+        self.model.eval()
+        loss_devs = []
+        for dev_sent in self.dataset_dev:
+            # 만약 spc_dropout이 1.0 이상이면 공백을 전혀 쓰지 않는 것이므로 평가 시에도 적용한다.
+            dev_labels, dev_contexts = dev_sent.to_tensor(self.cfg, self.rsc,
+                                                          self.cfg.spc_dropout >= 1.0)
+            if torch.cuda.is_available():
+                dev_labels = dev_labels.cuda()
+                dev_contexts = dev_contexts.cuda()
+            dev_outputs = self.model(dev_contexts)
+            loss_dev = self.criterion(dev_outputs, dev_labels)
+            loss_devs.append(loss_dev.item())
+            _, predicts = F.softmax(dev_outputs, dim=1).max(1)
+            pred_tags = [self.rsc.vocab_out[t.item()] for t in predicts]
+            pred_sent = copy.deepcopy(dev_sent)
+            pred_sent.set_pos_result(pred_tags, self.rsc.restore_dic)
+            self.evaler.count(dev_sent, pred_sent)
+        avg_loss_dev = sum(loss_devs) / len(loss_devs)
+        return (avg_loss_dev, ) + self.evaler.evaluate()
diff --git a/train/bin/train.py b/train/bin/train.py
new file mode 100755
index 0000000..fcd7998
--- /dev/null
+++ b/train/bin/train.py
@@ -0,0 +1,87 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+train part-of-speech model from data set
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+import logging
+import os
+
+from khaiii.train.trainer import Trainer
+
+
+#############
+# functions #
+#############
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments (config)
+    """
+    trner = Trainer(args)
+    trner.train()
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='train model from data')
+    parser.add_argument('-i', '--in-pfx', help='input data prefix', metavar='NAME', required=True)
+    parser.add_argument('--rsc-src', help='train resource dir <default: ../rsc/src>', metavar='DIR',
+                        default='../rsc/src')
+    parser.add_argument('--log-dir', help='tensorboard log dir', metavar='DIR',
+                        default='./logdir')
+    parser.add_argument('--window', help='left/right character window length', metavar='INT',
+                        type=int,
+                        default=3)
+    parser.add_argument('--spc-dropout', help='space(word delimiter) dropout rate', metavar='REAL',
+                        type=float,
+                        default=0.0)
+    parser.add_argument('--cutoff', help='cutoff', metavar='INT', type=int,
+                        default=2)
+    parser.add_argument('--embed-dim', help='embedding dimension', metavar='INT', type=int,
+                        default=30)
+    parser.add_argument('--learning-rate', help='learning rate', metavar='REAL', type=float,
+                        default=0.001)
+    parser.add_argument('--lr-decay', help='learning rate decay', metavar='REAL', type=float,
+                        default=0.9)
+    parser.add_argument('--batch-size', help='batch size', metavar='INT', type=int,
+                        default=500)
+    parser.add_argument('--check-iter', help='check iteration', metavar='INT', type=int,
+                        default=10000)
+    parser.add_argument('--batch-grow', help='batch growing iteration. 0 means no growing',
+                        type=int,
+                        default=10000)
+    parser.add_argument('--patience', help='maximum patience count to revert model', metavar='INT',
+                        type=int,
+                        default=10)
+    parser.add_argument('--gpu-num', help='GPU number to use', metavar='INT', type=int,
+                        default=0)
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_num)
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From f9d97d3d4bc7c5e5cb0da83acc490837d5a64c0f Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Wed, 23 Jan 2019 22:13:31 +0900
Subject: [PATCH 08/41] =?UTF-8?q?batch=20grow=20=ED=8C=8C=EB=9D=BC?=
 =?UTF-8?q?=EB=AF=B8=ED=84=B0,=20epoch=20=EA=B0=9C=EB=85=90=EC=9D=84=20?=
 =?UTF-8?q?=EC=97=86=EC=95=A0=EA=B3=A0=20=ED=95=99=EC=8A=B5=EA=B3=BC?=
 =?UTF-8?q?=EC=A0=95=EC=9D=84=20=EB=8B=A8=EC=88=9C=ED=99=94=ED=95=A8=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/main/python/khaiii/train/trainer.py | 219 ++++++++++++------------
 train/bin/train.py                      |  48 +++---
 train/requirements.txt                  |   3 +
 3 files changed, 131 insertions(+), 139 deletions(-)
 create mode 100644 train/requirements.txt

diff --git a/src/main/python/khaiii/train/trainer.py b/src/main/python/khaiii/train/trainer.py
index 140d558..e299fc1 100644
--- a/src/main/python/khaiii/train/trainer.py
+++ b/src/main/python/khaiii/train/trainer.py
@@ -19,8 +19,7 @@
 import os
 import pathlib
 import pprint
-import sys
-from typing import List, Tuple
+from typing import Iterator, List, Tuple
 
 from tensorboardX import SummaryWriter
 import torch
@@ -28,7 +27,7 @@
 import torch.nn.functional as F
 from tqdm import tqdm
 
-from khaiii.train.dataset import PosDataset
+from khaiii.train.dataset import PosDataset, PosSentTensor
 from khaiii.train.evaluator import Evaluator
 from khaiii.train.models import CnnModel
 from khaiii.resource.resource import Resource
@@ -48,7 +47,7 @@ def __init__(self, cfg: Namespace):
         """
         self.cfg = cfg
         setattr(cfg, 'model_id', self.model_id(cfg))
-        setattr(cfg, 'out_dir', '{}/{}'.format(cfg.log_dir, cfg.model_id))
+        setattr(cfg, 'out_dir', '{}/{}'.format(cfg.logdir, cfg.model_id))
         setattr(cfg, 'context_len', 2 * cfg.window + 1)
         self.rsc = Resource(cfg)
         self.model = CnnModel(cfg, self.rsc)
@@ -56,10 +55,9 @@ def __init__(self, cfg: Namespace):
         self.criterion = nn.CrossEntropyLoss()
         self.evaler = Evaluator()
         self._load_dataset()
-        if 'epoch' not in cfg.__dict__:
-            setattr(cfg, 'epoch', 0)
-            setattr(cfg, 'iteration', 0)
-            setattr(cfg, 'iter_best', 0)
+        if 'step' not in cfg.__dict__:
+            setattr(cfg, 'step', 0)
+            setattr(cfg, 'best_step', 0)
         self.log_file = None    # tab separated log file
         self.sum_wrt = None    # tensorboard summary writer
         self.loss_trains = []
@@ -68,7 +66,6 @@ def __init__(self, cfg: Namespace):
         self.acc_words = []
         self.f_scores = []
         self.learning_rates = []
-        self.batch_sizes = []
 
     @classmethod
     def model_id(cls, cfg: Namespace) -> str:
@@ -88,8 +85,7 @@ def model_id(cls, cfg: Namespace) -> str:
             'lr{}'.format(cfg.learning_rate),
             'lrd{}'.format(cfg.lr_decay),
             'bs{}'.format(cfg.batch_size),
-            'ci{}'.format(cfg.check_iter),
-            'bg{}'.format(cfg.batch_grow),
+            'cs{}'.format(cfg.check_step),
         ]
         return '.'.join(model_cfgs)
 
@@ -133,18 +129,19 @@ def _elapsed(cls, td_obj: timedelta) -> str:
         seconds -= minutes * 60
         return '{}:{:02d}:{:02d}'.format(hours, minutes, seconds)
 
-    def _check_continue(self):
+    def _restore_prev_train(self):
         """
         기존에 학습하다 중지한 경우 그 이후부터 계속해서 학습할 수 있도록 이전 상태를 복원한다.
         """
         out_path = pathlib.Path(self.cfg.out_dir)
-        if not out_path.is_dir():
+        cfg_path = pathlib.Path('{}/config.json'.format(self.cfg.out_dir))
+        if not out_path.is_dir() or not cfg_path.is_file():
             return
-        logging.info('==== continue learning: %s ====', self.cfg.model_id)
-        cfg = json.load(open('{}/config.json'.format(self.cfg.out_dir), 'r', encoding='UTF-8'))
+        logging.info('==== continue training: %s ====', self.cfg.model_id)
+        cfg = json.load(open(cfg_path, 'r', encoding='UTF-8'))
         for key, val in cfg.items():
             setattr(self.cfg, key, val)
-        self._revert(False)
+        self._revert_to_best(False)
 
         f_score_best = 0.0
         best_idx = -1
@@ -152,32 +149,43 @@ def _check_continue(self):
             line = line.rstrip('\r\n')
             if not line:
                 continue
-            (epoch, iteration, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate,
-             batch_size) = line.split('\t')
-            self.cfg.epoch = int(epoch) + 1
-            self.cfg.iteration = self.cfg.iter_best = int(iteration) * self.cfg.check_iter
+            (step, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate) \
+                    = line.split('\t')
+            self.cfg.step = self.cfg.best_step = int(step) * self.cfg.check_step
             self.loss_trains.append(float(loss_train))
             self.loss_devs.append(float(loss_dev))
             self.acc_chars.append(float(acc_char))
             self.acc_words.append(float(acc_word))
             self.f_scores.append(float(f_score))
             self.learning_rates.append(float(learning_rate))
-            self.batch_sizes.append(int(batch_size))
             if float(f_score) > f_score_best:
                 f_score_best = float(f_score)
                 best_idx = idx
-        logging.info('---- [%d|%d] loss(train/dev): %f / %f, acc(char/word): %f / %f, ' \
-                     'f-score: %f, lr: %f, bs: %d ----', self.cfg.epoch,
-                     self.cfg.iteration // self.cfg.check_iter, self.loss_trains[best_idx],
-                     self.loss_devs[best_idx], self.acc_chars[best_idx], self.acc_words[best_idx],
-                     self.f_scores[best_idx], self.learning_rates[-1], self.batch_sizes[-1])
+        logging.info('---- [%d] los(trn/dev): %.4f / %.4f, acc(chr/wrd): %.4f / %.4f, ' \
+                     'f-score: %.4f, lr: %.8f ----', self.cfg.step // self.cfg.check_step,
+                     self.loss_trains[best_idx], self.loss_devs[best_idx], self.acc_chars[best_idx],
+                     self.acc_words[best_idx], self.f_scores[best_idx], self.learning_rates[-1])
+
+    @classmethod
+    def _inf_data_iterator(cls, dataset: PosDataset) -> Iterator[PosSentTensor]:
+        """
+        데이터셋을 무한히 반복하여 문장을 출력하는 제너레이터
+        Args:
+            dataset:  데이터셋
+        Yields:
+            PosSentTensor 객체
+        """
+        for _ in range(1000000):
+            for sent in dataset:
+                yield sent
 
     def train(self):
         """
         train model with dataset
         """
-        self._check_continue()
+        self._restore_prev_train()
         logging.info('config: %s', pprint.pformat(self.cfg.__dict__))
+
         train_begin = datetime.now()
         logging.info('{{{{ training begin: %s {{{{', self._dt_str(train_begin))
         if torch.cuda.is_available():
@@ -185,26 +193,27 @@ def train(self):
         pathlib.Path(self.cfg.out_dir).mkdir(parents=True, exist_ok=True)
         self.log_file = open('{}/log.tsv'.format(self.cfg.out_dir), 'at')
         self.sum_wrt = SummaryWriter(self.cfg.out_dir)
+        check_start = (1 if self.cfg.step == 0 else (self.cfg.step // self.cfg.check_step + 1))
         patience = self.cfg.patience
-        for _ in range(10000):
-            has_best = self.train_epoch()
-            if has_best:
+        train_iter = self._inf_data_iterator(self.dataset_train)
+        for check_id in range(check_start, 1000000):
+            is_best = self._train_and_check(check_id, train_iter)
+            if is_best:
                 patience = self.cfg.patience
-            else:
-                if patience > 0:
-                    self._revert(True)
-                    patience -= 1
-                    logging.info('==== revert to iter: %d, f-score: %f, patience: %d ====',
-                                 self.cfg.iter_best, max(self.f_scores), patience)
-                else:
-                    break
+                continue
+            if patience <= 0:
+                break
+            self._revert_to_best(True)
+            patience -= 1
+            tqdm.write('==== revert to check: {}, f-score: {:.4f}, patience: {} ===='.format( \
+                    self.cfg.best_step // self.cfg.check_step, max(self.f_scores), patience))
+
         train_end = datetime.now()
         train_elapsed = self._elapsed(train_end - train_begin)
-        logging.info('}}}} training end: %s, elapsed: %s, epoch: %s, iter: %dk }}}}',
-                     self._dt_str(train_end), train_elapsed, self.cfg.epoch,
-                     self.cfg.iteration // 1000)
+        logging.info('}}}} training end: %s, elapsed: %s, step: %dk }}}}',
+                     self._dt_str(train_end), train_elapsed, self.cfg.step // 1000)
 
-    def _revert(self, is_decay_lr: bool):
+    def _revert_to_best(self, is_decay_lr: bool):
         """
         이전 best 모델로 되돌린다.
         Args:
@@ -215,79 +224,67 @@ def _revert(self, is_decay_lr: bool):
             self.cfg.learning_rate *= self.cfg.lr_decay
         self._load_optim('{}/optim.state'.format(self.cfg.out_dir), self.cfg.learning_rate)
 
-    def train_epoch(self) -> bool:
+    def _train_and_check(self, check_id: int, train_iter: Iterator[PosSentTensor]) -> bool:
         """
-        한 epoch을 학습한다. 배치 단위는 글자 단위
+        cfg.check_step 만큼의 step을 수행하고 evaluation을 수행한다.
+        Args:
+            check_id:  check ID
+            train_iter:  학습 데이터 iterator
         Returns:
-            best f-score를 기록한 iteration 여부
+            best f-score를 기록한 step 여부
         """
-        epoch_begin = datetime.now()
-        logging.info('{{{{ epoch: %d, begin: %s {{{{', self.cfg.epoch, self._dt_str(epoch_begin))
-        iter_lap1 = datetime.now()
-        has_best = False
-        batches = []
+        start_step = self.cfg.step
+        loss_batch = torch.tensor(0.0)    # pylint: disable=not-callable
+        batch_size = 0
         loss_trains = []
-        batch_size = self.cfg.batch_size
-        if self.cfg.batch_grow > 0:
-            batch_size = self.cfg.batch_size + self.cfg.iteration // self.cfg.batch_grow
-        for train_sent in tqdm(self.dataset_train, 'EPOCH [{}]'.format(self.cfg.epoch),
-                               len(self.dataset_train), mininterval=1, ncols=100):
+        train_sents = tqdm(train_iter, '[{}]'.format(check_id), mininterval=1, ncols=100)
+        for train_sent in train_sents:
             train_labels, train_contexts = train_sent.to_tensor(self.cfg, self.rsc, True)
             if torch.cuda.is_available():
                 train_labels = train_labels.cuda()
                 train_contexts = train_contexts.cuda()
+                loss_batch = loss_batch.cuda()
 
-            self.optimizer.zero_grad()
             self.model.train()
             train_outputs = self.model(train_contexts)
-            batches.append((train_labels, train_outputs))
-            if sum([batch[0].size(0) for batch in batches]) < batch_size:
+            train_outputs.requires_grad_()
+            loss_train = self.criterion(train_outputs, train_labels)
+            loss_train.backward()
+            loss_trains.append(loss_train.item())
+            loss_batch += loss_train
+            batch_size += len(train_labels)
+            if batch_size < self.cfg.batch_size:
                 continue
-            batch_label = torch.cat([x[0] for x in batches], 0)    # pylint: disable=no-member
-            batch_output = torch.cat([x[1] for x in batches], 0)    # pylint: disable=no-member
-            batches = []
 
-            batch_output.requires_grad_()
-            loss_train = self.criterion(batch_output, batch_label)
-            loss_trains.append(loss_train.item())
-            loss_train.backward()
             self.optimizer.step()
-            self.cfg.iteration += 1
+            self.optimizer.zero_grad()
+            self.sum_wrt.add_scalar('loss-batch', loss_batch.item(), self.cfg.step)
+            self.cfg.step += 1
+            loss_batch = torch.tensor(0.0)    # pylint: disable=not-callable
+            batch_size = 0
 
-            if self.cfg.iteration % self.cfg.check_iter == 0:
-                avg_loss_dev, acc_char, acc_word, f_score = self.evaluate()
-                iter_lap2 = datetime.now()
-                iter_elapsed = self._elapsed(iter_lap2 - iter_lap1)
-                iter_lap1 = iter_lap2
-                has_best |= self._check_iter(iter_elapsed, loss_trains, avg_loss_dev, acc_char,
-                                             acc_word, f_score, batch_size)
-            if self.cfg.batch_grow > 0 and self.cfg.iteration % self.cfg.batch_grow == 0:
-                batch_size = self.cfg.batch_size + self.cfg.iteration // self.cfg.batch_grow
+            if (self.cfg.step - start_step) >= self.cfg.check_step:
+                train_sents.close()
+                break
 
-        print(file=sys.stderr)
-        sys.stderr.flush()
-        epoch_end = datetime.now()
-        epoch_elapsed = self._elapsed(epoch_end - epoch_begin)
-        logging.info('}}}} epoch: %d, end: %s, elapsed: %s, %s best }}}}', self.cfg.epoch,
-                     self._dt_str(epoch_begin), epoch_elapsed, 'hit' if has_best else 'did not hit')
-        self.cfg.epoch += 1
-        return has_best
+        avg_loss_dev, acc_char, acc_word, f_score = self.evaluate()
+        return self._check(check_id, loss_trains, avg_loss_dev, acc_char, acc_word, f_score)
 
-    def _check_iter(self, elapsed: str, loss_trains: List[float], avg_loss_dev: float,
-                    acc_char: float, acc_word: float, f_score: float, batch_size: int) -> bool:
+    def _check(self, check_id: int, loss_trains: List[float], avg_loss_dev: float, acc_char: float,
+               acc_word: float, f_score: float) -> bool:
         """
-        cfg.check_iter번의 iteration마다 수행하는 체크
+        cfg.check_step번의 step마다 수행하는 체크
         Args:
-            elapsed:  경과 시간
-            loss_trains:   train 코퍼스에서 각 배치별 loss 리스트
+            check_id:  check ID
+            loss_trains:  train 코퍼스에서 각 배치별 loss 리스트
             avg_loss_dev:  dev 코퍼스 문장 별 평균 loss
             acc_char:  음절 정확도
             acc_word:  어절 정확도
             f_score:  f-score
-            batch_size:  batch size
         Returns:
-            현재 iteration이 best 성능을 나타냈는 지 여부
+            현재 step이 best 성능을 나타냈는 지 여부
         """
+        assert check_id == self.cfg.step // self.cfg.check_step
         avg_loss_train = sum(loss_trains) / len(loss_trains)
         loss_trains.clear()
         self.loss_trains.append(avg_loss_train)
@@ -296,36 +293,34 @@ def _check_iter(self, elapsed: str, loss_trains: List[float], avg_loss_dev: floa
         self.acc_words.append(acc_word)
         self.f_scores.append(f_score)
         self.learning_rates.append(self.cfg.learning_rate)
-        self.batch_sizes.append(batch_size)
-        check_num = self.cfg.iteration // self.cfg.check_iter
-        logging.info('---- [%d|%d] elapsed: %s, loss(train/dev): %f / %f, acc(char/word): %f / %f' \
-                     ', f-score: %f (max: %f), lr: %f, bs: %d ----', self.cfg.epoch, check_num,
-                     elapsed, avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score,
-                     max(self.f_scores), self.cfg.learning_rate, batch_size)
-        print('{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}'.format( \
-              self.cfg.epoch, check_num, avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score,
-              self.cfg.learning_rate, batch_size), file=self.log_file)
+        is_best = self._is_best()
+        is_best_str = 'BEST' if is_best else '< {:.4f}'.format(max(self.f_scores))
+        tqdm.write('    [Los trn]  [Los dev]  [Acc chr]  [Acc wrd]  [F-score]           [LR]')
+        tqdm.write('    {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f} {:8}  {:.8f}'.format( \
+                avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score, is_best_str,
+                self.cfg.learning_rate))
+        print('{}\t{}\t{}\t{}\t{}\t{}\t{}'.format(check_id, avg_loss_train, avg_loss_dev,
+                                                  acc_char, acc_word, f_score,
+                                                  self.cfg.learning_rate), file=self.log_file)
         self.log_file.flush()
-        self.sum_wrt.add_scalar('loss-train', avg_loss_train, check_num)
-        self.sum_wrt.add_scalar('loss-dev', avg_loss_dev, check_num)
-        self.sum_wrt.add_scalar('acc-char', acc_char, check_num)
-        self.sum_wrt.add_scalar('acc-word', acc_word, check_num)
-        self.sum_wrt.add_scalar('f-score', f_score, check_num)
-        self.sum_wrt.add_scalar('learning-rate', self.cfg.learning_rate, check_num)
-        self.sum_wrt.add_scalar('batch-size', batch_size, check_num)
-        return self._check_best()
+        self.sum_wrt.add_scalar('loss-train', avg_loss_train, check_id)
+        self.sum_wrt.add_scalar('loss-dev', avg_loss_dev, check_id)
+        self.sum_wrt.add_scalar('acc-char', acc_char, check_id)
+        self.sum_wrt.add_scalar('acc-word', acc_word, check_id)
+        self.sum_wrt.add_scalar('f-score', f_score, check_id)
+        self.sum_wrt.add_scalar('learning-rate', self.cfg.learning_rate, check_id)
+        return is_best
 
-    def _check_best(self) -> bool:
+    def _is_best(self) -> bool:
         """
-        이번 iteration에 가장 좋은 성능을 냈는 지 확인하고 그럴 경우 현재 상태를 저장한다.
+        이번 step에 가장 좋은 성능을 냈는 지 확인하고 그럴 경우 현재 상태를 저장한다.
         Returns:
             best 여부
         """
         if len(self.f_scores) > 1 and max(self.f_scores[:-1]) >= self.f_scores[-1]:
             return False
-        # this iteration hits new max value
-        logging.info('==== best model: %f ====', self.f_scores[-1])
-        self.cfg.iter_best = self.cfg.iteration
+        # this step hits new max value
+        self.cfg.best_step = self.cfg.step
         self.model.save('{}/model.state'.format(self.cfg.out_dir))
         self._save_optim('{}/optim.state'.format(self.cfg.out_dir))
         with open('{}/config.json'.format(self.cfg.out_dir), 'w', encoding='UTF-8') as fout:
diff --git a/train/bin/train.py b/train/bin/train.py
index fcd7998..b696d96 100755
--- a/train/bin/train.py
+++ b/train/bin/train.py
@@ -40,36 +40,30 @@ def main():
     main function processes only argument parsing
     """
     parser = ArgumentParser(description='train model from data')
-    parser.add_argument('-i', '--in-pfx', help='input data prefix', metavar='NAME', required=True)
+    parser.add_argument('-i', '--in-pfx', help='input data path prefix', metavar='NAME',
+                        required=True)
     parser.add_argument('--rsc-src', help='train resource dir <default: ../rsc/src>', metavar='DIR',
                         default='../rsc/src')
-    parser.add_argument('--log-dir', help='tensorboard log dir', metavar='DIR',
+    parser.add_argument('--logdir', help='tensorboard log dir <default: ./logdir>', metavar='DIR',
                         default='./logdir')
-    parser.add_argument('--window', help='left/right character window length', metavar='INT',
-                        type=int,
-                        default=3)
-    parser.add_argument('--spc-dropout', help='space(word delimiter) dropout rate', metavar='REAL',
-                        type=float,
-                        default=0.0)
-    parser.add_argument('--cutoff', help='cutoff', metavar='INT', type=int,
-                        default=2)
-    parser.add_argument('--embed-dim', help='embedding dimension', metavar='INT', type=int,
-                        default=30)
-    parser.add_argument('--learning-rate', help='learning rate', metavar='REAL', type=float,
-                        default=0.001)
-    parser.add_argument('--lr-decay', help='learning rate decay', metavar='REAL', type=float,
-                        default=0.9)
-    parser.add_argument('--batch-size', help='batch size', metavar='INT', type=int,
-                        default=500)
-    parser.add_argument('--check-iter', help='check iteration', metavar='INT', type=int,
-                        default=10000)
-    parser.add_argument('--batch-grow', help='batch growing iteration. 0 means no growing',
-                        type=int,
-                        default=10000)
-    parser.add_argument('--patience', help='maximum patience count to revert model', metavar='INT',
-                        type=int,
-                        default=10)
-    parser.add_argument('--gpu-num', help='GPU number to use', metavar='INT', type=int,
+    parser.add_argument('--window', help='left/right character window length <default: 3>',
+                        metavar='INT', type=int, default=3)
+    parser.add_argument('--spc-dropout', help='space(word delimiter) dropout rate <default: 0.0>',
+                        metavar='REAL', type=float, default=0.0)
+    parser.add_argument('--cutoff', help='cutoff <default: 2>', metavar='INT', type=int, default=2)
+    parser.add_argument('--embed-dim', help='embedding dimension <default: 30>', metavar='INT',
+                        type=int, default=30)
+    parser.add_argument('--learning-rate', help='learning rate <default: 0.001>', metavar='REAL',
+                        type=float, default=0.001)
+    parser.add_argument('--lr-decay', help='learning rate decay <default: 0.9>', metavar='REAL',
+                        type=float, default=0.9)
+    parser.add_argument('--batch-size', help='batch size <default: 1000>', metavar='INT', type=int,
+                        default=1000)
+    parser.add_argument('--check-step', help='check every N step <default: 10000>', metavar='INT',
+                        type=int, default=10000)
+    parser.add_argument('--patience', help='maximum patience count to revert model <default: 10>',
+                        metavar='INT', type=int, default=10)
+    parser.add_argument('--gpu-num', help='GPU number to use <default: 0>', metavar='INT', type=int,
                         default=0)
     parser.add_argument('--debug', help='enable debug', action='store_true')
     args = parser.parse_args()
diff --git a/train/requirements.txt b/train/requirements.txt
new file mode 100644
index 0000000..5627706
--- /dev/null
+++ b/train/requirements.txt
@@ -0,0 +1,3 @@
+tensorboardX
+torch==0.4.1
+tqdm

From cf8269b2cea3ca9b065b727fa9692b9061a55617 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Thu, 24 Jan 2019 14:34:44 +0900
Subject: [PATCH 09/41] =?UTF-8?q?=ED=95=99=EC=8A=B5=EC=9D=B4=20=EC=99=84?=
 =?UTF-8?q?=EB=A3=8C=EB=90=9C=20=EB=AA=A8=EB=8D=B8=20=ED=8C=8C=EC=9D=BC?=
 =?UTF-8?q?=EC=9D=84=20pickle=20=ED=98=95=ED=83=9C=EB=A1=9C=20=EC=A0=80?=
 =?UTF-8?q?=EC=9E=A5=ED=95=98=EB=8A=94=20=EC=8A=A4=ED=81=AC=EB=A6=BD?=
 =?UTF-8?q?=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 rsc/bin/compile_model.py  |   6 +-
 train/bin/pickle_model.py | 243 ++++++++++++++++++++++++++++++++++++++
 2 files changed, 246 insertions(+), 3 deletions(-)
 create mode 100755 train/bin/pickle_model.py

diff --git a/rsc/bin/compile_model.py b/rsc/bin/compile_model.py
index 1c10d1f..b06fff7 100755
--- a/rsc/bin/compile_model.py
+++ b/rsc/bin/compile_model.py
@@ -26,7 +26,7 @@
 #############
 # functions #
 #############
-def load_cfg_rsc(rsc_src: str, model_size: str) -> Tuple[Namespace, Resource]:
+def _load_cfg_rsc(rsc_src: str, model_size: str) -> Tuple[Namespace, Resource]:
     """
     load config and resource from source directory
     Args:
@@ -38,7 +38,7 @@ def load_cfg_rsc(rsc_src: str, model_size: str) -> Tuple[Namespace, Resource]:
     """
     file_path = '{}/{}.config.json'.format(rsc_src, model_size)
     cfg_dic = json.load(open(file_path, 'r', encoding='UTF-8'))
-    logging.info('config: %s', json.dumps(cfg_dic, indent=2))
+    logging.info('config: %s', json.dumps(cfg_dic, indent=4, sort_keys=True))
     cfg = Namespace()
     for key, val in cfg_dic.items():
         setattr(cfg, key, val)
@@ -145,7 +145,7 @@ def run(args: Namespace):
     Args:
         args:  program arguments
     """
-    cfg, rsc = load_cfg_rsc(args.rsc_src, args.model_size)
+    cfg, rsc = _load_cfg_rsc(args.rsc_src, args.model_size)
     data = pickle.load(open('{}/{}.model.pickle'.format(args.rsc_src, args.model_size), 'rb'))
     _write_config(cfg, rsc, args.rsc_dir)
     _write_data(data, args.rsc_dir)
diff --git a/train/bin/pickle_model.py b/train/bin/pickle_model.py
new file mode 100755
index 0000000..0eafbb1
--- /dev/null
+++ b/train/bin/pickle_model.py
@@ -0,0 +1,243 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+pickle trained model (state dict)
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+from array import array
+import json
+import logging
+import os
+import pickle
+import re
+
+import torch
+
+from khaiii.resource.resource import Resource
+
+
+#############
+# functions #
+#############
+def _validate_state_dict(cfg: Namespace, rsc: Resource, state_dict: dict):
+    """
+    validate state dict with config and resource
+    Args:
+        cfg:  config
+        rsc:  resource
+        state_dict:  state dict
+    """
+    def _assert(expected, actual, msg: str):
+        """
+        assert expected equals actual. if not raise exception
+        Args:
+            expected:  expected value
+            actual:  actual value
+            msg:  exception message if both are not equal
+        """
+        if expected != actual:
+            raise ValueError((msg + ': expected: {}, but actual: {}').format(expected, actual))
+
+    for name, tensor in state_dict.items():
+        logging.info('%s: %s', name, tensor.size())
+        if name == 'embedder.embedding.weight':
+            _assert(len(rsc.vocab_in), tensor.size(0), 'invalid input vocab')
+            _assert(cfg.embed_dim, tensor.size(1), 'invalid embedding dim')
+        if re.match(r'convs\..\.weight', name):
+            _assert(cfg.embed_dim, tensor.size(1), 'invalid conv input channel dim')
+        if name == 'hidden2tag.bias':
+            _assert(len(rsc.vocab_out), tensor.size(0), 'invalid output bias dim')
+
+
+def _get_embedding(rsc: Resource, state_dict: dict) -> dict:
+    """
+    get character embedding
+    Args:
+        rsc:  resource
+        state_dict:  state dictionary of model
+    Returns:
+        embedding data
+    """
+    data = {}
+    chars = [0, ] * 5    # 5 special characters
+    for idx in range(5, len(rsc.vocab_in)):
+        chars.append(ord(rsc.vocab_in[idx]))
+    data['chars'] = array('i', chars)    # [input vocab(char)] * 4(wchar_t)
+
+    embedding = state_dict['embedder.embedding.weight']
+    data['weights'] = []
+    for row in embedding:
+        data['weights'].append(array('f', row))    # [input vocab(char)] * embed_dim * 4(float)
+    return data
+
+
+def _get_linear(name_pfx: str, state_dict: dict) -> dict:
+    """
+    write weight and bias with given layer name prefix
+    Args:
+        name_pfx:  layer name prefix
+        state_dict:  state dictionary of model
+    Returns:
+        linear layer data
+    """
+    data = {}
+    data['weight'] = []
+    weight = state_dict['{}.weight'.format(name_pfx)]
+    for row in weight:
+        data['weight'].append(array('f', row))    # output * input * 4
+    bias_name = '{}.bias'.format(name_pfx)
+    if bias_name in state_dict:
+        bias = state_dict[bias_name]
+        data['bias'] = array('f', bias)    # output * 4
+    return data
+
+
+def _get_conv(name_pfx: str, kernel_size: int, state_dict: dict) -> dict:
+    """
+    get convolution module weight and bias with given layer name prefix and suffix(index)
+    Args:
+        name_pfx:  layer name prefix
+        kernel_size:  kernel size
+        state_dict:  state dictionary of model
+    Returns:
+        conv data with given kernel size
+    """
+    data = {}
+    weight = state_dict['{}.{}.weight'.format(name_pfx, kernel_size - 2)]
+    weight_t = weight.transpose(1, 2)    # [output chan] * [input chan] * kernel
+    #                                      => [output chan] * kernel * [input chan]
+    data['channel'] = []
+    for ch_out in weight_t:
+        data['channel'].append(array('f', ch_out.contiguous().view(-1)))
+    bias_name = '{}.{}.bias'.format(name_pfx, kernel_size - 2)
+    if bias_name in state_dict:
+        bias = state_dict[bias_name]
+        data['bias'] = array('f', bias)    # [output chan] * 4
+    return data
+
+
+def _get_data(rsc: Resource, state_dict: dict) -> dict:
+    """
+    get all data to pickle
+    Args:
+        rsc:  resource
+        state_dict:  state dictionary of model
+    Returns:
+        all data
+    """
+    data = {}
+    # key: [input vocab(char)] * 4(float)
+    # val: [input vocab(char)] * embed_dim * 4(float)
+    data['embedding'] = _get_embedding(rsc, state_dict)
+
+    data['convs'] = {}
+    for kernel in range(2, 6):
+        # weight: [output chan(embed_dim)] * kernel * [input chan(embed_dim)] * 4
+        # bias: [output chan] * 4
+        data['convs'][kernel] = _get_conv('convs', kernel, state_dict)
+
+    # weight: hidden_dim * [cnn layers * output chan(embed_dim)] * 4
+    # bias: hidden_dim * 4
+    data['conv2hidden'] = _get_linear('conv2hidden', state_dict)
+
+    # weight: [output vocab(tag)] * hidden_dim * 4
+    # bias: [output vocab(tag)] * 4
+    data['hidden2tag'] = _get_linear('hidden2tag', state_dict)
+    return data
+
+
+def _load_config(path: str) -> Namespace:
+    """
+    load config file
+    Args:
+        path:  path
+    Returns:
+        config
+    """
+    cfg_dic = json.load(open(path, 'r', encoding='UTF-8'))
+    logging.info('config: %s', json.dumps(cfg_dic, indent=4, sort_keys=True))
+    cfg = Namespace()
+    for key, val in cfg_dic.items():
+        if key not in {'cutoff', 'embed_dim', 'hidden_dim', 'model_id', 'model_name', 'rsc_src',
+                       'window'}:
+            continue
+        setattr(cfg, key, val)
+    return cfg
+
+
+def _load_resource(cfg: Namespace, rsc_src: str) -> Resource:
+    """
+    load resources
+    Args:
+        cfg:  config
+        rsc_src:  resource source dir
+    Returns:
+        Resource object
+    """
+    cwd = os.path.realpath(os.getcwd())
+    train_dir = os.path.realpath('{}/..'.format(rsc_src))
+    if cwd != train_dir:
+        os.chdir(train_dir)
+    rsc = Resource(cfg)
+    if cwd != train_dir:
+        os.chdir(cwd)
+    return rsc
+
+
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    cfg = _load_config('{}/config.json'.format(args.in_dir))
+    rsc = _load_resource(cfg, args.rsc_src)
+    state_dict = torch.load('{}/model.state'.format(args.in_dir),
+                            map_location=lambda storage, loc: storage)
+    _validate_state_dict(cfg, rsc, state_dict)
+    data = _get_data(rsc, state_dict)
+
+    config_path = '{}/{}.config.json'.format(args.rsc_src, args.model_size)
+    with open(config_path, 'w', encoding='UTF-8') as fout:
+        json.dump(vars(cfg), fout, indent=4, sort_keys=True)
+
+    pickle_path = '{}/{}.model.pickle'.format(args.rsc_src, args.model_size)
+    with open(pickle_path, 'wb') as fout:
+        pickle.dump(data, fout)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='pickle trained model (state dict)')
+    parser.add_argument('-i', '--in-dir', help='model dir', metavar='DIR', required=True)
+    parser.add_argument('--model-size', help='model size <default: base>',
+                        metavar='SIZE', default='base')
+    parser.add_argument('--rsc-src', help='resource source dir <default: ../rsc/src>',
+                        metavar='DIR', default='../rsc/src')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From 4f33102305e3799962572750a2060714f66cbcf6 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Fri, 25 Jan 2019 00:28:15 +0900
Subject: [PATCH 10/41] =?UTF-8?q?=ED=95=99=EC=8A=B5=EB=90=9C=20=EB=AA=A8?=
 =?UTF-8?q?=EB=8D=B8=EC=9D=84=20=EB=AA=85=EB=A0=B9=EC=B0=BD=EC=97=90?=
 =?UTF-8?q?=EC=84=9C=20=EC=8B=A4=ED=96=89=ED=95=B4=EB=B3=BC=20=EC=88=98=20?=
 =?UTF-8?q?=EC=9E=88=EB=8A=94=20=EB=8D=B0=EB=AA=A8=20=EC=8A=A4=ED=81=AC?=
 =?UTF-8?q?=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 rsc/bin/compile_model.py               |  2 +-
 src/main/python/khaiii/train/tagger.py |  9 ++-
 train/bin/pickle_model.py              |  2 +-
 train/bin/tag.py                       | 76 ++++++++++++++++++++++++++
 4 files changed, 82 insertions(+), 7 deletions(-)
 create mode 100755 train/bin/tag.py

diff --git a/rsc/bin/compile_model.py b/rsc/bin/compile_model.py
index b06fff7..767060c 100755
--- a/rsc/bin/compile_model.py
+++ b/rsc/bin/compile_model.py
@@ -43,7 +43,7 @@ def _load_cfg_rsc(rsc_src: str, model_size: str) -> Tuple[Namespace, Resource]:
     for key, val in cfg_dic.items():
         setattr(cfg, key, val)
     cwd = os.path.realpath(os.getcwd())
-    train_dir = os.path.realpath('{}/..'.format(rsc_src))
+    train_dir = os.path.realpath('{}/../../train'.format(rsc_src))
     if cwd != train_dir:
         os.chdir(train_dir)
     rsc = Resource(cfg)
diff --git a/src/main/python/khaiii/train/tagger.py b/src/main/python/khaiii/train/tagger.py
index 06fa0e2..c0997d8 100644
--- a/src/main/python/khaiii/train/tagger.py
+++ b/src/main/python/khaiii/train/tagger.py
@@ -31,19 +31,18 @@ class PosTagger:
     """
     part-of-speech tagger
     """
-    def __init__(self, rsc_src: str):
+    def __init__(self, model_dir: str):
         """
         Args:
-            rsc_src:  resource dir
+            model_dir:  model dir
         """
-        cfg_dict = json.load(open('{}/config.json'.format(rsc_src), 'r', encoding='UTF-8'))
+        cfg_dict = json.load(open('{}/config.json'.format(model_dir), 'r', encoding='UTF-8'))
         self.cfg = Namespace()
         for key, val in cfg_dict.items():
             setattr(self.cfg, key, val)
-        self.cfg.rsc_src = rsc_src
         self.rsc = Resource(self.cfg)
         self.model = CnnModel(self.cfg, self.rsc)
-        self.model.load('{}/model.state'.format(rsc_src))
+        self.model.load('{}/model.state'.format(model_dir))
         self.model.eval()
 
     def tag_raw(self, raw_sent: str, enable_restore: bool = True) -> PosSentTensor:
diff --git a/train/bin/pickle_model.py b/train/bin/pickle_model.py
index 0eafbb1..06f6620 100755
--- a/train/bin/pickle_model.py
+++ b/train/bin/pickle_model.py
@@ -184,7 +184,7 @@ def _load_resource(cfg: Namespace, rsc_src: str) -> Resource:
         Resource object
     """
     cwd = os.path.realpath(os.getcwd())
-    train_dir = os.path.realpath('{}/..'.format(rsc_src))
+    train_dir = os.path.realpath('{}/../../train'.format(rsc_src))
     if cwd != train_dir:
         os.chdir(train_dir)
     rsc = Resource(cfg)
diff --git a/train/bin/tag.py b/train/bin/tag.py
new file mode 100755
index 0000000..60bb2e9
--- /dev/null
+++ b/train/bin/tag.py
@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+command line part-of-speech tagger demo
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+import logging
+import os
+import sys
+
+from khaiii.train.tagger import PosTagger
+
+
+#############
+# functions #
+#############
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    tgr = PosTagger(args.model_dir)
+    for line_num, line in enumerate(sys.stdin, start=1):
+        if line_num % 100000 == 0:
+            logging.info('%d00k-th line..', (line_num // 100000))
+        line = line.rstrip('\r\n')
+        if not line:
+            print()
+            continue
+        pos_sent = tgr.tag_raw(line)
+        for pos_word in pos_sent.pos_tagged_words:
+            print(pos_word.raw, end='\t')
+            print(' + '.join([str(m) for m in pos_word.pos_tagged_morphs]))
+        print()
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='command line part-of-speech tagger demo')
+    parser.add_argument('-m', '--model-dir', help='model dir', metavar='DIR', required=True)
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--gpu-num', help='GPU number to use', metavar='INT', type=int, default=0)
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_num)
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From 817d6709b87b4c1bf78b6d9e1dbeb9bce26d7f8f Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Fri, 25 Jan 2019 00:49:26 +0900
Subject: [PATCH 11/41] =?UTF-8?q?=EA=B8=B0=EB=B6=84=EC=84=9D=20=EC=82=AC?=
 =?UTF-8?q?=EC=A0=84=EC=9D=84=20=EC=9E=90=EB=8F=99=EC=9C=BC=EB=A1=9C=20?=
 =?UTF-8?q?=EC=B6=94=EC=B6=9C=ED=95=98=EB=8A=94=20=EC=8A=A4=ED=81=AC?=
 =?UTF-8?q?=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 train/bin/extract_preanal.py | 209 +++++++++++++++++++++++++++++++++++
 1 file changed, 209 insertions(+)
 create mode 100755 train/bin/extract_preanal.py

diff --git a/train/bin/extract_preanal.py b/train/bin/extract_preanal.py
new file mode 100755
index 0000000..b23141a
--- /dev/null
+++ b/train/bin/extract_preanal.py
@@ -0,0 +1,209 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+기분석 사전 후보를 추출하는 스크립트
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+from collections import Counter, defaultdict
+import logging
+import sys
+from typing import Dict, List, Tuple
+
+from khaiii.munjong.sejong_corpus import Word
+from khaiii.resource.char_align import Aligner, AlignError
+
+import map_char_to_tag
+
+
+#########
+# types #
+#########
+class Entry:
+    """
+    preanalized dictionary entry
+    """
+    def __init__(self, freq: int, word: str, morph_str: str):
+        self.is_del = False
+        self.is_pfx = word[-1] == '\1'
+        self.freq = freq
+        self.word = word[:-1] if self.is_pfx else word
+        self.morph_str = morph_str
+
+    def __str__(self):
+        return '{}{}\t{}{}\t{}'.format('-' if self.is_del else '', self.freq, self.word,
+                                       '*' if self.is_pfx else '', self.morph_str)
+
+
+#############
+# variables #
+#############
+DIC_AMBIG = defaultdict(Counter)
+
+
+#############
+# functions #
+#############
+def _get_prefix(word: Word) -> Tuple[str, str]:
+    """
+    맨 뒤의 형태소 하나를 제외한 prefix와 그 분석 결과
+    Args:
+        word:  어절 객체
+    Returns:
+        prefix
+        분석 결과
+    """
+    if len(word.morphs) < 2:
+        return None, None
+    if not word.raw.endswith(word.morphs[-1].lex):
+        return None, None
+    prefix = '{}\1'.format(word.raw[:-len(word.morphs[-1].lex)])
+    if len(prefix) < 5:    # 4음절 미만은 버린다.
+        return None, None
+    morphs = ['{}/{}'.format(m.lex, m.tag) for m in word.morphs[:-1]]
+    return prefix, ' + '.join(morphs)
+
+
+def _count_ambig(args: Namespace):
+    """
+    count from courpus and make ambiguous dictionary
+    Args:
+        args:  program arguments
+    """
+    for num, sent in enumerate(map_char_to_tag.sent_iter(args), start=1):
+        if num % 10000 == 0:
+            logging.info('%dk-th sentence..', num // 1000)
+        for word in sent.words:
+            DIC_AMBIG[word.raw][word.morph_str()] += 1
+            prefix, morph_pfx = _get_prefix(word)
+            if prefix:
+                DIC_AMBIG[prefix][morph_pfx] += 1
+
+
+def _filter_no_ambig(min_freq: int) -> Dict[str, str]:
+    """
+    문맥과 상관 없이 중의성이 없는 엔트리를 출력
+    Args:
+        min_freq:  최소 빈도
+    Returns:
+        중의성이 있는 엔트리가 제거된 사전
+    """
+    dic_no_ambig = {}
+    for word, cnt in DIC_AMBIG.items():
+        if len(cnt) > 1:
+            continue
+        morph_str, freq = cnt.most_common(1)[0]
+        if freq < min_freq:
+            continue
+        dic_no_ambig[word] = morph_str
+    return dic_no_ambig
+
+
+def _make_entries(dic_no_ambig: Dict[str, str]) -> List[Entry]:
+    """
+    기분석 사전 엔트리를 생성한다.
+    Args:
+        dic_no_ambig:  중의성이 없는 엔트리 사전
+    Returns:
+        엔트리 리스트
+    """
+    entries = []
+    pfx_idx = -1
+    for idx, (word, morph_str) in enumerate(sorted(dic_no_ambig.items())):
+        _, freq = list(DIC_AMBIG[word].items())[0]
+        entry = Entry(freq, word, morph_str)
+        if entry.is_pfx:
+            if entries and entries[-1].word == entry.word:
+                # 이전 어절 exact가 현재 prefix와 같다면, 예: "제이미" vs "제이미*"
+                if entries[-1].morph_str == entry.morph_str:    # pylint: disable=simplifiable-if-statement
+                    # 분석 결과가 같다면 prefix를 남기고 exact를 제거한다.
+                    entries[-1].is_del = True
+                else:
+                    # 분석 결과가 다르면 안전하게 exact를 남기고 prefix를 제거한다.
+                    entry.is_del = True
+            pfx_idx = idx
+        elif (pfx_idx >= 0 and word.startswith(entries[pfx_idx].word) and
+              morph_str.startswith(entries[pfx_idx].morph_str) and
+              entries[pfx_idx].freq <= freq):
+            # 이전 prefix가 현재 어절 exact를 포함하면서 빈도는 같다면,
+            # 모두 동일한 어절에서 뽑힌 prefix이므로 prefix를 삭제한다.
+            # 예: 빈도가 같은 "강화된다*"(삭제) vs "강화된다."(남김)
+            entries[pfx_idx].is_del = True
+        entries.append(entry)
+    return entries
+
+
+def run(args: Namespace):
+    """
+    map characters with tags from eojeol and morphemes
+    Args:
+        args:  program arguments
+    """
+    aligner = Aligner(args.rsc_src)
+
+    _count_ambig(args)
+    dic_no_ambig = _filter_no_ambig(args.min_freq)
+    entries = _make_entries(dic_no_ambig)
+
+    del_word = 0
+    del_pfx = 0
+    for entry in entries:
+        if entry.is_del:
+            logging.debug(entry)
+            if entry.is_pfx:
+                del_pfx += 1
+            else:
+                del_word += 1
+        else:
+            word = Word.parse('\t'.join(['', entry.word, entry.morph_str]), '', 0)
+            try:
+                aligner.align(word)
+                entry_str = str(entry)
+                if not args.with_freq:
+                    entry_str = '\t'.join(entry_str.split('\t')[1:])
+                print(entry_str)
+            except AlignError as algn_err:
+                logging.error('%s: %s', algn_err, entry)
+
+    logging.info('deleted word: %d', del_word)
+    logging.info('deleted prefix: %d', del_pfx)
+
+
+########
+# main #
+########
+def main():
+    """
+    map characters with tags from eojeol and morphemes
+    """
+    parser = ArgumentParser(description='기분석 사전 후보를 추출하는 스크립트')
+    parser.add_argument('-c', '--corpus-dir', help='corpus dir', metavar='DIR', required=True)
+    parser.add_argument('--rsc-src', help='resource source dir <default: ../rsc/src>',
+                        metavar='DIR', default='../rsc/src')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--min-freq', help='minimum frequency <default: 10>', metavar='NUM',
+                        type=int, default=10)
+    parser.add_argument('--with-freq', help='print with frequency', action='store_true')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From 259ad5b032c441366d827b7d4f9c309a65b832b8 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Sat, 26 Jan 2019 15:08:47 +0900
Subject: [PATCH 12/41] =?UTF-8?q?=EC=98=A4=EB=B6=84=EC=84=9D=20=ED=8C=A8?=
 =?UTF-8?q?=EC=B9=98=20=ED=9B=84=EB=B3=B4=EB=A5=BC=20=EC=B6=94=EC=B6=9C?=
 =?UTF-8?q?=ED=95=98=EB=8A=94=20=EC=8A=A4=ED=81=AC=EB=A6=BD=ED=8A=B8=20?=
 =?UTF-8?q?=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 rsc/bin/compile_model.py      |   8 +-
 train/bin/extract_errpatch.py | 352 ++++++++++++++++++++++++++++++++++
 train/bin/pickle_model.py     |  23 +--
 3 files changed, 355 insertions(+), 28 deletions(-)
 create mode 100755 train/bin/extract_errpatch.py

diff --git a/rsc/bin/compile_model.py b/rsc/bin/compile_model.py
index 767060c..f40fc6b 100755
--- a/rsc/bin/compile_model.py
+++ b/rsc/bin/compile_model.py
@@ -15,7 +15,6 @@
 from argparse import ArgumentParser, Namespace
 import json
 import logging
-import os
 import pathlib
 import pickle
 from typing import Tuple
@@ -42,13 +41,8 @@ def _load_cfg_rsc(rsc_src: str, model_size: str) -> Tuple[Namespace, Resource]:
     cfg = Namespace()
     for key, val in cfg_dic.items():
         setattr(cfg, key, val)
-    cwd = os.path.realpath(os.getcwd())
-    train_dir = os.path.realpath('{}/../../train'.format(rsc_src))
-    if cwd != train_dir:
-        os.chdir(train_dir)
+    setattr(cfg, 'rsc_src', rsc_src)
     rsc = Resource(cfg)
-    if cwd != train_dir:
-        os.chdir(cwd)
     return cfg, rsc
 
 
diff --git a/train/bin/extract_errpatch.py b/train/bin/extract_errpatch.py
new file mode 100755
index 0000000..ff9dd8c
--- /dev/null
+++ b/train/bin/extract_errpatch.py
@@ -0,0 +1,352 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+오분석 패치 후보를 추출하는 스크립트
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+from collections import Counter
+import difflib
+import logging
+import multiprocessing
+import os
+import sys
+from typing import Iterator, List, Tuple
+
+from khaiii.munjong import sejong_corpus
+from khaiii.resource.morphs import WORD_DELIM_STR, SENT_DELIM_STR
+
+from khaiii.khaiii import KhaiiiApi
+
+
+#############
+# variables #
+#############
+_KHAIII = None    # khaiii API
+
+
+#############
+# functions #
+#############
+def _get_diff_sgmts(result_morphs: List[str], corpus_morphs: List[str], raw_sent: str,
+                    alignments: List[Tuple[int, int]]) -> List[Tuple[str, str, str]]:
+    """
+    패치 생성을 위해 자동 태깅 결과와 코퍼스의 정답이 다른 부분을 찾는다.
+    Args:
+        result_morphs:  자동 태깅 결과
+        corpus_morphs:  코퍼스의 정답
+        raw_sent:  문장 원문
+        alignments:  음절별 정렬 정보. 형태소별 (시작, 끝) 음절 포지션 pair
+    Returns:
+        (자동, 정답, 원문) triple의 리스트
+    """
+    def _get_origin(first_morph: int, last_morph: int) -> str:
+        """
+        첫번째 형태소와 마지막 형태소 인덱스를 바탕으로 원문을 얻는다.
+        Args:
+            first_morph:  첫번째 형태소 인덱스
+            last_morph:  마지막 형태소 인덱스
+        Returns:
+            원문
+        """
+        if first_morph < 0:
+            first_morph = 0
+        if last_morph >= len(alignments):
+            last_morph = len(alignments)
+        return raw_sent[alignments[first_morph][0]:alignments[last_morph-1][1]]
+
+    def _expand_begin(result_begin: int) -> int:
+        """
+        왼쪽 경계가 음소로 나눠진 경우 음절 경계가 나올 때까지 왼쪽으로 확장한다.
+        Args:
+            result_begin:  왼쪽 경계
+        Returns:
+            확장된 왼쪽 경계
+        """
+        if result_begin <= 0:
+            return result_begin
+        if alignments[result_begin - 1][1] > alignments[result_begin][0]:
+            return _expand_begin(result_begin - 1)
+        return result_begin
+
+    def _expand_end(result_end: int) -> int:
+        """
+        오른쪽 경계가 음소로 나눠진 경우 음절 경계가 나올 때까지 오른쪽으로 확장한다.
+        Args:
+            result_end:  오른쪽 경계.
+                         [result_begin, result_end)으로 오른쪽 경계는 open이므로,
+                         실제로 result_end-1이 마지막 형태소의 인덱스이다.
+        Returns:
+            확장된 오른쪽 경계
+        """
+        if result_end >= len(result_morphs):
+            return result_end
+        if alignments[result_end - 1][1] > alignments[result_end][0]:
+            return _expand_end(result_end + 1)
+        return result_end
+
+    def _expand_spc(diff_sgmts: List[Tuple[str, str, str]], result_begin: int, result_end: int,
+                    corpus_begin: int, corpus_end: int):
+        """
+        좌, 우 경계에 공백이 있을 경우 한 형태소 더 확장한다.
+        Args:
+            diff_sgmts:  diff를 추가할 리스트
+            result_begin:  오분석의 왼쪽 경계
+            result_end:  오분석의 오른쪽 경계
+            corpus_begin:  정분석(코퍼스)의 왼쪽 경계
+            corpus_begin:  정분석(코퍼스)의 오른쪽 경계
+        """
+        if (result_morphs[result_begin] != WORD_DELIM_STR and    # pylint: disable=consider-using-in
+                result_morphs[result_end-1] != WORD_DELIM_STR):
+            return
+        _result_begin = _expand_begin(result_begin-1) \
+            if result_morphs[result_begin] == WORD_DELIM_STR else result_begin
+        result_end_ = _expand_end(result_end+1) \
+            if result_morphs[result_end-1] == WORD_DELIM_STR else result_end
+        assert result_begin != _result_begin or result_end != result_end_
+        _left = result_morphs[_result_begin:result_end_]
+        _corpus_begin = corpus_begin - (result_begin - _result_begin)
+        corpus_end_ = corpus_end + (result_end_ - result_end)
+        right_ = corpus_morphs[_corpus_begin:corpus_end_]
+        origin = _get_origin(_result_begin, result_end_)
+        diff_sgmts.append((origin, _left, right_))
+
+    diff_sgmts = []
+    matches = difflib.SequenceMatcher(None, result_morphs, corpus_morphs).get_matching_blocks()
+    result_begin = 0
+    corpus_begin = 0
+    for match in matches:
+        result_end = match.a
+        left = []
+        if result_end > result_begin:
+            # _result_begin은 result_begin에서 왼쪽으로 같은 음절인 경우에 대해 확장한 시작 위치입니다.
+            # 가령: '된다'에 대해 오분석: '되+ㄴ+다', 정분석: '되+ㄴ다'일 경우 틀린 영역은 'ㄴ+다' vs 'ㄴ다' 이지만,
+            # 'ㄴ'이 포함된 음절이 '된'이므로 원문을 추출하면 '된'이 되어벼러 원문과 형태소간 불일치가 일어납니다.
+            # 이것을 보정해 주기 위해 음소단위 형태소인 경우 왼쪽으로 음절 경계까지 확장하는 것입니다.
+            # result_end_도 마찬가지로 result_end를 오른쪽으로 확장한 것입니다.
+            _result_begin = _expand_begin(result_begin)
+            result_end_ = _expand_end(result_end)
+            left = result_morphs[_result_begin:result_end_]
+        else:
+            _result_begin = result_begin
+            result_end_ = result_end
+
+        corpus_end = match.b
+        # _result_begin, result_end_가 각각 n, m개 확장된 경우,
+        # 동일한 갯수로 corpus_begin, corpus_end을 좌, 우로 n, m개 확장합니다.
+        # 틀린 부분의 좌, 우 형태소가 오분석과 정분석이 같다는 가정 하에 alignments 정보를 활용하지 않습니다.
+        _corpus_begin = corpus_begin - (result_begin - _result_begin)
+        corpus_end_ = corpus_end + (result_end_ - result_end)
+        right = corpus_morphs[_corpus_begin:corpus_end_]
+
+        if not left and not right:
+            result_begin = result_end + match.size
+            corpus_begin = corpus_end + match.size
+            continue
+        assert left != right, '{} == {}'.format(left, right)
+
+        if (len(left) >= 2 and right) or (left and len(right) >= 2):
+            # 결과가 다른 영역만
+            origin = _get_origin(_result_begin, result_end_)
+            diff_sgmts.append((origin, left, right))
+
+        # _left는 left에서 오분석인 영역에 왼쪽으로 하나의 형태소를 더해서 만든 것입니다.
+        # left_는 left에서 오른쪽으로 하나의 형태소, _left_는 양쪽에 하나씩 더해서 만든 것입니다.
+        # _right, right_, _right_도 left와 마찬가지입니다.
+
+        # 왼쪽 형태소를 추가
+        # 추가한 형태소가 음소로 나눠진 경우에도 역시 왼쪽으로 음절 경계까지 확장
+        __result_begin = _expand_begin(_result_begin-1)
+        _left = result_morphs[__result_begin:_result_begin] + left
+        __corpus_begin = _corpus_begin - (_result_begin - __result_begin)
+        _right = corpus_morphs[__corpus_begin:_corpus_begin] + right
+        left_first_lex = _left[0].rsplit('/', 1)[0]
+        right_first_lex = _right[0].rsplit('/', 1)[0]
+        if left_first_lex == right_first_lex and len(_left) >= 2 and len(_right) >= 2:
+            origin = _get_origin(__result_begin, result_end_)
+            diff_sgmts.append((origin, _left, _right))
+            _expand_spc(diff_sgmts, __result_begin, result_end_, __corpus_begin, corpus_end_)
+
+        # 오른쪽 형태소를 추가
+        # 추가한 형태소가 음소로 나눠진 경우에도 역시 오른쪽으로 음절 경계까지 확장
+        result_end__ = _expand_end(result_end_+1)
+        left_ = left + result_morphs[result_end_:result_end__]
+        corpus_end__ = corpus_end_ + (result_end__ - result_end_)
+        right_ = right + corpus_morphs[corpus_end_:corpus_end__]
+        left_last_lex = left_[-1].rsplit('/', 1)[0]
+        right_last_lex = right_[-1].rsplit('/', 1)[0]
+        if left_last_lex == right_last_lex and len(left_) >= 2 and len(right_) >= 2:
+            origin = _get_origin(_result_begin, result_end__)
+            diff_sgmts.append((origin, left_, right_))
+            _expand_spc(diff_sgmts, _result_begin, result_end__, _corpus_begin, corpus_end__)
+
+        # 양쪽에 형태소를 각각 추가
+        _left_ = _left + result_morphs[result_end_:result_end__]
+        _right_ = _right + corpus_morphs[corpus_end_:corpus_end__]
+        origin = _get_origin(__result_begin, result_end__)
+        diff_sgmts.append((origin, _left_, _right_))
+        _expand_spc(diff_sgmts, __result_begin, result_end__, __corpus_begin, corpus_end__)
+
+        result_begin = result_end + match.size
+        corpus_begin = corpus_end + match.size
+    return diff_sgmts
+
+
+def _count_error(args: Namespace, doc_path: str) -> Counter:
+    """
+    count from courpus and make ambiguous dictionary
+    Args:
+        args:  program arguments
+        doc_path:  document path
+    Returns:
+        오분석 패치 후보(원문, 오분석, 정분석 triple)의 카운터
+    """
+    global _KHAIII    # pylint: disable=global-statement
+    if not _KHAIII:
+        _KHAIII = KhaiiiApi(args.lib_path, args.rsc_dir)
+
+    cnt = Counter()
+    logging.info(doc_path)
+    for sent in sejong_corpus.sents(open(doc_path, 'r', encoding='UTF-8')):
+        raw_sent = sent.raw_str()
+        result_morphs = [SENT_DELIM_STR, ]
+        corpus_morphs = [SENT_DELIM_STR, ]
+        alignments = [(0, 0), ]
+        for result_word, corpus_word in zip(_KHAIII.analyze(raw_sent, ''), sent.words):
+            assert result_word.lex == corpus_word.raw, \
+                   '{}: "{}" != "{}"'.format(os.path.basename(doc_path), result_word, corpus_word)
+            if len(result_morphs) > 1:
+                result_morphs.append(WORD_DELIM_STR)
+                corpus_morphs.append(WORD_DELIM_STR)
+                alignments.append((alignments[-1][1], alignments[-1][1]+1))
+            result_morphs.extend(['{}/{}'.format(m.lex, m.tag) for m in result_word.morphs])
+            corpus_morphs.extend([str(m) for m in corpus_word.morphs])
+            alignments.extend([(m.begin, m.begin + m.length) for m in result_word.morphs])
+        result_morphs.append(SENT_DELIM_STR)
+        corpus_morphs.append(SENT_DELIM_STR)
+        alignments.append((alignments[-1][1], alignments[-1][1]))
+        if result_morphs != corpus_morphs:
+            diff_sgmts = _get_diff_sgmts(result_morphs, corpus_morphs, raw_sent, alignments)
+            for origin, left, right in diff_sgmts:
+                cnt[origin, ' + '.join(left), ' + '.join(right)] += 1
+    return cnt
+
+
+def _doc_iter(corpus_dir: str) -> Iterator[str]:
+    """
+    문종 코퍼스에서 문서의 경로를 리턴하는 generator
+    Args:
+        args:  program arguments
+    Yields:
+        document path
+    """
+    for name in sorted(os.listdir(corpus_dir)):
+        if not name.endswith('.txt'):
+            continue
+        yield '{}/{}'.format(corpus_dir, name)
+
+
+def _filter_cnt(args: Namespace, cnt: Counter) -> Counter:
+    """
+    규칙에 의해 필터링하여 빈도로 정렬한다.
+    Args:
+        args:  program arguments
+        cnt:  (원문, left, right) 별로 측정한 빈도 사전
+    Returns:
+        (빈도, (원문, left, right)) pair의 리스트
+    """
+    cnts = []
+    for (origin, left, right), freq in cnt.items():
+        if freq < args.min_freq:
+            continue
+        if len(origin.strip()) < args.min_len:
+            continue
+        cnts.append((freq, (origin, left, right)))
+    cnts.sort(key=lambda x: x[0], reverse=True)
+    return cnts
+
+
+def _print_cnt(args: Namespace, cnts: Counter):
+    """
+    최종 패치 후포를 출력한다.
+    Args:
+        args:  program arguments
+        cnts:  counts
+    """
+    for freq, (origin, left, right) in cnts:
+        freq_tab = ''
+        if args.with_freq:
+            freq_tab = '{}\t'.format(freq)
+        print('{}{}\t{}\t{}'.format(freq_tab, origin, left, right))
+
+
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    cnt = Counter()
+    if args.num_proc > 1:
+        pool = multiprocessing.Pool(args.num_proc)
+        results = ((doc, pool.apply_async(_count_error, (args, doc)))
+                   for doc in _doc_iter(args.corpus_dir))
+        for num, (doc, result) in enumerate(results, start=1):
+            if num % 100 == 0:
+                logging.info('%d-th document..', num)
+            try:
+                cnt.update(result.get(timeout=1000))
+            except multiprocessing.context.TimeoutError:
+                logging.error('timeout[%d]: %s', num, doc)
+    else:
+        for num, doc in enumerate(_doc_iter(args.corpus_dir), start=1):
+            if num % 10 == 0:
+                logging.info('%d-th document..', num)
+            cnt.update(_count_error(args, doc))
+
+    cnts = _filter_cnt(args, cnt)
+    _print_cnt(args, cnts)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='오분석 패치 후보를 추출하는 스크립트')
+    parser.add_argument('-c', '--corpus-dir', help='corpus dir', metavar='DIR', required=True)
+    parser.add_argument('--lib-path', help='khaiii shared library path', metavar='FILE', default='')
+    parser.add_argument('--rsc-dir', help='resource dir', metavar='DIR', default='')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--with-freq', help='print with frequency', action='store_true')
+    parser.add_argument('--min-freq', help='minimum frequency <default: 10>', metavar='FREQ',
+                        type=int, default=10)
+    parser.add_argument('--min-len', help='minimum original text length <default: 4>',
+                        metavar='LEN', type=int, default=4)
+    parser.add_argument('--num-proc', help='number of processes <default: 1>', metavar='NUM',
+                        type=int, default=1)
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/train/bin/pickle_model.py b/train/bin/pickle_model.py
index 06f6620..11bcddc 100755
--- a/train/bin/pickle_model.py
+++ b/train/bin/pickle_model.py
@@ -16,7 +16,6 @@
 from array import array
 import json
 import logging
-import os
 import pickle
 import re
 
@@ -174,25 +173,6 @@ def _load_config(path: str) -> Namespace:
     return cfg
 
 
-def _load_resource(cfg: Namespace, rsc_src: str) -> Resource:
-    """
-    load resources
-    Args:
-        cfg:  config
-        rsc_src:  resource source dir
-    Returns:
-        Resource object
-    """
-    cwd = os.path.realpath(os.getcwd())
-    train_dir = os.path.realpath('{}/../../train'.format(rsc_src))
-    if cwd != train_dir:
-        os.chdir(train_dir)
-    rsc = Resource(cfg)
-    if cwd != train_dir:
-        os.chdir(cwd)
-    return rsc
-
-
 def run(args: Namespace):
     """
     run function which is the start point of program
@@ -200,7 +180,8 @@ def run(args: Namespace):
         args:  program arguments
     """
     cfg = _load_config('{}/config.json'.format(args.in_dir))
-    rsc = _load_resource(cfg, args.rsc_src)
+    setattr(cfg, 'rsc_src', args.rsc_src)
+    rsc = Resource(cfg)
     state_dict = torch.load('{}/model.state'.format(args.in_dir),
                             map_location=lambda storage, loc: storage)
     _validate_state_dict(cfg, rsc, state_dict)

From 19f3083a73f11b6527efa260e505b66c3373e65b Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Sun, 27 Jan 2019 22:34:05 +0900
Subject: [PATCH 13/41] =?UTF-8?q?=EC=98=A4=EB=B6=84=EC=84=9D=20=ED=8C=A8?=
 =?UTF-8?q?=EC=B9=98=20=ED=9B=84=EB=B3=B4=EB=A5=BC=20=EA=B2=80=EC=A6=9D?=
 =?UTF-8?q?=ED=95=98=EB=8A=94=20=EC=8A=A4=ED=81=AC=EB=A6=BD=ED=8A=B8=20?=
 =?UTF-8?q?=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 rsc/bin/compile_errpatch.py                   | 106 +------
 rsc/bin/compile_preanal.py                    |  61 +---
 rsc/bin/compile_restore.py                    |  51 +---
 src/main/python/khaiii/resource/char_align.py | 128 ++++++++-
 src/main/python/khaiii/resource/morphs.py     |  27 ++
 src/main/python/khaiii/resource/resource.py   |  53 +++-
 train/bin/extract_errpatch.py                 |   7 +-
 train/bin/hd_validate_errpatch.bash           | 194 +++++++++++++
 train/bin/train.py                            |   4 +-
 train/bin/validate_errpatch.py                | 270 ++++++++++++++++++
 10 files changed, 688 insertions(+), 213 deletions(-)
 create mode 100755 train/bin/hd_validate_errpatch.bash
 create mode 100755 train/bin/validate_errpatch.py

diff --git a/rsc/bin/compile_errpatch.py b/rsc/bin/compile_errpatch.py
index caf7d8a..7e49fd9 100755
--- a/rsc/bin/compile_errpatch.py
+++ b/rsc/bin/compile_errpatch.py
@@ -13,23 +13,19 @@
 # imports #
 ###########
 from argparse import ArgumentParser, Namespace
-from collections import defaultdict
 import glob
-import itertools
 import logging
 import os
 import struct
 import sys
 from typing import Dict, List, Tuple
 
-from khaiii.munjong import sejong_corpus
-from khaiii.resource.char_align import Aligner, AlignError
-from khaiii.resource.morphs import Morph, ParseError
-from khaiii.resource.morphs import WORD_DELIM_STR, SENT_DELIM_STR, WORD_DELIM_NUM, SENT_DELIM_NUM
+from khaiii.resource.char_align import Aligner, align_patch
+from khaiii.resource.resource import load_restore_dic, load_vocab_out
+from khaiii.resource.morphs import Morph, ParseError, mix_char_tag
 from khaiii.resource.trie import Trie
 
-from compile_preanal import align_to_tag, print_errors
-from compile_restore import load_restore_dic, load_vocab_out
+from compile_preanal import print_errors
 
 
 #########
@@ -104,100 +100,6 @@ def _parse(self):
 #############
 # functions #
 #############
-def _split_list(lst: List[str], delim: str) -> List[List[str]]:
-    """
-    리스트를 delimiter로 split하는 함수
-
-    >>> _split_list(['가/JKS', '_', '너/NP'], '_')
-    [['가/JKS'], ['너/NP']]
-
-    Args:
-        lst:  리스트
-        delim:  delimiter
-    Returns:
-        list of sublists
-    """
-    sublists = []
-    while lst:
-        prefix = [x for x in itertools.takewhile(lambda x: x != delim, lst)]
-        sublists.append(prefix)
-        lst = lst[len(prefix):]
-        delims = [x for x in itertools.takewhile(lambda x: x == delim, lst)]
-        lst = lst[len(delims):]
-    return sublists
-
-
-def align_patch(rsc_src: Tuple[Aligner, Dict, Dict[str, int]], raw: str, morph_str: str) \
-        -> List[int]:
-    """
-    패치의 원문과 분석 결과를 음절단위 매핑(정렬)을 수행한다.
-    Args:
-        rsc_src:  (Aligner, restore dic, vocab out) resource triple
-        raw:  원문
-        morph_str:  형태소 분석 결과 (패치 기술 형식)
-    Returns:
-        정렬에 기반한 출력 태그 번호
-    """
-    aligner, restore_dic, vocab_out = rsc_src
-    raw_words = raw.strip().split()
-    morphs = morph_str.split(' + ')
-    morphs_strip = morphs
-    if morphs[0] in [WORD_DELIM_STR, SENT_DELIM_STR]:
-        morphs_strip = morphs_strip[1:]
-    if morphs[-1] in [WORD_DELIM_STR, SENT_DELIM_STR]:
-        morphs_strip = morphs_strip[:-1]
-    morph_words = _split_list(morphs_strip, WORD_DELIM_STR)
-    tag_nums = []
-    restore_new = defaultdict(dict)
-    vocab_new = defaultdict(list)
-    for raw_word, morph_word in zip(raw_words, morph_words):
-        word = sejong_corpus.Word.parse('\t'.join(['', raw_word, ' + '.join(morph_word)]), '', 0)
-        try:
-            word_align = aligner.align(word)
-            _, word_tag_nums = \
-                align_to_tag(raw_word, word_align, (restore_dic, restore_new),
-                             (vocab_out, vocab_new))
-            if restore_new or vocab_new:
-                logging.debug('needs dic update: %s', word)
-                return []
-        except AlignError as algn_err:
-            logging.debug('alignment error: %s', word)
-            logging.debug(str(algn_err))
-            return []
-        if tag_nums:
-            tag_nums.append(WORD_DELIM_NUM)
-        tag_nums.extend(word_tag_nums)
-    if morphs[0] in [WORD_DELIM_STR, SENT_DELIM_STR]:
-        tag_nums.insert(0, WORD_DELIM_NUM if morphs[0] == WORD_DELIM_STR else SENT_DELIM_NUM)
-    if morphs[-1] in [WORD_DELIM_STR, SENT_DELIM_STR]:
-        tag_nums.append(WORD_DELIM_NUM if morphs[-1] == WORD_DELIM_STR else SENT_DELIM_NUM)
-    return tag_nums
-
-
-def mix_char_tag(chars: str, tags: List[int]) -> List[int]:
-    """
-    음절과 출력 태그를 비트 연산으로 합쳐서 하나의 (32비트) 숫자로 표현한다.
-    Args:
-        chars:  음절 (유니코드) 리스트 (문자열)
-        tags:  출력 태그 번호의 리스트
-    Returns:
-        합쳐진 숫자의 리스트
-    """
-    char_nums = [ord(c) for c in chars]
-    if tags[0] == SENT_DELIM_NUM:
-        char_nums.insert(0, SENT_DELIM_NUM)
-    if tags[-1] == SENT_DELIM_NUM:
-        char_nums.append(SENT_DELIM_NUM)
-    for idx, char_num in enumerate(char_nums):
-        if char_num == ord(' '):
-            char_nums[idx] = WORD_DELIM_NUM
-            continue
-        elif tags[idx] == SENT_DELIM_NUM:
-            continue
-        char_nums[idx] = char_num << 12 | tags[idx]
-    return char_nums
-
-
 def _load_entries(args: Namespace) -> List[Entry]:
     """
     패치 엔트리를 파일로부터 로드한다.
diff --git a/rsc/bin/compile_preanal.py b/rsc/bin/compile_preanal.py
index 6874d2c..649e94c 100755
--- a/rsc/bin/compile_preanal.py
+++ b/rsc/bin/compile_preanal.py
@@ -19,14 +19,16 @@
 import os
 import struct
 import sys
-from typing import Dict, List, Tuple
+from typing import Dict, List
 
 from khaiii.munjong import sejong_corpus
-from khaiii.resource.char_align import Aligner, AlignError, MrpChr
+from khaiii.resource.char_align import Aligner, AlignError, align_to_tag
 from khaiii.resource.morphs import Morph, ParseError
+from khaiii.resource.resource import load_restore_dic, load_vocab_out
 from khaiii.resource.trie import Trie
 
-from compile_restore import load_restore_dic, load_vocab_out, append_new_entries
+from compile_restore import append_new_entries
+
 
 
 #########
@@ -175,59 +177,6 @@ def _set_align(aligner: Aligner, Word: type, entries: List[Entry]):    # pylint:
     print_errors(bad_entries)
 
 
-def align_to_tag(raw_word: str, alignment: List[List[str]], restore: Tuple[dict, dict],
-                 vocab: Tuple[Dict[str, int], Dict[str, int]]) \
-        -> Tuple[List[str], List[int]]:
-    """
-    어절의 원문과 정렬 정보를 활용해 음절과 매핑된 태그를 생성한다.
-    Args:
-        raw_word:  어절 원문
-        alignment:  정렬 정보
-        restore:  (원형복원 사전, 원형복원 사전에 추가할 엔트리) pair
-        vocab:  (출력 태그 사전, 출력 태그 사전에 추가할 새로운 태그) pair
-    Returns:
-        음절별 출력 태그
-        음절별 출력 태그의 번호
-    """
-    assert len(raw_word) == len(alignment)
-    restore_dic, restore_new = restore
-    vocab_out, vocab_new = vocab
-    tag_outs = []
-    tag_nums = []
-    for char, mrp_chrs in zip(raw_word, alignment):
-        if len(mrp_chrs) == 1 and mrp_chrs[0].char == char:
-            tag_outs.append(mrp_chrs[0].tag)
-        else:
-            tag_str = ':'.join([m.tag for m in mrp_chrs])
-            mrp_chr_str_key = MrpChr.to_str(mrp_chrs)
-            found = -1
-            max_num = -1
-            for num, mrp_chr_str_val in restore_dic[char, tag_str].items():
-                if num > max_num:
-                    max_num = num
-                if mrp_chr_str_key == mrp_chr_str_val:
-                    found = num
-                    break
-            if found >= 0:
-                tag_outs.append('{}:{}'.format(tag_str, found))
-            else:
-                new_num = max_num + 1
-                restore_dic[char, tag_str][new_num] = mrp_chr_str_key
-                restore_new[char, tag_str][new_num] = mrp_chr_str_key
-                tag_outs.append('{}:{}'.format(tag_str, new_num))
-        tag = tag_outs[-1]
-        if tag in vocab_out:
-            tag_nums.append(vocab_out[tag])
-        elif tag in vocab_new:
-            tag_nums.append(vocab_new[tag])
-        else:
-            new_tag_num = len(vocab_out) + len(vocab_new) + 1
-            logging.debug('new output tag: [%d] %s', new_tag_num, tag)
-            vocab_new[tag] = new_tag_num
-            tag_nums.append(new_tag_num)
-    return tag_outs, tag_nums
-
-
 def _set_tag_out(restore_dic: dict, restore_new: dict, vocab_out: Dict[str, int],
                  vocab_new: Dict[str, int], entries: List[Entry]):
     """
diff --git a/rsc/bin/compile_restore.py b/rsc/bin/compile_restore.py
index 132df3a..c5a0172 100755
--- a/rsc/bin/compile_restore.py
+++ b/rsc/bin/compile_restore.py
@@ -13,14 +13,14 @@
 # imports #
 ###########
 from argparse import ArgumentParser, Namespace
-from collections import defaultdict
 import logging
 import os
 import struct
 import sys
-from typing import Dict, Tuple
+from typing import Dict
 
 from khaiii.resource.morphs import TAG_SET
+from khaiii.resource.resource import load_restore_dic, load_vocab_out
 
 
 #############
@@ -32,53 +32,6 @@
 #############
 # functions #
 #############
-def load_restore_dic(file_path: str) -> Dict[Tuple[str, str], Dict[int, str]]:
-    """
-    원형복원 사전을 로드한다.
-    Args:
-        file_path:  파일 경로
-    Returns:
-        사전
-    """
-    file_name = os.path.basename(file_path)
-    restore_dic = defaultdict(dict)
-    for line_num, line in enumerate(open(file_path, 'r', encoding='UTF-8'), start=1):
-        line = line.rstrip()
-        if not line or line[0] == '#':
-            continue
-        char_tag_num, mrp_chr_str = line.split('\t')
-        char, tag_num = char_tag_num.rsplit('/', 1)
-        tag, num = tag_num.rsplit(':', 1)
-        num = int(num)
-        if (char, tag) in restore_dic:
-            num_mrp_chrs_dic = restore_dic[char, tag]
-            if num in num_mrp_chrs_dic:
-                logging.error('%s:%d: duplicated with %s: %s', file_name, line_num,
-                              num_mrp_chrs_dic[num], line)
-                return {}
-        restore_dic[char, tag][num] = mrp_chr_str
-    return restore_dic
-
-
-def load_vocab_out(rsc_src: str) -> Dict[str, int]:
-    """
-    출력 태그 vocabulary를 로드한다.
-    Args:
-        rsc_src:  리소스 디렉토리
-    Returns:
-        출력 태그 vocabulary
-    """
-    file_path = '{}/vocab.out'.format(rsc_src)
-    vocab_out = [line.strip() for line in open(file_path, 'r', encoding='UTF-8')
-                 if line.strip()]
-    vocab_out_more = []
-    file_path = '{}/vocab.out.more'.format(rsc_src)
-    if os.path.exists(file_path):
-        vocab_out_more = [line.strip() for line in open(file_path, 'r', encoding='UTF-8')
-                          if line.strip()]
-    return {tag: idx for idx, tag in enumerate(vocab_out + vocab_out_more, start=1)}
-
-
 def append_new_entries(rsc_src: str, restore_new: dict, vocab_new: Dict[str, int]):
     """
     기분석 사전 빌드 중에 새로 추가가 필요한 사전 엔트리를 해당 사전에 추가한다.
diff --git a/src/main/python/khaiii/resource/char_align.py b/src/main/python/khaiii/resource/char_align.py
index 1682606..a2052d8 100644
--- a/src/main/python/khaiii/resource/char_align.py
+++ b/src/main/python/khaiii/resource/char_align.py
@@ -12,13 +12,15 @@
 # imports #
 ###########
 from collections import Counter, defaultdict
+import itertools
 import logging
 import os
-from typing import List, Tuple
+from typing import Dict, List, Tuple
 
 from khaiii.munjong.sejong_corpus import Word
 from khaiii.resource import jaso
 from khaiii.resource.morphs import Morph
+from khaiii.resource.morphs import WORD_DELIM_NUM, WORD_DELIM_STR, SENT_DELIM_NUM, SENT_DELIM_STR
 
 
 #########
@@ -594,3 +596,127 @@ def add_msg(self, msg: str):
             msg:  에러 메세지
         """
         self._msgs.append(msg)
+
+
+#############
+# functions #
+#############
+def align_to_tag(raw_word: str, alignment: List[List[str]], restore: Tuple[dict, dict],
+                 vocab: Tuple[Dict[str, int], Dict[str, int]]) -> Tuple[List[str], List[int]]:
+    """
+    어절의 원문과 정렬 정보를 활용해 음절과 매핑된 태그를 생성한다.
+    Args:
+        raw_word:  어절 원문
+        alignment:  정렬 정보
+        restore:  (원형복원 사전, 원형복원 사전에 추가할 엔트리) pair
+        vocab:  (출력 태그 사전, 출력 태그 사전에 추가할 새로운 태그) pair
+    Returns:
+        음절별 출력 태그
+        음절별 출력 태그의 번호
+    """
+    assert len(raw_word) == len(alignment)
+    restore_dic, restore_new = restore
+    vocab_out, vocab_new = vocab
+    tag_outs = []
+    tag_nums = []
+    for char, mrp_chrs in zip(raw_word, alignment):
+        if len(mrp_chrs) == 1 and mrp_chrs[0].char == char:
+            tag_outs.append(mrp_chrs[0].tag)
+        else:
+            tag_str = ':'.join([m.tag for m in mrp_chrs])
+            mrp_chr_str_key = MrpChr.to_str(mrp_chrs)
+            found = -1
+            max_num = -1
+            for num, mrp_chr_str_val in restore_dic[char, tag_str].items():
+                if num > max_num:
+                    max_num = num
+                if mrp_chr_str_key == mrp_chr_str_val:
+                    found = num
+                    break
+            if found >= 0:
+                tag_outs.append('{}:{}'.format(tag_str, found))
+            else:
+                new_num = max_num + 1
+                restore_dic[char, tag_str][new_num] = mrp_chr_str_key
+                restore_new[char, tag_str][new_num] = mrp_chr_str_key
+                tag_outs.append('{}:{}'.format(tag_str, new_num))
+        tag = tag_outs[-1]
+        if tag in vocab_out:
+            tag_nums.append(vocab_out[tag])
+        elif tag in vocab_new:
+            tag_nums.append(vocab_new[tag])
+        else:
+            new_tag_num = len(vocab_out) + len(vocab_new) + 1
+            logging.debug('new output tag: [%d] %s', new_tag_num, tag)
+            vocab_new[tag] = new_tag_num
+            tag_nums.append(new_tag_num)
+    return tag_outs, tag_nums
+
+
+def _split_list(lst: List[str], delim: str) -> List[List[str]]:
+    """
+    리스트를 delimiter로 split하는 함수
+
+    >>> _split_list(['가/JKS', '_', '너/NP'], '_')
+    [['가/JKS'], ['너/NP']]
+
+    Args:
+        lst:  리스트
+        delim:  delimiter
+    Returns:
+        list of sublists
+    """
+    sublists = []
+    while lst:
+        prefix = [x for x in itertools.takewhile(lambda x: x != delim, lst)]
+        sublists.append(prefix)
+        lst = lst[len(prefix):]
+        delims = [x for x in itertools.takewhile(lambda x: x == delim, lst)]
+        lst = lst[len(delims):]
+    return sublists
+
+
+def align_patch(rsc_src: Tuple[Aligner, Dict, Dict[str, int]], raw: str, morph_str: str) \
+        -> List[int]:
+    """
+    패치의 원문과 분석 결과를 음절단위 매핑(정렬)을 수행한다.
+    Args:
+        rsc_src:  (Aligner, restore dic, vocab out) resource triple
+        raw:  원문
+        morph_str:  형태소 분석 결과 (패치 기술 형식)
+    Returns:
+        정렬에 기반한 출력 태그 번호
+    """
+    aligner, restore_dic, vocab_out = rsc_src
+    raw_words = raw.strip().split()
+    morphs = morph_str.split(' + ')
+    morphs_strip = morphs
+    if morphs[0] in [WORD_DELIM_STR, SENT_DELIM_STR]:
+        morphs_strip = morphs_strip[1:]
+    if morphs[-1] in [WORD_DELIM_STR, SENT_DELIM_STR]:
+        morphs_strip = morphs_strip[:-1]
+    morph_words = _split_list(morphs_strip, WORD_DELIM_STR)
+    tag_nums = []
+    restore_new = defaultdict(dict)
+    vocab_new = defaultdict(list)
+    for raw_word, morph_word in zip(raw_words, morph_words):
+        word = Word.parse('\t'.join(['', raw_word, ' + '.join(morph_word)]), '', 0)
+        try:
+            word_align = aligner.align(word)
+            _, word_tag_nums = align_to_tag(raw_word, word_align, (restore_dic, restore_new),
+                                            (vocab_out, vocab_new))
+            if restore_new or vocab_new:
+                logging.debug('needs dic update: %s', word)
+                return []
+        except AlignError as algn_err:
+            logging.debug('alignment error: %s', word)
+            logging.debug(str(algn_err))
+            return []
+        if tag_nums:
+            tag_nums.append(WORD_DELIM_NUM)
+        tag_nums.extend(word_tag_nums)
+    if morphs[0] in [WORD_DELIM_STR, SENT_DELIM_STR]:
+        tag_nums.insert(0, WORD_DELIM_NUM if morphs[0] == WORD_DELIM_STR else SENT_DELIM_NUM)
+    if morphs[-1] in [WORD_DELIM_STR, SENT_DELIM_STR]:
+        tag_nums.append(WORD_DELIM_NUM if morphs[-1] == WORD_DELIM_STR else SENT_DELIM_NUM)
+    return tag_nums
diff --git a/src/main/python/khaiii/resource/morphs.py b/src/main/python/khaiii/resource/morphs.py
index be6d232..80c85c9 100644
--- a/src/main/python/khaiii/resource/morphs.py
+++ b/src/main/python/khaiii/resource/morphs.py
@@ -127,3 +127,30 @@ def _parse_one(cls, morph_str: str) -> 'Morph':
         if tag not in TAG_SET:
             raise ParseError('invalid pos tag: {}'.format(tag))
         return Morph(lex, tag)
+
+
+#############
+# functions #
+#############
+def mix_char_tag(chars: str, tags: List[int]) -> List[int]:
+    """
+    음절과 출력 태그를 비트 연산으로 합쳐서 하나의 (32비트) 숫자로 표현한다.
+    Args:
+        chars:  음절 (유니코드) 리스트 (문자열)
+        tags:  출력 태그 번호의 리스트
+    Returns:
+        합쳐진 숫자의 리스트
+    """
+    char_nums = [ord(c) for c in chars]
+    if tags[0] == SENT_DELIM_NUM:
+        char_nums.insert(0, SENT_DELIM_NUM)
+    if tags[-1] == SENT_DELIM_NUM:
+        char_nums.append(SENT_DELIM_NUM)
+    for idx, char_num in enumerate(char_nums):
+        if char_num == ord(' '):
+            char_nums[idx] = WORD_DELIM_NUM
+            continue
+        elif tags[idx] == SENT_DELIM_NUM:
+            continue
+        char_nums[idx] = char_num << 12 | tags[idx]
+    return char_nums
diff --git a/src/main/python/khaiii/resource/resource.py b/src/main/python/khaiii/resource/resource.py
index 3514e81..110ee70 100644
--- a/src/main/python/khaiii/resource/resource.py
+++ b/src/main/python/khaiii/resource/resource.py
@@ -12,9 +12,10 @@
 # imports #
 ###########
 from argparse import Namespace
+from collections import defaultdict
 import logging
 import os
-from typing import Dict
+from typing import Dict, Tuple
 
 from khaiii.resource.vocabulary import Vocabulary
 
@@ -68,3 +69,53 @@ def _load_restore_dic(cls, path: str) -> Dict[str, str]:
             dic[key] = val
         logging.info('%s: %d entries', os.path.basename(path), len(dic))
         return dic
+
+
+#############
+# functions #
+#############
+def load_restore_dic(file_path: str) -> Dict[Tuple[str, str], Dict[int, str]]:
+    """
+    원형복원 사전을 로드한다.
+    Args:
+        file_path:  파일 경로
+    Returns:
+        사전
+    """
+    file_name = os.path.basename(file_path)
+    restore_dic = defaultdict(dict)
+    for line_num, line in enumerate(open(file_path, 'r', encoding='UTF-8'), start=1):
+        line = line.rstrip()
+        if not line or line[0] == '#':
+            continue
+        char_tag_num, mrp_chr_str = line.split('\t')
+        char, tag_num = char_tag_num.rsplit('/', 1)
+        tag, num = tag_num.rsplit(':', 1)
+        num = int(num)
+        if (char, tag) in restore_dic:
+            num_mrp_chrs_dic = restore_dic[char, tag]
+            if num in num_mrp_chrs_dic:
+                logging.error('%s:%d: duplicated with %s: %s', file_name, line_num,
+                              num_mrp_chrs_dic[num], line)
+                return {}
+        restore_dic[char, tag][num] = mrp_chr_str
+    return restore_dic
+
+
+def load_vocab_out(rsc_src: str) -> Dict[str, int]:
+    """
+    출력 태그 vocabulary를 로드한다.
+    Args:
+        rsc_src:  리소스 디렉토리
+    Returns:
+        출력 태그 vocabulary
+    """
+    file_path = '{}/vocab.out'.format(rsc_src)
+    vocab_out = [line.strip() for line in open(file_path, 'r', encoding='UTF-8')
+                 if line.strip()]
+    vocab_out_more = []
+    file_path = '{}/vocab.out.more'.format(rsc_src)
+    if os.path.exists(file_path):
+        vocab_out_more = [line.strip() for line in open(file_path, 'r', encoding='UTF-8')
+                          if line.strip()]
+    return {tag: idx for idx, tag in enumerate(vocab_out + vocab_out_more, start=1)}
diff --git a/train/bin/extract_errpatch.py b/train/bin/extract_errpatch.py
index ff9dd8c..7153eee 100755
--- a/train/bin/extract_errpatch.py
+++ b/train/bin/extract_errpatch.py
@@ -325,8 +325,11 @@ def main():
     """
     parser = ArgumentParser(description='오분석 패치 후보를 추출하는 스크립트')
     parser.add_argument('-c', '--corpus-dir', help='corpus dir', metavar='DIR', required=True)
-    parser.add_argument('--lib-path', help='khaiii shared library path', metavar='FILE', default='')
-    parser.add_argument('--rsc-dir', help='resource dir', metavar='DIR', default='')
+    parser.add_argument('--lib-path', help='khaiii shared library path' \
+                                           ' <default ../build/lib/libkhaiii.so>',
+                        metavar='FILE', default='../build/lib/libkhaiii.so')
+    parser.add_argument('--rsc-dir', help='resource dir <default: ../build/share/khaiii>',
+                        metavar='DIR', default='../build/share/khaiii')
     parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
     parser.add_argument('--with-freq', help='print with frequency', action='store_true')
     parser.add_argument('--min-freq', help='minimum frequency <default: 10>', metavar='FREQ',
diff --git a/train/bin/hd_validate_errpatch.bash b/train/bin/hd_validate_errpatch.bash
new file mode 100755
index 0000000..3042ef8
--- /dev/null
+++ b/train/bin/hd_validate_errpatch.bash
@@ -0,0 +1,194 @@
+#!/usr/bin/env bash
+set -e -u
+
+
+#############
+# functions #
+#############
+function print_usage() {
+    local msg=$1
+    echo "Usage: $(basename $0) [options]"
+    echo "Options:"
+    echo "  -h, --help        show this help message and exit"
+    echo "  -i FILE           input file"
+    echo "  -c DIR            corpus dir"
+    echo "  --rsc-src=DIR     <default: ../../rsc/src>"
+    echo "  --lib-path=FILE   <default: ../../build/lib/libkhaiii.so>"
+    echo "  --rsc-dir=DIR     <default: ../../build/share/khaiii>"
+    echo "  --num-mapper=NUM  <default: 1000>"
+    if [ -z "${msg}" ]; then
+        exit 0
+    else
+        echo
+        echo "${msg}"
+        exit 1
+    fi
+}
+
+
+function abspath() {
+    python3 -c "import os, sys; print(os.path.abspath(sys.argv[1]))" $1
+}
+
+
+function parse_args() {
+    INPUT_FILE=""
+    CORPUS_DIR=""
+    LIB_PATH=""
+    RSC_DIR=""
+    RSC_SRC=""
+    NUM_MAPPER=""
+
+    while [[ $# -ge 1 ]]; do
+        case $1 in
+            -h|--help)
+                print_usage ""
+                ;;
+            -i)
+                INPUT_FILE="$2"
+                shift
+                ;;
+            -c)
+                CORPUS_DIR="$2"
+                shift
+                ;;
+            --rsc-src)
+                RSC_SRC="$2"
+                shift
+                ;;
+            --lib-path)
+                LIB_PATH="$2"
+                shift
+                ;;
+            --rsc-dir)
+                RSC_DIR="$2"
+                shift
+                ;;
+            --num-mapper)
+                NUM_MAPPER="$2"
+                shift
+                ;;
+            --) break ;;
+        esac
+        shift
+    done
+
+    # input file 검사
+    if [ -z "${INPUT_FILE}" ]; then
+        print_usage "no input file"
+    fi
+
+    # corpus dir 검사
+    if [ -z "${CORPUS_DIR}" ]; then
+        print_usage "no corpus dir"
+    fi
+
+    if [ -z "${RSC_SRC}" ]; then
+        RSC_SRC=../../rsc/src
+    fi
+    if [ -z "${LIB_PATH}" ]; then
+        LIB_PATH=../../build/lib/libkhaiii.so
+    fi
+    if [ -z "${RSC_DIR}" ]; then
+        RSC_DIR=../../build/share/khaiii
+    fi
+    if [ -z "${NUM_MAPPER}" ]; then
+        NUM_MAPPER=1000
+    fi
+
+    INPUT_FILE=$(abspath ${INPUT_FILE})
+    LIB_PATH=$(abspath ${LIB_PATH})
+    RSC_DIR=$(abspath ${RSC_DIR})
+    RSC_SRC=$(abspath ${RSC_SRC})
+    CORPUS_DIR=$(abspath ${CORPUS_DIR})
+}
+
+
+function init_envs() {
+    # global variables
+    INPUT_DIR=errpatch.in
+    OUTPUT_DIR=errpatch.out
+    CACHE_DIR=errpatch.cache
+}
+
+
+function split_input() {
+    >&2 echo "{{{{{{{{{{ ${FUNCNAME[0]} {{{{{{{{{{"
+
+    local total_line
+    total_line=$(wc -l < ${INPUT_FILE})
+    local line_per_split=$((total_line / NUM_MAPPER))
+    rm -rf ${INPUT_DIR}
+    mkdir -p ${INPUT_DIR}
+    shuf ${INPUT_FILE} | split -d -a 5 -l ${line_per_split} - ${INPUT_DIR}/part-
+
+    hadoop fs -test -e ${INPUT_DIR} && hadoop fs -rm -skipTrash -r ${INPUT_DIR}
+    hadoop fs -put ${INPUT_DIR}
+
+    >&2 echo "}}}}}}}}}} ${FUNCNAME[0]} }}}}}}}}}}"
+}
+
+
+function cache_files() {
+    >&2 echo "{{{{{{{{{{ ${FUNCNAME[0]} {{{{{{{{{{"
+
+    hadoop fs -test -e ${CACHE_DIR} && hadoop fs -rm -skipTrash -r ${CACHE_DIR}
+    hadoop fs -mkdir -p ${CACHE_DIR}
+
+    hadoop fs -put ../../src/main/python/khaiii ${CACHE_DIR}
+    hadoop fs -mkdir -p ${CACHE_DIR}/khaiii/lib
+    hadoop fs -put ${LIB_PATH} ${CACHE_DIR}/khaiii/lib
+
+    hadoop fs -mkdir -p ${CACHE_DIR}/khaiii/share
+    hadoop fs -put ${RSC_DIR} ${CACHE_DIR}/khaiii/share/khaiii
+
+    hadoop fs -put ${RSC_SRC} ${CACHE_DIR}/rsc_src
+
+    hadoop fs -mkdir -p ${CACHE_DIR}/corpus
+    hadoop fs -put ${CORPUS_DIR}/*.txt ${CACHE_DIR}/corpus
+
+    >&2 echo "}}}}}}}}}} ${FUNCNAME[0]} }}}}}}}}}}"
+}
+
+
+function run_hadoop() {
+    >&2 echo "{{{{{{{{{{ ${FUNCNAME[0]} {{{{{{{{{{"
+
+    hadoop fs -test -e ${OUTPUT_DIR} && hadoop fs -rm -skipTrash -r ${OUTPUT_DIR}
+    yarn jar ${HADOOP_HOME}/share/hadoop/tools/lib/hadoop-streaming-*.jar \
+        -D mapred.job.name=validate_errpatch \
+        -D mapred.reduce.tasks=0 \
+        -cmdenv PYTHONPATH="./${CACHE_DIR}" \
+        -file ./validate_errpatch.py \
+        -input "${INPUT_DIR}" \
+        -output "${OUTPUT_DIR}" \
+        -cacheFile "${CACHE_DIR}#${CACHE_DIR}" \
+        -mapper "./validate_errpatch.py -c ./${CACHE_DIR}/corpus --rsc-src ./${CACHE_DIR}/rsc_src"
+
+    hadoop fs -text ${OUTPUT_DIR}/part-* > "$(dirname ${INPUT_FILE})/errpatch.valid"
+
+    >&2 echo "}}}}}}}}}} ${FUNCNAME[0]} }}}}}}}}}}"
+}
+
+
+function del_temp() {
+    >&2 echo "{{{{{{{{{{ ${FUNCNAME[0]} {{{{{{{{{{"
+
+    hadoop fs -rm -skipTrash -r ${INPUT_DIR} ${OUTPUT_DIR} ${CACHE_DIR}
+    rm -rf ${INPUT_DIR}
+
+    >&2 echo "}}}}}}}}}} ${FUNCNAME[0]} }}}}}}}}}}"
+}
+
+
+########
+# main #
+########
+parse_args $@
+cd "$(dirname $0)"
+init_envs
+
+# split_input
+# cache_files
+run_hadoop
+del_temp
diff --git a/train/bin/train.py b/train/bin/train.py
index b696d96..7684be4 100755
--- a/train/bin/train.py
+++ b/train/bin/train.py
@@ -42,8 +42,8 @@ def main():
     parser = ArgumentParser(description='train model from data')
     parser.add_argument('-i', '--in-pfx', help='input data path prefix', metavar='NAME',
                         required=True)
-    parser.add_argument('--rsc-src', help='train resource dir <default: ../rsc/src>', metavar='DIR',
-                        default='../rsc/src')
+    parser.add_argument('--rsc-src', help='resource source dir <default: ../rsc/src>',
+                        metavar='DIR', default='../rsc/src')
     parser.add_argument('--logdir', help='tensorboard log dir <default: ./logdir>', metavar='DIR',
                         default='./logdir')
     parser.add_argument('--window', help='left/right character window length <default: 3>',
diff --git a/train/bin/validate_errpatch.py b/train/bin/validate_errpatch.py
new file mode 100755
index 0000000..ed73271
--- /dev/null
+++ b/train/bin/validate_errpatch.py
@@ -0,0 +1,270 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+추출된 오분석 패치 후보를 검증하는 스크립트
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+from collections import defaultdict
+import io
+import logging
+import os
+import sys
+from typing import Dict, Iterator, List, Tuple
+
+from khaiii.khaiii import KhaiiiApi
+
+from khaiii.munjong.sejong_corpus import Sentence, sents
+from khaiii.resource.char_align import Aligner, AlignError, align_patch, align_to_tag
+from khaiii.resource.morphs import mix_char_tag, WORD_DELIM_NUM, SENT_DELIM_NUM
+from khaiii.resource.resource import load_restore_dic, load_vocab_out
+
+
+#########
+# types #
+#########
+class StrFile(io.StringIO):
+    """
+    StringIO 객체에 name 멤버를 추가해 파일 객체인 것 처럼 동작하도록 하기 위해
+    """
+    def __init__(self, name, buf):
+        """
+        Args:
+            name:  file name
+            buf:  buffer
+        """
+        super().__init__(buf)
+        self.name = name
+
+
+#############
+# variables #
+#############
+_CORPUS = []    # corpus cache
+
+
+##########
+# mapper #
+##########
+def _sent_iter(args: Namespace) -> Iterator[Sentence]:
+    """
+    sentence generator
+    Args:
+        args:  arguments
+    Yields:
+        sentence
+    """
+    if not _CORPUS:
+        for name in sorted(os.listdir(args.corpus_dir)):
+            if not name.endswith('.txt'):
+                continue
+            path = '{}/{}'.format(args.corpus_dir, name)
+            _CORPUS.append((name, open(path, 'r', encoding='UTF-8').read()))
+        _CORPUS.sort()
+
+    for name, lines in _CORPUS:
+        logging.info(name)
+        for sent in sents(StrFile(name, lines)):
+            yield sent
+
+
+def _find_list(haystack: list, needle: list) -> int:
+    """
+    Return the index at which the sequence needle appears in the sequence haystack,
+    or -1 if it is not found, using the Boyer-Moore-Horspool algorithm.
+    The elements of needle and haystack must be hashable.
+
+    >>> _find_list([10, 10, 20], [10, 20])
+    1
+
+    Args:
+        haystack:  list to find
+        needle:  pattern list
+    Returns:
+        start index. -1 if not found
+    """
+    h_len = len(haystack)
+    n_len = len(needle)
+    skip = {needle[i]: n_len - i - 1 for i in range(n_len - 1)}
+    idx = n_len - 1
+    while idx < h_len:
+        for jdx in range(n_len):
+            if haystack[idx - jdx] != needle[-jdx - 1]:
+                idx += skip.get(haystack[idx], n_len)
+                break
+        else:
+            return idx - n_len + 1
+    return -1
+
+
+def _align_sent(rsc_src: Tuple[Aligner, dict, Dict[str, str]], sent: Sentence) -> List[int]:
+    """
+    세종 문장 객체를 정렬하여 음절 별 출력 태그의 벡터로 표현한다.
+    Args:
+        rsc_src:  (Aligner, restore dic, vocab out) resource triple
+        sent:  sejong_corpus.Sentence 객체
+    Returns:
+        list of output tag numbers. empty list for alignment error
+    """
+    aligner, restore_dic, vocab_out = rsc_src
+    tag_nums = []
+    restore_new = defaultdict(dict)
+    vocab_new = defaultdict(list)
+    for word in sent.words:
+        try:
+            word_align = aligner.align(word)
+            _, word_tag_nums = align_to_tag(word.raw, word_align, (restore_dic, restore_new),
+                                            (vocab_out, vocab_new))
+        except AlignError as algn_err:
+            logging.debug('alignment error: %s', word)
+            logging.debug(str(algn_err))
+            return []
+        if tag_nums:
+            tag_nums.append(WORD_DELIM_NUM)
+        tag_nums.extend(word_tag_nums)
+    tag_nums.insert(0, SENT_DELIM_NUM)
+    tag_nums.append(SENT_DELIM_NUM)
+    return tag_nums
+
+
+def _analyze_sent(khaiii_api: KhaiiiApi, raw_sent: str) -> List[int]:
+    """
+    원시 문장에 대해 패치를 적용하지 않은 음절별 태깅 결과를 얻는다.
+    Args:
+        khaiii_api:  khaiii API 객체
+        raw_sent:  원시 문장
+    Returns:
+        list of output tag numbers
+    """
+    tag_nums = khaiii_api.analyze_bfr_errpatch(raw_sent, '')
+    logging.debug(tag_nums)
+    return tag_nums
+
+
+def _cnt_pos_neg(khaiii_api: KhaiiiApi, patch_raw: str, alignment: Tuple[list, list],
+                 rsc_src: Tuple[Aligner, dict, Dict[str, str]], sent: Sentence) -> Tuple[int, int]:
+    """
+    오분석을 정분석으로 바꾼 횟수와, 오분석을 다른 오분석으로 바꾼 횟수를 센다.
+    Args:
+        khaiii_api:  khaiii API object
+        patch_raw:  raw part of patch
+        alignment:  (left, right) alignment pair
+        rsc_src:  (Aligner, restore dic, vocab out) resource triple
+        sent:  Sentence object
+    Returns:
+        오분석 -> 정분석 횟수
+        오분석 -> 오분석 횟수
+    """
+    raw_sent = sent.raw_str()
+    if patch_raw not in raw_sent:
+        # 원문이 문장에서 발견되지 않으면 스킵
+        return 0, 0
+    aligner, restore_dic, vocab_out = rsc_src
+    sent_align = _align_sent((aligner, restore_dic, vocab_out), sent)
+    if not sent_align:
+        # 코퍼스 정답이 원문과 정렬이 되지 않고 오류가 발생하면 스킵
+        return 0, 0
+    left_align, right_align = alignment
+    left_needle = mix_char_tag(patch_raw, left_align)
+    sent_anal = khaiii_api.analyze_bfr_errpatch(raw_sent, '')
+    sent_haystack = mix_char_tag(raw_sent, sent_anal)
+    pos_cnt = 0
+    neg_cnt = 0
+    found = _find_list(sent_haystack, left_needle)
+    while found >= 0:
+        # 패치의 좌측 오분석 열이 분석 결과에서 나타난 경우 우측 정답 열과 코퍼스를 비교
+        right_corpus = sent_align[found:found + len(left_needle)]
+        if right_align == right_corpus:
+            pos_cnt += 1
+        else:
+            neg_cnt += 1
+        del sent_haystack[:found + len(left_needle)]
+        found = _find_list(sent_haystack, left_needle)
+    return pos_cnt, neg_cnt
+
+
+def run(args: Namespace):
+    """
+    actual function which is doing some task
+    Args:
+        args:  program arguments
+    """
+    aligner = Aligner(args.rsc_src)
+    restore_dic = load_restore_dic('{}/restore.dic'.format(args.rsc_src))
+    if not restore_dic:
+        sys.exit(1)
+    vocab_out = load_vocab_out(args.rsc_src)
+
+    khaiii_api = KhaiiiApi(args.lib_path, args.rsc_dir, '{"errpatch": false}')
+
+    for line_num, line in enumerate(sys.stdin, start=1):
+        line = line.rstrip('\r\n')
+        if not line or line[0] == '#':
+            continue
+        raw, left, right = line.split('\t')
+        left_align = align_patch((aligner, restore_dic, vocab_out), raw, left)
+        if not left_align:
+            logging.info('invalid %d-th line: left align: %s', line_num, line)
+            continue
+        right_align = align_patch((aligner, restore_dic, vocab_out), raw, right)
+        if not right_align:
+            logging.info('invalid %d-th line: right align: %s', line_num, line)
+            continue
+        if len(left_align) != len(right_align):
+            logging.info('invalid %d-th line: left/right diff: %s', line_num, line)
+            continue
+        pos_cnt = 0
+        neg_cnt = 0
+        for sent in _sent_iter(args):
+            pos_cnt_sent, neg_cnt_sent = _cnt_pos_neg(khaiii_api, raw, (left_align, right_align),
+                                                      (aligner, restore_dic, vocab_out), sent)
+            pos_cnt += pos_cnt_sent
+            neg_cnt += neg_cnt_sent
+            if neg_cnt > 0:
+                break
+        if neg_cnt > 0 or pos_cnt == 0:
+            logging.info('invalid %d-th line: +%d, -%d: %s', line_num, pos_cnt, neg_cnt, line)
+            continue
+        print('{}\t{}\t{}'.format(raw, left, right))
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='추출된 오분석 패치 후보를 검증하는 스크립트')
+    parser.add_argument('-c', '--corpus-dir', help='corpus dir', metavar='DIR', required=True)
+    parser.add_argument('--rsc-src', help='resource source dir <default: ../rsc/src>',
+                        metavar='DIR', default='../rsc/src')
+    parser.add_argument('--lib-path', help='khaiii shared library path', metavar='FILE', default='')
+    parser.add_argument('--rsc-dir', help='resource dir', metavar='DIR', default='')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From b351a8f702047b381f8f02ab25cf5c540da15871 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 28 Jan 2019 00:20:19 +0900
Subject: [PATCH 14/41] =?UTF-8?q?=EC=BD=94=ED=8D=BC=EC=8A=A4=EC=9D=98=20?=
 =?UTF-8?q?=ED=8C=A8=EC=B9=98=EB=A5=BC=20=EC=83=9D=EC=84=B1=ED=95=98?=
 =?UTF-8?q?=EA=B3=A0=20=EC=A0=81=EC=9A=A9=ED=95=98=EB=8A=94=20=EC=8A=A4?=
 =?UTF-8?q?=ED=81=AC=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#3?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 munjong/bin/apply_patch.py                 |  79 ++++
 munjong/bin/make_patch.py                  |  82 ++++
 src/main/python/khaiii/munjong/libpatch.py | 426 +++++++++++++++++++++
 3 files changed, 587 insertions(+)
 create mode 100755 munjong/bin/apply_patch.py
 create mode 100755 munjong/bin/make_patch.py
 create mode 100644 src/main/python/khaiii/munjong/libpatch.py

diff --git a/munjong/bin/apply_patch.py b/munjong/bin/apply_patch.py
new file mode 100755
index 0000000..44562ec
--- /dev/null
+++ b/munjong/bin/apply_patch.py
@@ -0,0 +1,79 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+apply patch to original Sejong corpus
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+import logging
+import os
+import shutil
+
+from khaiii.munjong import libpatch
+
+
+#############
+# functions #
+#############
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    if not os.path.exists(args.modified):
+        logging.info('creating modified corpus dir: %s', args.modified)
+        os.mkdir(args.modified)
+
+    for name in sorted(os.listdir(args.original)):
+        if not name.endswith('.txt'):
+            continue
+        org_path = '%s/%s' % (args.original, name)
+        mod_path = '%s/%s' % (args.modified, name)
+        patch_path = '%s/%s.patch' % (args.patch, name[:-len('.txt')])
+        if os.path.exists(patch_path):
+            logging.info('[%s] + [%s] = [%s]', org_path, patch_path, mod_path)
+            libpatch.apply(org_path, args.org_enc, patch_path, mod_path, args.mod_enc)
+        else:
+            logging.info('[%s] = [%s]', org_path, mod_path)
+            shutil.copyfile(org_path, mod_path)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='apply patch to original Sejong corpus')
+    parser.add_argument('-o', '--original', help='original corpus dir', metavar='DIR',
+                        required=True)
+    parser.add_argument('-p', '--patch', help='patch dir', metavar='DIR', required=True)
+    parser.add_argument('-m', '--modified', help='modified corpus output dir', metavar='DIR',
+                        required=True)
+    parser.add_argument('--org-enc', help='original corpus encoding <default: UTF-16>',
+                        metavar='ENCODING', default='UTF-16')
+    parser.add_argument('--mod-enc', help='modified corpus encoding <default: UTF-8>',
+                        metavar='ENCODING', default='UTF-8')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/make_patch.py b/munjong/bin/make_patch.py
new file mode 100755
index 0000000..25e3bd1
--- /dev/null
+++ b/munjong/bin/make_patch.py
@@ -0,0 +1,82 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+make patch from two Sejong corpora
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+import logging
+import os
+
+from khaiii.munjong import libpatch
+
+
+#############
+# functions #
+#############
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    if not os.path.exists(args.patch):
+        logging.info('creating patch dir: %s', args.patch)
+        os.mkdir(args.patch)
+
+    for name in sorted(os.listdir(args.original)):
+        if not name.endswith('.txt'):
+            continue
+        org_path = '%s/%s' % (args.original, name)
+        mod_path = '%s/%s' % (args.modified, name)
+        patch_path = '%s/%s.patch' % (args.patch, name[:-len('.txt')])
+        logging.info('[%s] - [%s] = [%s]', org_path, mod_path, patch_path)
+        patches = libpatch.make(org_path, args.org_enc, mod_path, args.mod_enc)
+        if patches:
+            logging.info('creating patch file: %s', patch_path)
+            with open(patch_path, 'w', encoding='UTF-8') as fout:
+                for patch in patches:
+                    print(patch, file=fout)
+        elif os.path.exists(patch_path):
+            logging.info('removing existing patch file: %s', patch_path)
+            os.remove(patch_path)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='make patch from two Sejong corpora')
+    parser.add_argument('-o', '--original', help='original corpus dir', metavar='DIR',
+                        required=True)
+    parser.add_argument('-m', '--modified', help='modified corpus dir', metavar='DIR',
+                        required=True)
+    parser.add_argument('-p', '--patch', help='patch output dir', metavar='DIR', required=True)
+    parser.add_argument('--org-enc', help='original corpus encoding <default: UTF-16>',
+                        metavar='ENCODING', default='UTF-16')
+    parser.add_argument('--mod-enc', help='modified corpus encoding <default: UTF-8>',
+                        metavar='ENCODING', default='UTF-8')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/src/main/python/khaiii/munjong/libpatch.py b/src/main/python/khaiii/munjong/libpatch.py
new file mode 100644
index 0000000..13e3ecf
--- /dev/null
+++ b/src/main/python/khaiii/munjong/libpatch.py
@@ -0,0 +1,426 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+patch library for Sejong corpus
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from collections import namedtuple
+import logging
+import os
+from typing import Dict, List, Tuple
+
+from khaiii.munjong.sejong_corpus import SENT_OPEN_TAGS, SENT_CLOSE_TAGS, WORD_ID_PTN
+
+
+#############
+# constatns #
+#############
+# line types
+WORD_TYPE = 1    # word (eojeol)
+BOS_TYPE = 2    # begin of sentence markup
+EOS_TYPE = 3    # end of sentence markup
+MISC_TYPE = 0    # other miscellaneous lines
+
+
+#########
+# types #
+#########
+Line = namedtuple('Line', ['type_', 'wid', 'content'])
+
+
+class Patch:    # pylint: disable=too-few-public-methods
+    """
+    patch line
+    """
+    # patch types (actions, operations)
+    SENT_SPLIT = 1
+    SENT_MERGE = 2
+    WORD_REPLACE = 3
+    WORD_INSERT = 4
+    WORD_DELETE = 5
+    # patch categories
+    SENT_CATE = 1
+    WORD_CATE = 2
+
+    def __init__(self, type_: int, wid: str, content: str):
+        """
+        Args:
+            type_:  patch type
+            wid:  word ID
+            content:  word content (tab separated second column)
+        """
+        self.type_ = type_
+        self.wid = wid
+        self.content = content
+
+    def __eq__(self, other: 'Patch') -> bool:
+        return self.type_ == other.type_ and self.wid == other.wid and self.content == other.content
+
+    def __cmp__(self, other: 'Patch') -> int:
+        if self.cate() == other.cate():
+            if self.wid < other.wid:
+                return -1
+            if self.wid > other.wid:
+                return 1
+            return 0
+        return self.cate() - other.cate()
+
+    def __lt__(self, other: 'Patch') -> bool:
+        return self.__cmp__(other) < 0
+
+    def __str__(self) -> str:
+        if self.type_ == self.WORD_REPLACE:
+            return '=\t%s\t%s' % (self.wid, self.content)
+        if self.type_ == self.WORD_INSERT:
+            return '+\t%s\t%s' % (self.wid, self.content)
+        if self.type_ == self.WORD_DELETE:
+            return '-\t%s' % self.wid
+        if self.type_ == self.SENT_SPLIT:
+            return 'S\t%s\t%s' % (self.wid, self.content)
+        if self.type_ == self.SENT_MERGE:
+            return 'M\t%s\t%s' % (self.wid, self.content)
+        raise RuntimeError('unknown patch type: %d' % self.type_)
+
+    def cate(self) -> int:
+        """
+        get patch category. EOS/BOS patches are 0. word patches are 1
+        Returns:
+            category number
+        """
+        return self.SENT_CATE if self.type_ in [self.SENT_SPLIT, self.SENT_MERGE] \
+                              else self.WORD_CATE
+
+    @classmethod
+    def parse(cls, line: str) -> 'Patch':
+        """
+        parse patch line
+        Args:
+            line:  patch line
+        Returns:
+            Patch object
+        """
+        cols = line.split('\t', 2)
+        if len(cols) == 3:
+            if cols[0] == '=':
+                return Patch(cls.WORD_REPLACE, cols[1], cols[2])
+            if cols[0] == '+':
+                return Patch(cls.WORD_INSERT, cols[1], cols[2])
+            if cols[0].upper() == 'M':
+                return Patch(cls.SENT_MERGE, cols[1], cols[2])
+            if cols[0].upper() == 'S':
+                return Patch(cls.SENT_SPLIT, cols[1], cols[2])
+        elif len(cols) == 2:
+            if cols[0] == '-':
+                return Patch(cls.WORD_DELETE, cols[1], '')
+        raise RuntimeError('invalid patch format: %s' % line)
+
+
+#############
+# functions #
+#############
+def _load_corpus(path: str, enc: str) -> Tuple[List[Line], Dict[str, int]]:
+    """
+    load corpus
+    Args:
+        path:  file path
+        enc:  file encoding
+    Returns:
+        list of lines
+        word ID dic
+    """
+    lines = []
+    wid_dic = {}
+    for line in open(path, 'r', encoding=enc):
+        line = line.rstrip('\r\n')
+        if WORD_ID_PTN.match(line):
+            wid, content = line.split('\t', 1)
+            if wid in wid_dic:
+                raise RuntimeError('duplicated word ID: %s' % line)
+            else:
+                wid_dic[wid] = len(lines)
+                lines.append(Line(WORD_TYPE, wid, content))
+        elif line in SENT_OPEN_TAGS:
+            lines.append(Line(BOS_TYPE, None, line))
+        elif line in SENT_CLOSE_TAGS:
+            lines.append(Line(EOS_TYPE, None, line))
+        else:
+            lines.append(Line(MISC_TYPE, None, line))
+            if line == '</tei.2>' and 'BTJO0443.txt' in path:
+                break
+    return lines, wid_dic
+
+
+def _make_sent_patch(org_lines: List[Line], mod_lines: List[Line]) -> List[Patch]:
+    """
+    make EOS/BOS patch
+    Args:
+        org_lines:  original lines
+        mod_lines:  modified lines
+    Returns:
+        EOS/BOS patches
+    """
+    def _get_eos_bos(lines: List[Line]) -> Tuple[str, str]:
+        """
+        get all EOS/BOS from lines
+        Args:
+            lines:  lines of corpus
+        Returns:
+            prev. word ID
+            next word ID
+        """
+        eos_bos = []
+        for idx, line in enumerate(lines):
+            if line.type_ == EOS_TYPE and (idx+2) < len(lines) and lines[idx+1].type_ == BOS_TYPE \
+                    and lines[idx-1].type_ == WORD_TYPE and lines[idx+2].type_ == WORD_TYPE:
+                prev_wid = lines[idx-1].wid
+                next_wid = lines[idx+2].wid
+                eos_bos.append((prev_wid[:17], next_wid[:17]))
+        return eos_bos
+
+    org_eos_bos = set(_get_eos_bos(org_lines))
+    mod_eos_bos = set(_get_eos_bos(mod_lines))
+    patches = []
+    for prev_wid, next_wid in org_eos_bos - mod_eos_bos:
+        patches.append(Patch(Patch.SENT_MERGE, prev_wid, next_wid))
+    for prev_wid, next_wid in mod_eos_bos - org_eos_bos:
+        patches.append(Patch(Patch.SENT_SPLIT, prev_wid, next_wid))
+    return sorted(patches)
+
+
+def _make_word_patch(org_lines: List[Line], org_wid_dic: Dict[str, int], mod_lines: List[Line],
+                     mod_wid_dic: Dict[str, int]) -> List[Patch]:
+    """
+    make word patch
+    Args:
+        org_lines:  original lines
+        org_wid_dic:  original word ID dictionary
+        mod_lines:  modified lines
+        mod_wid_dic:  modified word ID dictionary
+    Returns:
+        word patches
+    """
+    patches = []
+    for org_line in org_lines:
+        if org_line.type_ != WORD_TYPE:
+            continue
+        if org_line.wid in mod_wid_dic:
+            mod_line = mod_lines[mod_wid_dic[org_line.wid]]
+            if org_line.content != mod_line.content:
+                patches.append(Patch(Patch.WORD_REPLACE, mod_line.wid, mod_line.content))
+        else:
+            patches.append(Patch(Patch.WORD_DELETE, org_line.wid, ''))
+    for mod_line in mod_lines:
+        if mod_line.type_ != WORD_TYPE or mod_line.wid in org_wid_dic:
+            continue
+        patches.append(Patch(Patch.WORD_INSERT, mod_line.wid, mod_line.content))
+    return sorted(patches)
+
+
+def make(org_path: str, org_enc: str, mod_path: str, mod_enc: str) -> List[Patch]:
+    """
+    make patch from two file
+    Args:
+        org_path:  original file path
+        org_enc:  original file encoding
+        mod_path:  modified file path
+        mod_enc:  modified file encoding
+    Returns:
+        patch contents (list of patch lines)
+    """
+    org_lines, org_wid_dic = _load_corpus(org_path, org_enc)
+    mod_lines, mod_wid_dic = _load_corpus(mod_path, mod_enc)
+
+    sent_patches = _make_sent_patch(org_lines, mod_lines)
+    word_patches = _make_word_patch(org_lines, org_wid_dic, mod_lines, mod_wid_dic)
+    return sent_patches + word_patches
+
+
+def _load_patch(patch_path: str) -> List[Patch]:
+    """
+    load patch from file
+    Args:
+        patch_path:  patch path
+    Returns:
+        patches
+    """
+    patch_name = os.path.basename(patch_path)
+    wid_dic = set()
+    patches = []
+    for line_num, line in enumerate(open(patch_path, 'r', encoding='UTF-8'), start=1):
+        line = line.rstrip('\r\n')
+        if not line:
+            continue
+        patch = Patch.parse(line)
+        if patch.cate() == Patch.SENT_CATE:
+            if (patch.wid, patch.content) in wid_dic:
+                logging.error('%s(%d): patch conflict: %s', patch_name, line_num, line)
+            else:
+                patches.append(patch)
+                wid_dic.add((patch.wid, patch.content))
+        else:
+            if patch.wid in wid_dic:
+                logging.error('%s(%d): patch conflict: %s', patch_name, line_num, line)
+            else:
+                patches.append(patch)
+                wid_dic.add(patch.wid)
+    return patches
+
+
+def _apply_sent_merge_patch(org_lines: List[Line], patches: List[Patch]) -> List[Line]:
+    """
+    apply merge EOS/BOS patches
+    Args:
+        org_lines:  original lines
+        patches:  patches
+    Returns:
+        modified lines
+    """
+    merge_patches = {patch.wid: patch.content for patch in patches \
+                     if patch.type_ == Patch.SENT_MERGE}
+    mod_lines = []
+    idx = 0
+    while idx < len(org_lines):
+        org_line = org_lines[idx]
+        if org_line.type_ == EOS_TYPE and org_lines[idx+1].type_ == BOS_TYPE and \
+                        org_lines[idx-1].type_ == WORD_TYPE and org_lines[idx+2].type_ == WORD_TYPE:
+            prev_wid = org_lines[idx-1].wid[:17]
+            next_wid = org_lines[idx+2].wid[:17]
+            if prev_wid in merge_patches and merge_patches[prev_wid] == next_wid:
+                del merge_patches[prev_wid]
+                idx += 2
+                continue
+        mod_lines.append(org_line)
+        idx += 1
+    if merge_patches:
+        for prev_wid, next_wid in merge_patches.items():
+            logging.error('remaining merge sentence patches: %s\t%s', prev_wid, next_wid)
+    return mod_lines
+
+
+def _build_wid_dic(lines: List[Line]) -> Dict[str, int]:
+    """
+    build word ID dictionary
+    Args:
+        lines:  lines of corpus
+    Returns:
+        word ID dictionary
+    """
+    return {line.wid: idx for idx, line in enumerate(lines) if line.type_ == WORD_TYPE}
+
+
+def _apply_sent_split_patch(mod_lines: List[Line], patches: List[Patch]) -> List[Line]:
+    """
+    apply split EOS/BOS patches
+    Args:
+        mod_lines:  modified lines
+        patches:  patches
+    Returns:
+        modified lines
+    """
+    mod_wid_dic = _build_wid_dic(mod_lines)
+    split_patches = {patch.wid: patch.content for patch in patches \
+                     if patch.type_ == Patch.SENT_SPLIT}
+    for prev_wid, next_wid in sorted(split_patches.items(), key=lambda x: x[0], reverse=True):
+        idx = mod_wid_dic[next_wid]
+        assert mod_lines[idx].wid == next_wid
+        if mod_lines[idx-1].type_ == WORD_TYPE and mod_lines[idx-1].wid[:17] == prev_wid:
+            mod_lines.insert(idx, Line(BOS_TYPE, None, '<p>'))
+            mod_lines.insert(idx, Line(EOS_TYPE, None, '</p>'))
+            del split_patches[prev_wid]
+    if split_patches:
+        for prev_wid, next_wid in split_patches.items():
+            logging.error('remaining split sentence patches: %s\t%s', prev_wid, next_wid)
+    return mod_lines
+
+
+def _apply_word_del_rep_patch(mod_lines: List[Line], patches: List[Patch]) -> List[Line]:
+    """
+    apply word delete/replace patches
+    Args:
+        mod_lines:  modified lines
+        patches:  patches
+    Returns:
+        modified lines
+    """
+    delete_patches = {patch.wid for patch in patches if patch.type_ == Patch.WORD_DELETE}
+    replace_patches = {patch.wid: patch.content for patch in patches \
+                       if patch.type_ == Patch.WORD_REPLACE}
+    new_lines = []
+    for line in mod_lines:
+        if line.type_ == WORD_TYPE:
+            if line.wid in delete_patches:
+                delete_patches.remove(line.wid)
+                continue
+            elif line.wid in replace_patches:
+                new_lines.append(Line(WORD_TYPE, line.wid, replace_patches[line.wid]))
+                del replace_patches[line.wid]
+                continue
+        new_lines.append(line)
+    if delete_patches:
+        for wid in delete_patches:
+            logging.error('remaining delete word patches: %s', wid)
+    if replace_patches:
+        for wid, content in replace_patches.items():
+            logging.error('remaining replace word patches: %s\t%s', wid, content)
+    return new_lines
+
+
+def _apply_word_insert_patch(mod_lines: List[Line], patches: List[Patch]) -> List[Line]:
+    """
+    apply word insert patches
+    Args:
+        mod_lines:  modified lines
+        patches:  patches
+    Returns:
+        modified lines
+    """
+    insert_patches = sorted([(patch.wid, patch.content) for patch in patches \
+                             if patch.type_ == Patch.WORD_INSERT])
+    prev_idx = -1
+    curr_idx = 0
+    while curr_idx < len(mod_lines) and insert_patches:
+        curr_word = mod_lines[curr_idx]
+        if curr_word.type_ != WORD_TYPE:
+            curr_idx += 1
+            continue
+        wid, content = insert_patches[0]
+        if curr_word.wid < wid:
+            prev_idx = curr_idx
+            curr_idx += 1
+            continue
+        mod_lines.insert(prev_idx+1, Line(WORD_TYPE, wid, content))
+        del insert_patches[0]
+        prev_idx += 1
+        curr_idx += 1
+    return mod_lines
+
+
+def apply(org_path: str, org_enc: str, patch_path: str, mod_path: str, mod_enc: str):
+    """
+    apply path to original corpus then get modified corpus
+    Args:
+        org_path:  original file path
+        org_enc:  original file encoding
+        patch_path:  patch file path
+        mod_path:  modified file path
+        mod_enc:  modified file encoding
+    """
+    patches = _load_patch(patch_path)
+    org_lines, _ = _load_corpus(org_path, org_enc)
+    mod_lines = _apply_sent_merge_patch(org_lines, patches)
+    mod_lines = _apply_word_del_rep_patch(mod_lines, patches)
+    mod_lines = _apply_sent_split_patch(mod_lines, patches)
+    mod_lines = _apply_word_insert_patch(mod_lines, patches)
+    with open(mod_path, 'w', encoding=mod_enc) as fout:
+        for mod_line in mod_lines:
+            if mod_line.type_ == WORD_TYPE:
+                print('%s\t%s' % (mod_line.wid, mod_line.content), file=fout)
+            else:
+                print(mod_line.content, file=fout)

From 767d05734b845d89ed51e9d605d20217116842e4 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 28 Jan 2019 20:21:51 +0900
Subject: [PATCH 15/41] =?UTF-8?q?=EC=84=B8=EC=A2=85=20=EC=BD=94=ED=8D=BC?=
 =?UTF-8?q?=EC=8A=A4=EC=9D=98=20=EC=98=A4=EB=A5=98=EC=99=80=20=EA=B4=80?=
 =?UTF-8?q?=EB=A0=A8=EB=90=9C=20=EA=B0=81=EC=A2=85=20=EC=8A=A4=ED=81=AC?=
 =?UTF-8?q?=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#3?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 munjong/bin/convert_jamo_to_compat.py     |  79 ++++++++++++++
 munjong/bin/detect_sejong_period_error.py | 106 +++++++++++++++++++
 munjong/bin/fix_final_symbol_error.py     |  86 ++++++++++++++++
 munjong/bin/recover_english_case.py       |  96 +++++++++++++++++
 munjong/bin/recover_raw_morph_mismatch.py | 101 ++++++++++++++++++
 munjong/bin/recover_wide_quotation.py     | 108 +++++++++++++++++++
 munjong/bin/remove_sejong_period_error.py | 120 ++++++++++++++++++++++
 src/main/python/khaiii/resource/jaso.py   | 103 +++++++++++++++----
 8 files changed, 781 insertions(+), 18 deletions(-)
 create mode 100755 munjong/bin/convert_jamo_to_compat.py
 create mode 100755 munjong/bin/detect_sejong_period_error.py
 create mode 100755 munjong/bin/fix_final_symbol_error.py
 create mode 100755 munjong/bin/recover_english_case.py
 create mode 100755 munjong/bin/recover_raw_morph_mismatch.py
 create mode 100755 munjong/bin/recover_wide_quotation.py
 create mode 100755 munjong/bin/remove_sejong_period_error.py

diff --git a/munjong/bin/convert_jamo_to_compat.py b/munjong/bin/convert_jamo_to_compat.py
new file mode 100755
index 0000000..6ea1426
--- /dev/null
+++ b/munjong/bin/convert_jamo_to_compat.py
@@ -0,0 +1,79 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+한글 자모 영역의 코드를 호환 영역으로 변환
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import logging
+import sys
+
+from khaiii.munjong.sejong_corpus import WORD_ID_PTN
+from khaiii.resource.jaso import norm_compat
+
+
+#############
+# functions #
+#############
+def _norm(text: str) -> str:
+    """
+    정규화를 수행하는 함수
+    Args:
+        text:  입력 텍스트
+    Returns:
+        정규화된 텍스트
+    """
+    normalized = norm_compat(text)
+    normalized = normalized.replace('ᆞ', 'ㆍ')    # 0x119e -> 0x318d
+    normalized = normalized.replace('ᄝ', 'ㅱ')    # 0x111d -> 0x3171
+    return normalized
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    for line in sys.stdin:
+        line = line.rstrip('\r\n')
+        if not WORD_ID_PTN.match(line):
+            print(line)
+            continue
+        wid, word, morph = line.split('\t')
+        print('{}\t{}\t{}'.format(wid, _norm(word), _norm(morph)))
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='한글 자모 영역의 코드를 호환 영역으로 변환')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/detect_sejong_period_error.py b/munjong/bin/detect_sejong_period_error.py
new file mode 100755
index 0000000..7dfd4fb
--- /dev/null
+++ b/munjong/bin/detect_sejong_period_error.py
@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+detect period error of Sejong corpus
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import logging
+import os
+import re
+import sys
+from typing import Iterator, TextIO, Tuple
+
+from khaiii.munjong.sejong_corpus import Morph, WORD_ID_PTN
+
+
+#############
+# functions #
+#############
+def _get_two_lines(fin: TextIO) -> Iterator[Tuple[str, str]]:
+    """
+    get two lines tuple from file (generator)
+    Args:
+        fin:  input file
+    Yields:
+        current line
+        next line
+    """
+    curr_line = fin.readline().rstrip('\r\n')
+    for next_line in fin:
+        next_line = next_line.rstrip('\r\n')
+        yield curr_line, next_line
+        curr_line = next_line
+
+
+def _is_correct_eos(line: str) -> bool:
+    """
+    whether correct end of sentence or not
+    Args:
+        line:  line (word)
+    Returns:
+        whether correct or not
+    """
+    _, _, morphs_str = line.split('\t')
+    if re.match(r'.+/EF \+ ./SF$', morphs_str):
+        return True
+    if re.match(r'.+/SF \+ [\'"’”」\]]/SS$', morphs_str):
+        return True
+    morphs = [Morph.parse(_) for _ in morphs_str.split(' + ')]
+    tags_str = '+'.join([_.tag for _ in morphs])
+    if tags_str.endswith('+SF+SS+JKQ') or tags_str.endswith('+SF+SS+VCP+ETM'):
+        return True
+    return False
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    file_name = os.path.basename(sys.stdin.name)
+    for line_num, (curr_line, next_line) in enumerate(_get_two_lines(sys.stdin), start=1):
+        cols = curr_line.split('\t')
+        if len(cols) != 3 or not WORD_ID_PTN.match(cols[0]):
+            continue
+        if '/SF + ' not in cols[2] or not next_line.startswith('</'):
+            continue
+        if _is_correct_eos(curr_line):
+            continue
+        print('{}:{}\t{}'.format(file_name, line_num, curr_line))
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='detect period error of Sejong corpus')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'rt')
+    if args.output:
+        sys.stdout = open(args.output, 'wt')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/fix_final_symbol_error.py b/munjong/bin/fix_final_symbol_error.py
new file mode 100755
index 0000000..5ba1644
--- /dev/null
+++ b/munjong/bin/fix_final_symbol_error.py
@@ -0,0 +1,86 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+fix final symbol errors on Sejong corpus
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import logging
+import os
+import sys
+
+from khaiii.munjong.sejong_corpus import Morph, Word, WORD_ID_PTN
+
+
+#############
+# functions #
+#############
+def _attach_missing_symbol(word: Word):
+    """
+    attach missing symbol
+    Args:
+        word:  Word object
+    """
+    raw_word = word.raw
+    raw_morph = ''.join([_.lex for _ in word.morphs])
+    if not raw_word.startswith(raw_morph) or len(raw_word) != len(raw_morph)+1:
+        return
+    last_symbol = raw_word[-1]
+    if last_symbol == '.' and word.morphs[-1].tag == 'EC':
+        word.morphs.append(Morph('.', 'SF'))
+    elif last_symbol == ',':
+        word.morphs.append(Morph(',', 'SP'))
+    elif last_symbol == '"':
+        word.morphs.append(Morph('"', 'SS'))
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    file_name = os.path.basename(sys.stdin.name)
+    for line_num, line in enumerate(sys.stdin, start=1):
+        line = line.rstrip('\r\n')
+        if not WORD_ID_PTN.match(line):
+            print(line)
+            continue
+        word = Word.parse(line, file_name, line_num)
+        _attach_missing_symbol(word)
+        print(word)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='fix final symbol errors on Sejong corpus')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'rt')
+    if args.output:
+        sys.stdout = open(args.output, 'wt')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/recover_english_case.py b/munjong/bin/recover_english_case.py
new file mode 100755
index 0000000..5a0dd9d
--- /dev/null
+++ b/munjong/bin/recover_english_case.py
@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+recover cases of English letters in Sejong corpus
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import copy
+import logging
+import os
+import re
+import sys
+
+from khaiii.munjong.sejong_corpus import Word, WORD_ID_PTN
+
+
+#############
+# functions #
+#############
+def _recover(word: Word):
+    """
+    recover cases
+    Args:
+        word:  Word object
+    """
+    word_letters = [_ for _ in word.raw if re.match(r'[a-zA-Z]', _)]
+    letter_idx = -1
+    is_recovered = False
+    word_copy = copy.deepcopy(word)
+    for morph in word_copy.morphs:
+        for idx, char in enumerate(morph.lex):
+            if not re.match(r'[a-zA-Z]', char):
+                continue
+            letter_idx += 1
+            if word_letters[letter_idx] == char:
+                continue
+            morph.lex = morph.lex[:idx] + word_letters[letter_idx] + morph.lex[idx+1:]
+            is_recovered = True
+    if is_recovered:
+        logging.info('%s  =>  %s', str(word), word_copy.morph_str())
+        word.morphs = word_copy.morphs
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    file_name = os.path.basename(sys.stdin.name)
+    for line_num, line in enumerate(sys.stdin, start=1):
+        line = line.rstrip('\r\n')
+        if not WORD_ID_PTN.match(line):
+            print(line)
+            continue
+        word = Word.parse(line, file_name, line_num)
+        try:
+            _recover(word)
+        except IndexError as idx_err:
+            logging.error('%s(%d): %s: %s', file_name, line_num, idx_err, word)
+        print(word)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='recover cases of English letters in Sejong corpus')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/recover_raw_morph_mismatch.py b/munjong/bin/recover_raw_morph_mismatch.py
new file mode 100755
index 0000000..2bcb136
--- /dev/null
+++ b/munjong/bin/recover_raw_morph_mismatch.py
@@ -0,0 +1,101 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+어절의 원문과 형태소 분석 결과의 문자가 정규화하면 같지만 코드가 다른 경우 원문의 문자로 복원
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import logging
+import os
+import sys
+
+from khaiii.munjong.sejong_corpus import Morph, ParseError, Word, WORD_ID_PTN
+
+
+#############
+# functions #
+#############
+def _recover(line: str) -> str:
+    """
+    문자를 복원한다.
+    Args:
+        line:  어절 라인
+    Returns:
+        복원된 라인
+    """
+    wid, raw, morphs_str = line.split('\t')
+    raw_idx = 0
+    morphs = []
+    for token_str in morphs_str.split(' + '):
+        morph = Morph.parse(token_str)
+        lex = []
+        for _ in range(len(morph.lex)):
+            try:
+                lex.append(raw[raw_idx])
+                raw_idx += 1
+            except IndexError as idx_err:
+                logging.error(line)
+                raise idx_err
+        morph.lex = ''.join(lex)
+        morphs.append(morph)
+    morphs_new = ' + '.join([str(m) for m in morphs])
+    logging.debug('%s\t%s\t%s  =>  %s', wid, raw, morphs_str, morphs_new)
+    return '{}\t{}\t{}'.format(wid, raw, morphs_new)
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    file_name = os.path.basename(sys.stdin.name)
+    for line_num, line in enumerate(sys.stdin, start=1):
+        line = line.rstrip('\r\n')
+        if not WORD_ID_PTN.match(line):
+            print(line)
+            continue
+        try:
+            Word.parse(line, file_name, line_num)
+        except ParseError as par_err:
+            if 'raw-morph mismatch' in str(par_err):
+                line = _recover(line)
+            else:
+                raise par_err
+        print(line)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='어절의 원문과 형태소 분석 결과의 문자가 정규화하면 같지만 코드가 다른 경우'
+                                        ' 원문의 문자로 복원')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/recover_wide_quotation.py b/munjong/bin/recover_wide_quotation.py
new file mode 100755
index 0000000..0ee0792
--- /dev/null
+++ b/munjong/bin/recover_wide_quotation.py
@@ -0,0 +1,108 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+recover wide char quotations in Sejong corpus
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import logging
+import os
+import sys
+
+from khaiii.munjong.sejong_corpus import Word, WORD_ID_PTN
+
+
+#############
+# constants #
+#############
+_QUOT_NORM = {
+    '"': '"',
+    '“': '"',
+    '”': '"',
+    "'": "'",
+    "‘": "'",
+    "’": "'",
+    "`": "'",
+}
+
+
+#############
+# functions #
+#############
+def _recover(word: Word):
+    """
+    recover wide char quotations
+    Args:
+        word:  Word object
+    """
+    word_quots = [_ for _ in word.raw if _ in _QUOT_NORM]
+    morph_quots = []
+    for idx, morph in enumerate(word.morphs):
+        if morph.tag != 'SS' or morph.lex not in _QUOT_NORM:
+            continue
+        morph_quots.append((idx, morph))
+        quot_idx = len(morph_quots)-1
+        if len(word_quots) <= quot_idx or _QUOT_NORM[word_quots[quot_idx]] != _QUOT_NORM[morph.lex]:
+            logging.error('%d-th quots are different: %s', quot_idx+1, word)
+            return
+    if len(word_quots) != len(morph_quots):
+        morph_quots = [_ for _ in word.morph_str() if _ in _QUOT_NORM]
+        if word_quots != morph_quots:
+            logging.error('number of quots are different: %s', word)
+        return
+    for word_char, (idx, morph) in zip(word_quots, morph_quots):
+        if word_char == morph.lex:
+            continue
+        morph.lex = word_char
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    file_name = os.path.basename(sys.stdin.name)
+    for line_num, line in enumerate(sys.stdin, start=1):
+        line = line.rstrip('\r\n')
+        if not WORD_ID_PTN.match(line):
+            print(line)
+            continue
+        word = Word.parse(line, file_name, line_num)
+        _recover(word)
+        print(word)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='recover wide char quotations in Sejong corpus')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'rt')
+    if args.output:
+        sys.stdout = open(args.output, 'wt')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/remove_sejong_period_error.py b/munjong/bin/remove_sejong_period_error.py
new file mode 100755
index 0000000..59ffd86
--- /dev/null
+++ b/munjong/bin/remove_sejong_period_error.py
@@ -0,0 +1,120 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+remove wrong sentence breaking marks after period error eojeol
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2017-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import logging
+import os
+import re
+import sys
+from typing import TextIO, Tuple
+
+from khaiii.munjong.sejong_corpus import Morph, WORD_ID_PTN
+
+
+#############
+# functions #
+#############
+def _get_three_lines(fin: TextIO) -> Tuple[str, str, str]:
+    """
+    get three lines tuple from file (generator)
+    Args:
+        fin:  input file
+    Yields:
+        prev. prev. line
+        prev. line
+        curr. line
+    """
+    prev_prev_line = fin.readline().rstrip('\r\n')
+    prev_line = fin.readline().rstrip('\r\n')
+    # print first two lines
+    print(prev_prev_line)
+    print(prev_line)
+    for curr_line in fin:
+        curr_line = curr_line.rstrip('\r\n')
+        yield prev_prev_line, prev_line, curr_line
+        prev_prev_line = prev_line
+        prev_line = curr_line
+
+
+def _is_known_period_error_eojeol(line: str) -> bool:
+    """
+    알려진 특정 문장분리 오류를 포함하는 어절인 지 여부
+    Args:
+        line:  line (eojeol)
+    Returns:
+        whether has error or not
+    """
+    cols = line.split('\t')
+    if len(cols) != 3 or not WORD_ID_PTN.match(cols[0]):
+        return False
+    if '/SF + ' not in cols[2] or re.match(r'.+/EF \+ ./SF$', cols[2]):
+        return False
+    if re.match(r'.+/SF \+ [\'"’”]/SS$', cols[2]):
+        return False
+    morphs = [Morph.parse(_) for _ in cols[2].split(' + ')]
+    tags_str = '+'.join([_.tag for _ in morphs])
+    if 'SN+SF+SN' in tags_str and not tags_str.endswith('+SF'):
+        # 4.6판: 4/SN + ./SF + 6/SN + 판/NNB
+        if 'XSN+SF+SN' not in tags_str:
+            return True
+    elif 'SL+SF+SL' in tags_str and not tags_str.endswith('+SF'):
+        # S.M.오너: S/SL + ./SF + M/SL + ./SF + 오너/NNG
+        return True
+    return False
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    file_name = os.path.basename(sys.stdin.name)
+    for line_num, (prev_prev_line, prev_line, curr_line) in enumerate(_get_three_lines(sys.stdin),
+                                                                      start=1):
+        if curr_line == '</p>' and _is_known_period_error_eojeol(prev_line):
+            continue
+        elif prev_line == '</p>' and curr_line == '<p>' and \
+                _is_known_period_error_eojeol(prev_prev_line):
+            logging.info('%s:%d\t%s', file_name, line_num, prev_prev_line)
+            continue
+        print(curr_line)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='remove wrong sentence breaking marks after'
+                                        ' period error eojeol')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'rt')
+    if args.output:
+        sys.stdout = open(args.output, 'wt')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/src/main/python/khaiii/resource/jaso.py b/src/main/python/khaiii/resource/jaso.py
index 1f5e184..7b667b8 100644
--- a/src/main/python/khaiii/resource/jaso.py
+++ b/src/main/python/khaiii/resource/jaso.py
@@ -17,21 +17,63 @@
 #############
 # constants #
 #############
-_FIRST = ['\u3131', '\u3132', '\u3134', '\u3137', '\u3138',    # 초성
-          '\u3139', '\u3141', '\u3142', '\u3143', '\u3145',
-          '\u3146', '\u3147', '\u3148', '\u3149', '\u314a',
-          '\u314b', '\u314c', '\u314d', '\u314e']
-_MIDDLE = ['\u314f', '\u3150', '\u3151', '\u3152', '\u3153',    # 중성
-           '\u3154', '\u3155', '\u3156', '\u3157', '\u3158',
-           '\u3159', '\u315a', '\u315b', '\u315c', '\u315d',
-           '\u315e', '\u315f', '\u3160', '\u3161', '\u3162',
-           '\u3163']
-_LAST = ['\u3131', '\u3132', '\u3133', '\u3134', '\u3135',    # 종성
-         '\u3136', '\u3137', '\u3139', '\u313a', '\u313b',
-         '\u313c', '\u313d', '\u313e', '\u313f', '\u3140',
-         '\u3141', '\u3142', '\u3144', '\u3145', '\u3146',
-         '\u3147', '\u3148', '\u314a', '\u314b', '\u314c',
-         '\u314d', '\u314e']
+# 한글 자모 호환 영역 (초성과 종성이 같음. 두벌식 키보드로 입력할 때 들어가는 코드)
+_FIRST_COMPAT = ['\u3131', '\u3132', '\u3134', '\u3137', '\u3138',    # 초성
+                 '\u3139', '\u3141', '\u3142', '\u3143', '\u3145',
+                 '\u3146', '\u3147', '\u3148', '\u3149', '\u314a',
+                 '\u314b', '\u314c', '\u314d', '\u314e']
+_MIDDLE_COMPAT = ['\u314f', '\u3150', '\u3151', '\u3152', '\u3153',    # 중성
+                  '\u3154', '\u3155', '\u3156', '\u3157', '\u3158',
+                  '\u3159', '\u315a', '\u315b', '\u315c', '\u315d',
+                  '\u315e', '\u315f', '\u3160', '\u3161', '\u3162',
+                  '\u3163']
+_LAST_COMPAT = ['\u3131', '\u3132', '\u3133', '\u3134', '\u3135',    # 종성
+                '\u3136', '\u3137', '\u3139', '\u313a', '\u313b',
+                '\u313c', '\u313d', '\u313e', '\u313f', '\u3140',
+                '\u3141', '\u3142', '\u3144', '\u3145', '\u3146',
+                '\u3147', '\u3148', '\u314a', '\u314b', '\u314c',
+                '\u314d', '\u314e']
+_ALL_COMPAT = _FIRST_COMPAT + _MIDDLE_COMPAT + _LAST_COMPAT
+
+# 한글 자모 영역 (초성과 종성이 다름. 세종 코퍼스에서 사용한 코드)
+_FIRST_JAMO = ['\u1100', '\u1101', '\u1102', '\u1103', '\u1104',    # 초성
+               '\u1105', '\u1106', '\u1107', '\u1108', '\u1109',
+               '\u110a', '\u110b', '\u110c', '\u110d', '\u110e',
+               '\u110f', '\u1110', '\u1111', '\u1112']
+_MIDDLE_JAMO = ['\u1161', '\u1162', '\u1163', '\u1164', '\u1165',    # 중성
+                '\u1166', '\u1167', '\u1168', '\u1169', '\u116a',
+                '\u116b', '\u116c', '\u116d', '\u116e', '\u116f',
+                '\u1170', '\u1171', '\u1172', '\u1173', '\u1174',
+                '\u1175']
+_LAST_JAMO = ['\u11a8', '\u11a9', '\u11aa', '\u11ab', '\u11ac',    # 종성
+              '\u11ad', '\u11ae', '\u11af', '\u11b0', '\u11b1',
+              '\u11b2', '\u11b3', '\u11b4', '\u11b5', '\u11b6',
+              '\u11b7', '\u11b8', '\u11b9', '\u11ba', '\u11bb',
+              '\u11bc', '\u11bd', '\u11be', '\u11bf', '\u11c0',
+              '\u11c1', '\u11c2']
+_ALL_JAMO = _FIRST_JAMO + _MIDDLE_JAMO + _LAST_JAMO
+_ALL_JAMO_SET = set(_ALL_JAMO)
+_JAMO_TO_COMPAT = dict(zip(_ALL_JAMO, _ALL_COMPAT))
+
+# 반각 자모 영역 (호환 영역과 비슷하게 초성과 종성이 같으나 글자 폭이 절반인 코드)
+_FIRST_HALFWIDTH = ['\uffa1', '\uffa2', '\uffa4', '\uffa7', '\uffa8',    # 초성
+                    '\uffa9', '\uffb1', '\uffb2', '\uffb3', '\uffb5',
+                    '\uffb6', '\uffb7', '\uffb8', '\uffb9', '\uffba',
+                    '\uffbb', '\uffbc', '\uffbd', '\uffbe']
+_MIDDLE_HALFWIDTH = ['\uffc2', '\uffc3', '\uffc4', '\uffc5', '\uffc6',    # 중성
+                     '\uffc7', '\uffca', '\uffcb', '\uffcc', '\uffcd',
+                     '\uffce', '\uffcf', '\uffd2', '\uffd3', '\uffd4',
+                     '\uffd5', '\uffd6', '\uffd7', '\uffda', '\uffdb',
+                     '\uffdc']
+_LAST_HALFWIDTH = ['\uffa1', '\uffa2', '\uffa3', '\uffa4', '\uffa5',    # 종성
+                   '\uffa6', '\uffa7', '\uffa9', '\uffaa', '\uffab',
+                   '\uffac', '\uffad', '\uffae', '\uffaf', '\uffb0',
+                   '\uffb1', '\uffb2', '\uffb4', '\uffb5', '\uffb6',
+                   '\uffb7', '\uffb8', '\uffba', '\uffbb', '\uffbc',
+                   '\uffbd', '\uffbe']
+_ALL_HALFWIDTH = _FIRST_HALFWIDTH + _MIDDLE_HALFWIDTH + _LAST_HALFWIDTH
+_ALL_HALFWIDTH_SET = set(_ALL_HALFWIDTH)
+_HALFWIDTH_TO_COMPAT = dict(zip(_ALL_HALFWIDTH, _ALL_COMPAT))
 
 
 #############
@@ -54,12 +96,12 @@ def _decomp_char(char: str) -> Tuple[str, str, str]:
     first_idx = first_start // 21
     middle_idx = first_start % 21
 
-    first = _FIRST[first_idx]
-    middle = _MIDDLE[middle_idx]
+    first = _FIRST_COMPAT[first_idx]
+    middle = _MIDDLE_COMPAT[middle_idx]
     if not last_idx:
         return first, middle
 
-    last = _LAST[last_idx-1]
+    last = _LAST_COMPAT[last_idx-1]
     return first, middle, last
 
 
@@ -71,6 +113,9 @@ def decompose(text: str) -> str:
     Returns:
         자소 분해된 텍스트
     """
+    if not text:
+        return text
+
     decomposed = []
     for char in text:
         code = ord(char)
@@ -79,3 +124,25 @@ def decompose(text: str) -> str:
         else:
             decomposed.extend(_decomp_char(char))
     return ''.join(decomposed)
+
+
+def norm_compat(text: str) -> str:
+    """
+    유니코드 내 한글 자소를 호환 영역으로 정규화한다.
+    Args:
+        text:  한글 텍스트
+    Returns:
+        자소가 호환 영역으로 정규화된 텍스트
+    """
+    if not text:
+        return text
+
+    normalized = []
+    for char in text:
+        if char in _ALL_JAMO_SET:
+            normalized.append(_JAMO_TO_COMPAT[char])
+        elif char in _ALL_HALFWIDTH_SET:
+            normalized.append(_HALFWIDTH_TO_COMPAT[char])
+        else:
+            normalized.append(char)
+    return ''.join(normalized)

From 0d6a398a0f90e2ed2f8188f396f0e36df343ddf8 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 28 Jan 2019 20:57:56 +0900
Subject: [PATCH 16/41] =?UTF-8?q?train/bin=20=EC=95=84=EB=9E=98=20?=
 =?UTF-8?q?=EC=8A=A4=ED=81=AC=EB=A6=BD=ED=8A=B8=EB=A5=BC=20train=EC=9C=BC?=
 =?UTF-8?q?=EB=A1=9C=20=EC=9D=B4=EB=8F=99=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 train/{bin => }/extract_errpatch.py       |  0
 train/{bin => }/extract_preanal.py        |  0
 train/{bin => }/hd_validate_errpatch.bash | 18 +++++++++---------
 train/{bin => }/make_vocab.py             |  0
 train/{bin => }/map_char_to_tag.py        |  0
 train/{bin => }/pickle_model.py           |  0
 train/{bin => }/split_corpus.py           |  0
 train/{bin => }/tag.py                    |  0
 train/{bin => }/train.py                  |  0
 train/{bin => }/validate_errpatch.py      |  0
 10 files changed, 9 insertions(+), 9 deletions(-)
 rename train/{bin => }/extract_errpatch.py (100%)
 rename train/{bin => }/extract_preanal.py (100%)
 rename train/{bin => }/hd_validate_errpatch.bash (91%)
 rename train/{bin => }/make_vocab.py (100%)
 rename train/{bin => }/map_char_to_tag.py (100%)
 rename train/{bin => }/pickle_model.py (100%)
 rename train/{bin => }/split_corpus.py (100%)
 rename train/{bin => }/tag.py (100%)
 rename train/{bin => }/train.py (100%)
 rename train/{bin => }/validate_errpatch.py (100%)

diff --git a/train/bin/extract_errpatch.py b/train/extract_errpatch.py
similarity index 100%
rename from train/bin/extract_errpatch.py
rename to train/extract_errpatch.py
diff --git a/train/bin/extract_preanal.py b/train/extract_preanal.py
similarity index 100%
rename from train/bin/extract_preanal.py
rename to train/extract_preanal.py
diff --git a/train/bin/hd_validate_errpatch.bash b/train/hd_validate_errpatch.bash
similarity index 91%
rename from train/bin/hd_validate_errpatch.bash
rename to train/hd_validate_errpatch.bash
index 3042ef8..c4b4bb2 100755
--- a/train/bin/hd_validate_errpatch.bash
+++ b/train/hd_validate_errpatch.bash
@@ -12,9 +12,9 @@ function print_usage() {
     echo "  -h, --help        show this help message and exit"
     echo "  -i FILE           input file"
     echo "  -c DIR            corpus dir"
-    echo "  --rsc-src=DIR     <default: ../../rsc/src>"
-    echo "  --lib-path=FILE   <default: ../../build/lib/libkhaiii.so>"
-    echo "  --rsc-dir=DIR     <default: ../../build/share/khaiii>"
+    echo "  --rsc-src=DIR     <default: ../rsc/src>"
+    echo "  --lib-path=FILE   <default: ../build/lib/libkhaiii.so>"
+    echo "  --rsc-dir=DIR     <default: ../build/share/khaiii>"
     echo "  --num-mapper=NUM  <default: 1000>"
     if [ -z "${msg}" ]; then
         exit 0
@@ -84,13 +84,13 @@ function parse_args() {
     fi
 
     if [ -z "${RSC_SRC}" ]; then
-        RSC_SRC=../../rsc/src
+        RSC_SRC=../rsc/src
     fi
     if [ -z "${LIB_PATH}" ]; then
-        LIB_PATH=../../build/lib/libkhaiii.so
+        LIB_PATH=../build/lib/libkhaiii.so
     fi
     if [ -z "${RSC_DIR}" ]; then
-        RSC_DIR=../../build/share/khaiii
+        RSC_DIR=../build/share/khaiii
     fi
     if [ -z "${NUM_MAPPER}" ]; then
         NUM_MAPPER=1000
@@ -135,7 +135,7 @@ function cache_files() {
     hadoop fs -test -e ${CACHE_DIR} && hadoop fs -rm -skipTrash -r ${CACHE_DIR}
     hadoop fs -mkdir -p ${CACHE_DIR}
 
-    hadoop fs -put ../../src/main/python/khaiii ${CACHE_DIR}
+    hadoop fs -put ../src/main/python/khaiii ${CACHE_DIR}
     hadoop fs -mkdir -p ${CACHE_DIR}/khaiii/lib
     hadoop fs -put ${LIB_PATH} ${CACHE_DIR}/khaiii/lib
 
@@ -188,7 +188,7 @@ parse_args $@
 cd "$(dirname $0)"
 init_envs
 
-# split_input
-# cache_files
+split_input
+cache_files
 run_hadoop
 del_temp
diff --git a/train/bin/make_vocab.py b/train/make_vocab.py
similarity index 100%
rename from train/bin/make_vocab.py
rename to train/make_vocab.py
diff --git a/train/bin/map_char_to_tag.py b/train/map_char_to_tag.py
similarity index 100%
rename from train/bin/map_char_to_tag.py
rename to train/map_char_to_tag.py
diff --git a/train/bin/pickle_model.py b/train/pickle_model.py
similarity index 100%
rename from train/bin/pickle_model.py
rename to train/pickle_model.py
diff --git a/train/bin/split_corpus.py b/train/split_corpus.py
similarity index 100%
rename from train/bin/split_corpus.py
rename to train/split_corpus.py
diff --git a/train/bin/tag.py b/train/tag.py
similarity index 100%
rename from train/bin/tag.py
rename to train/tag.py
diff --git a/train/bin/train.py b/train/train.py
similarity index 100%
rename from train/bin/train.py
rename to train/train.py
diff --git a/train/bin/validate_errpatch.py b/train/validate_errpatch.py
similarity index 100%
rename from train/bin/validate_errpatch.py
rename to train/validate_errpatch.py

From 2744ba5dbe3e44a76223ed248e97affb109bf080 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 28 Jan 2019 20:59:15 +0900
Subject: [PATCH 17/41] =?UTF-8?q?munjong/bin=20=EC=95=84=EB=9E=98=20?=
 =?UTF-8?q?=EC=8A=A4=ED=81=AC=EB=A6=BD=ED=8A=B8=EB=A5=BC=20munjong?=
 =?UTF-8?q?=EC=9C=BC=EB=A1=9C=20=EC=9D=B4=EB=8F=99=20#3?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 munjong/{bin => }/apply_patch.py                | 0
 munjong/{bin => }/convert_jamo_to_compat.py     | 0
 munjong/{bin => }/detect_sejong_period_error.py | 0
 munjong/{bin => }/fix_final_symbol_error.py     | 0
 munjong/{bin => }/make_patch.py                 | 0
 munjong/{bin => }/recover_english_case.py       | 0
 munjong/{bin => }/recover_raw_morph_mismatch.py | 0
 munjong/{bin => }/recover_wide_quotation.py     | 0
 munjong/{bin => }/remove_sejong_period_error.py | 0
 9 files changed, 0 insertions(+), 0 deletions(-)
 rename munjong/{bin => }/apply_patch.py (100%)
 rename munjong/{bin => }/convert_jamo_to_compat.py (100%)
 rename munjong/{bin => }/detect_sejong_period_error.py (100%)
 rename munjong/{bin => }/fix_final_symbol_error.py (100%)
 rename munjong/{bin => }/make_patch.py (100%)
 rename munjong/{bin => }/recover_english_case.py (100%)
 rename munjong/{bin => }/recover_raw_morph_mismatch.py (100%)
 rename munjong/{bin => }/recover_wide_quotation.py (100%)
 rename munjong/{bin => }/remove_sejong_period_error.py (100%)

diff --git a/munjong/bin/apply_patch.py b/munjong/apply_patch.py
similarity index 100%
rename from munjong/bin/apply_patch.py
rename to munjong/apply_patch.py
diff --git a/munjong/bin/convert_jamo_to_compat.py b/munjong/convert_jamo_to_compat.py
similarity index 100%
rename from munjong/bin/convert_jamo_to_compat.py
rename to munjong/convert_jamo_to_compat.py
diff --git a/munjong/bin/detect_sejong_period_error.py b/munjong/detect_sejong_period_error.py
similarity index 100%
rename from munjong/bin/detect_sejong_period_error.py
rename to munjong/detect_sejong_period_error.py
diff --git a/munjong/bin/fix_final_symbol_error.py b/munjong/fix_final_symbol_error.py
similarity index 100%
rename from munjong/bin/fix_final_symbol_error.py
rename to munjong/fix_final_symbol_error.py
diff --git a/munjong/bin/make_patch.py b/munjong/make_patch.py
similarity index 100%
rename from munjong/bin/make_patch.py
rename to munjong/make_patch.py
diff --git a/munjong/bin/recover_english_case.py b/munjong/recover_english_case.py
similarity index 100%
rename from munjong/bin/recover_english_case.py
rename to munjong/recover_english_case.py
diff --git a/munjong/bin/recover_raw_morph_mismatch.py b/munjong/recover_raw_morph_mismatch.py
similarity index 100%
rename from munjong/bin/recover_raw_morph_mismatch.py
rename to munjong/recover_raw_morph_mismatch.py
diff --git a/munjong/bin/recover_wide_quotation.py b/munjong/recover_wide_quotation.py
similarity index 100%
rename from munjong/bin/recover_wide_quotation.py
rename to munjong/recover_wide_quotation.py
diff --git a/munjong/bin/remove_sejong_period_error.py b/munjong/remove_sejong_period_error.py
similarity index 100%
rename from munjong/bin/remove_sejong_period_error.py
rename to munjong/remove_sejong_period_error.py

From 586fc01d0be0530e3670f9fce577efa65a6a94cc Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Tue, 29 Jan 2019 10:40:55 +0900
Subject: [PATCH 18/41] =?UTF-8?q?=ED=95=99=EC=8A=B5=20=EC=BD=94=EB=93=9C?=
 =?UTF-8?q?=EB=A5=BC=20=EA=B8=B0=EC=A1=B4=20=EB=B0=A9=EC=8B=9D=EC=9C=BC?=
 =?UTF-8?q?=EB=A1=9C=20=ED=9A=8C=EA=B7=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/main/python/khaiii/train/trainer.py | 111 ++++++++++++------------
 train/train.py                          |   6 +-
 2 files changed, 58 insertions(+), 59 deletions(-)

diff --git a/src/main/python/khaiii/train/trainer.py b/src/main/python/khaiii/train/trainer.py
index e299fc1..038baee 100644
--- a/src/main/python/khaiii/train/trainer.py
+++ b/src/main/python/khaiii/train/trainer.py
@@ -19,7 +19,7 @@
 import os
 import pathlib
 import pprint
-from typing import Iterator, List, Tuple
+from typing import List, Tuple
 
 from tensorboardX import SummaryWriter
 import torch
@@ -27,7 +27,7 @@
 import torch.nn.functional as F
 from tqdm import tqdm
 
-from khaiii.train.dataset import PosDataset, PosSentTensor
+from khaiii.train.dataset import PosDataset
 from khaiii.train.evaluator import Evaluator
 from khaiii.train.models import CnnModel
 from khaiii.resource.resource import Resource
@@ -56,6 +56,7 @@ def __init__(self, cfg: Namespace):
         self.evaler = Evaluator()
         self._load_dataset()
         if 'step' not in cfg.__dict__:
+            setattr(cfg, 'epoch', 0)
             setattr(cfg, 'step', 0)
             setattr(cfg, 'best_step', 0)
         self.log_file = None    # tab separated log file
@@ -66,6 +67,7 @@ def __init__(self, cfg: Namespace):
         self.acc_words = []
         self.f_scores = []
         self.learning_rates = []
+        self.batch_sizes = []
 
     @classmethod
     def model_id(cls, cfg: Namespace) -> str:
@@ -86,6 +88,7 @@ def model_id(cls, cfg: Namespace) -> str:
             'lrd{}'.format(cfg.lr_decay),
             'bs{}'.format(cfg.batch_size),
             'cs{}'.format(cfg.check_step),
+            'bg{}'.format(cfg.batch_grow),
         ]
         return '.'.join(model_cfgs)
 
@@ -149,35 +152,25 @@ def _restore_prev_train(self):
             line = line.rstrip('\r\n')
             if not line:
                 continue
-            (step, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate) \
-                    = line.split('\t')
-            self.cfg.step = self.cfg.best_step = int(step) * self.cfg.check_step
+            (epoch, step, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate,
+             batch_size) = line.split('\t')
+            self.cfg.epoch = int(epoch) + 1
+            self.cfg.step = self.cfg.best_step = (int(step)+1) * self.cfg.check_step
             self.loss_trains.append(float(loss_train))
             self.loss_devs.append(float(loss_dev))
             self.acc_chars.append(float(acc_char))
             self.acc_words.append(float(acc_word))
             self.f_scores.append(float(f_score))
             self.learning_rates.append(float(learning_rate))
+            self.batch_sizes.append(int(batch_size))
             if float(f_score) > f_score_best:
                 f_score_best = float(f_score)
                 best_idx = idx
-        logging.info('---- [%d] los(trn/dev): %.4f / %.4f, acc(chr/wrd): %.4f / %.4f, ' \
-                     'f-score: %.4f, lr: %.8f ----', self.cfg.step // self.cfg.check_step,
-                     self.loss_trains[best_idx], self.loss_devs[best_idx], self.acc_chars[best_idx],
-                     self.acc_words[best_idx], self.f_scores[best_idx], self.learning_rates[-1])
-
-    @classmethod
-    def _inf_data_iterator(cls, dataset: PosDataset) -> Iterator[PosSentTensor]:
-        """
-        데이터셋을 무한히 반복하여 문장을 출력하는 제너레이터
-        Args:
-            dataset:  데이터셋
-        Yields:
-            PosSentTensor 객체
-        """
-        for _ in range(1000000):
-            for sent in dataset:
-                yield sent
+        logging.info('---- [%d|%d] los(trn/dev): %.4f / %.4f, acc(chr/wrd): %.4f / %.4f, ' \
+                     'f-score: %.4f, lr: %.8f, bs: %d ----', self.cfg.epoch,
+                     self.cfg.step // self.cfg.check_step, self.loss_trains[best_idx],
+                     self.loss_devs[best_idx], self.acc_chars[best_idx], self.acc_words[best_idx],
+                     self.f_scores[best_idx], self.learning_rates[-1], self.batch_sizes[-1])
 
     def train(self):
         """
@@ -193,12 +186,10 @@ def train(self):
         pathlib.Path(self.cfg.out_dir).mkdir(parents=True, exist_ok=True)
         self.log_file = open('{}/log.tsv'.format(self.cfg.out_dir), 'at')
         self.sum_wrt = SummaryWriter(self.cfg.out_dir)
-        check_start = (1 if self.cfg.step == 0 else (self.cfg.step // self.cfg.check_step + 1))
         patience = self.cfg.patience
-        train_iter = self._inf_data_iterator(self.dataset_train)
-        for check_id in range(check_start, 1000000):
-            is_best = self._train_and_check(check_id, train_iter)
-            if is_best:
+        for _ in range(1000000):
+            has_best = self._train_epoch()
+            if has_best:
                 patience = self.cfg.patience
                 continue
             if patience <= 0:
@@ -210,8 +201,8 @@ def train(self):
 
         train_end = datetime.now()
         train_elapsed = self._elapsed(train_end - train_begin)
-        logging.info('}}}} training end: %s, elapsed: %s, step: %dk }}}}',
-                     self._dt_str(train_end), train_elapsed, self.cfg.step // 1000)
+        logging.info('}}}} training end: %s, elapsed: %s, epoch: %d, step: %dk }}}}',
+                     self._dt_str(train_end), train_elapsed, self.cfg.epoch, self.cfg.step // 1000)
 
     def _revert_to_best(self, is_decay_lr: bool):
         """
@@ -224,21 +215,21 @@ def _revert_to_best(self, is_decay_lr: bool):
             self.cfg.learning_rate *= self.cfg.lr_decay
         self._load_optim('{}/optim.state'.format(self.cfg.out_dir), self.cfg.learning_rate)
 
-    def _train_and_check(self, check_id: int, train_iter: Iterator[PosSentTensor]) -> bool:
+    def _train_epoch(self) -> bool:
         """
-        cfg.check_step 만큼의 step을 수행하고 evaluation을 수행한다.
-        Args:
-            check_id:  check ID
-            train_iter:  학습 데이터 iterator
+        한 epoch을 학습한다. 배치 단위는 글자 단위
         Returns:
             best f-score를 기록한 step 여부
         """
-        start_step = self.cfg.step
+        has_best = False
         loss_batch = torch.tensor(0.0)    # pylint: disable=not-callable
-        batch_size = 0
+        num_in_batch = 0
+        batch_size = self.cfg.batch_size
+        if self.cfg.batch_grow > 0:
+            batch_size = self.cfg.batch_size + self.cfg.step // self.cfg.batch_grow
         loss_trains = []
-        train_sents = tqdm(train_iter, '[{}]'.format(check_id), mininterval=1, ncols=100)
-        for train_sent in train_sents:
+        for train_sent in tqdm(self.dataset_train, 'EPOCH[{}]'.format(self.cfg.epoch),
+                               mininterval=1, ncols=100):
             train_labels, train_contexts = train_sent.to_tensor(self.cfg, self.rsc, True)
             if torch.cuda.is_available():
                 train_labels = train_labels.cuda()
@@ -252,8 +243,8 @@ def _train_and_check(self, check_id: int, train_iter: Iterator[PosSentTensor]) -
             loss_train.backward()
             loss_trains.append(loss_train.item())
             loss_batch += loss_train
-            batch_size += len(train_labels)
-            if batch_size < self.cfg.batch_size:
+            num_in_batch += len(train_labels)
+            if num_in_batch < batch_size:
                 continue
 
             self.optimizer.step()
@@ -261,30 +252,32 @@ def _train_and_check(self, check_id: int, train_iter: Iterator[PosSentTensor]) -
             self.sum_wrt.add_scalar('loss-batch', loss_batch.item(), self.cfg.step)
             self.cfg.step += 1
             loss_batch = torch.tensor(0.0)    # pylint: disable=not-callable
-            batch_size = 0
+            num_in_batch = 0
 
-            if (self.cfg.step - start_step) >= self.cfg.check_step:
-                train_sents.close()
-                break
+            if self.cfg.step % self.cfg.check_step == 0:
+                avg_loss_dev, acc_char, acc_word, f_score = self.evaluate()
+                has_best |= self._check(loss_trains, avg_loss_dev, acc_char, acc_word, f_score,
+                                        batch_size)
+            if self.cfg.batch_grow > 0 and self.cfg.step % self.cfg.batch_grow == 0:
+                batch_size = self.cfg.batch_size + self.cfg.step // self.cfg.batch_grow
 
-        avg_loss_dev, acc_char, acc_word, f_score = self.evaluate()
-        return self._check(check_id, loss_trains, avg_loss_dev, acc_char, acc_word, f_score)
+        self.cfg.epoch += 1
+        return has_best
 
-    def _check(self, check_id: int, loss_trains: List[float], avg_loss_dev: float, acc_char: float,
-               acc_word: float, f_score: float) -> bool:
+    def _check(self, loss_trains: List[float], avg_loss_dev: float, acc_char: float,
+               acc_word: float, f_score: float, batch_size: int) -> bool:
         """
         cfg.check_step번의 step마다 수행하는 체크
         Args:
-            check_id:  check ID
             loss_trains:  train 코퍼스에서 각 배치별 loss 리스트
             avg_loss_dev:  dev 코퍼스 문장 별 평균 loss
             acc_char:  음절 정확도
             acc_word:  어절 정확도
             f_score:  f-score
+            batch_size:  배치 크기
         Returns:
             현재 step이 best 성능을 나타냈는 지 여부
         """
-        assert check_id == self.cfg.step // self.cfg.check_step
         avg_loss_train = sum(loss_trains) / len(loss_trains)
         loss_trains.clear()
         self.loss_trains.append(avg_loss_train)
@@ -293,15 +286,18 @@ def _check(self, check_id: int, loss_trains: List[float], avg_loss_dev: float, a
         self.acc_words.append(acc_word)
         self.f_scores.append(f_score)
         self.learning_rates.append(self.cfg.learning_rate)
+        self.batch_sizes.append(batch_size)
+        check_id = self.cfg.step // self.cfg.check_step - 1
         is_best = self._is_best()
         is_best_str = 'BEST' if is_best else '< {:.4f}'.format(max(self.f_scores))
-        tqdm.write('    [Los trn]  [Los dev]  [Acc chr]  [Acc wrd]  [F-score]           [LR]')
-        tqdm.write('    {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f} {:8}  {:.8f}'.format( \
-                avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score, is_best_str,
-                self.cfg.learning_rate))
-        print('{}\t{}\t{}\t{}\t{}\t{}\t{}'.format(check_id, avg_loss_train, avg_loss_dev,
-                                                  acc_char, acc_word, f_score,
-                                                  self.cfg.learning_rate), file=self.log_file)
+        tqdm.write('    [{:3d}|{:5d}]  [Los trn]  [Los dev]  [Acc chr]  [Acc wrd]  [F-score]'
+                   '           [LR]        [BS]'.format(self.cfg.epoch, check_id))
+        tqdm.write('                 {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f} {:8}  {:.8f}  {}' \
+                .format(avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score, is_best_str,
+                        self.cfg.learning_rate, batch_size))
+        print('{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}'.format( \
+                self.cfg.epoch, check_id, avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score,
+                self.cfg.learning_rate, batch_size), file=self.log_file)
         self.log_file.flush()
         self.sum_wrt.add_scalar('loss-train', avg_loss_train, check_id)
         self.sum_wrt.add_scalar('loss-dev', avg_loss_dev, check_id)
@@ -309,6 +305,7 @@ def _check(self, check_id: int, loss_trains: List[float], avg_loss_dev: float, a
         self.sum_wrt.add_scalar('acc-word', acc_word, check_id)
         self.sum_wrt.add_scalar('f-score', f_score, check_id)
         self.sum_wrt.add_scalar('learning-rate', self.cfg.learning_rate, check_id)
+        self.sum_wrt.add_scalar('batch-size', batch_size, check_id)
         return is_best
 
     def _is_best(self) -> bool:
diff --git a/train/train.py b/train/train.py
index 7684be4..ce4b4e5 100755
--- a/train/train.py
+++ b/train/train.py
@@ -57,10 +57,12 @@ def main():
                         type=float, default=0.001)
     parser.add_argument('--lr-decay', help='learning rate decay <default: 0.9>', metavar='REAL',
                         type=float, default=0.9)
-    parser.add_argument('--batch-size', help='batch size <default: 1000>', metavar='INT', type=int,
-                        default=1000)
+    parser.add_argument('--batch-size', help='batch size <default: 500>', metavar='INT', type=int,
+                        default=500)
     parser.add_argument('--check-step', help='check every N step <default: 10000>', metavar='INT',
                         type=int, default=10000)
+    parser.add_argument('--batch-grow', help='grow batch size by 1 per N step <default: 10000>',
+                        metavar='INT', type=int, default=10000)
     parser.add_argument('--patience', help='maximum patience count to revert model <default: 10>',
                         metavar='INT', type=int, default=10)
     parser.add_argument('--gpu-num', help='GPU number to use <default: 0>', metavar='INT', type=int,

From d673ca92b96791839156cf9ed24e9504b663bee2 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Thu, 31 Jan 2019 17:58:21 +0900
Subject: [PATCH 19/41] =?UTF-8?q?=EC=98=A4=ED=94=88=EC=86=8C=EC=8A=A4=20?=
 =?UTF-8?q?=EA=B3=A0=EC=A7=80=EB=AC=B8=20=EC=97=85=EB=8D=B0=EC=9D=B4?=
 =?UTF-8?q?=ED=8A=B8=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 NOTICE.md | 20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/NOTICE.md b/NOTICE.md
index e384491..d6d5505 100644
--- a/NOTICE.md
+++ b/NOTICE.md
@@ -97,16 +97,6 @@ Copyright 2013-2018 Niels Lohmann
 MIT License 
 
 
- **NumPy**
-
-https://github.com/numpy/numpy
-
-Copyright 2005-2018, NumPy Developers.
-
-
-BSD 3-Clause "New" or "Revised" License 
-
-
  **PyTorch**
 
 https://github.com/pytorch/pytorch
@@ -124,6 +114,16 @@ https://github.com/gabime/spdlog
 Copyright 2016 Gabi Melman
 
 
+MIT License 
+
+
+ **tensorboardX**
+
+https://github.com/lanpa/tensorboardX
+
+Copyright 2017 Tzu-Wei Huang
+
+
 MIT License 
 
 

From adef433f267a4f2c09693229dec47ba3c9ad9706 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Sun, 3 Feb 2019 13:08:44 +0900
Subject: [PATCH 20/41] =?UTF-8?q?=EC=84=B1=EB=8A=A5=EC=A0=80=ED=95=98?=
 =?UTF-8?q?=EC=9D=98=20=EC=9B=90=EC=9D=B8=EC=9D=84=20=ED=95=B4=EA=B2=B0?=
 =?UTF-8?q?=ED=95=98=EA=B3=A0=20epoch=20=EB=8B=A8=EC=9C=84=EB=A1=9C=20?=
 =?UTF-8?q?=EB=8B=A8=EC=88=9C=ED=99=94=ED=95=9C=20=EC=BD=94=EB=93=9C=20?=
 =?UTF-8?q?=EB=8B=A4=EC=8B=9C=20=EC=A0=81=EC=9A=A9=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/main/python/khaiii/train/trainer.py | 166 +++++++++++-------------
 1 file changed, 76 insertions(+), 90 deletions(-)

diff --git a/src/main/python/khaiii/train/trainer.py b/src/main/python/khaiii/train/trainer.py
index 038baee..6bc2f37 100644
--- a/src/main/python/khaiii/train/trainer.py
+++ b/src/main/python/khaiii/train/trainer.py
@@ -55,10 +55,9 @@ def __init__(self, cfg: Namespace):
         self.criterion = nn.CrossEntropyLoss()
         self.evaler = Evaluator()
         self._load_dataset()
-        if 'step' not in cfg.__dict__:
+        if 'epoch' not in cfg.__dict__:
             setattr(cfg, 'epoch', 0)
-            setattr(cfg, 'step', 0)
-            setattr(cfg, 'best_step', 0)
+            setattr(cfg, 'best_epoch', 0)
         self.log_file = None    # tab separated log file
         self.sum_wrt = None    # tensorboard summary writer
         self.loss_trains = []
@@ -67,7 +66,6 @@ def __init__(self, cfg: Namespace):
         self.acc_words = []
         self.f_scores = []
         self.learning_rates = []
-        self.batch_sizes = []
 
     @classmethod
     def model_id(cls, cfg: Namespace) -> str:
@@ -87,8 +85,6 @@ def model_id(cls, cfg: Namespace) -> str:
             'lr{}'.format(cfg.learning_rate),
             'lrd{}'.format(cfg.lr_decay),
             'bs{}'.format(cfg.batch_size),
-            'cs{}'.format(cfg.check_step),
-            'bg{}'.format(cfg.batch_grow),
         ]
         return '.'.join(model_cfgs)
 
@@ -99,6 +95,9 @@ def _load_dataset(self):
         dataset_dev_path = '{}.dev'.format(self.cfg.in_pfx)
         self.dataset_dev = PosDataset(self.cfg, self.rsc.restore_dic,
                                       open(dataset_dev_path, 'r', encoding='UTF-8'))
+        dataset_test_path = '{}.test'.format(self.cfg.in_pfx)
+        self.dataset_test = PosDataset(self.cfg, self.rsc.restore_dic,
+                                       open(dataset_test_path, 'r', encoding='UTF-8'))
         dataset_train_path = '{}.train'.format(self.cfg.in_pfx)
         self.dataset_train = PosDataset(self.cfg, self.rsc.restore_dic,
                                         open(dataset_train_path, 'r', encoding='UTF-8'))
@@ -152,25 +151,23 @@ def _restore_prev_train(self):
             line = line.rstrip('\r\n')
             if not line:
                 continue
-            (epoch, step, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate,
-             batch_size) = line.split('\t')
+            (epoch, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate) = \
+                    line.split('\t')
             self.cfg.epoch = int(epoch) + 1
-            self.cfg.step = self.cfg.best_step = (int(step)+1) * self.cfg.check_step
+            self.cfg.best_epoch = self.cfg.epoch
             self.loss_trains.append(float(loss_train))
             self.loss_devs.append(float(loss_dev))
             self.acc_chars.append(float(acc_char))
             self.acc_words.append(float(acc_word))
             self.f_scores.append(float(f_score))
             self.learning_rates.append(float(learning_rate))
-            self.batch_sizes.append(int(batch_size))
             if float(f_score) > f_score_best:
                 f_score_best = float(f_score)
                 best_idx = idx
-        logging.info('---- [%d|%d] los(trn/dev): %.4f / %.4f, acc(chr/wrd): %.4f / %.4f, ' \
-                     'f-score: %.4f, lr: %.8f, bs: %d ----', self.cfg.epoch,
-                     self.cfg.step // self.cfg.check_step, self.loss_trains[best_idx],
-                     self.loss_devs[best_idx], self.acc_chars[best_idx], self.acc_words[best_idx],
-                     self.f_scores[best_idx], self.learning_rates[-1], self.batch_sizes[-1])
+        logging.info('---- [%d] los(trn/dev): %.4f / %.4f, acc(chr/wrd): %.4f / %.4f, ' \
+                     'f-score: %.4f, lr: %.8f ----', self.cfg.epoch,
+                     self.loss_trains[best_idx], self.loss_devs[best_idx], self.acc_chars[best_idx],
+                     self.acc_words[best_idx], self.f_scores[best_idx], self.learning_rates[-1])
 
     def train(self):
         """
@@ -188,21 +185,25 @@ def train(self):
         self.sum_wrt = SummaryWriter(self.cfg.out_dir)
         patience = self.cfg.patience
         for _ in range(1000000):
-            has_best = self._train_epoch()
-            if has_best:
+            is_best = self._train_epoch()
+            if is_best:
                 patience = self.cfg.patience
                 continue
             if patience <= 0:
                 break
             self._revert_to_best(True)
             patience -= 1
-            tqdm.write('==== revert to check: {}, f-score: {:.4f}, patience: {} ===='.format( \
-                    self.cfg.best_step // self.cfg.check_step, max(self.f_scores), patience))
+            logging.info('==== revert to EPOCH[%d], f-score: %.4f, patience: %d ====',
+                         self.cfg.best_epoch, max(self.f_scores), patience)
 
         train_end = datetime.now()
         train_elapsed = self._elapsed(train_end - train_begin)
-        logging.info('}}}} training end: %s, elapsed: %s, epoch: %d, step: %dk }}}}',
-                     self._dt_str(train_end), train_elapsed, self.cfg.epoch, self.cfg.step // 1000)
+        logging.info('}}}} training end: %s, elapsed: %s, epoch: %s }}}}',
+                     self._dt_str(train_end), train_elapsed, self.cfg.epoch)
+
+        avg_loss, acc_char, acc_word, f_score = self.evaluate(False)
+        logging.info('==== test loss: %.4f, char acc: %.4f, word acc: %.4f, f-score: %.4f ====',
+                     avg_loss, acc_char, acc_word, f_score)
 
     def _revert_to_best(self, is_decay_lr: bool):
         """
@@ -219,64 +220,51 @@ def _train_epoch(self) -> bool:
         """
         한 epoch을 학습한다. 배치 단위는 글자 단위
         Returns:
-            best f-score를 기록한 step 여부
+            현재 epoch이 best 성능을 나타냈는 지 여부
         """
-        has_best = False
-        loss_batch = torch.tensor(0.0)    # pylint: disable=not-callable
-        num_in_batch = 0
-        batch_size = self.cfg.batch_size
-        if self.cfg.batch_grow > 0:
-            batch_size = self.cfg.batch_size + self.cfg.step // self.cfg.batch_grow
+        batches = []
         loss_trains = []
         for train_sent in tqdm(self.dataset_train, 'EPOCH[{}]'.format(self.cfg.epoch),
-                               mininterval=1, ncols=100):
+                               len(self.dataset_train), mininterval=1, ncols=100):
             train_labels, train_contexts = train_sent.to_tensor(self.cfg, self.rsc, True)
             if torch.cuda.is_available():
                 train_labels = train_labels.cuda()
                 train_contexts = train_contexts.cuda()
-                loss_batch = loss_batch.cuda()
 
             self.model.train()
             train_outputs = self.model(train_contexts)
-            train_outputs.requires_grad_()
-            loss_train = self.criterion(train_outputs, train_labels)
-            loss_train.backward()
-            loss_trains.append(loss_train.item())
-            loss_batch += loss_train
-            num_in_batch += len(train_labels)
-            if num_in_batch < batch_size:
+            batches.append((train_labels, train_outputs))
+            if sum([batch[0].size(0) for batch in batches]) < self.cfg.batch_size:
                 continue
 
+            batch_label = torch.cat([x[0] for x in batches], 0)    # pylint: disable=no-member
+            batch_output = torch.cat([x[1] for x in batches], 0)    # pylint: disable=no-member
+            batches = []
+
+            batch_output.requires_grad_()
+            loss_train = self.criterion(batch_output, batch_label)
+            loss_trains.append(loss_train.item())
+            loss_train.backward()
             self.optimizer.step()
             self.optimizer.zero_grad()
-            self.sum_wrt.add_scalar('loss-batch', loss_batch.item(), self.cfg.step)
-            self.cfg.step += 1
-            loss_batch = torch.tensor(0.0)    # pylint: disable=not-callable
-            num_in_batch = 0
-
-            if self.cfg.step % self.cfg.check_step == 0:
-                avg_loss_dev, acc_char, acc_word, f_score = self.evaluate()
-                has_best |= self._check(loss_trains, avg_loss_dev, acc_char, acc_word, f_score,
-                                        batch_size)
-            if self.cfg.batch_grow > 0 and self.cfg.step % self.cfg.batch_grow == 0:
-                batch_size = self.cfg.batch_size + self.cfg.step // self.cfg.batch_grow
 
+        avg_loss_dev, acc_char, acc_word, f_score = self.evaluate(True)
+        is_best = self._check_epoch(loss_trains, avg_loss_dev, acc_char, acc_word, f_score)
         self.cfg.epoch += 1
-        return has_best
+        return is_best
 
-    def _check(self, loss_trains: List[float], avg_loss_dev: float, acc_char: float,
-               acc_word: float, f_score: float, batch_size: int) -> bool:
+    def _check_epoch(self, loss_trains: List[float], avg_loss_dev: float, acc_char: float,
+                     acc_word: float, f_score: float) -> bool:
         """
-        cfg.check_step번의 step마다 수행하는 체크
+        매 epoch마다 수행하는 체크
         Args:
-            loss_trains:  train 코퍼스에서 각 배치별 loss 리스트
+            loss_trains:   train 코퍼스에서 각 배치별 loss 리스트
             avg_loss_dev:  dev 코퍼스 문장 별 평균 loss
             acc_char:  음절 정확도
             acc_word:  어절 정확도
             f_score:  f-score
-            batch_size:  배치 크기
         Returns:
-            현재 step이 best 성능을 나타냈는 지 여부
+            현재 epoch이 best 성능을 나타냈는 지 여부
         """
         avg_loss_train = sum(loss_trains) / len(loss_trains)
         loss_trains.clear()
@@ -286,38 +274,34 @@ def _check(self, loss_trains: List[float], avg_loss_dev: float, acc_char: float,
         self.acc_words.append(acc_word)
         self.f_scores.append(f_score)
         self.learning_rates.append(self.cfg.learning_rate)
-        self.batch_sizes.append(batch_size)
-        check_id = self.cfg.step // self.cfg.check_step - 1
         is_best = self._is_best()
         is_best_str = 'BEST' if is_best else '< {:.4f}'.format(max(self.f_scores))
-        tqdm.write('    [{:3d}|{:5d}]  [Los trn]  [Los dev]  [Acc chr]  [Acc wrd]  [F-score]'
-                   '           [LR]        [BS]'.format(self.cfg.epoch, check_id))
-        tqdm.write('                 {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f} {:8}  {:.8f}  {}' \
+        logging.info('[Los trn]  [Los dev]  [Acc chr]  [Acc wrd]  [F-score]           [LR]')
+        logging.info('{:9.4f}  {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f} {:8}  {:.8f}' \
                 .format(avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score, is_best_str,
-                        self.cfg.learning_rate, batch_size))
-        print('{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}'.format( \
-                self.cfg.epoch, check_id, avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score,
-                self.cfg.learning_rate, batch_size), file=self.log_file)
+                        self.cfg.learning_rate))
+        print('{}\t{}\t{}\t{}\t{}\t{}\t{}'.format(self.cfg.epoch, avg_loss_train, avg_loss_dev,
+                                                  acc_char, acc_word, f_score,
+                                                  self.cfg.learning_rate), file=self.log_file)
         self.log_file.flush()
-        self.sum_wrt.add_scalar('loss-train', avg_loss_train, check_id)
-        self.sum_wrt.add_scalar('loss-dev', avg_loss_dev, check_id)
-        self.sum_wrt.add_scalar('acc-char', acc_char, check_id)
-        self.sum_wrt.add_scalar('acc-word', acc_word, check_id)
-        self.sum_wrt.add_scalar('f-score', f_score, check_id)
-        self.sum_wrt.add_scalar('learning-rate', self.cfg.learning_rate, check_id)
-        self.sum_wrt.add_scalar('batch-size', batch_size, check_id)
+        self.sum_wrt.add_scalar('loss-train', avg_loss_train, self.cfg.epoch)
+        self.sum_wrt.add_scalar('loss-dev', avg_loss_dev, self.cfg.epoch)
+        self.sum_wrt.add_scalar('acc-char', acc_char, self.cfg.epoch)
+        self.sum_wrt.add_scalar('acc-word', acc_word, self.cfg.epoch)
+        self.sum_wrt.add_scalar('f-score', f_score, self.cfg.epoch)
+        self.sum_wrt.add_scalar('learning-rate', self.cfg.learning_rate, self.cfg.epoch)
         return is_best
 
     def _is_best(self) -> bool:
         """
-        이번 step에 가장 좋은 성능을 냈는 지 확인하고 그럴 경우 현재 상태를 저장한다.
+        이번 epoch에 가장 좋은 성능을 냈는 지 확인하고 그럴 경우 현재 상태를 저장한다.
         Returns:
-            best 여부
+            마지막 f-score의 best 여부
         """
         if len(self.f_scores) > 1 and max(self.f_scores[:-1]) >= self.f_scores[-1]:
             return False
-        # this step hits new max value
-        self.cfg.best_step = self.cfg.step
+        # this epoch hits new max value
+        self.cfg.best_epoch = self.cfg.epoch
         self.model.save('{}/model.state'.format(self.cfg.out_dir))
         self._save_optim('{}/optim.state'.format(self.cfg.out_dir))
         with open('{}/config.json'.format(self.cfg.out_dir), 'w', encoding='UTF-8') as fout:
@@ -347,31 +331,33 @@ def _load_optim(self, path: str, learning_rate: float):
         self.optimizer.load_state_dict(state_dict)
         self.optimizer.param_groups[0]['lr'] = learning_rate
 
-    def evaluate(self) -> Tuple[float, float, float, float]:
+    def evaluate(self, is_dev: bool) -> Tuple[float, float, float, float]:
         """
         evaluate f-score
+        Args:
+            is_dev:  whether evaluate on dev set or not
         Returns:
             average dev loss
             character accuracy
             word accuracy
             f-score
         """
+        dataset = self.dataset_dev if is_dev else self.dataset_test
         self.model.eval()
-        loss_devs = []
-        for dev_sent in self.dataset_dev:
+        losses = []
+        for sent in dataset:
             # 만약 spc_dropout이 1.0 이상이면 공백을 전혀 쓰지 않는 것이므로 평가 시에도 적용한다.
-            dev_labels, dev_contexts = dev_sent.to_tensor(self.cfg, self.rsc,
-                                                          self.cfg.spc_dropout >= 1.0)
+            labels, contexts = sent.to_tensor(self.cfg, self.rsc, self.cfg.spc_dropout >= 1.0)
             if torch.cuda.is_available():
-                dev_labels = dev_labels.cuda()
-                dev_contexts = dev_contexts.cuda()
-            dev_outputs = self.model(dev_contexts)
-            loss_dev = self.criterion(dev_outputs, dev_labels)
-            loss_devs.append(loss_dev.item())
-            _, predicts = F.softmax(dev_outputs, dim=1).max(1)
+                labels = labels.cuda()
+                contexts = contexts.cuda()
+            outputs = self.model(contexts)
+            loss = self.criterion(outputs, labels)
+            losses.append(loss.item())
+            _, predicts = F.softmax(outputs, dim=1).max(1)
             pred_tags = [self.rsc.vocab_out[t.item()] for t in predicts]
-            pred_sent = copy.deepcopy(dev_sent)
+            pred_sent = copy.deepcopy(sent)
             pred_sent.set_pos_result(pred_tags, self.rsc.restore_dic)
-            self.evaler.count(dev_sent, pred_sent)
-        avg_loss_dev = sum(loss_devs) / len(loss_devs)
-        return (avg_loss_dev, ) + self.evaler.evaluate()
+            self.evaler.count(sent, pred_sent)
+        avg_loss = sum(losses) / len(losses)
+        return (avg_loss, ) + self.evaler.evaluate()

From 28df3f2fb245a3c6306b408c95a37da275224aa6 Mon Sep 17 00:00:00 2001
From: krikit <krikit@naver.com>
Date: Tue, 5 Feb 2019 22:28:49 +0900
Subject: [PATCH 21/41] =?UTF-8?q?--check-step,=20--batch-grow=20=EC=98=B5?=
 =?UTF-8?q?=EC=85=98=20=EC=82=AD=EC=A0=9C=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 train/train.py | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/train/train.py b/train/train.py
index ce4b4e5..8943041 100755
--- a/train/train.py
+++ b/train/train.py
@@ -28,8 +28,7 @@ def run(args: Namespace):
     Args:
         args:  program arguments (config)
     """
-    trner = Trainer(args)
-    trner.train()
+    Trainer(args).train()
 
 
 ########
@@ -59,10 +58,6 @@ def main():
                         type=float, default=0.9)
     parser.add_argument('--batch-size', help='batch size <default: 500>', metavar='INT', type=int,
                         default=500)
-    parser.add_argument('--check-step', help='check every N step <default: 10000>', metavar='INT',
-                        type=int, default=10000)
-    parser.add_argument('--batch-grow', help='grow batch size by 1 per N step <default: 10000>',
-                        metavar='INT', type=int, default=10000)
     parser.add_argument('--patience', help='maximum patience count to revert model <default: 10>',
                         metavar='INT', type=int, default=10)
     parser.add_argument('--gpu-num', help='GPU number to use <default: 0>', metavar='INT', type=int,

From 918955f276c6cdddcc1a32e692618470cb571fa4 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Fri, 18 Jan 2019 17:41:14 +0900
Subject: [PATCH 22/41] =?UTF-8?q?=ED=95=99=EC=8A=B5=20=EC=8A=A4=ED=81=AC?=
 =?UTF-8?q?=EB=A6=BD=ED=8A=B8=EB=A5=BC=20=EC=B6=94=EA=B0=80=ED=95=98?=
 =?UTF-8?q?=EA=B8=B0=20=EC=9C=84=ED=95=B4=20rsc/lib=20=EC=95=84=EB=9E=98?=
 =?UTF-8?q?=EC=9D=98=20python=20=EB=AA=A8=EB=93=88=EB=93=A4=EC=9D=84=20src?=
 =?UTF-8?q?/main/python/khaiii=20=EC=95=84=EB=9E=98=EB=A1=9C=20=EC=A0=95?=
 =?UTF-8?q?=EB=A6=AC.=20type=20hint=20=EA=B4=80=EB=A0=A8=20=EC=BD=94?=
 =?UTF-8?q?=EB=93=9C=20=EB=A6=AC=ED=8C=A9=ED=86=A0=EB=A7=81=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 rsc/Makefile                                  |  10 +-
 rsc/bin/compile_errpatch.py                   |  41 +--
 rsc/bin/compile_model.py                      |  11 +-
 rsc/bin/compile_preanal.py                    |  41 +--
 rsc/bin/compile_restore.py                    |  26 +-
 src/main/python/khaiii/__init__.py            |   0
 src/main/python/khaiii/khaiii.py              |  12 +-
 src/main/python/khaiii/munjong/__init__.py    |   0
 .../python/khaiii/munjong}/sejong_corpus.py   | 272 +++++++++---------
 src/main/python/khaiii/resource/__init__.py   |   0
 .../python/khaiii/resource}/char_align.py     | 140 ++++-----
 .../main/python/khaiii/resource}/jaso.py      |  10 +-
 .../main/python/khaiii/resource}/morphs.py    |  26 +-
 .../main/python/khaiii/resource}/resource.py  |  24 +-
 .../main/python/khaiii/resource}/trie.py      | 149 +++++-----
 .../python/khaiii/resource}/vocabulary.py     |  14 +-
 16 files changed, 404 insertions(+), 372 deletions(-)
 create mode 100644 src/main/python/khaiii/__init__.py
 create mode 100644 src/main/python/khaiii/munjong/__init__.py
 rename {rsc/lib => src/main/python/khaiii/munjong}/sejong_corpus.py (76%)
 create mode 100644 src/main/python/khaiii/resource/__init__.py
 rename {rsc/lib => src/main/python/khaiii/resource}/char_align.py (90%)
 rename {rsc/lib => src/main/python/khaiii/resource}/jaso.py (92%)
 rename {rsc/lib => src/main/python/khaiii/resource}/morphs.py (87%)
 rename {rsc/lib => src/main/python/khaiii/resource}/resource.py (72%)
 rename {rsc/lib => src/main/python/khaiii/resource}/trie.py (90%)
 rename {rsc/lib => src/main/python/khaiii/resource}/vocabulary.py (89%)

diff --git a/rsc/Makefile b/rsc/Makefile
index a99039b..36d1a3a 100644
--- a/rsc/Makefile
+++ b/rsc/Makefile
@@ -1,6 +1,6 @@
 HOME_DIR = .
 BIN_DIR = $(HOME_DIR)/bin
-LIB_DIR = $(HOME_DIR)/lib
+SRC_PYTHON = $(HOME_DIR)/../src/main/python
 RSC_SRC = $(HOME_DIR)/src
 PREFIX = /usr/local
 RSC_DIR = $(PREFIX)/share/khaiii
@@ -35,22 +35,22 @@ all: $(MODEL) $(PREANAL) $(RESTORE) $(ERRPATCH)
 $(wordlist 2,100,$(MODEL)): $(firstword $(MODEL))
 $(firstword $(MODEL)): $(RSC_SRC)/$(MODEL_SIZE).config.json $(RSC_SRC)/$(MODEL_SIZE).model.pickle
 	mkdir -p $(RSC_DIR)
-	PYTHONPATH=$(LIB_DIR) python3 $(BIN_DIR)/compile_model.py --model-size $(MODEL_SIZE) --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
+	PYTHONPATH=$(SRC_PYTHON) python3 $(BIN_DIR)/compile_model.py --model-size $(MODEL_SIZE) --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
 
 $(wordlist 2,100,$(PREANAL)): $(firstword $(PREANAL))
 $(firstword $(PREANAL)): $(RSC_SRC)/preanal.auto $(RSC_SRC)/preanal.manual
 	mkdir -p $(RSC_DIR)
-	PYTHONPATH=$(LIB_DIR) python3 $(BIN_DIR)/compile_preanal.py --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
+	PYTHONPATH=$(SRC_PYTHON) python3 $(BIN_DIR)/compile_preanal.py --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
 
 $(wordlist 2,100,$(RESTORE)): $(firstword $(RESTORE))
 $(firstword $(RESTORE)): $(RSC_SRC)/restore.dic $(RSC_SRC)/vocab.out $(RSC_SRC)/vocab.out.more
 	mkdir -p $(RSC_DIR)
-	PYTHONPATH=$(LIB_DIR) python3 $(BIN_DIR)/compile_restore.py --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
+	PYTHONPATH=$(SRC_PYTHON) python3 $(BIN_DIR)/compile_restore.py --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
 
 $(wordlist 2,100,$(ERRPATCH)): $(firstword $(ERRPATCH))
 $(firstword $(ERRPATCH)): $(RSC_SRC)/$(MODEL_SIZE).errpatch.auto $(RSC_SRC)/$(MODEL_SIZE).errpatch.manual
 	mkdir -p $(RSC_DIR)
-	PYTHONPATH=$(LIB_DIR) python3 $(BIN_DIR)/compile_errpatch.py --model-size $(MODEL_SIZE) --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
+	PYTHONPATH=$(SRC_PYTHON) python3 $(BIN_DIR)/compile_errpatch.py --model-size $(MODEL_SIZE) --rsc-src $(RSC_SRC) --rsc-dir $(RSC_DIR)
 
 clean:
 	rm -rf $(RSC_DIR)
diff --git a/rsc/bin/compile_errpatch.py b/rsc/bin/compile_errpatch.py
index 34e4486..caf7d8a 100755
--- a/rsc/bin/compile_errpatch.py
+++ b/rsc/bin/compile_errpatch.py
@@ -5,14 +5,14 @@
 """
 오분석 패치를 빌드하는 스크립트
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import argparse
+from argparse import ArgumentParser, Namespace
 from collections import defaultdict
 import glob
 import itertools
@@ -20,14 +20,16 @@
 import os
 import struct
 import sys
+from typing import Dict, List, Tuple
+
+from khaiii.munjong import sejong_corpus
+from khaiii.resource.char_align import Aligner, AlignError
+from khaiii.resource.morphs import Morph, ParseError
+from khaiii.resource.morphs import WORD_DELIM_STR, SENT_DELIM_STR, WORD_DELIM_NUM, SENT_DELIM_NUM
+from khaiii.resource.trie import Trie
 
-from char_align import Aligner, AlignError
 from compile_preanal import align_to_tag, print_errors
 from compile_restore import load_restore_dic, load_vocab_out
-from morphs import Morph, ParseError
-from morphs import WORD_DELIM_STR, SENT_DELIM_STR, WORD_DELIM_NUM, SENT_DELIM_NUM
-import sejong_corpus
-from trie import Trie
 
 
 #########
@@ -37,7 +39,7 @@ class Entry:
     """
     error patch entry
     """
-    def __init__(self, file_path, line_num, line):
+    def __init__(self, file_path: str, line_num: int, line: str):
         """
         Args:
             file_path:  파일 경로
@@ -65,7 +67,7 @@ def __str__(self):
             return '{}: "{}"'.format(file_num, line)
         return '{}\t{}\t{}'.format(self.raw, Morph.to_str(self.left), Morph.to_str(self.right))
 
-    def key_str(self):
+    def key_str(self) -> str:
         """
         패치의 중복 검사를 하기 위해 원문과 left를 이용하여 키를 생성
         Returns:
@@ -102,7 +104,7 @@ def _parse(self):
 #############
 # functions #
 #############
-def _split_list(lst, delim):
+def _split_list(lst: List[str], delim: str) -> List[List[str]]:
     """
     리스트를 delimiter로 split하는 함수
 
@@ -125,7 +127,8 @@ def _split_list(lst, delim):
     return sublists
 
 
-def align_patch(rsc_src, raw, morph_str):
+def align_patch(rsc_src: Tuple[Aligner, Dict, Dict[str, int]], raw: str, morph_str: str) \
+        -> List[int]:
     """
     패치의 원문과 분석 결과를 음절단위 매핑(정렬)을 수행한다.
     Args:
@@ -171,7 +174,7 @@ def align_patch(rsc_src, raw, morph_str):
     return tag_nums
 
 
-def mix_char_tag(chars, tags):
+def mix_char_tag(chars: str, tags: List[int]) -> List[int]:
     """
     음절과 출력 태그를 비트 연산으로 합쳐서 하나의 (32비트) 숫자로 표현한다.
     Args:
@@ -195,11 +198,11 @@ def mix_char_tag(chars, tags):
     return char_nums
 
 
-def _load_entries(args):
+def _load_entries(args: Namespace) -> List[Entry]:
     """
     패치 엔트리를 파일로부터 로드한다.
     Args:
-        args:  arguments
+        args:  program arguments
     Returns:
         엔트리 리스트
     """
@@ -221,7 +224,7 @@ def _load_entries(args):
     return good_entries
 
 
-def _check_dup(entries):
+def _check_dup(entries: List[Entry]):
     """
     중복된 엔트리가 없는 지 확인한다.
     Args:
@@ -239,7 +242,7 @@ def _check_dup(entries):
     print_errors(bad_entries)
 
 
-def _set_align(rsc_src, entries):    # pylint: disable=invalid-name
+def _set_align(rsc_src: Tuple[Aligner, dict, Dict[str, int]], entries: List[Entry]):
     """
     음절과 형태소 분석 결과를 정렬한다.
     Args:
@@ -265,7 +268,7 @@ def _set_align(rsc_src, entries):    # pylint: disable=invalid-name
     print_errors(bad_entries)
 
 
-def _save_trie(rsc_dir, entries):
+def _save_trie(rsc_dir: str, entries: List[Entry]):
     """
     트라이를 저장한다.
     Args:
@@ -309,7 +312,7 @@ def _save_trie(rsc_dir, entries):
                  (sum([len(r) for r in rights])+1) * struct.Struct('h').size)
 
 
-def run(args):
+def run(args: Namespace):
     """
     run function which is the start point of program
     Args:
@@ -338,7 +341,7 @@ def main():
     """
     main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='기분석 사전을 빌드하는 스크립트')
+    parser = ArgumentParser(description='기분석 사전을 빌드하는 스크립트')
     parser.add_argument('--model-size', help='model size <default: base>',
                         metavar='SIZE', default='base')
     parser.add_argument('--rsc-src', help='source directory (text) <default: ./src>',
diff --git a/rsc/bin/compile_model.py b/rsc/bin/compile_model.py
index fa7b122..1c10d1f 100755
--- a/rsc/bin/compile_model.py
+++ b/rsc/bin/compile_model.py
@@ -5,15 +5,14 @@
 """
 compile trained model for C/C++ decoder
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import argparse
-from argparse import Namespace
+from argparse import ArgumentParser, Namespace
 import json
 import logging
 import os
@@ -21,7 +20,7 @@
 import pickle
 from typing import Tuple
 
-from resource import Resource    # pylint: disable=wrong-import-order
+from khaiii.resource.resource import Resource
 
 
 #############
@@ -40,7 +39,7 @@ def load_cfg_rsc(rsc_src: str, model_size: str) -> Tuple[Namespace, Resource]:
     file_path = '{}/{}.config.json'.format(rsc_src, model_size)
     cfg_dic = json.load(open(file_path, 'r', encoding='UTF-8'))
     logging.info('config: %s', json.dumps(cfg_dic, indent=2))
-    cfg = argparse.Namespace()
+    cfg = Namespace()
     for key, val in cfg_dic.items():
         setattr(cfg, key, val)
     cwd = os.path.realpath(os.getcwd())
@@ -159,7 +158,7 @@ def main():
     """
     main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='part-of-speech tagger')
+    parser = ArgumentParser(description='part-of-speech tagger')
     parser.add_argument('--model-size', help='model size <default: base>',
                         metavar='SIZE', default='base')
     parser.add_argument('--rsc-src', help='source directory (model) <default: ./src>',
diff --git a/rsc/bin/compile_preanal.py b/rsc/bin/compile_preanal.py
index 556f5ea..6874d2c 100755
--- a/rsc/bin/compile_preanal.py
+++ b/rsc/bin/compile_preanal.py
@@ -5,36 +5,38 @@
 """
 기분석 사전을 빌드하는 스크립트
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import argparse
+from argparse import ArgumentParser, Namespace
 from collections import defaultdict
 import glob
 import logging
 import os
 import struct
 import sys
+from typing import Dict, List, Tuple
+
+from khaiii.munjong import sejong_corpus
+from khaiii.resource.char_align import Aligner, AlignError, MrpChr
+from khaiii.resource.morphs import Morph, ParseError
+from khaiii.resource.trie import Trie
 
 from compile_restore import load_restore_dic, load_vocab_out, append_new_entries
-from char_align import Aligner, AlignError, MrpChr
-from morphs import Morph, ParseError
-import sejong_corpus
-from trie import Trie
 
 
 #########
 # types #
 #########
-class Entry(object):
+class Entry:
     """
     pre-analyzed dictionary entry
     """
-    def __init__(self, file_path, line_num, line):
+    def __init__(self, file_path: str, line_num: int, line: str):
         """
         Args:
             file_path:  파일 경로
@@ -59,7 +61,7 @@ def __str__(self):
         line = '# {}'.format(self.line) if self.is_sharp else self.line
         if self.err_msg:
             return '{}{}: "{}"'.format(file_num, self.err_msg, line)
-        elif self.is_sharp:
+        if self.is_sharp:
             return '{}: "{}"'.format(file_num, line)
         return '{}{}\t{}'.format(self.word, '*' if self.is_pfx else '', Morph.to_str(self.morphs))
 
@@ -94,7 +96,7 @@ def _parse(self):
 #############
 # functions #
 #############
-def print_errors(entries):
+def print_errors(entries: List[Entry]):
     """
     에러가 발생한 엔트리를 출력하고 프로그램을 종료한다.
     Args:
@@ -107,7 +109,7 @@ def print_errors(entries):
         sys.exit(1)
 
 
-def _load_entries(args):
+def _load_entries(args: Namespace) -> List[Entry]:
     """
     사전 엔트리를 파일로부터 로드한다.
     Args:
@@ -133,7 +135,7 @@ def _load_entries(args):
     return good_entries
 
 
-def _check_dup(entries):
+def _check_dup(entries: List[Entry]):
     """
     중복된 엔트리가 없는 지 확인한다.
     Args:
@@ -151,7 +153,7 @@ def _check_dup(entries):
     print_errors(bad_entries)
 
 
-def _set_align(aligner, Word, entries):    # pylint: disable=invalid-name
+def _set_align(aligner: Aligner, Word: type, entries: List[Entry]):    # pylint: disable=invalid-name
     """
     음절과 형태소 분석 결과를 정렬한다.
     Args:
@@ -173,7 +175,9 @@ def _set_align(aligner, Word, entries):    # pylint: disable=invalid-name
     print_errors(bad_entries)
 
 
-def align_to_tag(raw_word, alignment, restore, vocab):
+def align_to_tag(raw_word: str, alignment: List[List[str]], restore: Tuple[dict, dict],
+                 vocab: Tuple[Dict[str, int], Dict[str, int]]) \
+        -> Tuple[List[str], List[int]]:
     """
     어절의 원문과 정렬 정보를 활용해 음절과 매핑된 태그를 생성한다.
     Args:
@@ -224,7 +228,8 @@ def align_to_tag(raw_word, alignment, restore, vocab):
     return tag_outs, tag_nums
 
 
-def _set_tag_out(restore_dic, restore_new, vocab_out, vocab_new, entries):
+def _set_tag_out(restore_dic: dict, restore_new: dict, vocab_out: Dict[str, int],
+                 vocab_new: Dict[str, int], entries: List[Entry]):
     """
     음절 정렬로부터 출력 태그를 결정하고 출력 태그의 번호를 매핑한다.
     Args:
@@ -240,7 +245,7 @@ def _set_tag_out(restore_dic, restore_new, vocab_out, vocab_new, entries):
                                                       (vocab_out, vocab_new))
 
 
-def _save_trie(rsc_dir, entries):
+def _save_trie(rsc_dir: str, entries: List[Entry]):
     """
     트라이를 저장한다.
     Args:
@@ -270,7 +275,7 @@ def _save_trie(rsc_dir, entries):
                  (sum([len(e.tag_nums) for e in entries])+1) * struct.Struct('H').size)
 
 
-def run(args):
+def run(args: Namespace):
     """
     run function which is the start point of program
     Args:
@@ -301,7 +306,7 @@ def main():
     """
     main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='기분석 사전을 빌드하는 스크립트')
+    parser = ArgumentParser(description='기분석 사전을 빌드하는 스크립트')
     parser.add_argument('--rsc-src', help='source directory (text) <default: ./src>',
                         metavar='DIR', default='./src')
     parser.add_argument('--rsc-dir', help='target directory (binary) <default: ./share/khaiii>',
diff --git a/rsc/bin/compile_restore.py b/rsc/bin/compile_restore.py
index f602096..132df3a 100755
--- a/rsc/bin/compile_restore.py
+++ b/rsc/bin/compile_restore.py
@@ -5,23 +5,22 @@
 """
 원형복원 사전을 빌드하는 스크립트
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-from __future__ import print_function
-
-import argparse
+from argparse import ArgumentParser, Namespace
 from collections import defaultdict
 import logging
 import os
 import struct
 import sys
+from typing import Dict, Tuple
 
-from morphs import TAG_SET
+from khaiii.resource.morphs import TAG_SET
 
 
 #############
@@ -33,7 +32,7 @@
 #############
 # functions #
 #############
-def load_restore_dic(file_path):
+def load_restore_dic(file_path: str) -> Dict[Tuple[str, str], Dict[int, str]]:
     """
     원형복원 사전을 로드한다.
     Args:
@@ -61,7 +60,7 @@ def load_restore_dic(file_path):
     return restore_dic
 
 
-def load_vocab_out(rsc_src):
+def load_vocab_out(rsc_src: str) -> Dict[str, int]:
     """
     출력 태그 vocabulary를 로드한다.
     Args:
@@ -80,7 +79,7 @@ def load_vocab_out(rsc_src):
     return {tag: idx for idx, tag in enumerate(vocab_out + vocab_out_more, start=1)}
 
 
-def append_new_entries(rsc_src, restore_new, vocab_new):
+def append_new_entries(rsc_src: str, restore_new: dict, vocab_new: Dict[str, int]):
     """
     기분석 사전 빌드 중에 새로 추가가 필요한 사전 엔트리를 해당 사전에 추가한다.
     Args:
@@ -103,7 +102,7 @@ def append_new_entries(rsc_src, restore_new, vocab_new):
                 print(tag, file=fout)
 
 
-def _make_bin(restore_dic, vocab_out, vocab_new):
+def _make_bin(restore_dic: dict, vocab_out: Dict[str, int], vocab_new: Dict[str, int]) -> dict:
     """
     두 텍스트 사전을 읽어들여 바이너리 형태의 key-value 사전을 만든다.
     Args:
@@ -140,7 +139,7 @@ def _make_bin(restore_dic, vocab_out, vocab_new):
     return bin_dic
 
 
-def _save_restore_dic(rsc_dir, bin_dic):
+def _save_restore_dic(rsc_dir: str, bin_dic: dict):
     """
     원형복원 바이너리 사전을 저장한다.
     Args:
@@ -158,14 +157,13 @@ def _save_restore_dic(rsc_dir, bin_dic):
     logging.info('restore.val: %d', 4 * sum([len(vals) for vals in bin_dic.values()]))
 
 
-def _save_restore_one(rsc_dir, vocab_out, vocab_new):
+def _save_restore_one(rsc_dir: str, vocab_out: Dict[str, int], vocab_new: Dict[str, int]):
     """
     출력 태그 번호 별 원형복원을 하지 않는 비복원 사전을 저장한다.
     Args:
         rsc_dir:  resource directory
         vocab_out:  출력 태그 사전
         vocab_new:  출력 태그 사전에 추가할 새로운 태그
-    :return:
     """
     idx_tags = sorted([(idx, tag) for tag, idx
                        in list(vocab_out.items()) + list(vocab_new.items())])
@@ -182,7 +180,7 @@ def _save_restore_one(rsc_dir, vocab_out, vocab_new):
     logging.info('restore.one: %d', 1 + len(idx_tags))
 
 
-def run(args):
+def run(args: Namespace):
     """
     run function which is the start point of program
     Args:
@@ -210,7 +208,7 @@ def main():
     """
     main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='기분석 사전을 빌드하는 스크립트')
+    parser = ArgumentParser(description='기분석 사전을 빌드하는 스크립트')
     parser.add_argument('--rsc-src', help='source directory (text) <default: ./src>',
                         metavar='DIR', default='./src')
     parser.add_argument('--rsc-dir', help='target directory (binary) <default: ./share/khaiii>',
diff --git a/src/main/python/khaiii/__init__.py b/src/main/python/khaiii/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/src/main/python/khaiii/khaiii.py b/src/main/python/khaiii/khaiii.py
index 5f174e2..4bdb193 100644
--- a/src/main/python/khaiii/khaiii.py
+++ b/src/main/python/khaiii/khaiii.py
@@ -12,8 +12,7 @@
 ###########
 # imports #
 ###########
-import argparse
-from argparse import Namespace
+from argparse import ArgumentParser, Namespace
 import ctypes
 from ctypes.util import find_library
 import logging
@@ -26,11 +25,10 @@
 #########
 # types #
 #########
-class _khaiii_morph_t(ctypes.Structure):    # pylint: disable=invalid-name
+class _khaiii_morph_t(ctypes.Structure):    # pylint: disable=invalid-name,too-few-public-methods
     """
     khaiii_morph_t structure
     """
-    pass
 
 
 _khaiii_morph_t._fields_ = [    # pylint: disable=protected-access
@@ -43,11 +41,10 @@ class _khaiii_morph_t(ctypes.Structure):    # pylint: disable=invalid-name
 ]
 
 
-class _khaiii_word_t(ctypes.Structure):    # pylint: disable=invalid-name
+class _khaiii_word_t(ctypes.Structure):    # pylint: disable=invalid-name,too-few-public-methods
     """
     khaiii_word_t structure
     """
-    pass
 
 
 _khaiii_word_t._fields_ = [    # pylint: disable=protected-access
@@ -63,7 +60,6 @@ class KhaiiiExcept(Exception):
     """
     khaiii API를 위한 표준 예외 클래스
     """
-    pass
 
 
 class KhaiiiMorph:
@@ -384,7 +380,7 @@ def main():
     """
     main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='khaiii API module test program')
+    parser = ArgumentParser(description='khaiii API module test program')
     parser.add_argument('--lib-path', help='library path', metavar='FILE', default='')
     parser.add_argument('--rsc-dir', help='resource directory', metavar='DIR', default='')
     parser.add_argument('--opt-str', help='option string (JSON format)', metavar='JSON', default='')
diff --git a/src/main/python/khaiii/munjong/__init__.py b/src/main/python/khaiii/munjong/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/rsc/lib/sejong_corpus.py b/src/main/python/khaiii/munjong/sejong_corpus.py
similarity index 76%
rename from rsc/lib/sejong_corpus.py
rename to src/main/python/khaiii/munjong/sejong_corpus.py
index c8b9681..fb19fe3 100755
--- a/rsc/lib/sejong_corpus.py
+++ b/src/main/python/khaiii/munjong/sejong_corpus.py
@@ -5,19 +5,19 @@
 """
 Sejong corpus parser
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import argparse
-import codecs
+from argparse import ArgumentParser
 import logging
 import os
 import re
 import sys
+from typing import Iterator, TextIO
 from unicodedata import normalize as norm
 
 
@@ -45,7 +45,7 @@
 _OPEN_TAGS_IN_WRITTEN_SENT = {'<date>', }    # open tags in sentence at written corpus
 
 # close tags in sentence at written corpus
-_CLOSE_TAGS_IN_WRITTEN_SENT = set([tag[0] + '/' + tag[1:] for tag in _OPEN_TAGS_IN_WRITTEN_SENT])
+_CLOSE_TAGS_IN_WRITTEN_SENT = {tag[0] + '/' + tag[1:] for tag in _OPEN_TAGS_IN_WRITTEN_SENT}
 
 # open and close tags in sentence at written corpus
 _TAGS_IN_WRITTEN_SENT = _OPEN_TAGS_IN_WRITTEN_SENT | _CLOSE_TAGS_IN_WRITTEN_SENT
@@ -58,84 +58,67 @@ class ParseError(Exception):
     """
     error occurred while parsing corpus
     """
-    pass
 
 
-class Sentence(object):
+class Morph:    # pylint: disable=too-few-public-methods
     """
-    sentence
+    morpheme
     """
-    def __init__(self):
-        self.words = []    # word list
-        self._wids = set()    # word ID set
+    def __init__(self, lex: str = '', tag: str = ''):
+        if ' ' in lex:
+            raise ParseError('space in raw morph: %s' % lex)
+        self.lex = lex    # lexical form
+        self.tag = tag    # part-of-speech tag
 
     def __str__(self):
-        words_str = '\n'.join([str(_) for _ in self.words])
-        return '# %s\n%s' % (self.raw_str(), words_str)
-
-    def raw_str(self):
-        """
-        raw sentence (words)
-        :return:  raw sentence
-        """
-        return ' '.join([_.raw for _ in self.words])
-
-    def morph_str(self):
-        """
-        make morpheme string
-        :return:  morpheme string
-        """
-        return ' + '.join([_.morph_str() for _ in self.words])
+        return '{}/{}'.format(self.lex, self.tag)
 
-    @classmethod
-    def is_opening(cls, line):
-        """
-        whether sentence opening or not
-        :param  line:  line
-        :return:  whether opening or not
-        """
-        return line in SENT_OPEN_TAGS
+    def __eq__(self, other: 'Morph'):
+        return self.lex == other.lex and self.tag == other.tag
 
-    @classmethod
-    def is_closing(cls, line):
-        """
-        whether sentence closing or not
-        :param  line:  line
-        :return:  whether closing or not
-        """
-        return line in SENT_CLOSE_TAGS
+    def __hash__(self):
+        return hash(str(self))
 
     @classmethod
-    def is_tag_in_sent(cls, line):
-        """
-        whether tag in sentence or not
-        :param  line:  line
-        :return:  whether tag or not
+    def parse(cls, token_str: str, file_name: str = '', line_num: int = 0):
         """
-        return line in _TAGS_IN_WRITTEN_SENT
-
-    def is_good_tags(self):
-        """
-        whether all tags in sentence are good(correct) or not
-        :return:  whether all tags are good
+        parse token string
+        Args:
+            token_str:  morpheme/tag string
+            file_name:  file name
+            line_num:  line number
+        Returns:
+            Morph object
         """
-        return not [morph.tag for word in self.words for morph in word.morphs
-                    if morph.tag not in TAG_SET]
+        def _raise(err_msg: str, file_name: str = None, line_num: int = 0):
+            """
+            에러 메세지를 이용해 파싱 에러를 발생시킨다.
+            Arguments:
+                err_msg:  에러 메세지
+                file_name:  파일명
+                line_num:  라인 번호
+            """
+            file_pfx = '{}({}) '.format(file_name, line_num) if file_name and line_num else ''
+            raise ParseError('{}{}'.format(file_pfx, err_msg))
 
-    def append(self, word):
-        """
-        append word
-        :param word:  Word object
-        """
-        if word.wid and word.wid in self._wids:
-            raise ParseError('duplicated word ID: %s' % word)
-        self.words.append(word)
-        self._wids.add(word.wid)
+        morph = Morph()
+        morph.lex, morph.tag = token_str.rsplit('/', 1)
+        if not morph.lex:
+            _raise('no text in morpheme: {}'.format(token_str), file_name, line_num)
+        if ' ' in morph.lex:
+            _raise('space in raw morph: {}'.format(token_str), file_name, line_num)
+        if morph.tag not in TAG_SET:
+            _raise('invalid tag: {} in {}'.format(morph.tag, token_str), file_name, line_num)
+        for char in morph.lex:
+            if 0x1100 <= ord(char) < 0x1200:
+                _raise('Hangul Jamo character: "{}" in {}'.format(char, token_str), file_name,
+                       line_num)
+        return morph
 
 
-class Word(object):
+class Word:
     """
-    word(EoJeol)
+    word(어절)
     """
     def __init__(self):
         self.wid = ''    # word ID
@@ -143,10 +126,9 @@ def __init__(self):
         self.morphs = []    # morpheme list
 
     def __str__(self):
-        return '%s\t%s\t%s' % (self.wid, self.raw,
-                               ' + '.join([str(morph) for morph in self.morphs]))
+        return '{}\t{}\t{}'.format(self.wid, self.raw, ' + '.join([str(m) for m in self.morphs]))
 
-    def __eq__(self, other):
+    def __eq__(self, other: 'Word'):
         if self.raw != other.raw:
             return False
         if len(self.morphs) != len(other.morphs):
@@ -156,20 +138,24 @@ def __eq__(self, other):
                 return False
         return True
 
-    def morph_str(self):
+    def morph_str(self) -> str:
         """
         make morpheme string
-        :return:  morpheme string
+        Returns:
+            morpheme string
         """
         return ' + '.join([str(_) for _ in self.morphs])
 
     @classmethod
-    def parse(cls, line, file_name, line_num):
+    def parse(cls, line: str, file_name: str, line_num: int) -> 'Word':
         """
         parse word(EoJeol) with single line
-        :param  line:  line
-        :param  file_name:  file name
-        :param  line_num:  line number
+        Args:
+            line:  line
+            file_name:  file name
+            line_num:  line number
+        Returns:
+            Word object
         """
         cols = line.split('\t')
         if len(cols) != 3:
@@ -192,68 +178,98 @@ def parse(cls, line, file_name, line_num):
         return word
 
 
-class Morph(object):    # pylint: disable=too-few-public-methods
+class Sentence:
     """
-    morpheme
+    sentence
     """
-    def __init__(self, lex='', tag=''):
-        if ' ' in lex:
-            raise ParseError('space in raw morph: %s' % lex)
-        self.lex = lex    # lexical form
-        self.tag = tag    # part-of-speech tag
+    def __init__(self):
+        self.words = []    # word list
+        self._wids = set()    # word ID set
 
     def __str__(self):
-        return '%s/%s' % (self.lex, self.tag)
+        words_str = '\n'.join([str(_) for _ in self.words])
+        return '# %s\n%s' % (self.raw_str(), words_str)
 
-    def __eq__(self, other):
-        return self.lex == other.lex and self.tag == other.tag
+    def raw_str(self) -> str:
+        """
+        raw sentence (words)
+        Returns:
+            raw sentence
+        """
+        return ' '.join([_.raw for _ in self.words])
 
-    def __hash__(self):
-        return hash(str(self))
+    def morph_str(self) -> str:
+        """
+        make morpheme string
+        Returns:
+            morpheme string
+        """
+        return ' + '.join([_.morph_str() for _ in self.words])
 
     @classmethod
-    def parse(cls, token_str, file_name='', line_num=0):
+    def is_opening(cls, line: str) -> bool:
         """
-        parse token string
-        :param  token_str:  morpheme/tag string
-        :param  file_name:  file name
-        :param  line_num:  line number
-        :return:  Morph object
+        whether sentence opening or not
+        Args:
+            line:  line
+        Returns:
+            whether opening or not
         """
-        def _raise(err_msg, file_name=None, line_num=0):
-            """
-            에러 메세지를 이용해 파싱 에러를 발생시킨다.
-            Arguments:
-                err_msg:  에러 메세지
-                file_name:  파일명
-                line_num:  라인 번호
-            """
-            file_pfx = '{}({}) '.format(file_name, line_num) if file_name and line_num else ''
-            raise ParseError('{}{}'.format(file_pfx, err_msg))
+        return line in SENT_OPEN_TAGS
 
-        morph = Morph()
-        morph.lex, morph.tag = token_str.rsplit('/', 1)
-        if not morph.lex:
-            _raise('no text in morpheme: {}'.format(token_str), file_name, line_num)
-        if ' ' in morph.lex:
-            _raise('space in raw morph: {}'.format(token_str), file_name, line_num)
-        if morph.tag not in TAG_SET:
-            _raise('invalid tag: {} in {}'.format(morph.tag, token_str), file_name, line_num)
-        for char in morph.lex:
-            if 0x1100 <= ord(char) < 0x1200:
-                _raise('Hangul Jamo character: "{}" in {}'.format(char, token_str), file_name,
-                       line_num)
-        return morph
+    @classmethod
+    def is_closing(cls, line: str) -> bool:
+        """
+        whether sentence closing or not
+        Args:
+            line:  line
+        Returns:
+            whether closing or not
+        """
+        return line in SENT_CLOSE_TAGS
+
+    @classmethod
+    def is_tag_in_sent(cls, line: str) -> bool:
+        """
+        whether tag in sentence or not
+        Args:
+            line:  line
+        Returns:
+            whether tag or not
+        """
+        return line in _TAGS_IN_WRITTEN_SENT
+
+    def is_good_tags(self) -> bool:
+        """
+        whether all tags in sentence are good(correct) or not
+        Returns:
+            whether all tags are good
+        """
+        return not [morph.tag for word in self.words for morph in word.morphs
+                    if morph.tag not in TAG_SET]
+
+    def append(self, word: Word):
+        """
+        append word
+        Args:
+            word:  Word object
+        """
+        if word.wid and word.wid in self._wids:
+            raise ParseError('duplicated word ID: %s' % word)
+        self.words.append(word)
+        self._wids.add(word.wid)
 
 
 #############
 # functions #
 #############
-def sents(fin):
+def sents(fin: TextIO) -> Iterator[Sentence]:
     """
     load from file and return sentences (generator)
-    :param  fin:  input file object
-    :yeild:  Sentence object
+    Args:
+        fin:  input file object
+    Yields:
+        Sentence object
     """
     file_name = os.path.basename(fin.name)
     par_errs = []
@@ -292,14 +308,10 @@ def sents(fin):
 
 def run():
     """
-    load Sejong corpus and print
+    run function which is the start point of program
     """
-    try:
-        for sent in sents(sys.stdin):
-            print(sent)
-    except ParseError as par_err:
-        logging.error(par_err)
-        sys.exit(1)
+    for sent in sents(sys.stdin):
+        print(sent)
 
 
 ########
@@ -307,18 +319,18 @@ def run():
 ########
 def main():
     """
-    load Sejong corpus and print
+    main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='load Sejong corpus and print')
+    parser = ArgumentParser(description='load Sejong corpus and print')
     parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
     parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
     parser.add_argument('--debug', help='enable debug', action='store_true')
     args = parser.parse_args()
 
     if args.input:
-        sys.stdin = codecs.open(args.input, 'r', encoding='UTF-8')
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
     if args.output:
-        sys.stdout = codecs.open(args.output, 'w', encoding='UTF-8')
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
     if args.debug:
         logging.basicConfig(level=logging.DEBUG)
     else:
diff --git a/src/main/python/khaiii/resource/__init__.py b/src/main/python/khaiii/resource/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/rsc/lib/char_align.py b/src/main/python/khaiii/resource/char_align.py
similarity index 90%
rename from rsc/lib/char_align.py
rename to src/main/python/khaiii/resource/char_align.py
index 65af965..1682606 100644
--- a/rsc/lib/char_align.py
+++ b/src/main/python/khaiii/resource/char_align.py
@@ -4,29 +4,71 @@
 """
 형태소 분석 결과와 원문의 음절을 정렬하는 모듈
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import codecs
 from collections import Counter, defaultdict
 import logging
 import os
+from typing import List, Tuple
 
-import jaso
+from khaiii.munjong.sejong_corpus import Word
+from khaiii.resource import jaso
+from khaiii.resource.morphs import Morph
 
 
 #########
 # types #
 #########
-class Aligner(object):
+class MrpChr:    # pylint: disable=too-few-public-methods
+    """
+    음절과 태그 pair
+    """
+    def __init__(self, char: str, tag: str):
+        """
+        Args:
+            char:  음절
+            tag:  태그
+        """
+        self.char = char
+        self.tag = tag
+
+    def __str__(self):
+        return '{}/{}'.format(self.char, self.tag)
+
+    def __hash__(self):
+        return hash(str(self))
+
+    def __eq__(self, other: 'MrpChr'):
+        """
+        Args:
+            other:  다른 객체
+        Returns:
+            같을 경우 True
+        """
+        return self.char == other.char and self.tag == other.tag
+
+    @classmethod
+    def to_str(cls, mrp_chrs: List['MrpChr']):
+        """
+        MrpChr 객체 리스트를 문자열로 변환하는 메소드
+        Args:
+            mrp_chrs:  MrpChr 객체 리스트
+        Returns:
+            변환된 문자열
+        """
+        return ' '.join([str(m) for m in mrp_chrs])
+
+
+class Aligner:
     """
     음절과 형태소 분석 결과의 정렬을 수행하는 클래스
     """
-    def __init__(self, rsc_src):
+    def __init__(self, rsc_src: str):
         """
         리소스를 오픈하고 초기화한다.
         Args:
@@ -37,7 +79,7 @@ def __init__(self, rsc_src):
         self.middle_unmapped = defaultdict(Counter)
         self._open(rsc_src)
 
-    def align(self, word):
+    def align(self, word: Word) -> List[List[MrpChr]]:
         """
         어절의 원문과 분석 결과를 음절 단위로 정렬(매핑)한다.
         Args:
@@ -78,7 +120,7 @@ def print_middle_cnt(self):
         logging.info('total number of unmapped pairs: %d',
                      sum([sum(cnt.values()) for cnt in self.middle_unmapped.values()]))
 
-    def _open(self, rsc_dir):
+    def _open(self, rsc_dir: str):
         """
         initialize resources
         Args:
@@ -86,7 +128,7 @@ def _open(self, rsc_dir):
         """
         file_path = '{}/char_align.map'.format(rsc_dir)
         file_name = os.path.basename(file_path)
-        for line_num, line in enumerate(codecs.open(file_path, 'r', encoding='UTF-8'), start=1):
+        for line_num, line in enumerate(open(file_path, 'r', encoding='UTF-8'), start=1):
             line = line.rstrip('\r\n')
             if not line or line[0] == '#':
                 continue
@@ -105,7 +147,7 @@ def _open(self, rsc_dir):
             self.align_map[key] = map_nums
 
     @classmethod
-    def _get_morph_raw(cls, word):
+    def _get_morph_raw(cls, word: Word) -> str:
         """
         get raw string from morphemes
         Args:
@@ -116,7 +158,7 @@ def _get_morph_raw(cls, word):
         return ''.join([m.lex for m in word.morphs])
 
     @classmethod
-    def _norm(cls, text):
+    def _norm(cls, text: str) -> str:
         """
         unicode normalization of text
         Args:
@@ -127,7 +169,7 @@ def _norm(cls, text):
         return jaso.decompose(text)
 
     @classmethod
-    def _align_phoneme(cls, raw_word, mrp_chrs):
+    def _align_phoneme(cls, raw_word: str, mrp_chrs: List[MrpChr]) -> List[List[MrpChr]]:
         """
         align word with morpheme which is same phoneme
         Args:
@@ -163,7 +205,7 @@ def _align_phoneme(cls, raw_word, mrp_chrs):
         return maps
 
     @classmethod
-    def _align_forward(cls, raw_word, mrp_chrs):
+    def _align_forward(cls, raw_word: str, mrp_chrs: List[MrpChr]) -> Tuple[int, int]:
         """
         align from front of word
         Args:
@@ -192,7 +234,7 @@ def _align_forward(cls, raw_word, mrp_chrs):
         return word_idx, mrp_chrs_idx
 
     @classmethod
-    def _align_backward(cls, raw_word, mrp_chrs):
+    def _align_backward(cls, raw_word: str, mrp_chrs: List[MrpChr]) -> Tuple[int, int]:
         """
         align from back of word
         Args:
@@ -226,7 +268,7 @@ def _align_backward(cls, raw_word, mrp_chrs):
         return word_idx+1, mrp_chrs_idx+1
 
     @classmethod
-    def _is_verb_ending(cls, verb, ending):
+    def _is_verb_ending(cls, verb: Morph, ending: Morph) -> bool:
         """
         whether is verb + ending pattern or not
         Args:
@@ -241,7 +283,7 @@ def _is_verb_ending(cls, verb, ending):
             ending_tag in {'EC', 'EP', 'EF', 'ETN', 'ETM'}
 
     @classmethod
-    def _are_first_last_phoneme_same(cls, raw_word, mrp_chrs):
+    def _are_first_last_phoneme_same(cls, raw_word: str, mrp_chrs: List[MrpChr]) -> bool:
         """
         whether are same the first phoneme and last phoneme
         Args:
@@ -255,7 +297,7 @@ def _are_first_last_phoneme_same(cls, raw_word, mrp_chrs):
         return word_norm[0] == morph_norm[0] and word_norm[-1] == morph_norm[-1]
 
     @classmethod
-    def _is_ah_ending_verb(cls, mrp_chr):
+    def _is_ah_ending_verb(cls, mrp_chr: MrpChr) -> bool:
         """
         whether 'ㅏ' ending verb or not
         Args:
@@ -269,7 +311,7 @@ def _is_ah_ending_verb(cls, mrp_chr):
         return len(norm_char) == 2 and norm_char[1] == 'ᅡ'    # code is 4449, not 12623
 
     @classmethod
-    def _is_eo_ending_verb(cls, mrp_chr):
+    def _is_eo_ending_verb(cls, mrp_chr: MrpChr) -> bool:
         """
         whether 'ㅓ', 'ㅐ' ending verb or not
         Args:
@@ -284,7 +326,8 @@ def _is_eo_ending_verb(cls, mrp_chr):
                 norm_char[1] in ['ᅥ', 'ᅧ', 'ᅢ'])    # code is 4453, 4455, 4450
 
     @classmethod
-    def _align_middle_zero2one(cls, pfx_word, pfx_map, mdl_mrp_chrs, sfx_word, sfx_map):
+    def _align_middle_zero2one(cls, pfx_word: Word, pfx_map: List[MrpChr],
+                               mdl_mrp_chrs: List[MrpChr], sfx_word: Word, sfx_map: List[MrpChr]):
         """
         align middle chunks after forward/backward aligning which has no middle raw character,
             but has a remaining middle single morpheme character
@@ -322,7 +365,7 @@ def _attach_to_pfx():
             raise RuntimeError('nowhere attach to')
 
     @classmethod
-    def _is_share_phoneme(cls, mdl_word, mdl_mrp_chrs):
+    def _is_share_phoneme(cls, mdl_word: str, mdl_mrp_chrs: List[MrpChr]) -> bool:
         """
         whether middle word characters and morpheme characters share same phoneme
         Args:
@@ -342,7 +385,7 @@ def _is_share_phoneme(cls, mdl_word, mdl_mrp_chrs):
                 return False
         return True
 
-    def _align_middle_by_dic(self, mdl_word, mdl_mrp_chrs):
+    def _align_middle_by_dic(self, mdl_word: str, mdl_mrp_chrs: List[MrpChr]) -> List[List[MrpChr]]:
         """
         align middle chunks after forward/backward aligning with mapping dictionary
         Args:
@@ -364,7 +407,8 @@ def _align_middle_by_dic(self, mdl_word, mdl_mrp_chrs):
                 idx += int(map_num)
         return maps
 
-    def _align_middle(self, mdl_word, mdl_mrp_chrs, raw_word, mrp_chrs):
+    def _align_middle(self, mdl_word: str, mdl_mrp_chrs: List[MrpChr], raw_word: str,
+                      mrp_chrs: List[MrpChr]) -> List[List[MrpChr]]:
         """
         align middle chunks after forward/backward aligning
         Args:
@@ -417,7 +461,9 @@ def _align_middle(self, mdl_word, mdl_mrp_chrs, raw_word, mrp_chrs):
         return maps
 
     @classmethod
-    def _align_middle_preproc(cls, pfx_mrp_chrs, pfx_map, mdl_mrp_chrs, sfx_mrp_chrs, sfx_map):
+    def _align_middle_preproc(cls, pfx_mrp_chrs: List[MrpChr], pfx_map: List[List[MrpChr]],
+                              mdl_mrp_chrs: List[MrpChr], sfx_mrp_chrs: List[MrpChr],
+                              sfx_map: List[List[MrpChr]]):
         """
         pre-processing middle part after forward/backward mapping before applying rules
         Args:
@@ -454,7 +500,7 @@ def _align_middle_preproc(cls, pfx_mrp_chrs, pfx_map, mdl_mrp_chrs, sfx_mrp_chrs
                 sfx_map[0].insert(0, mdl_mrp_chrs[-1])
                 del mdl_mrp_chrs[-1]
 
-    def _get_pfx_mdl_sfx(self, raw_word, mrp_chrs):
+    def _get_pfx_mdl_sfx(self, raw_word: str, mrp_chrs: List[MrpChr]) -> Tuple:
         """
         get prefix, middle, suffix after forward/backward align
         Args:
@@ -481,7 +527,7 @@ def _get_pfx_mdl_sfx(self, raw_word, mrp_chrs):
                (pfx_mrp_chrs, mdl_mrp_chrs, sfx_mrp_chrs), \
                (pfx_map, sfx_map)
 
-    def _align_forward_backward(self, raw_word, mrp_chrs):
+    def _align_forward_backward(self, raw_word: str, mrp_chrs: List[MrpChr]) -> List[str]:
         """
         align word with morpheme which is same phoneme
         Args:
@@ -497,7 +543,7 @@ def _align_forward_backward(self, raw_word, mrp_chrs):
 
         if not mdl_word and not mdl_mrp_chrs:
             return pfx_map + sfx_map
-        elif not mdl_word:
+        if not mdl_word:
             if len(mdl_mrp_chrs) == 1:
                 self._align_middle_zero2one(pfx_word, pfx_map, mdl_mrp_chrs, sfx_word, sfx_map)
                 return pfx_map + sfx_map
@@ -529,7 +575,7 @@ class AlignError(Exception):
     """
     음절 정렬 과정에서 나타나는 예외
     """
-    def __init__(self, pfx):
+    def __init__(self, pfx: str):
         """
         Args:
             pfx:  예외 출력 시 보여줄 prefix (카테고리)
@@ -541,50 +587,10 @@ def __init__(self, pfx):
     def __str__(self):
         return '\n'.join(['%s %s' % (self._pfx, _) for _ in self._msgs] + ['', ])
 
-    def add_msg(self, msg):
+    def add_msg(self, msg: str):
         """
         메세제를 추가한다.
         Args:
             msg:  에러 메세지
         """
         self._msgs.append(msg)
-
-
-class MrpChr(object):    # pylint: disable=too-few-public-methods
-    """
-    음절과 태그 pair
-    """
-    def __init__(self, char, tag):
-        """
-        Args:
-            char:  음절
-            tag:  태그
-        """
-        self.char = char
-        self.tag = tag
-
-    def __str__(self):
-        return '%s/%s' % (self.char, self.tag)
-
-    def __hash__(self):
-        return hash(str(self))
-
-    def __eq__(self, other):
-        """
-        Args:
-            other:  다른 객체
-        Returns:
-            같을 경우 True
-        """
-        return self.char == other.char and self.tag == other.tag
-
-    @classmethod
-    def to_str(cls, mrp_chrs):
-        """
-        MrpChr 객체 리스트를 문자열로 변환하는 메소드
-        Args:
-            mrp_chrs:  MrpChr 객체 리스트
-        Returns:
-            변환된 문자열
-        """
-        return ' '.join([str(_) for _ in mrp_chrs])
diff --git a/rsc/lib/jaso.py b/src/main/python/khaiii/resource/jaso.py
similarity index 92%
rename from rsc/lib/jaso.py
rename to src/main/python/khaiii/resource/jaso.py
index a93a64a..1f5e184 100644
--- a/rsc/lib/jaso.py
+++ b/src/main/python/khaiii/resource/jaso.py
@@ -4,10 +4,16 @@
 """
 한글 자소 관련 유틸리티 모듈
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
+###########
+# imports #
+###########
+from typing import Tuple
+
+
 #############
 # constants #
 #############
@@ -31,7 +37,7 @@
 #############
 # functions #
 #############
-def _decomp_char(char):
+def _decomp_char(char: str) -> Tuple[str, str, str]:
     """
     한글 음절 하나를 자소로 분해한다.
     Args:
diff --git a/rsc/lib/morphs.py b/src/main/python/khaiii/resource/morphs.py
similarity index 87%
rename from rsc/lib/morphs.py
rename to src/main/python/khaiii/resource/morphs.py
index 69184f1..be6d232 100644
--- a/rsc/lib/morphs.py
+++ b/src/main/python/khaiii/resource/morphs.py
@@ -2,12 +2,19 @@
 
 
 """
-형태소 분석 결과를 기술한 문자열을 파싱하는 모듈
+형태소 분석 결과를 기술한 문자열을 파싱하는 모듈.
+TODO(jamie): sejong_corpus 모듈의 Morph 클래스와 중복되므로 정리 필요
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
+###########
+# imports #
+###########
+from typing import List
+
+
 #############
 # constants #
 #############
@@ -36,14 +43,13 @@ class ParseError(Exception):
     """
     형태소 분석 결과 문자열을 파싱하면서 발생하는 오류
     """
-    pass
 
 
-class Morph(object):
+class Morph:
     """
     형태소
     """
-    def __init__(self, lex, tag):
+    def __init__(self, lex: str, tag: str):
         """
         Arguments:
             lex:  형태소(어휘)
@@ -57,7 +63,7 @@ def __str__(self):
             return self.lex
         return '{}/{}'.format(self.lex, self.tag)
 
-    def is_word_delim(self):
+    def is_word_delim(self) -> bool:
         """
         어절의 경계를 나타태는 지 여부
         Returns:
@@ -65,7 +71,7 @@ def is_word_delim(self):
         """
         return not self.tag and self.lex == WORD_DELIM_STR
 
-    def is_sent_delim(self):
+    def is_sent_delim(self) -> bool:
         """
         문장의 경계를 나타태는 지 여부
         Returns:
@@ -74,7 +80,7 @@ def is_sent_delim(self):
         return not self.tag and self.lex == SENT_DELIM_STR
 
     @classmethod
-    def to_str(cls, morphs):
+    def to_str(cls, morphs: List['Morph']) -> str:
         """
         Morph 객체 리스트를 문자열로 변환한다.
         Arguments:
@@ -85,7 +91,7 @@ def to_str(cls, morphs):
         return ' + '.join([str(m) for m in morphs])
 
     @classmethod
-    def parse(cls, morphs_str):
+    def parse(cls, morphs_str: str) -> List['Morph']:
         """
         형태소 분석 결과 형태의 문자열을 파싱하여 Morph 객체 리스트를 반환하는 파싱 함수
         Arguments:
@@ -98,7 +104,7 @@ def parse(cls, morphs_str):
         return [cls._parse_one(m) for m in morphs_str.split(' + ')]
 
     @classmethod
-    def _parse_one(cls, morph_str):
+    def _parse_one(cls, morph_str: str) -> 'Morph':
         """
         하나의 형태소 객체를 기술한 문자열을 파싱한다.
         Arguments:
diff --git a/rsc/lib/resource.py b/src/main/python/khaiii/resource/resource.py
similarity index 72%
rename from rsc/lib/resource.py
rename to src/main/python/khaiii/resource/resource.py
index a2ebc9e..3514e81 100644
--- a/rsc/lib/resource.py
+++ b/src/main/python/khaiii/resource/resource.py
@@ -4,18 +4,19 @@
 """
 resources for training and tagging
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import codecs
+from argparse import Namespace
 import logging
 import os
+from typing import Dict
 
-from vocabulary import Vocabulary
+from khaiii.resource.vocabulary import Vocabulary
 
 
 #############
@@ -33,13 +34,14 @@
 #########
 # types #
 #########
-class Resource(object):
+class Resource:
     """
     resources
     """
-    def __init__(self, cfg):
+    def __init__(self, cfg: Namespace):
         """
-        :param  cfg:  config
+        Args:
+            cfg:  config
         """
         vocab_in_path = '{}/vocab.in'.format(cfg.rsc_src)
         self.vocab_in = Vocabulary(vocab_in_path, cfg.cutoff, SPECIAL_CHARS)
@@ -49,14 +51,16 @@ def __init__(self, cfg):
         self.restore_dic = self._load_restore_dic(restore_dic_path)
 
     @classmethod
-    def _load_restore_dic(cls, path):
+    def _load_restore_dic(cls, path: str) -> Dict[str, str]:
         """
         load character to output tag mapping
-        :param  path:  file path
-        :return:  dictionary
+        Args:
+            path:  file path
+        Returns:
+            dictionary
         """
         dic = {}
-        for line in codecs.open(path, 'r', encoding='UTF-8'):
+        for line in open(path, 'r', encoding='UTF-8'):
             line = line.rstrip('\r\n')
             if not line:
                 continue
diff --git a/rsc/lib/trie.py b/src/main/python/khaiii/resource/trie.py
similarity index 90%
rename from rsc/lib/trie.py
rename to src/main/python/khaiii/resource/trie.py
index b380ee9..f843285 100755
--- a/rsc/lib/trie.py
+++ b/src/main/python/khaiii/resource/trie.py
@@ -5,18 +5,18 @@
 """
 TRIE 모듈
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import argparse
-import codecs
+from argparse import ArgumentParser, Namespace
 import logging
 import struct
 import sys
+from typing import List
 
 
 #############
@@ -28,14 +28,73 @@
 #########
 # types #
 #########
-class Trie(object):
+class Node:
+    """
+    TRIE 노드
+    """
+    def __init__(self, char: str = ''):
+        """
+        Args:
+            char:  (유니코드) 입력 문자
+        """
+        self.depth = -1    # 노드를 배열로 펼치기 위해 필요한 현재 노드의 루트로부터의 깊이
+        self.char = char    # 입력 문자
+        self.value = 0    # 단말 노드일 경우 그 값 (단말 노드가 아닌 경우 0)
+        self.children = {}    # 자식 노드들
+        self.child_start = -1    # 현재 노드로부터 자식 노드들의 시작점 사이에 있는 노드의 갯수
+
+    def __str__(self):
+        node_str = self.char if self.char else 'ROOT'
+        if self.value > 0:
+            node_str = '{%s: %d}' % (node_str, self.value)
+        pfx = ''
+        if self.depth > 0:
+            pfx = '·' * self.depth
+        child_start_str = ''
+        if self.child_start > 0:
+            child_start_str = ', {}'.format(self.child_start)
+        return '{}({}, [{}]{})'.format(pfx, node_str, ', '.join(self.children.keys()),
+                                       child_start_str)
+
+    def insert(self, key: str, val: int):
+        """
+        문자열 키와 그 값을 현재 노드로부터 내려가며 적절한 위치에 삽입한다.
+        Args:
+            key:  키 문자열
+            val:  값 (양수)
+        """
+        if val <= 0:
+            raise ValueError('value must be greater than zero')
+        if not key:
+            self.value = val
+        elif key[0] in self.children:
+            self.children[key[0]].insert(key[1:], val)
+        else:
+            new_node = Node(key[0])
+            self.children[key[0]] = new_node
+            new_node.insert(key[1:], val)
+        logging.debug('INSERT {%s: %d} INTO %s', key, val, self)
+
+    def pack(self) -> bytes:
+        """
+        구조체로 packing한다.
+        Returns:
+            packing된 구조체
+        """
+        char = 0 if not self.char else self.char
+        if isinstance(char, str):
+            char = ord(char)
+        return _NODE_STRUCT.pack(char, self.value, self.child_start, len(self.children))
+
+
+class Trie:
     """
     TRIE 인터페이스
     """
     def __init__(self):
         self.root = Node()
 
-    def insert(self, key, val):
+    def insert(self, key: str, val: int):
         """
         하나의 문자열 키와 그 값을 삽입한다.
         Args:
@@ -44,7 +103,7 @@ def insert(self, key, val):
         """
         self.root.insert(key, val)
 
-    def update(self, keys, vals=None):
+    def update(self, keys: List[str], vals: List[int] = None):
         """
         문자열 키와 그 값의 리스트를 차례로 삽입한다.
         값이 없을 경우 키 목록의 인덱스(1부터 시작)를 값으로 설정한다.
@@ -59,7 +118,7 @@ def update(self, keys, vals=None):
         for key, val in zip(keys, vals):
             self.insert(key, val)
 
-    def save(self, file_path):
+    def save(self, file_path: str):
         """
         TRIE 자료구조를 파일로 저장한다.
         Args:
@@ -74,7 +133,7 @@ def save(self, file_path):
         logging.info('total nodes: %d', len(nodes))
         logging.info('expected size: %d', len(nodes) * _NODE_STRUCT.size)
 
-    def find(self, key):
+    def find(self, key: str) -> int:
         """
         키를 이용하여 값을 얻는다.
         Args:
@@ -86,11 +145,10 @@ def find(self, key):
         for char in key:
             if char not in node.children:
                 return 0
-            else:
-                node = node.children[char]
+            node = node.children[char]
         return node.value
 
-    def _breadth_first_traverse(self):
+    def _breadth_first_traverse(self) -> List[Node]:
         """
         너비우선으로 루트로부터 전체 노드를 탐색한다.
         Returns:
@@ -113,7 +171,7 @@ def _breadth_first_traverse(self):
         return nodes
 
     @classmethod
-    def _set_child_start(cls, nodes):
+    def _set_child_start(cls, nodes: List[Node]):
         """
         child_start 필드를 세팅한다.
         Args:
@@ -135,69 +193,10 @@ def _set_child_start(cls, nodes):
                                                                                 else -1
 
 
-class Node(object):
-    """
-    TRIE 노드
-    """
-    def __init__(self, char=''):
-        """
-        Args:
-            char:  (유니코드) 입력 문자
-        """
-        self.depth = -1    # 노드를 배열로 펼치기 위해 필요한 현재 노드의 루트로부터의 깊이
-        self.char = char    # 입력 문자
-        self.value = 0    # 단말 노드일 경우 그 값 (단말 노드가 아닌 경우 0)
-        self.children = {}    # 자식 노드들
-        self.child_start = -1    # 현재 노드로부터 자식 노드들의 시작점 사이에 있는 노드의 갯수
-
-    def __str__(self):
-        node_str = self.char if self.char else 'ROOT'
-        if self.value > 0:
-            node_str = '{%s: %d}' % (node_str, self.value)
-        pfx = ''
-        if self.depth > 0:
-            pfx = '·' * self.depth
-        child_start_str = ''
-        if self.child_start > 0:
-            child_start_str = ', {}'.format(self.child_start)
-        return '{}({}, [{}]{})'.format(pfx, node_str, ', '.join(self.children.keys()),
-                                       child_start_str)
-
-    def insert(self, key, val):
-        """
-        문자열 키와 그 값을 현재 노드로부터 내려가며 적절한 위치에 삽입한다.
-        Args:
-            key:  키 문자열
-            val:  값 (양수)
-        """
-        if val <= 0:
-            raise ValueError('value must be greater than zero')
-        if not key:
-            self.value = val
-        elif key[0] in self.children:
-            self.children[key[0]].insert(key[1:], val)
-        else:
-            new_node = Node(key[0])
-            self.children[key[0]] = new_node
-            new_node.insert(key[1:], val)
-        logging.debug('INSERT {%s: %d} INTO %s', key, val, self)
-
-    def pack(self):
-        """
-        구조체로 packing한다.
-        Returns:
-            packing된 구조체
-        """
-        char = 0 if not self.char else self.char
-        if isinstance(char, str):
-            char = ord(char)
-        return _NODE_STRUCT.pack(char, self.value, self.child_start, len(self.children))
-
-
 #############
 # functions #
 #############
-def run(args):
+def run(args: Namespace):
     """
     run function which is the start point of program
     Args:
@@ -233,7 +232,7 @@ def main():
     """
     main function processes only argument parsing
     """
-    parser = argparse.ArgumentParser(description='트라이를 빌드합니다.')
+    parser = ArgumentParser(description='트라이를 빌드합니다.')
     parser.add_argument('-o', '--output', help='output file', metavar='FILE', required=True)
     parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
     parser.add_argument('--val', dest='has_val', help='탭으로 구분된 마지막 컬럼이 값일 경우',
@@ -242,9 +241,7 @@ def main():
     args = parser.parse_args()
 
     if args.input:
-        sys.stdin = codecs.open(args.input, 'r', encoding='UTF-8')
-    else:
-        sys.stdin = codecs.getreader('UTF-8')(sys.stdin.detach())
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
     if args.debug:
         logging.basicConfig(level=logging.DEBUG)
     else:
diff --git a/rsc/lib/vocabulary.py b/src/main/python/khaiii/resource/vocabulary.py
similarity index 89%
rename from rsc/lib/vocabulary.py
rename to src/main/python/khaiii/resource/vocabulary.py
index f07fb1c..6e9ba87 100644
--- a/rsc/lib/vocabulary.py
+++ b/src/main/python/khaiii/resource/vocabulary.py
@@ -4,17 +4,17 @@
 """
 vocabulary library
 __author__ = 'Jamie (jamie.lim@kakaocorp.com)'
-__copyright__ = 'Copyright (C) 2018-, Kakao Corp. All rights reserved.'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
 """
 
 
 ###########
 # imports #
 ###########
-import codecs
 import copy
 import logging
 import os
+from typing import List
 
 
 #########
@@ -24,7 +24,7 @@ class Vocabulary:
     """
     vocabulary class
     """
-    def __init__(self, path, cutoff=1, special=None, padding=''):
+    def __init__(self, path: str, cutoff: int = 1, special: List[str] = None, padding: str = ''):
         """
         Args:
             path:  file path
@@ -60,7 +60,7 @@ def __getitem__(self, key):
         except KeyError:
             return 0    # unknown word number
 
-    def __len__(self):
+    def __len__(self) -> int:
         return len(self.dic)
 
     '''
@@ -78,7 +78,7 @@ def get_embedding(self, dim, padding_idx=None):
         return nn.Embedding(len(self), dim)
     '''    # pylint: disable=pointless-string-statement
 
-    def padding_idx(self):
+    def padding_idx(self) -> int:
         """
         맨 마지막에 추가한 패딩의 인덱스를 리턴한다.
         Returns:
@@ -88,7 +88,7 @@ def padding_idx(self):
             raise RuntimeError('vocabulary has no padding')
         return self.dic[self.padding]
 
-    def _load(self, path, cutoff=1):
+    def _load(self, path: str, cutoff: int = 1):
         """
         load vocabulary from file
         Args:
@@ -97,7 +97,7 @@ def _load(self, path, cutoff=1):
         """
         append_num = 0
         cutoff_num = 0
-        for line in codecs.open(path, 'r', encoding='UTF-8'):
+        for line in open(path, 'r', encoding='UTF-8'):
             line = line.rstrip('\r\n')
             if not line:
                 continue

From eb69cf39777d45b999cf43ba33fbbb4c3c7d5c83 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 21 Jan 2019 16:59:24 +0900
Subject: [PATCH 23/41] =?UTF-8?q?=EC=84=B8=EC=A2=85=20=EC=BD=94=ED=8D=BC?=
 =?UTF-8?q?=EC=8A=A4=EB=A1=9C=EB=B6=80=ED=84=B0=20=EC=9D=8C=EC=A0=88?=
 =?UTF-8?q?=EB=8B=A8=EC=9C=84=20=EC=A0=95=EB=A0=AC=EC=9D=84=20=EC=88=98?=
 =?UTF-8?q?=ED=96=89=ED=95=98=EB=8A=94=20=EC=8A=A4=ED=81=AC=EB=A6=BD?=
 =?UTF-8?q?=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 train/bin/map_char_to_tag.py | 184 +++++++++++++++++++++++++++++++++++
 1 file changed, 184 insertions(+)
 create mode 100755 train/bin/map_char_to_tag.py

diff --git a/train/bin/map_char_to_tag.py b/train/bin/map_char_to_tag.py
new file mode 100755
index 0000000..b762ab3
--- /dev/null
+++ b/train/bin/map_char_to_tag.py
@@ -0,0 +1,184 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+어절 내에서 원문과 형태소 사이에 정렬을 수행하고 음절기반 학습 코퍼스를 생성
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+from collections import defaultdict
+import logging
+import os
+import random
+import sys
+from typing import Iterator, List
+
+from khaiii.munjong.sejong_corpus import Sentence, sents, Word
+from khaiii.resource.char_align import Aligner, AlignError, MrpChr
+
+
+#############
+# variables #
+#############
+_MAP_DIC = defaultdict(list)    # char/tag => morpheme result mapping dictionary
+_MAP_CASE = defaultdict(list)    # char/tag => case(word) dictionary
+
+
+#############
+# functions #
+#############
+def _print_restore_dic(args: Namespace):
+    """
+    원형복원 사전을 출력한다.
+    Args:
+        args:  program arguments
+    """
+    if not args.restore_dic:
+        return
+    with open(args.restore_dic, 'w', encoding='UTF-8') as fout:
+        with open('{}.case'.format(args.restore_dic), 'w', encoding='UTF-8') as fcase:
+            for (char, tag), vals in _MAP_DIC.items():
+                for idx, val in enumerate(vals):
+                    print('{}/{}:{}\t{}'.format(char, tag, idx, MrpChr.to_str(val)), file=fout)
+                    word = _MAP_CASE[char, tag][idx]
+                    print('{}/{}:{}\t{}'.format(char, tag, idx, word), file=fcase)
+
+
+def sent_iter(args: Namespace) -> Iterator[Sentence]:
+    """
+    sentence generator
+    Args:
+        args:  program arguments
+    Yields:
+        sentence
+    """
+    for name in sorted(os.listdir(args.corpus_dir)):
+        if not name.endswith('.txt'):
+            continue
+        logging.info(name)
+        path = '{}/{}'.format(args.corpus_dir, name)
+        for sent in sents(open(path, 'r', encoding='UTF-8')):
+            yield sent
+
+
+def _maps_to_tag(word: Word, maps: List[List[MrpChr]]) -> List[str]:
+    """
+    매핑 정보를 이용해 어절 내 각 음절 별로 출력 태그를 생성한다.
+    Args:
+        word:  Word object
+        maps:  음절 별 매핑 정보
+    Returns:
+        출력 태그 리스트
+    """
+    tags = []
+    for char, align in zip(word.raw, maps):
+        if len(align) == 1 and align[0].char == char:
+            tags.append(align[0].tag)
+            continue
+        tag = ':'.join([_.tag for _ in align])
+        if (char, tag) in _MAP_DIC:
+            vals = _MAP_DIC[char, tag]
+            try:
+                idx = vals.index(align)
+            except ValueError:
+                idx = len(vals)
+                vals.append(align)
+                _MAP_CASE[char, tag].append(word)
+            tag = '{}:{}'.format(tag, idx)
+        else:
+            _MAP_DIC[char, tag].append(align)
+            _MAP_CASE[char, tag].append(word)
+            tag = '{}:0'.format(tag)
+        tags.append(tag)
+    return tags
+
+
+def _print_sent(sent: Sentence, word_per_maps: List[List[List[MrpChr]]]):
+    """
+    각 어절 별 매핑 정보 리스트를 이용해 한 문장을 출력한다.
+    Args:
+        sent:  Sentence object
+        word_per_maps:  각 어절별로 음절 매핑 정보를 포함하는 리스트
+    """
+    lines = []
+    has_error = False
+    for word, maps in zip(sent.words, word_per_maps):
+        if maps:
+            if len(word.raw) == len(maps):
+                lines.append('{}\t{}'.format(word.raw, ' '.join(_maps_to_tag(word, maps))))
+            else:
+                raise RuntimeError('length of maps is different from length of word')
+        else:
+            has_error = True
+            logging.debug(word)
+    if not has_error:
+        print('\n'.join(lines))
+        print()
+
+
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    aligner = Aligner(args.rsc_src)
+    funmap = open(args.unmapped, 'w', encoding='UTF-8') if args.unmapped else None
+
+    for sent in sent_iter(args):
+        if 0.0 < args.sample < 1.0 and random.random() >= args.sample:
+            continue
+        word_per_maps = []
+        for word in sent.words:
+            try:
+                maps = aligner.align(word)
+            except AlignError as algn_err:
+                if funmap:
+                    algn_err.add_msg(str(word))
+                    print(algn_err, file=funmap)
+                maps = []
+            word_per_maps.append(maps)
+        _print_sent(sent, word_per_maps)
+
+    _print_restore_dic(args)
+    aligner.print_middle_cnt()
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='어절 내에서 원문과 형태소 사이에 정렬을 수행하고'
+                                        ' 음절기반 학습 코퍼스를 생성')
+    parser.add_argument('-c', '--corpus-dir', help='corpus dir', metavar='DIR', required=True)
+    parser.add_argument('--rsc-src', help='train resource dir <default: ../rsc/src>', metavar='DIR',
+                        default='../rsc/src')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--restore-dic', help='restore dic output file', metavar='FILE')
+    parser.add_argument('--unmapped', help='unmapped log file', metavar='FILE')
+    parser.add_argument('--sample', help='sampling ratio', metavar='REAL', type=float, default=1.0)
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From 24055856f3c28cda1357024381957745453d438a Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 21 Jan 2019 17:25:43 +0900
Subject: [PATCH 24/41] =?UTF-8?q?=ED=95=99=EC=8A=B5=20=EC=BD=94=ED=8D=BC?=
 =?UTF-8?q?=EC=8A=A4=EB=A5=BC=20dev/test/train=EC=9C=BC=EB=A1=9C=20?=
 =?UTF-8?q?=EB=B6=84=ED=95=A0=ED=95=98=EB=8A=94=20=EC=8A=A4=ED=81=AC?=
 =?UTF-8?q?=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 train/bin/split_corpus.py | 106 ++++++++++++++++++++++++++++++++++++++
 1 file changed, 106 insertions(+)
 create mode 100755 train/bin/split_corpus.py

diff --git a/train/bin/split_corpus.py b/train/bin/split_corpus.py
new file mode 100755
index 0000000..4dda3b8
--- /dev/null
+++ b/train/bin/split_corpus.py
@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+코퍼스를 train/dev/test로 분할한다.
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+import logging
+import random
+import sys
+from typing import Iterator, List, TextIO
+
+
+#############
+# functions #
+#############
+def _sents(fin: TextIO) -> Iterator[List[str]]:
+    """
+    read from file and yield a sentence (generator)
+    Args:
+        fin:  input file
+    Yields:
+        sentence (list of lines)
+    """
+    sent = []
+    for line in fin:
+        line = line.rstrip('\r\n')
+        if not line:
+            if sent:
+                yield sent
+                sent = []
+            continue
+        sent.append(line)
+    if sent:
+        yield sent
+
+
+def _write_to_file(path: str, sents: List[List[str]]):
+    """
+    파일에 쓴다.
+    Args:
+        path:  path
+        sents:  sentences
+    """
+    with open(path, 'w', encoding='UTF-8') as fout:
+        for sent in sents:
+            print('\n'.join(sent), file=fout)
+            print(file=fout)
+
+
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    sents = []
+    for num, sent in enumerate(_sents(sys.stdin), start=1):
+        if num % 100000 == 0:
+            logging.info('%d00k-th sent..', num // 100000)
+        sents.append(sent)
+    random.shuffle(sents)
+    _write_to_file('{}.dev'.format(args.out_pfx), sents[:args.dev])
+    _write_to_file('{}.test'.format(args.out_pfx), sents[args.dev:args.dev+args.test])
+    _write_to_file('{}.train'.format(args.out_pfx), sents[args.dev+args.test:])
+    logging.info('dev / test / train: %d / %d / %d', args.dev, args.test,
+                 len(sents[args.dev+args.test:]))
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='코퍼스를 train/dev/test로 분할한다.')
+    parser.add_argument('-o', '--out-pfx', help='output file prefix', metavar='NAME', required=True)
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--dev', help='number of sentence in dev set', metavar='NUM', type=int,
+                        default=5000)
+    parser.add_argument('--test', help='number of sentence in test set', metavar='NUM', type=int,
+                        default=5000)
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From 281f0d49c80703d816aaa9485c97ff9512775344 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 21 Jan 2019 18:02:41 +0900
Subject: [PATCH 25/41] =?UTF-8?q?=EC=9E=85,=20=EC=B6=9C=EB=A0=A5=20vocab?=
 =?UTF-8?q?=EC=9D=84=20=EC=83=9D=EC=84=B1=ED=95=98=EB=8A=94=20=EC=8A=A4?=
 =?UTF-8?q?=ED=81=AC=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 train/bin/make_vocab.py      | 96 ++++++++++++++++++++++++++++++++++++
 train/bin/map_char_to_tag.py |  4 +-
 2 files changed, 98 insertions(+), 2 deletions(-)
 create mode 100755 train/bin/make_vocab.py

diff --git a/train/bin/make_vocab.py b/train/bin/make_vocab.py
new file mode 100755
index 0000000..c5b0fd4
--- /dev/null
+++ b/train/bin/make_vocab.py
@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+입력(음절) 및 출력(태그) vocabulary를 생성한다.
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+from collections import Counter
+import logging
+import os
+import sys
+from typing import TextIO
+
+from khaiii.resource.morphs import TAGS
+
+
+#############
+# functions #
+#############
+def _print(cnt: Counter, fout: TextIO, is_with_freq: bool = True):
+    """
+    vocabulary 사전을 출력한다.
+    Args:
+        cnt:  Counter object
+        fout:  출력 파일
+        is_with_freq:  빈도를 함께 출력할 지 여부
+    """
+    for char, freq in sorted(cnt.items(), key=lambda x: x[0]):
+        if is_with_freq and freq < 2:
+            continue
+        if is_with_freq:
+            print('{}\t{}'.format(char, freq), file=fout)
+        else:
+            print(char, file=fout)
+
+
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    in_cnt = Counter()
+    out_cnt = Counter()
+    for line_num, line in enumerate(sys.stdin, start=1):
+        if line_num % 1000000 == 0:
+            logging.info('%dm-th line', line_num // 1000000)
+        line = line.rstrip('\r\n')
+        if not line:
+            continue
+        raw, tagged = line.split('\t')
+        in_cnt.update(list(raw))
+        out_cnt.update([tag for tag in tagged.split() if tag[2:] not in TAGS])
+    os.makedirs(args.rsc_src, exist_ok=True)
+    with open('{}/vocab.in'.format(args.rsc_src), 'w', encoding='UTF-8') as fout:
+        _print(in_cnt, fout)
+    with open('{}/vocab.out'.format(args.rsc_src), 'w', encoding='UTF-8') as fout:
+        print('\n'.join(['B-{}'.format(tag) for tag in TAGS]), file=fout)
+        print('\n'.join(['I-{}'.format(tag) for tag in TAGS]), file=fout)
+        _print(out_cnt, fout, is_with_freq=False)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='입력(음절) 및 출력(태그) vocabulary를 생성한다.')
+    parser.add_argument('--rsc-src', help='resource source dir <default: ../rsc/src>',
+                        metavar='DIR', default='../rsc/src')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/train/bin/map_char_to_tag.py b/train/bin/map_char_to_tag.py
index b762ab3..d868eb5 100755
--- a/train/bin/map_char_to_tag.py
+++ b/train/bin/map_char_to_tag.py
@@ -161,8 +161,8 @@ def main():
     parser = ArgumentParser(description='어절 내에서 원문과 형태소 사이에 정렬을 수행하고'
                                         ' 음절기반 학습 코퍼스를 생성')
     parser.add_argument('-c', '--corpus-dir', help='corpus dir', metavar='DIR', required=True)
-    parser.add_argument('--rsc-src', help='train resource dir <default: ../rsc/src>', metavar='DIR',
-                        default='../rsc/src')
+    parser.add_argument('--rsc-src', help='resource source dir <default: ../rsc/src>',
+                        metavar='DIR', default='../rsc/src')
     parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
     parser.add_argument('--restore-dic', help='restore dic output file', metavar='FILE')
     parser.add_argument('--unmapped', help='unmapped log file', metavar='FILE')

From ad99b4f2d424a3fca6d2870e34fc3ec6a0497fbb Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Tue, 22 Jan 2019 13:49:21 +0900
Subject: [PATCH 26/41] =?UTF-8?q?=EB=AA=A8=EB=8D=B8=20=ED=95=99=EC=8A=B5?=
 =?UTF-8?q?=EC=9D=84=20=EC=9C=84=ED=95=9C=20=EB=AA=A8=EB=93=88=EA=B3=BC=20?=
 =?UTF-8?q?=EC=8A=A4=ED=81=AC=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/main/python/khaiii/resource/vocabulary.py |  17 +-
 src/main/python/khaiii/train/dataset.py       | 196 +++++++++
 src/main/python/khaiii/train/embedder.py      | 102 +++++
 src/main/python/khaiii/train/evaluator.py     | 116 ++++++
 src/main/python/khaiii/train/models.py        | 115 ++++++
 src/main/python/khaiii/train/sentence.py      | 278 +++++++++++++
 src/main/python/khaiii/train/tagger.py        |  71 ++++
 src/main/python/khaiii/train/trainer.py       | 385 ++++++++++++++++++
 train/bin/train.py                            |  87 ++++
 9 files changed, 1351 insertions(+), 16 deletions(-)
 create mode 100644 src/main/python/khaiii/train/dataset.py
 create mode 100644 src/main/python/khaiii/train/embedder.py
 create mode 100644 src/main/python/khaiii/train/evaluator.py
 create mode 100644 src/main/python/khaiii/train/models.py
 create mode 100644 src/main/python/khaiii/train/sentence.py
 create mode 100644 src/main/python/khaiii/train/tagger.py
 create mode 100644 src/main/python/khaiii/train/trainer.py
 create mode 100755 train/bin/train.py

diff --git a/src/main/python/khaiii/resource/vocabulary.py b/src/main/python/khaiii/resource/vocabulary.py
index 6e9ba87..eb7a67c 100644
--- a/src/main/python/khaiii/resource/vocabulary.py
+++ b/src/main/python/khaiii/resource/vocabulary.py
@@ -60,24 +60,9 @@ def __getitem__(self, key):
         except KeyError:
             return 0    # unknown word number
 
-    def __len__(self) -> int:
+    def __len__(self):
         return len(self.dic)
 
-    '''
-    # 리소스 빌드 시 pytorch 의존성 제거를 위해 임시로 메서드를 제거합니다.
-    # 추후 학습 코드를 추가할 때 이 부분을 리팩토링 합니다.
-    def get_embedding(self, dim, padding_idx=None):
-        """
-        embedding을 리턴합니다.
-        Args:
-            dim:  embedding dimension
-            padding_idx:  padding index
-        """
-        if padding_idx:
-            return nn.Embedding(len(self), dim, padding_idx=padding_idx)
-        return nn.Embedding(len(self), dim)
-    '''    # pylint: disable=pointless-string-statement
-
     def padding_idx(self) -> int:
         """
         맨 마지막에 추가한 패딩의 인덱스를 리턴한다.
diff --git a/src/main/python/khaiii/train/dataset.py b/src/main/python/khaiii/train/dataset.py
new file mode 100644
index 0000000..7aee785
--- /dev/null
+++ b/src/main/python/khaiii/train/dataset.py
@@ -0,0 +1,196 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+corpus for training
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import Namespace
+import logging
+import os
+import random
+from typing import List, TextIO, Tuple
+
+from torch import LongTensor, Tensor    # pylint: disable=no-member, no-name-in-module
+from tqdm import tqdm
+
+from khaiii.resource.resource import PAD_CHR, Resource
+from khaiii.train.sentence import PosSentence, PosWord
+
+
+#########
+# types #
+#########
+class PosSentTensor(PosSentence):
+    """
+    tensor transformable sentence
+    """
+    def __init__(self, raw_sent: str = ''):
+        super().__init__(raw_sent)
+        if raw_sent:
+            self.init_pos_tags()
+
+    def __len__(self):
+        # RNN에서 길이가 긴 것으로 정렬하기위한 문장의 길이 = 음절 갯수 + 문장 경계 + 어절 경계
+        if self.words:
+            return sum([len(w) for w in self.words]) + len(self.words) + 1
+        if self.pos_tagged_words:
+            return sum([len(w.raw) for w in self.pos_tagged_words]) + len(self.pos_tagged_words) + 1
+        return 0
+
+    def make_labels(self, with_spc: bool) -> List[str]:
+        """
+        각 음절별로 출력 레이블(태그)를 생성한다.
+        Args:
+            with_spc:  공백(어절 경계) 포함 여부
+        Returns:
+            레이블 리스트
+        """
+        if not with_spc:
+            # 문장 경계, 어절 경계 등 가상 음절을 제외하고 순수한 음절들의 레이블
+            return [tag for pos_word in self.pos_tagged_words for tag in pos_word.tags]
+        labels = [PAD_CHR, ]    # 문장 시작
+        for pos_word in self.pos_tagged_words:
+            if len(labels) > 1:
+                labels.append(PAD_CHR)    # 어절 경계
+            labels.extend(pos_word.tags)
+        labels.append(PAD_CHR)    # 문장 종료
+        return labels
+
+    def make_contexts(self, window: int, spc_dropout: float) -> List[str]:
+        """
+        각 음절 별로 좌/우 window 크기 만큼 context를 만든다.
+        Args:
+            window:  left/right window size
+            spc_dropout:  space(word delimiter) dropout rate
+        Returns:
+            contexts
+        """
+        contexts = []
+        for wrd_idx, word in enumerate(self.words):
+            for chr_idx, char in enumerate(word):
+                left_context = list(reversed(word[:chr_idx]))
+                if random.random() >= spc_dropout:
+                    left_context.append('<w>')
+                for left_word in reversed(self.words[:wrd_idx]):
+                    left_context.extend(reversed(left_word))
+                    if len(left_context) >= window:
+                        break
+                if len(left_context) < window:
+                    left_context.extend(['<s>', ] * (window - len(left_context)))
+                left_context = list(reversed(left_context[:window]))
+                assert len(left_context) == window
+
+                right_context = list(word[chr_idx+1:])
+                if random.random() >= spc_dropout:
+                    right_context.append('</w>')
+                for right_word in self.words[wrd_idx+1:]:
+                    right_context.extend(list(right_word))
+                    if len(right_context) >= window:
+                        break
+                if len(right_context) < window:
+                    right_context.extend(['</s>', ] * (window - len(right_context)))
+                right_context = right_context[:window]
+                assert len(right_context) == window
+                contexts.append(left_context + [char, ] + right_context)
+        return contexts
+
+    def to_tensor(self, cfg: Namespace, rsc: Resource, is_train: bool) -> Tuple[Tensor, Tensor]:
+        """
+        문장 내에 포함된 전체 음절들과 태그를 모델의 forward 메소드에 넣을 수 있는 텐서로 변환한다.
+        Args:
+            cfg:  config
+            rsc:  Resource object
+            is_train:  whether is train or not
+        Returns:
+            labels tensor
+            contexts tensor
+        """
+        # 차원: [문장내 음절 갯수, ]
+        label_nums = [rsc.vocab_out[l] for l in self.make_labels(False)]
+        labels_tensor = LongTensor(label_nums)
+        # 차원: [문장내 음절 갯수 x context 크기]
+        spc_dropout = cfg.spc_dropout if is_train else 0.0
+        context_nums = [[rsc.vocab_in[c] for c in context] \
+                        for context in self.make_contexts(cfg.window, spc_dropout)]
+        contexts_tensor = LongTensor(context_nums)
+        return labels_tensor, contexts_tensor
+
+    def make_chars(self) -> List[str]:
+        """
+        문장 내 포함된 음절들을 만든다. 문장 경계 및 어절 경계를 포함한다.
+        Returns:
+            음절의 리스트
+        """
+        chars = ['<s>', ]    # 문장 시작
+        for word in self.words:
+            if len(chars) > 1:
+                chars.append('<w>')    # 어절 경계
+            chars.extend(word)
+        chars.append('</s>')    # 문장 종료
+        return chars
+
+
+class PosDataset:
+    """
+    part-of-speech tag dataset
+    """
+    def __init__(self, cfg: Namespace, restore_dic: dict, fin: TextIO):
+        """
+        Args:
+            cfg:  config
+            restore_dic:  restore dictionary
+            fin:  input file
+        """
+        self.cfg = cfg
+        self.fin = fin
+        self.sents = []
+        self.sent_idx = -1
+        self._load(restore_dic)
+
+    def __str__(self):
+        return '<PosDataset: file: {}, sents: {}, sent_idx: {}>'.format(
+            os.path.basename(self.fin.name), len(self.sents), self.sent_idx)
+
+    def _load(self, restore_dic: dict):
+        """
+        load data file
+        Args:
+            restore_dic:  restore dictionary
+        """
+        sent = PosSentTensor()
+        lines = self.fin.readlines()
+        for line in tqdm(lines, os.path.basename(self.fin.name), len(lines), mininterval=1,
+                         ncols=100):
+            line = line.rstrip('\r\n')
+            if not line:
+                if sent and sent.pos_tagged_words:
+                    sent.set_raw_by_words()
+                    self.sents.append(sent)
+                sent = PosSentTensor()
+                continue
+            raw, tags = line.split('\t')
+            pos_word = PosWord(raw)
+            pos_word.set_pos_result(tags.split(), restore_dic)
+            sent.pos_tagged_words.append(pos_word)
+        logging.info('%s: %d sentences', os.path.basename(self.fin.name), len(self.sents))
+
+    def __iter__(self):
+        self.sent_idx = -1
+        random.shuffle(self.sents)
+        return self
+
+    def __next__(self):
+        self.sent_idx += 1
+        if self.sent_idx >= len(self.sents):
+            raise StopIteration()
+        return self.sents[self.sent_idx]
+
+    def __len__(self):
+        return len(self.sents)
diff --git a/src/main/python/khaiii/train/embedder.py b/src/main/python/khaiii/train/embedder.py
new file mode 100644
index 0000000..61acf2d
--- /dev/null
+++ b/src/main/python/khaiii/train/embedder.py
@@ -0,0 +1,102 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+making embedding models
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import Namespace
+import math
+
+import torch
+from torch import nn, Tensor
+
+from khaiii.resource.resource import Resource
+
+
+class Embedder(nn.Module):
+    """
+    embedder class
+    """
+    def __init__(self, cfg: Namespace, rsc: Resource):
+        """
+        Args:
+            cfg:  config
+            rsc:  Resource object
+        """
+        super().__init__()
+        self.cfg = cfg
+        self.rsc = rsc
+        self.embedding = nn.Embedding(len(rsc.vocab_in), cfg.embed_dim)
+
+    def forward(self, inputs):    # pylint: disable=arguments-differ
+        """
+        임베딩을 생성하는 메소드
+        Args:
+            inputs:  contexts of batch size
+        Returns:
+            embedding
+        """
+        embeds = self.embedding(inputs)
+        embeds += positional_encoding(self.cfg.context_len, self.cfg.context_len,
+                                      self.cfg.embed_dim, 1)
+        return embeds
+
+
+#############
+# functions #
+#############
+def memoize(func):
+    """
+    memoize decorator
+    """
+    class Memodict(dict):
+        """
+        Memoization decorator for a function taking one or more arguments.
+        """
+        def __getitem__(self, *key):
+            return dict.__getitem__(self, key)
+
+        def __missing__(self, key):
+            ret = self[key] = func(*key)
+            return ret
+
+    return Memodict().__getitem__
+
+
+@memoize
+def positional_encoding(sent_len: int, max_dim: int, embed_dim: int, method: int = 1) -> Tensor:
+    """
+    positional encoding Tensor 출력.
+    embeds [batch_size, context_len, embed_dim]에 Broadcasting 으로 더해짐
+    Args:
+        sent_len:  actual sentence length
+        max_dim:  maximum dimension
+        embed_dim:  embedding dimension
+        method:  method number (1. end-to-end memory networks or 2. attention is all you need)
+    Returns:
+        pe [context_len, embed_dim]
+    """
+    pe_tensor = torch.zeros([max_dim, embed_dim])    # pylint: disable=no-member
+    for pos in range(1, sent_len + 1):
+        for i in range(1, embed_dim+1):
+            if method == 1:
+                # end-to-end memory networks
+                pe_tensor[pos-1, i-1] = 1 - pos / sent_len - ((i / embed_dim) *
+                                                              (1 - 2 * pos / sent_len))
+            elif method == 2:
+                # attention is all you need
+                if i % 2 == 0:
+                    pe_tensor[pos-1, i-1] = math.sin(pos / 10000 ** (2*i / embed_dim))
+                else:
+                    pe_tensor[pos-1, i-1] = math.cos(pos / 10000 ** (2*i / embed_dim))
+    if torch.cuda.is_available():
+        pe_tensor = pe_tensor.cuda()
+    pe_tensor.detach()
+    return pe_tensor
diff --git a/src/main/python/khaiii/train/evaluator.py b/src/main/python/khaiii/train/evaluator.py
new file mode 100644
index 0000000..33f5fe3
--- /dev/null
+++ b/src/main/python/khaiii/train/evaluator.py
@@ -0,0 +1,116 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+evaluation related module
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from collections import Counter
+import logging
+from typing import List, TextIO, Tuple
+
+from khaiii.train.sentence import PosMorph, PosSentence, PosWord
+
+
+#########
+# types #
+#########
+class Evaluator:
+    """
+    evauator
+    """
+    def __init__(self):
+        self.cnt = Counter()
+
+    def evaluate(self) -> Tuple[float, float, float]:
+        """
+        char/word accuracy, f-score(recall/precision)를 측정한다.
+        Returns:
+            character accuracy
+            word accuracy
+            f-score
+        """
+        char_acc = self.cnt['match_chars'] / self.cnt['total_chars']
+        word_acc = self.cnt['match_words'] / self.cnt['total_words']
+        recall = self.cnt['match_morphs'] / self.cnt['total_gold_morphs']
+        precision = self.cnt['match_morphs'] / self.cnt['total_pred_morphs']
+        f_score = 2.0 * recall * precision / (recall + precision)
+        self.cnt.clear()
+        return char_acc, word_acc, f_score
+
+    def count(self, correct_sent: PosSentence, predict_sent: PosSentence):
+        """
+        정답 문장과 비교하여 맞춘 갯수를 샌다.
+        Args:
+            correct_sent:  정답 문장
+            predict_sent:  예측한 문장
+        """
+        assert len(correct_sent.words) == len(predict_sent.words)
+        for gold, pred in zip(correct_sent.pos_tagged_words, predict_sent.pos_tagged_words):
+            self.cnt['total_chars'] += len(gold.res_tags)
+            self.cnt['match_chars'] += len([1 for x, y in zip(gold.res_tags, pred.res_tags)
+                                            if x == y])
+            self._count_word(gold, pred)
+
+    def _count_word(self, gold: PosWord, pred: PosWord):
+        """
+        count with gold standard and predicted (will update counter)
+        Args:
+            gold:  gold standard word
+            pred:  predicted word
+        """
+        self.cnt['total_words'] += 1
+        gold_morphs = gold.pos_tagged_morphs
+        pred_morphs = pred.pos_tagged_morphs
+        if gold == pred:
+            self.cnt['match_words'] += 1
+            num_match = len(gold_morphs)
+            self.cnt['total_gold_morphs'] += num_match
+            self.cnt['total_pred_morphs'] += num_match
+            self.cnt['match_morphs'] += num_match
+            return
+        logging.debug('gold: %s', ' '.join([str(_) for _ in gold_morphs]))
+        logging.debug('pred: %s', ' '.join([str(_) for _ in pred_morphs]))
+        self.cnt['total_gold_morphs'] += len(gold_morphs)
+        self.cnt['total_pred_morphs'] += len(pred_morphs)
+        gold_set = self.morphs_to_set(gold_morphs)
+        pred_set = self.morphs_to_set(pred_morphs)
+        self.cnt['match_morphs'] += len(gold_set & pred_set)
+
+    @classmethod
+    def morphs_to_set(cls, morphs: List[PosMorph]) -> set:
+        """
+        make set from morpheme list
+        Args:
+            morphs:  morpheme list
+        Returns:
+            morphemes set
+        """
+        morph_cnt = Counter([(morph.morph, morph.pos_tag) for morph in morphs])
+        morph_set = set()
+        for (lex, tag), freq in morph_cnt.items():
+            if freq == 1:
+                morph_set.add((lex, tag))
+            else:
+                morph_set.update([(lex, tag, _) for _ in range(1, freq+1)])
+        return morph_set
+
+    def report(self, fout: TextIO):
+        """
+        report recall/precision to file
+        :param  fout:  output file
+        """
+        print('word accuracy: %d / %d = %.4f' % (self.cnt['match_words'], self.cnt['total_words'],
+                                                 self.cnt['match_words'] / self.cnt['total_words']),
+              file=fout)
+        recall = self.cnt['match_morphs'] / self.cnt['total_gold_morphs']
+        precision = self.cnt['match_morphs'] / self.cnt['total_pred_morphs']
+        f_score = 2.0 * recall * precision / (recall + precision)
+        print('f-score / (recall, precision): %.4f / (%.4f, %.4f)' % (f_score, recall, precision),
+              file=fout)
diff --git a/src/main/python/khaiii/train/models.py b/src/main/python/khaiii/train/models.py
new file mode 100644
index 0000000..c1b3222
--- /dev/null
+++ b/src/main/python/khaiii/train/models.py
@@ -0,0 +1,115 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+Pytorch models
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import Namespace
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from khaiii.resource.resource import Resource
+from khaiii.train.embedder import Embedder
+
+
+#########
+# types #
+#########
+class PosModel(nn.Module):
+    """
+    part-of-speech tagger pytorch model
+    """
+    def __init__(self, cfg: Namespace, rsc: Resource):
+        """
+        Args:
+            cfg (Namespace):  config
+            rsc (Resource):  Resource object
+        """
+        super().__init__()
+        self.cfg = cfg
+        self.rsc = rsc
+        self.embedder = Embedder(cfg, rsc)
+
+    def forward(self, *inputs):
+        raise NotImplementedError
+
+    def save(self, path: str):
+        """
+        모델을 저장하는 메소드
+        Args:
+            path (str):  경로
+        """
+        torch.save(self.state_dict(), path)
+
+    def load(self, path: str):
+        """
+        저장된 모델을 로드하는 메소드
+        Args:
+            path (str):  경로
+        """
+        if torch.cuda.is_available():
+            state_dict = torch.load(path)
+        else:
+            state_dict = torch.load(path, map_location=lambda storage, loc: storage)
+        self.load_state_dict(state_dict)
+        if torch.cuda.is_available():
+            self.cuda()
+
+
+class CnnModel(PosModel):
+    """
+    convolutional neural network based part-of-speech tagger
+    """
+    def __init__(self, cfg: Namespace, rsc: Resource):
+        """
+        Args:
+            cfg (Namespace):  config
+            rsc (Resource):  Resource object
+        """
+        super().__init__(cfg, rsc)
+
+        ngram = min(5, cfg.window * 2 + 1)
+        self.convs = nn.ModuleList([nn.Conv1d(cfg.embed_dim, cfg.embed_dim, kernel_size)
+                                    for kernel_size in range(2, ngram+1)])
+
+        # conv => hidden
+        setattr(cfg, 'hidden_dim', (cfg.embed_dim * len(self.convs) + len(rsc.vocab_out)) // 2)
+        self.conv2hidden = nn.Linear(cfg.embed_dim * len(self.convs), cfg.hidden_dim)
+
+        # hidden => tag
+        self.hidden2tag = nn.Linear(cfg.hidden_dim, len(rsc.vocab_out))
+
+    def forward(self, contexts):    # pylint: disable=arguments-differ
+        """
+        forward path
+        Args:
+            contexts:  batch size list of character and context
+        Returns:
+            output score
+        """
+        embeds = self.embedder(contexts)
+        embeds_t = embeds.transpose(1, 2)
+
+        pool_outs = []
+        for conv in self.convs:
+            conv_out = F.relu(conv(embeds_t))
+            pool_outs.append(F.max_pool1d(conv_out, conv_out.size(2)))
+
+        # conv => hidden
+        features = torch.cat([p.view(contexts.size(0), -1) for p in pool_outs], dim=1)    # pylint: disable=no-member
+        features_drop = F.dropout(features)
+        hidden_out = F.relu(self.conv2hidden(features_drop))
+
+        # hidden => tag
+        hidden_out_drop = F.dropout(hidden_out)
+        tag_out = self.hidden2tag(hidden_out_drop)
+        return tag_out
diff --git a/src/main/python/khaiii/train/sentence.py b/src/main/python/khaiii/train/sentence.py
new file mode 100644
index 0000000..c36113c
--- /dev/null
+++ b/src/main/python/khaiii/train/sentence.py
@@ -0,0 +1,278 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+sentence, word, morph, ...
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+import logging
+import re
+
+from typing import List, Tuple
+
+
+#########
+# types #
+#########
+class Sentence:
+    """
+    raw sentence
+    """
+    def __init__(self, raw: str):
+        """
+        :param  raw:  raw sentence
+        """
+        self.words = raw.split()
+
+
+class PosMorph:
+    """
+    morpheme
+    """
+    def __init__(self, morph: str, pos_tag: str = 'O', beg: int = -1, end: int = -1):
+        self.morph = morph
+        self.pos_tag = pos_tag
+        self.beg = beg
+        self.end = end
+
+    def __str__(self):
+        return '{}/{}'.format(self.morph, self.pos_tag)
+
+    def __len__(self):
+        return self.end - self.beg
+
+
+class PosWord:
+    """
+    part-of-speech tagged word
+    """
+    def __init__(self, raw: str):
+        """
+        Args:
+            raw:  raw word
+        """
+        self.raw = raw
+        self.tags = []    # output tags for each character
+        self.res_chrs = raw    # 원형 복원된 형태소들의 음절들을 합친 것
+        self.res_tags = []    # 원형 복원된 형태소 음절들의 음절별 태그 (IOB)
+        self.pos_tagged_morphs = [PosMorph(x, 'O', idx, idx+1) for idx, x in enumerate(raw)]
+
+    def __str__(self):
+        return '{}\t{}'.format(self.raw, ' '.join([str(x) for x in self.pos_tagged_morphs]))
+
+    def for_pretrain(self) -> str:
+        """
+        pre-training을 위한 출력
+        Returns:
+            pre-training을 위한 문장
+        """
+        morph_strs = []
+        morph = ''
+        prev_tag = ''
+        for char, iob_tag in zip(self.raw, self.tags):
+            try:
+                iob_tag, _ = iob_tag.split(':', 1)
+            except ValueError:
+                pass
+            try:
+                iob, tag = iob_tag.split('-')
+            except ValueError as val_err:
+                logging.error('raw: %s', self.raw)
+                logging.error('tags: %s', self.tags)
+                logging.error('iob_tag: %s', iob_tag)
+                raise val_err
+            if iob == 'B':
+                if morph:
+                    morph_strs.append('%s/%s' % (re.sub(r'\d', '0', morph), prev_tag))
+                morph = char
+                prev_tag = tag
+            elif iob == 'I':
+                if prev_tag == tag:
+                    morph += char
+                else:
+                    if morph:
+                        morph_strs.append('%s/%s' % (re.sub(r'\d', '0', morph), prev_tag))
+                    morph = char
+                    prev_tag = tag
+        if morph:
+            morph_strs.append('%s/%s' % (re.sub(r'\d', '0', morph), prev_tag))
+        return ' '.join(morph_strs)
+
+    def __eq__(self, other: 'PosWord'):
+        """
+        어절의 형태소 분석 결과가 일치할 경우 같다고 간주한다. (평가 프로그램에서 어절 단위 일치 여부 판단 시 사용)
+        Args:
+            other:  other object
+        """
+        return self.res_chrs == other.res_chrs and self.res_tags == other.res_tags
+
+    def set_pos_result(self, tags: List[str], restore_dic: dict = None):
+        """
+        외부에서 생성된 PosWord객체의 정보를 현재 인스턴스에 설정합니다.
+        Args:
+            tags:  파일로 부터 읽은 형태소 태그(음절단위)
+            restore_dic:  원형 복원 사전
+        """
+        if not restore_dic:
+            tags = [x.split(':', 1)[0] for x in tags]
+        self.tags = tags
+        assert len(self.raw) == len(self.tags)    # 음절수와 태그수는 동일해야 한다.
+        self.pos_tagged_morphs = self._make_pos_morphs(restore_dic)
+
+    def _make_pos_morphs(self, restore_dic: dict = None):
+        """
+        형태소 태그리스트를 대상으로 B/I 로 병합되는 위치를 구합니다.
+        Args:
+            restore_dic:  원형 복원 사전
+        """
+        if not self.tags:
+            return []
+
+        self._restore(restore_dic)
+
+        pos_morphs = []
+        for beg, (lex, iob_tag) in enumerate(zip(self.res_chrs, self.res_tags)):
+            try:
+                iob, pos_tag = iob_tag.rsplit('-', 1)
+            except ValueError as val_err:
+                logging.error('invalid char/tag: %s/%s in [%s] %s', lex, iob_tag, self.res_chrs,
+                              self.res_tags)
+                raise val_err
+            if iob == 'B' or not pos_morphs or pos_morphs[-1].pos_tag != pos_tag:
+                pos_morphs.append(PosMorph(lex, pos_tag, beg, beg+1))
+            elif iob == 'I':
+                if pos_morphs[-1].pos_tag == pos_tag:
+                    pos_morphs[-1].morph += lex
+                    pos_morphs[-1].end += len(lex)
+                else:
+                    logging.debug('tag is different between B and I: %s vs %s',
+                                  pos_morphs[-1].pos_tag, pos_tag)
+                    pos_morphs.append(PosMorph(lex, pos_tag, beg, beg+1))
+            else:
+                raise ValueError('invalid IOB tag: {}/{} in [{}] {}'.format \
+                                     (lex, iob_tag, self.res_chrs, self.res_tags))
+        return pos_morphs
+
+    def _restore(self, restore_dic: dict):
+        """
+        원형 복원 사전을 이용하여 형태소의 원형을 복원한다.
+        Args:
+            restore_dic:  원형 복원 사전
+        """
+        if not restore_dic:
+            self.res_chrs = self.raw
+            self.res_tags = self.tags
+            return
+
+        res_chrs = []
+        self.res_tags = []
+        for char, tag in zip(self.raw, self.tags):
+            if ':' in tag:
+                key = '{}/{}'.format(char, tag)
+                if key in restore_dic:
+                    for char_tag in restore_dic[key].split():
+                        res_chr, res_tag = char_tag.rsplit('/', 1)
+                        res_chrs.append(res_chr)
+                        self.res_tags.append(res_tag)
+                    continue
+                else:
+                    logging.debug('mapping not found: %s/%s', char, tag)
+                    tag, _ = tag.split(':', 1)
+            res_chrs.append(char)
+            self.res_tags.append(tag)
+        self.res_chrs = ''.join(res_chrs)
+
+
+class PosSentence(Sentence):
+    """
+    part-of-speech tagged sentence
+    """
+    def __init__(self, raw: str):
+        """
+        Args:
+            raw:  raw sentence
+        """
+        super().__init__(raw)
+        self.pos_tagged_words = [] # list of PosWord
+
+    def __str__(self):
+        return '\n'.join([str(pos_word) for pos_word in self.pos_tagged_words])
+
+    def get_beg_end_list(self) -> Tuple[List[int], List[int]]:
+        """
+        모든 형태소의 시작위치를 담는 리스트와, 끝 위치를 담는 리스트를 구합니다.
+        Returns:
+            list of begin positions
+            list of end positions
+        """
+        begs = []
+        ends = []
+        for word in self.pos_tagged_words:
+            for morph in word.pos_tagged_morphs:
+                begs.append(morph.beg)
+                ends.append(morph.end)
+        return begs, ends
+
+    def set_raw_by_words(self):
+        """
+        Sentence 객체의 'words' 멤버를 PosWords 객체의 raw를 이용하여 채운다.
+        """
+        self.words = [pos_word.raw for pos_word in self.pos_tagged_words]
+
+    def init_pos_tags(self):
+        """
+        PosWord 객체를 생성하고 태그를 'O'로 세팅한다.
+        """
+        if self.pos_tagged_words:
+            raise RuntimeError('PoS tagged words are already initialized')
+        for word in self.words:
+            self.pos_tagged_words.append(PosWord(word))
+
+    def set_pos_result(self, tags: List[str], restore_dic: dict = None):
+        """
+        문장 전체에 대한 형태소 태그 출력 레이블 정보를 세팅하고 형태소를 복원한다.
+        Args:
+            tags:  문장 전체 태그 출력 레이블 정보
+            restore_dic:  원형 복원 사전
+        """
+        total_char_num = 0
+        for pos_word in self.pos_tagged_words:
+            pos_word.set_pos_result(tags[total_char_num:total_char_num + len(pos_word.raw)],
+                                    restore_dic)
+            total_char_num += len(pos_word.raw)
+        assert total_char_num == len(tags)
+
+    def get_sequence(self, morph: bool = True, tag: bool = True, simple: bool = False) -> List[str]:
+        """
+        태그를 포함한 형태소 문자열을 생성하여 리턴합니다.
+        Args:
+            morph:  형태소 출력
+            tag:  태그 표시 방법
+            simple:  tag를 1byte만 출력
+        Returns:
+            문자열의 리스트
+        """
+        sequence = []
+        for word in self.pos_tagged_words:
+            for pos_morph in word.pos_tagged_morphs:
+                morphs = []
+                if morph:
+                    morphs.append(pos_morph.morph)
+                if tag:
+                    morphs.append(pos_morph.pos_tag if not simple else pos_morph.pos_tag[0])
+                sequence.append('/'.join(morphs))
+        return sequence
+
+    def get_all_morphs(self) -> List[str]:
+        """
+        문장을 구성하는 모든 PosMorph의 리스트를 리턴합니다.
+        Returns:
+            모든 형태소 리스트
+        """
+        return [morph for word in self.pos_tagged_words for morph in word.pos_tagged_morphs]
diff --git a/src/main/python/khaiii/train/tagger.py b/src/main/python/khaiii/train/tagger.py
new file mode 100644
index 0000000..06fa0e2
--- /dev/null
+++ b/src/main/python/khaiii/train/tagger.py
@@ -0,0 +1,71 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+part-of-speech tagger
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import Namespace
+import json
+import logging
+import re
+
+import torch
+import torch.nn.functional as F
+
+from khaiii.resource.resource import Resource
+from khaiii.train.dataset import PosSentTensor
+from khaiii.train.models import CnnModel
+
+
+#########
+# types #
+#########
+class PosTagger:
+    """
+    part-of-speech tagger
+    """
+    def __init__(self, rsc_src: str):
+        """
+        Args:
+            rsc_src:  resource dir
+        """
+        cfg_dict = json.load(open('{}/config.json'.format(rsc_src), 'r', encoding='UTF-8'))
+        self.cfg = Namespace()
+        for key, val in cfg_dict.items():
+            setattr(self.cfg, key, val)
+        self.cfg.rsc_src = rsc_src
+        self.rsc = Resource(self.cfg)
+        self.model = CnnModel(self.cfg, self.rsc)
+        self.model.load('{}/model.state'.format(rsc_src))
+        self.model.eval()
+
+    def tag_raw(self, raw_sent: str, enable_restore: bool = True) -> PosSentTensor:
+        """
+        part-of-speech tagging at raw sentence
+        Args:
+            raw_sent:  raw input sentence
+        Returns:
+            PosSentTensor object
+        """
+        pos_sent = PosSentTensor(raw_sent)
+        _, contexts = pos_sent.to_tensor(self.cfg, self.rsc, False)
+        if torch.cuda.is_available():
+            contexts = contexts.cuda()
+        outputs = self.model(contexts)
+        _, predicts = F.softmax(outputs, dim=1).max(1)
+        tags = [self.rsc.vocab_out[t.item()] for t in predicts]
+        pos_sent.set_pos_result(tags, self.rsc.restore_dic if enable_restore else None)
+
+        if logging.getLogger().isEnabledFor(logging.DEBUG):
+            raw_nospc = re.sub(r'\s+', '', raw_sent)
+            for idx, (tag, pred) in enumerate(zip(tags, predicts)):
+                logging.debug('[%2d]%s: %5s(%d)', idx, raw_nospc[idx], tag, pred.data[0])
+
+        return pos_sent
diff --git a/src/main/python/khaiii/train/trainer.py b/src/main/python/khaiii/train/trainer.py
new file mode 100644
index 0000000..140d558
--- /dev/null
+++ b/src/main/python/khaiii/train/trainer.py
@@ -0,0 +1,385 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+training related library
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import Namespace
+import copy
+from datetime import datetime, timedelta
+import json
+import logging
+import os
+import pathlib
+import pprint
+import sys
+from typing import List, Tuple
+
+from tensorboardX import SummaryWriter
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from tqdm import tqdm
+
+from khaiii.train.dataset import PosDataset
+from khaiii.train.evaluator import Evaluator
+from khaiii.train.models import CnnModel
+from khaiii.resource.resource import Resource
+
+
+#############
+# functions #
+#############
+class Trainer:
+    """
+    trainer class
+    """
+    def __init__(self, cfg: Namespace):
+        """
+        Args:
+            cfg:  config
+        """
+        self.cfg = cfg
+        setattr(cfg, 'model_id', self.model_id(cfg))
+        setattr(cfg, 'out_dir', '{}/{}'.format(cfg.log_dir, cfg.model_id))
+        setattr(cfg, 'context_len', 2 * cfg.window + 1)
+        self.rsc = Resource(cfg)
+        self.model = CnnModel(cfg, self.rsc)
+        self.optimizer = torch.optim.Adam(self.model.parameters(), cfg.learning_rate)
+        self.criterion = nn.CrossEntropyLoss()
+        self.evaler = Evaluator()
+        self._load_dataset()
+        if 'epoch' not in cfg.__dict__:
+            setattr(cfg, 'epoch', 0)
+            setattr(cfg, 'iteration', 0)
+            setattr(cfg, 'iter_best', 0)
+        self.log_file = None    # tab separated log file
+        self.sum_wrt = None    # tensorboard summary writer
+        self.loss_trains = []
+        self.loss_devs = []
+        self.acc_chars = []
+        self.acc_words = []
+        self.f_scores = []
+        self.learning_rates = []
+        self.batch_sizes = []
+
+    @classmethod
+    def model_id(cls, cfg: Namespace) -> str:
+        """
+        get model ID
+        Args:
+            cfg:  config
+        Returns:
+            model ID
+        """
+        model_cfgs = [
+            os.path.basename(cfg.in_pfx),
+            'cut{}'.format(cfg.cutoff),
+            'win{}'.format(cfg.window),
+            'sdo{}'.format(cfg.spc_dropout),
+            'emb{}'.format(cfg.embed_dim),
+            'lr{}'.format(cfg.learning_rate),
+            'lrd{}'.format(cfg.lr_decay),
+            'bs{}'.format(cfg.batch_size),
+            'ci{}'.format(cfg.check_iter),
+            'bg{}'.format(cfg.batch_grow),
+        ]
+        return '.'.join(model_cfgs)
+
+    def _load_dataset(self):
+        """
+        load training dataset
+        """
+        dataset_dev_path = '{}.dev'.format(self.cfg.in_pfx)
+        self.dataset_dev = PosDataset(self.cfg, self.rsc.restore_dic,
+                                      open(dataset_dev_path, 'r', encoding='UTF-8'))
+        dataset_train_path = '{}.train'.format(self.cfg.in_pfx)
+        self.dataset_train = PosDataset(self.cfg, self.rsc.restore_dic,
+                                        open(dataset_train_path, 'r', encoding='UTF-8'))
+
+    @classmethod
+    def _dt_str(cls, dt_obj: datetime) -> str:
+        """
+        string formatting for datetime object
+        Args:
+            dt_obj:  datetime object
+        Returns:
+            string
+        """
+        return dt_obj.strftime('%m/%d %H:%M:%S')
+
+    @classmethod
+    def _elapsed(cls, td_obj: timedelta) -> str:
+        """
+        string formatting for timedelta object
+        Args:
+            td_obj:  timedelta object
+        Returns:
+            string
+        """
+        seconds = td_obj.seconds
+        if td_obj.days > 0:
+            seconds += td_obj.days * 24 * 3600
+        hours = seconds // 3600
+        seconds -= hours * 3600
+        minutes = seconds // 60
+        seconds -= minutes * 60
+        return '{}:{:02d}:{:02d}'.format(hours, minutes, seconds)
+
+    def _check_continue(self):
+        """
+        기존에 학습하다 중지한 경우 그 이후부터 계속해서 학습할 수 있도록 이전 상태를 복원한다.
+        """
+        out_path = pathlib.Path(self.cfg.out_dir)
+        if not out_path.is_dir():
+            return
+        logging.info('==== continue learning: %s ====', self.cfg.model_id)
+        cfg = json.load(open('{}/config.json'.format(self.cfg.out_dir), 'r', encoding='UTF-8'))
+        for key, val in cfg.items():
+            setattr(self.cfg, key, val)
+        self._revert(False)
+
+        f_score_best = 0.0
+        best_idx = -1
+        for idx, line in enumerate(open('{}/log.tsv'.format(self.cfg.out_dir))):
+            line = line.rstrip('\r\n')
+            if not line:
+                continue
+            (epoch, iteration, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate,
+             batch_size) = line.split('\t')
+            self.cfg.epoch = int(epoch) + 1
+            self.cfg.iteration = self.cfg.iter_best = int(iteration) * self.cfg.check_iter
+            self.loss_trains.append(float(loss_train))
+            self.loss_devs.append(float(loss_dev))
+            self.acc_chars.append(float(acc_char))
+            self.acc_words.append(float(acc_word))
+            self.f_scores.append(float(f_score))
+            self.learning_rates.append(float(learning_rate))
+            self.batch_sizes.append(int(batch_size))
+            if float(f_score) > f_score_best:
+                f_score_best = float(f_score)
+                best_idx = idx
+        logging.info('---- [%d|%d] loss(train/dev): %f / %f, acc(char/word): %f / %f, ' \
+                     'f-score: %f, lr: %f, bs: %d ----', self.cfg.epoch,
+                     self.cfg.iteration // self.cfg.check_iter, self.loss_trains[best_idx],
+                     self.loss_devs[best_idx], self.acc_chars[best_idx], self.acc_words[best_idx],
+                     self.f_scores[best_idx], self.learning_rates[-1], self.batch_sizes[-1])
+
+    def train(self):
+        """
+        train model with dataset
+        """
+        self._check_continue()
+        logging.info('config: %s', pprint.pformat(self.cfg.__dict__))
+        train_begin = datetime.now()
+        logging.info('{{{{ training begin: %s {{{{', self._dt_str(train_begin))
+        if torch.cuda.is_available():
+            self.model.cuda()
+        pathlib.Path(self.cfg.out_dir).mkdir(parents=True, exist_ok=True)
+        self.log_file = open('{}/log.tsv'.format(self.cfg.out_dir), 'at')
+        self.sum_wrt = SummaryWriter(self.cfg.out_dir)
+        patience = self.cfg.patience
+        for _ in range(10000):
+            has_best = self.train_epoch()
+            if has_best:
+                patience = self.cfg.patience
+            else:
+                if patience > 0:
+                    self._revert(True)
+                    patience -= 1
+                    logging.info('==== revert to iter: %d, f-score: %f, patience: %d ====',
+                                 self.cfg.iter_best, max(self.f_scores), patience)
+                else:
+                    break
+        train_end = datetime.now()
+        train_elapsed = self._elapsed(train_end - train_begin)
+        logging.info('}}}} training end: %s, elapsed: %s, epoch: %s, iter: %dk }}}}',
+                     self._dt_str(train_end), train_elapsed, self.cfg.epoch,
+                     self.cfg.iteration // 1000)
+
+    def _revert(self, is_decay_lr: bool):
+        """
+        이전 best 모델로 되돌린다.
+        Args:
+            is_decay_lr:  whether multiply decay factor or not
+        """
+        self.model.load('{}/model.state'.format(self.cfg.out_dir))
+        if is_decay_lr:
+            self.cfg.learning_rate *= self.cfg.lr_decay
+        self._load_optim('{}/optim.state'.format(self.cfg.out_dir), self.cfg.learning_rate)
+
+    def train_epoch(self) -> bool:
+        """
+        한 epoch을 학습한다. 배치 단위는 글자 단위
+        Returns:
+            best f-score를 기록한 iteration 여부
+        """
+        epoch_begin = datetime.now()
+        logging.info('{{{{ epoch: %d, begin: %s {{{{', self.cfg.epoch, self._dt_str(epoch_begin))
+        iter_lap1 = datetime.now()
+        has_best = False
+        batches = []
+        loss_trains = []
+        batch_size = self.cfg.batch_size
+        if self.cfg.batch_grow > 0:
+            batch_size = self.cfg.batch_size + self.cfg.iteration // self.cfg.batch_grow
+        for train_sent in tqdm(self.dataset_train, 'EPOCH [{}]'.format(self.cfg.epoch),
+                               len(self.dataset_train), mininterval=1, ncols=100):
+            train_labels, train_contexts = train_sent.to_tensor(self.cfg, self.rsc, True)
+            if torch.cuda.is_available():
+                train_labels = train_labels.cuda()
+                train_contexts = train_contexts.cuda()
+
+            self.optimizer.zero_grad()
+            self.model.train()
+            train_outputs = self.model(train_contexts)
+            batches.append((train_labels, train_outputs))
+            if sum([batch[0].size(0) for batch in batches]) < batch_size:
+                continue
+            batch_label = torch.cat([x[0] for x in batches], 0)    # pylint: disable=no-member
+            batch_output = torch.cat([x[1] for x in batches], 0)    # pylint: disable=no-member
+            batches = []
+
+            batch_output.requires_grad_()
+            loss_train = self.criterion(batch_output, batch_label)
+            loss_trains.append(loss_train.item())
+            loss_train.backward()
+            self.optimizer.step()
+            self.cfg.iteration += 1
+
+            if self.cfg.iteration % self.cfg.check_iter == 0:
+                avg_loss_dev, acc_char, acc_word, f_score = self.evaluate()
+                iter_lap2 = datetime.now()
+                iter_elapsed = self._elapsed(iter_lap2 - iter_lap1)
+                iter_lap1 = iter_lap2
+                has_best |= self._check_iter(iter_elapsed, loss_trains, avg_loss_dev, acc_char,
+                                             acc_word, f_score, batch_size)
+            if self.cfg.batch_grow > 0 and self.cfg.iteration % self.cfg.batch_grow == 0:
+                batch_size = self.cfg.batch_size + self.cfg.iteration // self.cfg.batch_grow
+
+        print(file=sys.stderr)
+        sys.stderr.flush()
+        epoch_end = datetime.now()
+        epoch_elapsed = self._elapsed(epoch_end - epoch_begin)
+        logging.info('}}}} epoch: %d, end: %s, elapsed: %s, %s best }}}}', self.cfg.epoch,
+                     self._dt_str(epoch_begin), epoch_elapsed, 'hit' if has_best else 'did not hit')
+        self.cfg.epoch += 1
+        return has_best
+
+    def _check_iter(self, elapsed: str, loss_trains: List[float], avg_loss_dev: float,
+                    acc_char: float, acc_word: float, f_score: float, batch_size: int) -> bool:
+        """
+        cfg.check_iter번의 iteration마다 수행하는 체크
+        Args:
+            elapsed:  경과 시간
+            loss_trains:   train 코퍼스에서 각 배치별 loss 리스트
+            avg_loss_dev:  dev 코퍼스 문장 별 평균 loss
+            acc_char:  음절 정확도
+            acc_word:  어절 정확도
+            f_score:  f-score
+            batch_size:  batch size
+        Returns:
+            현재 iteration이 best 성능을 나타냈는 지 여부
+        """
+        avg_loss_train = sum(loss_trains) / len(loss_trains)
+        loss_trains.clear()
+        self.loss_trains.append(avg_loss_train)
+        self.loss_devs.append(avg_loss_dev)
+        self.acc_chars.append(acc_char)
+        self.acc_words.append(acc_word)
+        self.f_scores.append(f_score)
+        self.learning_rates.append(self.cfg.learning_rate)
+        self.batch_sizes.append(batch_size)
+        check_num = self.cfg.iteration // self.cfg.check_iter
+        logging.info('---- [%d|%d] elapsed: %s, loss(train/dev): %f / %f, acc(char/word): %f / %f' \
+                     ', f-score: %f (max: %f), lr: %f, bs: %d ----', self.cfg.epoch, check_num,
+                     elapsed, avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score,
+                     max(self.f_scores), self.cfg.learning_rate, batch_size)
+        print('{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}'.format( \
+              self.cfg.epoch, check_num, avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score,
+              self.cfg.learning_rate, batch_size), file=self.log_file)
+        self.log_file.flush()
+        self.sum_wrt.add_scalar('loss-train', avg_loss_train, check_num)
+        self.sum_wrt.add_scalar('loss-dev', avg_loss_dev, check_num)
+        self.sum_wrt.add_scalar('acc-char', acc_char, check_num)
+        self.sum_wrt.add_scalar('acc-word', acc_word, check_num)
+        self.sum_wrt.add_scalar('f-score', f_score, check_num)
+        self.sum_wrt.add_scalar('learning-rate', self.cfg.learning_rate, check_num)
+        self.sum_wrt.add_scalar('batch-size', batch_size, check_num)
+        return self._check_best()
+
+    def _check_best(self) -> bool:
+        """
+        이번 iteration에 가장 좋은 성능을 냈는 지 확인하고 그럴 경우 현재 상태를 저장한다.
+        Returns:
+            best 여부
+        """
+        if len(self.f_scores) > 1 and max(self.f_scores[:-1]) >= self.f_scores[-1]:
+            return False
+        # this iteration hits new max value
+        logging.info('==== best model: %f ====', self.f_scores[-1])
+        self.cfg.iter_best = self.cfg.iteration
+        self.model.save('{}/model.state'.format(self.cfg.out_dir))
+        self._save_optim('{}/optim.state'.format(self.cfg.out_dir))
+        with open('{}/config.json'.format(self.cfg.out_dir), 'w', encoding='UTF-8') as fout:
+            json.dump(vars(self.cfg), fout, indent=2, sort_keys=True)
+        return True
+
+    def _save_optim(self, path: str):
+        """
+        save optimizer parameters
+        Args:
+            path:  path
+        """
+        torch.save(self.optimizer.state_dict(), path)
+
+    def _load_optim(self, path: str, learning_rate: float):
+        """
+        load optimizer parameters
+        Args:
+            path:  path
+            learning_rate:  learning rate
+        """
+        if torch.cuda.is_available():
+            state_dict = torch.load(path)
+        else:
+            state_dict = torch.load(path, map_location=lambda storage, loc: storage)
+        self.optimizer = torch.optim.Adam(self.model.parameters(), learning_rate)
+        self.optimizer.load_state_dict(state_dict)
+        self.optimizer.param_groups[0]['lr'] = learning_rate
+
+    def evaluate(self) -> Tuple[float, float, float, float]:
+        """
+        evaluate f-score
+        Returns:
+            average dev loss
+            character accuracy
+            word accuracy
+            f-score
+        """
+        self.model.eval()
+        loss_devs = []
+        for dev_sent in self.dataset_dev:
+            # 만약 spc_dropout이 1.0 이상이면 공백을 전혀 쓰지 않는 것이므로 평가 시에도 적용한다.
+            dev_labels, dev_contexts = dev_sent.to_tensor(self.cfg, self.rsc,
+                                                          self.cfg.spc_dropout >= 1.0)
+            if torch.cuda.is_available():
+                dev_labels = dev_labels.cuda()
+                dev_contexts = dev_contexts.cuda()
+            dev_outputs = self.model(dev_contexts)
+            loss_dev = self.criterion(dev_outputs, dev_labels)
+            loss_devs.append(loss_dev.item())
+            _, predicts = F.softmax(dev_outputs, dim=1).max(1)
+            pred_tags = [self.rsc.vocab_out[t.item()] for t in predicts]
+            pred_sent = copy.deepcopy(dev_sent)
+            pred_sent.set_pos_result(pred_tags, self.rsc.restore_dic)
+            self.evaler.count(dev_sent, pred_sent)
+        avg_loss_dev = sum(loss_devs) / len(loss_devs)
+        return (avg_loss_dev, ) + self.evaler.evaluate()
diff --git a/train/bin/train.py b/train/bin/train.py
new file mode 100755
index 0000000..fcd7998
--- /dev/null
+++ b/train/bin/train.py
@@ -0,0 +1,87 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+train part-of-speech model from data set
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+import logging
+import os
+
+from khaiii.train.trainer import Trainer
+
+
+#############
+# functions #
+#############
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments (config)
+    """
+    trner = Trainer(args)
+    trner.train()
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='train model from data')
+    parser.add_argument('-i', '--in-pfx', help='input data prefix', metavar='NAME', required=True)
+    parser.add_argument('--rsc-src', help='train resource dir <default: ../rsc/src>', metavar='DIR',
+                        default='../rsc/src')
+    parser.add_argument('--log-dir', help='tensorboard log dir', metavar='DIR',
+                        default='./logdir')
+    parser.add_argument('--window', help='left/right character window length', metavar='INT',
+                        type=int,
+                        default=3)
+    parser.add_argument('--spc-dropout', help='space(word delimiter) dropout rate', metavar='REAL',
+                        type=float,
+                        default=0.0)
+    parser.add_argument('--cutoff', help='cutoff', metavar='INT', type=int,
+                        default=2)
+    parser.add_argument('--embed-dim', help='embedding dimension', metavar='INT', type=int,
+                        default=30)
+    parser.add_argument('--learning-rate', help='learning rate', metavar='REAL', type=float,
+                        default=0.001)
+    parser.add_argument('--lr-decay', help='learning rate decay', metavar='REAL', type=float,
+                        default=0.9)
+    parser.add_argument('--batch-size', help='batch size', metavar='INT', type=int,
+                        default=500)
+    parser.add_argument('--check-iter', help='check iteration', metavar='INT', type=int,
+                        default=10000)
+    parser.add_argument('--batch-grow', help='batch growing iteration. 0 means no growing',
+                        type=int,
+                        default=10000)
+    parser.add_argument('--patience', help='maximum patience count to revert model', metavar='INT',
+                        type=int,
+                        default=10)
+    parser.add_argument('--gpu-num', help='GPU number to use', metavar='INT', type=int,
+                        default=0)
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_num)
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From d995b529fc33b0910bb2b0385f05211c27c6cdc3 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Wed, 23 Jan 2019 22:13:31 +0900
Subject: [PATCH 27/41] =?UTF-8?q?batch=20grow=20=ED=8C=8C=EB=9D=BC?=
 =?UTF-8?q?=EB=AF=B8=ED=84=B0,=20epoch=20=EA=B0=9C=EB=85=90=EC=9D=84=20?=
 =?UTF-8?q?=EC=97=86=EC=95=A0=EA=B3=A0=20=ED=95=99=EC=8A=B5=EA=B3=BC?=
 =?UTF-8?q?=EC=A0=95=EC=9D=84=20=EB=8B=A8=EC=88=9C=ED=99=94=ED=95=A8=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/main/python/khaiii/train/trainer.py | 219 ++++++++++++------------
 train/bin/train.py                      |  48 +++---
 train/requirements.txt                  |   3 +
 3 files changed, 131 insertions(+), 139 deletions(-)
 create mode 100644 train/requirements.txt

diff --git a/src/main/python/khaiii/train/trainer.py b/src/main/python/khaiii/train/trainer.py
index 140d558..e299fc1 100644
--- a/src/main/python/khaiii/train/trainer.py
+++ b/src/main/python/khaiii/train/trainer.py
@@ -19,8 +19,7 @@
 import os
 import pathlib
 import pprint
-import sys
-from typing import List, Tuple
+from typing import Iterator, List, Tuple
 
 from tensorboardX import SummaryWriter
 import torch
@@ -28,7 +27,7 @@
 import torch.nn.functional as F
 from tqdm import tqdm
 
-from khaiii.train.dataset import PosDataset
+from khaiii.train.dataset import PosDataset, PosSentTensor
 from khaiii.train.evaluator import Evaluator
 from khaiii.train.models import CnnModel
 from khaiii.resource.resource import Resource
@@ -48,7 +47,7 @@ def __init__(self, cfg: Namespace):
         """
         self.cfg = cfg
         setattr(cfg, 'model_id', self.model_id(cfg))
-        setattr(cfg, 'out_dir', '{}/{}'.format(cfg.log_dir, cfg.model_id))
+        setattr(cfg, 'out_dir', '{}/{}'.format(cfg.logdir, cfg.model_id))
         setattr(cfg, 'context_len', 2 * cfg.window + 1)
         self.rsc = Resource(cfg)
         self.model = CnnModel(cfg, self.rsc)
@@ -56,10 +55,9 @@ def __init__(self, cfg: Namespace):
         self.criterion = nn.CrossEntropyLoss()
         self.evaler = Evaluator()
         self._load_dataset()
-        if 'epoch' not in cfg.__dict__:
-            setattr(cfg, 'epoch', 0)
-            setattr(cfg, 'iteration', 0)
-            setattr(cfg, 'iter_best', 0)
+        if 'step' not in cfg.__dict__:
+            setattr(cfg, 'step', 0)
+            setattr(cfg, 'best_step', 0)
         self.log_file = None    # tab separated log file
         self.sum_wrt = None    # tensorboard summary writer
         self.loss_trains = []
@@ -68,7 +66,6 @@ def __init__(self, cfg: Namespace):
         self.acc_words = []
         self.f_scores = []
         self.learning_rates = []
-        self.batch_sizes = []
 
     @classmethod
     def model_id(cls, cfg: Namespace) -> str:
@@ -88,8 +85,7 @@ def model_id(cls, cfg: Namespace) -> str:
             'lr{}'.format(cfg.learning_rate),
             'lrd{}'.format(cfg.lr_decay),
             'bs{}'.format(cfg.batch_size),
-            'ci{}'.format(cfg.check_iter),
-            'bg{}'.format(cfg.batch_grow),
+            'cs{}'.format(cfg.check_step),
         ]
         return '.'.join(model_cfgs)
 
@@ -133,18 +129,19 @@ def _elapsed(cls, td_obj: timedelta) -> str:
         seconds -= minutes * 60
         return '{}:{:02d}:{:02d}'.format(hours, minutes, seconds)
 
-    def _check_continue(self):
+    def _restore_prev_train(self):
         """
         기존에 학습하다 중지한 경우 그 이후부터 계속해서 학습할 수 있도록 이전 상태를 복원한다.
         """
         out_path = pathlib.Path(self.cfg.out_dir)
-        if not out_path.is_dir():
+        cfg_path = pathlib.Path('{}/config.json'.format(self.cfg.out_dir))
+        if not out_path.is_dir() or not cfg_path.is_file():
             return
-        logging.info('==== continue learning: %s ====', self.cfg.model_id)
-        cfg = json.load(open('{}/config.json'.format(self.cfg.out_dir), 'r', encoding='UTF-8'))
+        logging.info('==== continue training: %s ====', self.cfg.model_id)
+        cfg = json.load(open(cfg_path, 'r', encoding='UTF-8'))
         for key, val in cfg.items():
             setattr(self.cfg, key, val)
-        self._revert(False)
+        self._revert_to_best(False)
 
         f_score_best = 0.0
         best_idx = -1
@@ -152,32 +149,43 @@ def _check_continue(self):
             line = line.rstrip('\r\n')
             if not line:
                 continue
-            (epoch, iteration, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate,
-             batch_size) = line.split('\t')
-            self.cfg.epoch = int(epoch) + 1
-            self.cfg.iteration = self.cfg.iter_best = int(iteration) * self.cfg.check_iter
+            (step, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate) \
+                    = line.split('\t')
+            self.cfg.step = self.cfg.best_step = int(step) * self.cfg.check_step
             self.loss_trains.append(float(loss_train))
             self.loss_devs.append(float(loss_dev))
             self.acc_chars.append(float(acc_char))
             self.acc_words.append(float(acc_word))
             self.f_scores.append(float(f_score))
             self.learning_rates.append(float(learning_rate))
-            self.batch_sizes.append(int(batch_size))
             if float(f_score) > f_score_best:
                 f_score_best = float(f_score)
                 best_idx = idx
-        logging.info('---- [%d|%d] loss(train/dev): %f / %f, acc(char/word): %f / %f, ' \
-                     'f-score: %f, lr: %f, bs: %d ----', self.cfg.epoch,
-                     self.cfg.iteration // self.cfg.check_iter, self.loss_trains[best_idx],
-                     self.loss_devs[best_idx], self.acc_chars[best_idx], self.acc_words[best_idx],
-                     self.f_scores[best_idx], self.learning_rates[-1], self.batch_sizes[-1])
+        logging.info('---- [%d] los(trn/dev): %.4f / %.4f, acc(chr/wrd): %.4f / %.4f, ' \
+                     'f-score: %.4f, lr: %.8f ----', self.cfg.step // self.cfg.check_step,
+                     self.loss_trains[best_idx], self.loss_devs[best_idx], self.acc_chars[best_idx],
+                     self.acc_words[best_idx], self.f_scores[best_idx], self.learning_rates[-1])
+
+    @classmethod
+    def _inf_data_iterator(cls, dataset: PosDataset) -> Iterator[PosSentTensor]:
+        """
+        데이터셋을 무한히 반복하여 문장을 출력하는 제너레이터
+        Args:
+            dataset:  데이터셋
+        Yields:
+            PosSentTensor 객체
+        """
+        for _ in range(1000000):
+            for sent in dataset:
+                yield sent
 
     def train(self):
         """
         train model with dataset
         """
-        self._check_continue()
+        self._restore_prev_train()
         logging.info('config: %s', pprint.pformat(self.cfg.__dict__))
+
         train_begin = datetime.now()
         logging.info('{{{{ training begin: %s {{{{', self._dt_str(train_begin))
         if torch.cuda.is_available():
@@ -185,26 +193,27 @@ def train(self):
         pathlib.Path(self.cfg.out_dir).mkdir(parents=True, exist_ok=True)
         self.log_file = open('{}/log.tsv'.format(self.cfg.out_dir), 'at')
         self.sum_wrt = SummaryWriter(self.cfg.out_dir)
+        check_start = (1 if self.cfg.step == 0 else (self.cfg.step // self.cfg.check_step + 1))
         patience = self.cfg.patience
-        for _ in range(10000):
-            has_best = self.train_epoch()
-            if has_best:
+        train_iter = self._inf_data_iterator(self.dataset_train)
+        for check_id in range(check_start, 1000000):
+            is_best = self._train_and_check(check_id, train_iter)
+            if is_best:
                 patience = self.cfg.patience
-            else:
-                if patience > 0:
-                    self._revert(True)
-                    patience -= 1
-                    logging.info('==== revert to iter: %d, f-score: %f, patience: %d ====',
-                                 self.cfg.iter_best, max(self.f_scores), patience)
-                else:
-                    break
+                continue
+            if patience <= 0:
+                break
+            self._revert_to_best(True)
+            patience -= 1
+            tqdm.write('==== revert to check: {}, f-score: {:.4f}, patience: {} ===='.format( \
+                    self.cfg.best_step // self.cfg.check_step, max(self.f_scores), patience))
+
         train_end = datetime.now()
         train_elapsed = self._elapsed(train_end - train_begin)
-        logging.info('}}}} training end: %s, elapsed: %s, epoch: %s, iter: %dk }}}}',
-                     self._dt_str(train_end), train_elapsed, self.cfg.epoch,
-                     self.cfg.iteration // 1000)
+        logging.info('}}}} training end: %s, elapsed: %s, step: %dk }}}}',
+                     self._dt_str(train_end), train_elapsed, self.cfg.step // 1000)
 
-    def _revert(self, is_decay_lr: bool):
+    def _revert_to_best(self, is_decay_lr: bool):
         """
         이전 best 모델로 되돌린다.
         Args:
@@ -215,79 +224,67 @@ def _revert(self, is_decay_lr: bool):
             self.cfg.learning_rate *= self.cfg.lr_decay
         self._load_optim('{}/optim.state'.format(self.cfg.out_dir), self.cfg.learning_rate)
 
-    def train_epoch(self) -> bool:
+    def _train_and_check(self, check_id: int, train_iter: Iterator[PosSentTensor]) -> bool:
         """
-        한 epoch을 학습한다. 배치 단위는 글자 단위
+        cfg.check_step 만큼의 step을 수행하고 evaluation을 수행한다.
+        Args:
+            check_id:  check ID
+            train_iter:  학습 데이터 iterator
         Returns:
-            best f-score를 기록한 iteration 여부
+            best f-score를 기록한 step 여부
         """
-        epoch_begin = datetime.now()
-        logging.info('{{{{ epoch: %d, begin: %s {{{{', self.cfg.epoch, self._dt_str(epoch_begin))
-        iter_lap1 = datetime.now()
-        has_best = False
-        batches = []
+        start_step = self.cfg.step
+        loss_batch = torch.tensor(0.0)    # pylint: disable=not-callable
+        batch_size = 0
         loss_trains = []
-        batch_size = self.cfg.batch_size
-        if self.cfg.batch_grow > 0:
-            batch_size = self.cfg.batch_size + self.cfg.iteration // self.cfg.batch_grow
-        for train_sent in tqdm(self.dataset_train, 'EPOCH [{}]'.format(self.cfg.epoch),
-                               len(self.dataset_train), mininterval=1, ncols=100):
+        train_sents = tqdm(train_iter, '[{}]'.format(check_id), mininterval=1, ncols=100)
+        for train_sent in train_sents:
             train_labels, train_contexts = train_sent.to_tensor(self.cfg, self.rsc, True)
             if torch.cuda.is_available():
                 train_labels = train_labels.cuda()
                 train_contexts = train_contexts.cuda()
+                loss_batch = loss_batch.cuda()
 
-            self.optimizer.zero_grad()
             self.model.train()
             train_outputs = self.model(train_contexts)
-            batches.append((train_labels, train_outputs))
-            if sum([batch[0].size(0) for batch in batches]) < batch_size:
+            train_outputs.requires_grad_()
+            loss_train = self.criterion(train_outputs, train_labels)
+            loss_train.backward()
+            loss_trains.append(loss_train.item())
+            loss_batch += loss_train
+            batch_size += len(train_labels)
+            if batch_size < self.cfg.batch_size:
                 continue
-            batch_label = torch.cat([x[0] for x in batches], 0)    # pylint: disable=no-member
-            batch_output = torch.cat([x[1] for x in batches], 0)    # pylint: disable=no-member
-            batches = []
 
-            batch_output.requires_grad_()
-            loss_train = self.criterion(batch_output, batch_label)
-            loss_trains.append(loss_train.item())
-            loss_train.backward()
             self.optimizer.step()
-            self.cfg.iteration += 1
+            self.optimizer.zero_grad()
+            self.sum_wrt.add_scalar('loss-batch', loss_batch.item(), self.cfg.step)
+            self.cfg.step += 1
+            loss_batch = torch.tensor(0.0)    # pylint: disable=not-callable
+            batch_size = 0
 
-            if self.cfg.iteration % self.cfg.check_iter == 0:
-                avg_loss_dev, acc_char, acc_word, f_score = self.evaluate()
-                iter_lap2 = datetime.now()
-                iter_elapsed = self._elapsed(iter_lap2 - iter_lap1)
-                iter_lap1 = iter_lap2
-                has_best |= self._check_iter(iter_elapsed, loss_trains, avg_loss_dev, acc_char,
-                                             acc_word, f_score, batch_size)
-            if self.cfg.batch_grow > 0 and self.cfg.iteration % self.cfg.batch_grow == 0:
-                batch_size = self.cfg.batch_size + self.cfg.iteration // self.cfg.batch_grow
+            if (self.cfg.step - start_step) >= self.cfg.check_step:
+                train_sents.close()
+                break
 
-        print(file=sys.stderr)
-        sys.stderr.flush()
-        epoch_end = datetime.now()
-        epoch_elapsed = self._elapsed(epoch_end - epoch_begin)
-        logging.info('}}}} epoch: %d, end: %s, elapsed: %s, %s best }}}}', self.cfg.epoch,
-                     self._dt_str(epoch_begin), epoch_elapsed, 'hit' if has_best else 'did not hit')
-        self.cfg.epoch += 1
-        return has_best
+        avg_loss_dev, acc_char, acc_word, f_score = self.evaluate()
+        return self._check(check_id, loss_trains, avg_loss_dev, acc_char, acc_word, f_score)
 
-    def _check_iter(self, elapsed: str, loss_trains: List[float], avg_loss_dev: float,
-                    acc_char: float, acc_word: float, f_score: float, batch_size: int) -> bool:
+    def _check(self, check_id: int, loss_trains: List[float], avg_loss_dev: float, acc_char: float,
+               acc_word: float, f_score: float) -> bool:
         """
-        cfg.check_iter번의 iteration마다 수행하는 체크
+        cfg.check_step번의 step마다 수행하는 체크
         Args:
-            elapsed:  경과 시간
-            loss_trains:   train 코퍼스에서 각 배치별 loss 리스트
+            check_id:  check ID
+            loss_trains:  train 코퍼스에서 각 배치별 loss 리스트
             avg_loss_dev:  dev 코퍼스 문장 별 평균 loss
             acc_char:  음절 정확도
             acc_word:  어절 정확도
             f_score:  f-score
-            batch_size:  batch size
         Returns:
-            현재 iteration이 best 성능을 나타냈는 지 여부
+            현재 step이 best 성능을 나타냈는 지 여부
         """
+        assert check_id == self.cfg.step // self.cfg.check_step
         avg_loss_train = sum(loss_trains) / len(loss_trains)
         loss_trains.clear()
         self.loss_trains.append(avg_loss_train)
@@ -296,36 +293,34 @@ def _check_iter(self, elapsed: str, loss_trains: List[float], avg_loss_dev: floa
         self.acc_words.append(acc_word)
         self.f_scores.append(f_score)
         self.learning_rates.append(self.cfg.learning_rate)
-        self.batch_sizes.append(batch_size)
-        check_num = self.cfg.iteration // self.cfg.check_iter
-        logging.info('---- [%d|%d] elapsed: %s, loss(train/dev): %f / %f, acc(char/word): %f / %f' \
-                     ', f-score: %f (max: %f), lr: %f, bs: %d ----', self.cfg.epoch, check_num,
-                     elapsed, avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score,
-                     max(self.f_scores), self.cfg.learning_rate, batch_size)
-        print('{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}'.format( \
-              self.cfg.epoch, check_num, avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score,
-              self.cfg.learning_rate, batch_size), file=self.log_file)
+        is_best = self._is_best()
+        is_best_str = 'BEST' if is_best else '< {:.4f}'.format(max(self.f_scores))
+        tqdm.write('    [Los trn]  [Los dev]  [Acc chr]  [Acc wrd]  [F-score]           [LR]')
+        tqdm.write('    {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f} {:8}  {:.8f}'.format( \
+                avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score, is_best_str,
+                self.cfg.learning_rate))
+        print('{}\t{}\t{}\t{}\t{}\t{}\t{}'.format(check_id, avg_loss_train, avg_loss_dev,
+                                                  acc_char, acc_word, f_score,
+                                                  self.cfg.learning_rate), file=self.log_file)
         self.log_file.flush()
-        self.sum_wrt.add_scalar('loss-train', avg_loss_train, check_num)
-        self.sum_wrt.add_scalar('loss-dev', avg_loss_dev, check_num)
-        self.sum_wrt.add_scalar('acc-char', acc_char, check_num)
-        self.sum_wrt.add_scalar('acc-word', acc_word, check_num)
-        self.sum_wrt.add_scalar('f-score', f_score, check_num)
-        self.sum_wrt.add_scalar('learning-rate', self.cfg.learning_rate, check_num)
-        self.sum_wrt.add_scalar('batch-size', batch_size, check_num)
-        return self._check_best()
+        self.sum_wrt.add_scalar('loss-train', avg_loss_train, check_id)
+        self.sum_wrt.add_scalar('loss-dev', avg_loss_dev, check_id)
+        self.sum_wrt.add_scalar('acc-char', acc_char, check_id)
+        self.sum_wrt.add_scalar('acc-word', acc_word, check_id)
+        self.sum_wrt.add_scalar('f-score', f_score, check_id)
+        self.sum_wrt.add_scalar('learning-rate', self.cfg.learning_rate, check_id)
+        return is_best
 
-    def _check_best(self) -> bool:
+    def _is_best(self) -> bool:
         """
-        이번 iteration에 가장 좋은 성능을 냈는 지 확인하고 그럴 경우 현재 상태를 저장한다.
+        이번 step에 가장 좋은 성능을 냈는 지 확인하고 그럴 경우 현재 상태를 저장한다.
         Returns:
             best 여부
         """
         if len(self.f_scores) > 1 and max(self.f_scores[:-1]) >= self.f_scores[-1]:
             return False
-        # this iteration hits new max value
-        logging.info('==== best model: %f ====', self.f_scores[-1])
-        self.cfg.iter_best = self.cfg.iteration
+        # this step hits new max value
+        self.cfg.best_step = self.cfg.step
         self.model.save('{}/model.state'.format(self.cfg.out_dir))
         self._save_optim('{}/optim.state'.format(self.cfg.out_dir))
         with open('{}/config.json'.format(self.cfg.out_dir), 'w', encoding='UTF-8') as fout:
diff --git a/train/bin/train.py b/train/bin/train.py
index fcd7998..b696d96 100755
--- a/train/bin/train.py
+++ b/train/bin/train.py
@@ -40,36 +40,30 @@ def main():
     main function processes only argument parsing
     """
     parser = ArgumentParser(description='train model from data')
-    parser.add_argument('-i', '--in-pfx', help='input data prefix', metavar='NAME', required=True)
+    parser.add_argument('-i', '--in-pfx', help='input data path prefix', metavar='NAME',
+                        required=True)
     parser.add_argument('--rsc-src', help='train resource dir <default: ../rsc/src>', metavar='DIR',
                         default='../rsc/src')
-    parser.add_argument('--log-dir', help='tensorboard log dir', metavar='DIR',
+    parser.add_argument('--logdir', help='tensorboard log dir <default: ./logdir>', metavar='DIR',
                         default='./logdir')
-    parser.add_argument('--window', help='left/right character window length', metavar='INT',
-                        type=int,
-                        default=3)
-    parser.add_argument('--spc-dropout', help='space(word delimiter) dropout rate', metavar='REAL',
-                        type=float,
-                        default=0.0)
-    parser.add_argument('--cutoff', help='cutoff', metavar='INT', type=int,
-                        default=2)
-    parser.add_argument('--embed-dim', help='embedding dimension', metavar='INT', type=int,
-                        default=30)
-    parser.add_argument('--learning-rate', help='learning rate', metavar='REAL', type=float,
-                        default=0.001)
-    parser.add_argument('--lr-decay', help='learning rate decay', metavar='REAL', type=float,
-                        default=0.9)
-    parser.add_argument('--batch-size', help='batch size', metavar='INT', type=int,
-                        default=500)
-    parser.add_argument('--check-iter', help='check iteration', metavar='INT', type=int,
-                        default=10000)
-    parser.add_argument('--batch-grow', help='batch growing iteration. 0 means no growing',
-                        type=int,
-                        default=10000)
-    parser.add_argument('--patience', help='maximum patience count to revert model', metavar='INT',
-                        type=int,
-                        default=10)
-    parser.add_argument('--gpu-num', help='GPU number to use', metavar='INT', type=int,
+    parser.add_argument('--window', help='left/right character window length <default: 3>',
+                        metavar='INT', type=int, default=3)
+    parser.add_argument('--spc-dropout', help='space(word delimiter) dropout rate <default: 0.0>',
+                        metavar='REAL', type=float, default=0.0)
+    parser.add_argument('--cutoff', help='cutoff <default: 2>', metavar='INT', type=int, default=2)
+    parser.add_argument('--embed-dim', help='embedding dimension <default: 30>', metavar='INT',
+                        type=int, default=30)
+    parser.add_argument('--learning-rate', help='learning rate <default: 0.001>', metavar='REAL',
+                        type=float, default=0.001)
+    parser.add_argument('--lr-decay', help='learning rate decay <default: 0.9>', metavar='REAL',
+                        type=float, default=0.9)
+    parser.add_argument('--batch-size', help='batch size <default: 1000>', metavar='INT', type=int,
+                        default=1000)
+    parser.add_argument('--check-step', help='check every N step <default: 10000>', metavar='INT',
+                        type=int, default=10000)
+    parser.add_argument('--patience', help='maximum patience count to revert model <default: 10>',
+                        metavar='INT', type=int, default=10)
+    parser.add_argument('--gpu-num', help='GPU number to use <default: 0>', metavar='INT', type=int,
                         default=0)
     parser.add_argument('--debug', help='enable debug', action='store_true')
     args = parser.parse_args()
diff --git a/train/requirements.txt b/train/requirements.txt
new file mode 100644
index 0000000..5627706
--- /dev/null
+++ b/train/requirements.txt
@@ -0,0 +1,3 @@
+tensorboardX
+torch==0.4.1
+tqdm

From 3097e6e8da1da05d92a320a991d580fb398d5638 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Thu, 24 Jan 2019 14:34:44 +0900
Subject: [PATCH 28/41] =?UTF-8?q?=ED=95=99=EC=8A=B5=EC=9D=B4=20=EC=99=84?=
 =?UTF-8?q?=EB=A3=8C=EB=90=9C=20=EB=AA=A8=EB=8D=B8=20=ED=8C=8C=EC=9D=BC?=
 =?UTF-8?q?=EC=9D=84=20pickle=20=ED=98=95=ED=83=9C=EB=A1=9C=20=EC=A0=80?=
 =?UTF-8?q?=EC=9E=A5=ED=95=98=EB=8A=94=20=EC=8A=A4=ED=81=AC=EB=A6=BD?=
 =?UTF-8?q?=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 rsc/bin/compile_model.py  |   6 +-
 train/bin/pickle_model.py | 243 ++++++++++++++++++++++++++++++++++++++
 2 files changed, 246 insertions(+), 3 deletions(-)
 create mode 100755 train/bin/pickle_model.py

diff --git a/rsc/bin/compile_model.py b/rsc/bin/compile_model.py
index 1c10d1f..b06fff7 100755
--- a/rsc/bin/compile_model.py
+++ b/rsc/bin/compile_model.py
@@ -26,7 +26,7 @@
 #############
 # functions #
 #############
-def load_cfg_rsc(rsc_src: str, model_size: str) -> Tuple[Namespace, Resource]:
+def _load_cfg_rsc(rsc_src: str, model_size: str) -> Tuple[Namespace, Resource]:
     """
     load config and resource from source directory
     Args:
@@ -38,7 +38,7 @@ def load_cfg_rsc(rsc_src: str, model_size: str) -> Tuple[Namespace, Resource]:
     """
     file_path = '{}/{}.config.json'.format(rsc_src, model_size)
     cfg_dic = json.load(open(file_path, 'r', encoding='UTF-8'))
-    logging.info('config: %s', json.dumps(cfg_dic, indent=2))
+    logging.info('config: %s', json.dumps(cfg_dic, indent=4, sort_keys=True))
     cfg = Namespace()
     for key, val in cfg_dic.items():
         setattr(cfg, key, val)
@@ -145,7 +145,7 @@ def run(args: Namespace):
     Args:
         args:  program arguments
     """
-    cfg, rsc = load_cfg_rsc(args.rsc_src, args.model_size)
+    cfg, rsc = _load_cfg_rsc(args.rsc_src, args.model_size)
     data = pickle.load(open('{}/{}.model.pickle'.format(args.rsc_src, args.model_size), 'rb'))
     _write_config(cfg, rsc, args.rsc_dir)
     _write_data(data, args.rsc_dir)
diff --git a/train/bin/pickle_model.py b/train/bin/pickle_model.py
new file mode 100755
index 0000000..0eafbb1
--- /dev/null
+++ b/train/bin/pickle_model.py
@@ -0,0 +1,243 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+pickle trained model (state dict)
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+from array import array
+import json
+import logging
+import os
+import pickle
+import re
+
+import torch
+
+from khaiii.resource.resource import Resource
+
+
+#############
+# functions #
+#############
+def _validate_state_dict(cfg: Namespace, rsc: Resource, state_dict: dict):
+    """
+    validate state dict with config and resource
+    Args:
+        cfg:  config
+        rsc:  resource
+        state_dict:  state dict
+    """
+    def _assert(expected, actual, msg: str):
+        """
+        assert expected equals actual. if not raise exception
+        Args:
+            expected:  expected value
+            actual:  actual value
+            msg:  exception message if both are not equal
+        """
+        if expected != actual:
+            raise ValueError((msg + ': expected: {}, but actual: {}').format(expected, actual))
+
+    for name, tensor in state_dict.items():
+        logging.info('%s: %s', name, tensor.size())
+        if name == 'embedder.embedding.weight':
+            _assert(len(rsc.vocab_in), tensor.size(0), 'invalid input vocab')
+            _assert(cfg.embed_dim, tensor.size(1), 'invalid embedding dim')
+        if re.match(r'convs\..\.weight', name):
+            _assert(cfg.embed_dim, tensor.size(1), 'invalid conv input channel dim')
+        if name == 'hidden2tag.bias':
+            _assert(len(rsc.vocab_out), tensor.size(0), 'invalid output bias dim')
+
+
+def _get_embedding(rsc: Resource, state_dict: dict) -> dict:
+    """
+    get character embedding
+    Args:
+        rsc:  resource
+        state_dict:  state dictionary of model
+    Returns:
+        embedding data
+    """
+    data = {}
+    chars = [0, ] * 5    # 5 special characters
+    for idx in range(5, len(rsc.vocab_in)):
+        chars.append(ord(rsc.vocab_in[idx]))
+    data['chars'] = array('i', chars)    # [input vocab(char)] * 4(wchar_t)
+
+    embedding = state_dict['embedder.embedding.weight']
+    data['weights'] = []
+    for row in embedding:
+        data['weights'].append(array('f', row))    # [input vocab(char)] * embed_dim * 4(float)
+    return data
+
+
+def _get_linear(name_pfx: str, state_dict: dict) -> dict:
+    """
+    write weight and bias with given layer name prefix
+    Args:
+        name_pfx:  layer name prefix
+        state_dict:  state dictionary of model
+    Returns:
+        linear layer data
+    """
+    data = {}
+    data['weight'] = []
+    weight = state_dict['{}.weight'.format(name_pfx)]
+    for row in weight:
+        data['weight'].append(array('f', row))    # output * input * 4
+    bias_name = '{}.bias'.format(name_pfx)
+    if bias_name in state_dict:
+        bias = state_dict[bias_name]
+        data['bias'] = array('f', bias)    # output * 4
+    return data
+
+
+def _get_conv(name_pfx: str, kernel_size: int, state_dict: dict) -> dict:
+    """
+    get convolution module weight and bias with given layer name prefix and suffix(index)
+    Args:
+        name_pfx:  layer name prefix
+        kernel_size:  kernel size
+        state_dict:  state dictionary of model
+    Returns:
+        conv data with given kernel size
+    """
+    data = {}
+    weight = state_dict['{}.{}.weight'.format(name_pfx, kernel_size - 2)]
+    weight_t = weight.transpose(1, 2)    # [output chan] * [input chan] * kernel
+    #                                      => [output chan] * kernel * [input chan]
+    data['channel'] = []
+    for ch_out in weight_t:
+        data['channel'].append(array('f', ch_out.contiguous().view(-1)))
+    bias_name = '{}.{}.bias'.format(name_pfx, kernel_size - 2)
+    if bias_name in state_dict:
+        bias = state_dict[bias_name]
+        data['bias'] = array('f', bias)    # [output chan] * 4
+    return data
+
+
+def _get_data(rsc: Resource, state_dict: dict) -> dict:
+    """
+    get all data to pickle
+    Args:
+        rsc:  resource
+        state_dict:  state dictionary of model
+    Returns:
+        all data
+    """
+    data = {}
+    # key: [input vocab(char)] * 4(float)
+    # val: [input vocab(char)] * embed_dim * 4(float)
+    data['embedding'] = _get_embedding(rsc, state_dict)
+
+    data['convs'] = {}
+    for kernel in range(2, 6):
+        # weight: [output chan(embed_dim)] * kernel * [input chan(embed_dim)] * 4
+        # bias: [output chan] * 4
+        data['convs'][kernel] = _get_conv('convs', kernel, state_dict)
+
+    # weight: hidden_dim * [cnn layers * output chan(embed_dim)] * 4
+    # bias: hidden_dim * 4
+    data['conv2hidden'] = _get_linear('conv2hidden', state_dict)
+
+    # weight: [output vocab(tag)] * hidden_dim * 4
+    # bias: [output vocab(tag)] * 4
+    data['hidden2tag'] = _get_linear('hidden2tag', state_dict)
+    return data
+
+
+def _load_config(path: str) -> Namespace:
+    """
+    load config file
+    Args:
+        path:  path
+    Returns:
+        config
+    """
+    cfg_dic = json.load(open(path, 'r', encoding='UTF-8'))
+    logging.info('config: %s', json.dumps(cfg_dic, indent=4, sort_keys=True))
+    cfg = Namespace()
+    for key, val in cfg_dic.items():
+        if key not in {'cutoff', 'embed_dim', 'hidden_dim', 'model_id', 'model_name', 'rsc_src',
+                       'window'}:
+            continue
+        setattr(cfg, key, val)
+    return cfg
+
+
+def _load_resource(cfg: Namespace, rsc_src: str) -> Resource:
+    """
+    load resources
+    Args:
+        cfg:  config
+        rsc_src:  resource source dir
+    Returns:
+        Resource object
+    """
+    cwd = os.path.realpath(os.getcwd())
+    train_dir = os.path.realpath('{}/..'.format(rsc_src))
+    if cwd != train_dir:
+        os.chdir(train_dir)
+    rsc = Resource(cfg)
+    if cwd != train_dir:
+        os.chdir(cwd)
+    return rsc
+
+
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    cfg = _load_config('{}/config.json'.format(args.in_dir))
+    rsc = _load_resource(cfg, args.rsc_src)
+    state_dict = torch.load('{}/model.state'.format(args.in_dir),
+                            map_location=lambda storage, loc: storage)
+    _validate_state_dict(cfg, rsc, state_dict)
+    data = _get_data(rsc, state_dict)
+
+    config_path = '{}/{}.config.json'.format(args.rsc_src, args.model_size)
+    with open(config_path, 'w', encoding='UTF-8') as fout:
+        json.dump(vars(cfg), fout, indent=4, sort_keys=True)
+
+    pickle_path = '{}/{}.model.pickle'.format(args.rsc_src, args.model_size)
+    with open(pickle_path, 'wb') as fout:
+        pickle.dump(data, fout)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='pickle trained model (state dict)')
+    parser.add_argument('-i', '--in-dir', help='model dir', metavar='DIR', required=True)
+    parser.add_argument('--model-size', help='model size <default: base>',
+                        metavar='SIZE', default='base')
+    parser.add_argument('--rsc-src', help='resource source dir <default: ../rsc/src>',
+                        metavar='DIR', default='../rsc/src')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From 5753e171e4888f97a0fb81ccbcdf369a6c81994d Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Fri, 25 Jan 2019 00:28:15 +0900
Subject: [PATCH 29/41] =?UTF-8?q?=ED=95=99=EC=8A=B5=EB=90=9C=20=EB=AA=A8?=
 =?UTF-8?q?=EB=8D=B8=EC=9D=84=20=EB=AA=85=EB=A0=B9=EC=B0=BD=EC=97=90?=
 =?UTF-8?q?=EC=84=9C=20=EC=8B=A4=ED=96=89=ED=95=B4=EB=B3=BC=20=EC=88=98=20?=
 =?UTF-8?q?=EC=9E=88=EB=8A=94=20=EB=8D=B0=EB=AA=A8=20=EC=8A=A4=ED=81=AC?=
 =?UTF-8?q?=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 rsc/bin/compile_model.py               |  2 +-
 src/main/python/khaiii/train/tagger.py |  9 ++-
 train/bin/pickle_model.py              |  2 +-
 train/bin/tag.py                       | 76 ++++++++++++++++++++++++++
 4 files changed, 82 insertions(+), 7 deletions(-)
 create mode 100755 train/bin/tag.py

diff --git a/rsc/bin/compile_model.py b/rsc/bin/compile_model.py
index b06fff7..767060c 100755
--- a/rsc/bin/compile_model.py
+++ b/rsc/bin/compile_model.py
@@ -43,7 +43,7 @@ def _load_cfg_rsc(rsc_src: str, model_size: str) -> Tuple[Namespace, Resource]:
     for key, val in cfg_dic.items():
         setattr(cfg, key, val)
     cwd = os.path.realpath(os.getcwd())
-    train_dir = os.path.realpath('{}/..'.format(rsc_src))
+    train_dir = os.path.realpath('{}/../../train'.format(rsc_src))
     if cwd != train_dir:
         os.chdir(train_dir)
     rsc = Resource(cfg)
diff --git a/src/main/python/khaiii/train/tagger.py b/src/main/python/khaiii/train/tagger.py
index 06fa0e2..c0997d8 100644
--- a/src/main/python/khaiii/train/tagger.py
+++ b/src/main/python/khaiii/train/tagger.py
@@ -31,19 +31,18 @@ class PosTagger:
     """
     part-of-speech tagger
     """
-    def __init__(self, rsc_src: str):
+    def __init__(self, model_dir: str):
         """
         Args:
-            rsc_src:  resource dir
+            model_dir:  model dir
         """
-        cfg_dict = json.load(open('{}/config.json'.format(rsc_src), 'r', encoding='UTF-8'))
+        cfg_dict = json.load(open('{}/config.json'.format(model_dir), 'r', encoding='UTF-8'))
         self.cfg = Namespace()
         for key, val in cfg_dict.items():
             setattr(self.cfg, key, val)
-        self.cfg.rsc_src = rsc_src
         self.rsc = Resource(self.cfg)
         self.model = CnnModel(self.cfg, self.rsc)
-        self.model.load('{}/model.state'.format(rsc_src))
+        self.model.load('{}/model.state'.format(model_dir))
         self.model.eval()
 
     def tag_raw(self, raw_sent: str, enable_restore: bool = True) -> PosSentTensor:
diff --git a/train/bin/pickle_model.py b/train/bin/pickle_model.py
index 0eafbb1..06f6620 100755
--- a/train/bin/pickle_model.py
+++ b/train/bin/pickle_model.py
@@ -184,7 +184,7 @@ def _load_resource(cfg: Namespace, rsc_src: str) -> Resource:
         Resource object
     """
     cwd = os.path.realpath(os.getcwd())
-    train_dir = os.path.realpath('{}/..'.format(rsc_src))
+    train_dir = os.path.realpath('{}/../../train'.format(rsc_src))
     if cwd != train_dir:
         os.chdir(train_dir)
     rsc = Resource(cfg)
diff --git a/train/bin/tag.py b/train/bin/tag.py
new file mode 100755
index 0000000..60bb2e9
--- /dev/null
+++ b/train/bin/tag.py
@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+command line part-of-speech tagger demo
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+import logging
+import os
+import sys
+
+from khaiii.train.tagger import PosTagger
+
+
+#############
+# functions #
+#############
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    tgr = PosTagger(args.model_dir)
+    for line_num, line in enumerate(sys.stdin, start=1):
+        if line_num % 100000 == 0:
+            logging.info('%d00k-th line..', (line_num // 100000))
+        line = line.rstrip('\r\n')
+        if not line:
+            print()
+            continue
+        pos_sent = tgr.tag_raw(line)
+        for pos_word in pos_sent.pos_tagged_words:
+            print(pos_word.raw, end='\t')
+            print(' + '.join([str(m) for m in pos_word.pos_tagged_morphs]))
+        print()
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='command line part-of-speech tagger demo')
+    parser.add_argument('-m', '--model-dir', help='model dir', metavar='DIR', required=True)
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--gpu-num', help='GPU number to use', metavar='INT', type=int, default=0)
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_num)
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From 9f72716380c519c5744de630f66e1351d453ad45 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Fri, 25 Jan 2019 00:49:26 +0900
Subject: [PATCH 30/41] =?UTF-8?q?=EA=B8=B0=EB=B6=84=EC=84=9D=20=EC=82=AC?=
 =?UTF-8?q?=EC=A0=84=EC=9D=84=20=EC=9E=90=EB=8F=99=EC=9C=BC=EB=A1=9C=20?=
 =?UTF-8?q?=EC=B6=94=EC=B6=9C=ED=95=98=EB=8A=94=20=EC=8A=A4=ED=81=AC?=
 =?UTF-8?q?=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 train/bin/extract_preanal.py | 209 +++++++++++++++++++++++++++++++++++
 1 file changed, 209 insertions(+)
 create mode 100755 train/bin/extract_preanal.py

diff --git a/train/bin/extract_preanal.py b/train/bin/extract_preanal.py
new file mode 100755
index 0000000..b23141a
--- /dev/null
+++ b/train/bin/extract_preanal.py
@@ -0,0 +1,209 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+기분석 사전 후보를 추출하는 스크립트
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+from collections import Counter, defaultdict
+import logging
+import sys
+from typing import Dict, List, Tuple
+
+from khaiii.munjong.sejong_corpus import Word
+from khaiii.resource.char_align import Aligner, AlignError
+
+import map_char_to_tag
+
+
+#########
+# types #
+#########
+class Entry:
+    """
+    preanalized dictionary entry
+    """
+    def __init__(self, freq: int, word: str, morph_str: str):
+        self.is_del = False
+        self.is_pfx = word[-1] == '\1'
+        self.freq = freq
+        self.word = word[:-1] if self.is_pfx else word
+        self.morph_str = morph_str
+
+    def __str__(self):
+        return '{}{}\t{}{}\t{}'.format('-' if self.is_del else '', self.freq, self.word,
+                                       '*' if self.is_pfx else '', self.morph_str)
+
+
+#############
+# variables #
+#############
+DIC_AMBIG = defaultdict(Counter)
+
+
+#############
+# functions #
+#############
+def _get_prefix(word: Word) -> Tuple[str, str]:
+    """
+    맨 뒤의 형태소 하나를 제외한 prefix와 그 분석 결과
+    Args:
+        word:  어절 객체
+    Returns:
+        prefix
+        분석 결과
+    """
+    if len(word.morphs) < 2:
+        return None, None
+    if not word.raw.endswith(word.morphs[-1].lex):
+        return None, None
+    prefix = '{}\1'.format(word.raw[:-len(word.morphs[-1].lex)])
+    if len(prefix) < 5:    # 4음절 미만은 버린다.
+        return None, None
+    morphs = ['{}/{}'.format(m.lex, m.tag) for m in word.morphs[:-1]]
+    return prefix, ' + '.join(morphs)
+
+
+def _count_ambig(args: Namespace):
+    """
+    count from courpus and make ambiguous dictionary
+    Args:
+        args:  program arguments
+    """
+    for num, sent in enumerate(map_char_to_tag.sent_iter(args), start=1):
+        if num % 10000 == 0:
+            logging.info('%dk-th sentence..', num // 1000)
+        for word in sent.words:
+            DIC_AMBIG[word.raw][word.morph_str()] += 1
+            prefix, morph_pfx = _get_prefix(word)
+            if prefix:
+                DIC_AMBIG[prefix][morph_pfx] += 1
+
+
+def _filter_no_ambig(min_freq: int) -> Dict[str, str]:
+    """
+    문맥과 상관 없이 중의성이 없는 엔트리를 출력
+    Args:
+        min_freq:  최소 빈도
+    Returns:
+        중의성이 있는 엔트리가 제거된 사전
+    """
+    dic_no_ambig = {}
+    for word, cnt in DIC_AMBIG.items():
+        if len(cnt) > 1:
+            continue
+        morph_str, freq = cnt.most_common(1)[0]
+        if freq < min_freq:
+            continue
+        dic_no_ambig[word] = morph_str
+    return dic_no_ambig
+
+
+def _make_entries(dic_no_ambig: Dict[str, str]) -> List[Entry]:
+    """
+    기분석 사전 엔트리를 생성한다.
+    Args:
+        dic_no_ambig:  중의성이 없는 엔트리 사전
+    Returns:
+        엔트리 리스트
+    """
+    entries = []
+    pfx_idx = -1
+    for idx, (word, morph_str) in enumerate(sorted(dic_no_ambig.items())):
+        _, freq = list(DIC_AMBIG[word].items())[0]
+        entry = Entry(freq, word, morph_str)
+        if entry.is_pfx:
+            if entries and entries[-1].word == entry.word:
+                # 이전 어절 exact가 현재 prefix와 같다면, 예: "제이미" vs "제이미*"
+                if entries[-1].morph_str == entry.morph_str:    # pylint: disable=simplifiable-if-statement
+                    # 분석 결과가 같다면 prefix를 남기고 exact를 제거한다.
+                    entries[-1].is_del = True
+                else:
+                    # 분석 결과가 다르면 안전하게 exact를 남기고 prefix를 제거한다.
+                    entry.is_del = True
+            pfx_idx = idx
+        elif (pfx_idx >= 0 and word.startswith(entries[pfx_idx].word) and
+              morph_str.startswith(entries[pfx_idx].morph_str) and
+              entries[pfx_idx].freq <= freq):
+            # 이전 prefix가 현재 어절 exact를 포함하면서 빈도는 같다면,
+            # 모두 동일한 어절에서 뽑힌 prefix이므로 prefix를 삭제한다.
+            # 예: 빈도가 같은 "강화된다*"(삭제) vs "강화된다."(남김)
+            entries[pfx_idx].is_del = True
+        entries.append(entry)
+    return entries
+
+
+def run(args: Namespace):
+    """
+    map characters with tags from eojeol and morphemes
+    Args:
+        args:  program arguments
+    """
+    aligner = Aligner(args.rsc_src)
+
+    _count_ambig(args)
+    dic_no_ambig = _filter_no_ambig(args.min_freq)
+    entries = _make_entries(dic_no_ambig)
+
+    del_word = 0
+    del_pfx = 0
+    for entry in entries:
+        if entry.is_del:
+            logging.debug(entry)
+            if entry.is_pfx:
+                del_pfx += 1
+            else:
+                del_word += 1
+        else:
+            word = Word.parse('\t'.join(['', entry.word, entry.morph_str]), '', 0)
+            try:
+                aligner.align(word)
+                entry_str = str(entry)
+                if not args.with_freq:
+                    entry_str = '\t'.join(entry_str.split('\t')[1:])
+                print(entry_str)
+            except AlignError as algn_err:
+                logging.error('%s: %s', algn_err, entry)
+
+    logging.info('deleted word: %d', del_word)
+    logging.info('deleted prefix: %d', del_pfx)
+
+
+########
+# main #
+########
+def main():
+    """
+    map characters with tags from eojeol and morphemes
+    """
+    parser = ArgumentParser(description='기분석 사전 후보를 추출하는 스크립트')
+    parser.add_argument('-c', '--corpus-dir', help='corpus dir', metavar='DIR', required=True)
+    parser.add_argument('--rsc-src', help='resource source dir <default: ../rsc/src>',
+                        metavar='DIR', default='../rsc/src')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--min-freq', help='minimum frequency <default: 10>', metavar='NUM',
+                        type=int, default=10)
+    parser.add_argument('--with-freq', help='print with frequency', action='store_true')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From a2df2247f3971d6db130379eeebf118f9a55df92 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Sat, 26 Jan 2019 15:08:47 +0900
Subject: [PATCH 31/41] =?UTF-8?q?=EC=98=A4=EB=B6=84=EC=84=9D=20=ED=8C=A8?=
 =?UTF-8?q?=EC=B9=98=20=ED=9B=84=EB=B3=B4=EB=A5=BC=20=EC=B6=94=EC=B6=9C?=
 =?UTF-8?q?=ED=95=98=EB=8A=94=20=EC=8A=A4=ED=81=AC=EB=A6=BD=ED=8A=B8=20?=
 =?UTF-8?q?=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 rsc/bin/compile_model.py      |   8 +-
 train/bin/extract_errpatch.py | 352 ++++++++++++++++++++++++++++++++++
 train/bin/pickle_model.py     |  23 +--
 3 files changed, 355 insertions(+), 28 deletions(-)
 create mode 100755 train/bin/extract_errpatch.py

diff --git a/rsc/bin/compile_model.py b/rsc/bin/compile_model.py
index 767060c..f40fc6b 100755
--- a/rsc/bin/compile_model.py
+++ b/rsc/bin/compile_model.py
@@ -15,7 +15,6 @@
 from argparse import ArgumentParser, Namespace
 import json
 import logging
-import os
 import pathlib
 import pickle
 from typing import Tuple
@@ -42,13 +41,8 @@ def _load_cfg_rsc(rsc_src: str, model_size: str) -> Tuple[Namespace, Resource]:
     cfg = Namespace()
     for key, val in cfg_dic.items():
         setattr(cfg, key, val)
-    cwd = os.path.realpath(os.getcwd())
-    train_dir = os.path.realpath('{}/../../train'.format(rsc_src))
-    if cwd != train_dir:
-        os.chdir(train_dir)
+    setattr(cfg, 'rsc_src', rsc_src)
     rsc = Resource(cfg)
-    if cwd != train_dir:
-        os.chdir(cwd)
     return cfg, rsc
 
 
diff --git a/train/bin/extract_errpatch.py b/train/bin/extract_errpatch.py
new file mode 100755
index 0000000..ff9dd8c
--- /dev/null
+++ b/train/bin/extract_errpatch.py
@@ -0,0 +1,352 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+오분석 패치 후보를 추출하는 스크립트
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+from collections import Counter
+import difflib
+import logging
+import multiprocessing
+import os
+import sys
+from typing import Iterator, List, Tuple
+
+from khaiii.munjong import sejong_corpus
+from khaiii.resource.morphs import WORD_DELIM_STR, SENT_DELIM_STR
+
+from khaiii.khaiii import KhaiiiApi
+
+
+#############
+# variables #
+#############
+_KHAIII = None    # khaiii API
+
+
+#############
+# functions #
+#############
+def _get_diff_sgmts(result_morphs: List[str], corpus_morphs: List[str], raw_sent: str,
+                    alignments: List[Tuple[int, int]]) -> List[Tuple[str, str, str]]:
+    """
+    패치 생성을 위해 자동 태깅 결과와 코퍼스의 정답이 다른 부분을 찾는다.
+    Args:
+        result_morphs:  자동 태깅 결과
+        corpus_morphs:  코퍼스의 정답
+        raw_sent:  문장 원문
+        alignments:  음절별 정렬 정보. 형태소별 (시작, 끝) 음절 포지션 pair
+    Returns:
+        (자동, 정답, 원문) triple의 리스트
+    """
+    def _get_origin(first_morph: int, last_morph: int) -> str:
+        """
+        첫번째 형태소와 마지막 형태소 인덱스를 바탕으로 원문을 얻는다.
+        Args:
+            first_morph:  첫번째 형태소 인덱스
+            last_morph:  마지막 형태소 인덱스
+        Returns:
+            원문
+        """
+        if first_morph < 0:
+            first_morph = 0
+        if last_morph >= len(alignments):
+            last_morph = len(alignments)
+        return raw_sent[alignments[first_morph][0]:alignments[last_morph-1][1]]
+
+    def _expand_begin(result_begin: int) -> int:
+        """
+        왼쪽 경계가 음소로 나눠진 경우 음절 경계가 나올 때까지 왼쪽으로 확장한다.
+        Args:
+            result_begin:  왼쪽 경계
+        Returns:
+            확장된 왼쪽 경계
+        """
+        if result_begin <= 0:
+            return result_begin
+        if alignments[result_begin - 1][1] > alignments[result_begin][0]:
+            return _expand_begin(result_begin - 1)
+        return result_begin
+
+    def _expand_end(result_end: int) -> int:
+        """
+        오른쪽 경계가 음소로 나눠진 경우 음절 경계가 나올 때까지 오른쪽으로 확장한다.
+        Args:
+            result_end:  오른쪽 경계.
+                         [result_begin, result_end)으로 오른쪽 경계는 open이므로,
+                         실제로 result_end-1이 마지막 형태소의 인덱스이다.
+        Returns:
+            확장된 오른쪽 경계
+        """
+        if result_end >= len(result_morphs):
+            return result_end
+        if alignments[result_end - 1][1] > alignments[result_end][0]:
+            return _expand_end(result_end + 1)
+        return result_end
+
+    def _expand_spc(diff_sgmts: List[Tuple[str, str, str]], result_begin: int, result_end: int,
+                    corpus_begin: int, corpus_end: int):
+        """
+        좌, 우 경계에 공백이 있을 경우 한 형태소 더 확장한다.
+        Args:
+            diff_sgmts:  diff를 추가할 리스트
+            result_begin:  오분석의 왼쪽 경계
+            result_end:  오분석의 오른쪽 경계
+            corpus_begin:  정분석(코퍼스)의 왼쪽 경계
+            corpus_begin:  정분석(코퍼스)의 오른쪽 경계
+        """
+        if (result_morphs[result_begin] != WORD_DELIM_STR and    # pylint: disable=consider-using-in
+                result_morphs[result_end-1] != WORD_DELIM_STR):
+            return
+        _result_begin = _expand_begin(result_begin-1) \
+            if result_morphs[result_begin] == WORD_DELIM_STR else result_begin
+        result_end_ = _expand_end(result_end+1) \
+            if result_morphs[result_end-1] == WORD_DELIM_STR else result_end
+        assert result_begin != _result_begin or result_end != result_end_
+        _left = result_morphs[_result_begin:result_end_]
+        _corpus_begin = corpus_begin - (result_begin - _result_begin)
+        corpus_end_ = corpus_end + (result_end_ - result_end)
+        right_ = corpus_morphs[_corpus_begin:corpus_end_]
+        origin = _get_origin(_result_begin, result_end_)
+        diff_sgmts.append((origin, _left, right_))
+
+    diff_sgmts = []
+    matches = difflib.SequenceMatcher(None, result_morphs, corpus_morphs).get_matching_blocks()
+    result_begin = 0
+    corpus_begin = 0
+    for match in matches:
+        result_end = match.a
+        left = []
+        if result_end > result_begin:
+            # _result_begin은 result_begin에서 왼쪽으로 같은 음절인 경우에 대해 확장한 시작 위치입니다.
+            # 가령: '된다'에 대해 오분석: '되+ㄴ+다', 정분석: '되+ㄴ다'일 경우 틀린 영역은 'ㄴ+다' vs 'ㄴ다' 이지만,
+            # 'ㄴ'이 포함된 음절이 '된'이므로 원문을 추출하면 '된'이 되어벼러 원문과 형태소간 불일치가 일어납니다.
+            # 이것을 보정해 주기 위해 음소단위 형태소인 경우 왼쪽으로 음절 경계까지 확장하는 것입니다.
+            # result_end_도 마찬가지로 result_end를 오른쪽으로 확장한 것입니다.
+            _result_begin = _expand_begin(result_begin)
+            result_end_ = _expand_end(result_end)
+            left = result_morphs[_result_begin:result_end_]
+        else:
+            _result_begin = result_begin
+            result_end_ = result_end
+
+        corpus_end = match.b
+        # _result_begin, result_end_가 각각 n, m개 확장된 경우,
+        # 동일한 갯수로 corpus_begin, corpus_end을 좌, 우로 n, m개 확장합니다.
+        # 틀린 부분의 좌, 우 형태소가 오분석과 정분석이 같다는 가정 하에 alignments 정보를 활용하지 않습니다.
+        _corpus_begin = corpus_begin - (result_begin - _result_begin)
+        corpus_end_ = corpus_end + (result_end_ - result_end)
+        right = corpus_morphs[_corpus_begin:corpus_end_]
+
+        if not left and not right:
+            result_begin = result_end + match.size
+            corpus_begin = corpus_end + match.size
+            continue
+        assert left != right, '{} == {}'.format(left, right)
+
+        if (len(left) >= 2 and right) or (left and len(right) >= 2):
+            # 결과가 다른 영역만
+            origin = _get_origin(_result_begin, result_end_)
+            diff_sgmts.append((origin, left, right))
+
+        # _left는 left에서 오분석인 영역에 왼쪽으로 하나의 형태소를 더해서 만든 것입니다.
+        # left_는 left에서 오른쪽으로 하나의 형태소, _left_는 양쪽에 하나씩 더해서 만든 것입니다.
+        # _right, right_, _right_도 left와 마찬가지입니다.
+
+        # 왼쪽 형태소를 추가
+        # 추가한 형태소가 음소로 나눠진 경우에도 역시 왼쪽으로 음절 경계까지 확장
+        __result_begin = _expand_begin(_result_begin-1)
+        _left = result_morphs[__result_begin:_result_begin] + left
+        __corpus_begin = _corpus_begin - (_result_begin - __result_begin)
+        _right = corpus_morphs[__corpus_begin:_corpus_begin] + right
+        left_first_lex = _left[0].rsplit('/', 1)[0]
+        right_first_lex = _right[0].rsplit('/', 1)[0]
+        if left_first_lex == right_first_lex and len(_left) >= 2 and len(_right) >= 2:
+            origin = _get_origin(__result_begin, result_end_)
+            diff_sgmts.append((origin, _left, _right))
+            _expand_spc(diff_sgmts, __result_begin, result_end_, __corpus_begin, corpus_end_)
+
+        # 오른쪽 형태소를 추가
+        # 추가한 형태소가 음소로 나눠진 경우에도 역시 오른쪽으로 음절 경계까지 확장
+        result_end__ = _expand_end(result_end_+1)
+        left_ = left + result_morphs[result_end_:result_end__]
+        corpus_end__ = corpus_end_ + (result_end__ - result_end_)
+        right_ = right + corpus_morphs[corpus_end_:corpus_end__]
+        left_last_lex = left_[-1].rsplit('/', 1)[0]
+        right_last_lex = right_[-1].rsplit('/', 1)[0]
+        if left_last_lex == right_last_lex and len(left_) >= 2 and len(right_) >= 2:
+            origin = _get_origin(_result_begin, result_end__)
+            diff_sgmts.append((origin, left_, right_))
+            _expand_spc(diff_sgmts, _result_begin, result_end__, _corpus_begin, corpus_end__)
+
+        # 양쪽에 형태소를 각각 추가
+        _left_ = _left + result_morphs[result_end_:result_end__]
+        _right_ = _right + corpus_morphs[corpus_end_:corpus_end__]
+        origin = _get_origin(__result_begin, result_end__)
+        diff_sgmts.append((origin, _left_, _right_))
+        _expand_spc(diff_sgmts, __result_begin, result_end__, __corpus_begin, corpus_end__)
+
+        result_begin = result_end + match.size
+        corpus_begin = corpus_end + match.size
+    return diff_sgmts
+
+
+def _count_error(args: Namespace, doc_path: str) -> Counter:
+    """
+    count from courpus and make ambiguous dictionary
+    Args:
+        args:  program arguments
+        doc_path:  document path
+    Returns:
+        오분석 패치 후보(원문, 오분석, 정분석 triple)의 카운터
+    """
+    global _KHAIII    # pylint: disable=global-statement
+    if not _KHAIII:
+        _KHAIII = KhaiiiApi(args.lib_path, args.rsc_dir)
+
+    cnt = Counter()
+    logging.info(doc_path)
+    for sent in sejong_corpus.sents(open(doc_path, 'r', encoding='UTF-8')):
+        raw_sent = sent.raw_str()
+        result_morphs = [SENT_DELIM_STR, ]
+        corpus_morphs = [SENT_DELIM_STR, ]
+        alignments = [(0, 0), ]
+        for result_word, corpus_word in zip(_KHAIII.analyze(raw_sent, ''), sent.words):
+            assert result_word.lex == corpus_word.raw, \
+                   '{}: "{}" != "{}"'.format(os.path.basename(doc_path), result_word, corpus_word)
+            if len(result_morphs) > 1:
+                result_morphs.append(WORD_DELIM_STR)
+                corpus_morphs.append(WORD_DELIM_STR)
+                alignments.append((alignments[-1][1], alignments[-1][1]+1))
+            result_morphs.extend(['{}/{}'.format(m.lex, m.tag) for m in result_word.morphs])
+            corpus_morphs.extend([str(m) for m in corpus_word.morphs])
+            alignments.extend([(m.begin, m.begin + m.length) for m in result_word.morphs])
+        result_morphs.append(SENT_DELIM_STR)
+        corpus_morphs.append(SENT_DELIM_STR)
+        alignments.append((alignments[-1][1], alignments[-1][1]))
+        if result_morphs != corpus_morphs:
+            diff_sgmts = _get_diff_sgmts(result_morphs, corpus_morphs, raw_sent, alignments)
+            for origin, left, right in diff_sgmts:
+                cnt[origin, ' + '.join(left), ' + '.join(right)] += 1
+    return cnt
+
+
+def _doc_iter(corpus_dir: str) -> Iterator[str]:
+    """
+    문종 코퍼스에서 문서의 경로를 리턴하는 generator
+    Args:
+        args:  program arguments
+    Yields:
+        document path
+    """
+    for name in sorted(os.listdir(corpus_dir)):
+        if not name.endswith('.txt'):
+            continue
+        yield '{}/{}'.format(corpus_dir, name)
+
+
+def _filter_cnt(args: Namespace, cnt: Counter) -> Counter:
+    """
+    규칙에 의해 필터링하여 빈도로 정렬한다.
+    Args:
+        args:  program arguments
+        cnt:  (원문, left, right) 별로 측정한 빈도 사전
+    Returns:
+        (빈도, (원문, left, right)) pair의 리스트
+    """
+    cnts = []
+    for (origin, left, right), freq in cnt.items():
+        if freq < args.min_freq:
+            continue
+        if len(origin.strip()) < args.min_len:
+            continue
+        cnts.append((freq, (origin, left, right)))
+    cnts.sort(key=lambda x: x[0], reverse=True)
+    return cnts
+
+
+def _print_cnt(args: Namespace, cnts: Counter):
+    """
+    최종 패치 후포를 출력한다.
+    Args:
+        args:  program arguments
+        cnts:  counts
+    """
+    for freq, (origin, left, right) in cnts:
+        freq_tab = ''
+        if args.with_freq:
+            freq_tab = '{}\t'.format(freq)
+        print('{}{}\t{}\t{}'.format(freq_tab, origin, left, right))
+
+
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    cnt = Counter()
+    if args.num_proc > 1:
+        pool = multiprocessing.Pool(args.num_proc)
+        results = ((doc, pool.apply_async(_count_error, (args, doc)))
+                   for doc in _doc_iter(args.corpus_dir))
+        for num, (doc, result) in enumerate(results, start=1):
+            if num % 100 == 0:
+                logging.info('%d-th document..', num)
+            try:
+                cnt.update(result.get(timeout=1000))
+            except multiprocessing.context.TimeoutError:
+                logging.error('timeout[%d]: %s', num, doc)
+    else:
+        for num, doc in enumerate(_doc_iter(args.corpus_dir), start=1):
+            if num % 10 == 0:
+                logging.info('%d-th document..', num)
+            cnt.update(_count_error(args, doc))
+
+    cnts = _filter_cnt(args, cnt)
+    _print_cnt(args, cnts)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='오분석 패치 후보를 추출하는 스크립트')
+    parser.add_argument('-c', '--corpus-dir', help='corpus dir', metavar='DIR', required=True)
+    parser.add_argument('--lib-path', help='khaiii shared library path', metavar='FILE', default='')
+    parser.add_argument('--rsc-dir', help='resource dir', metavar='DIR', default='')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--with-freq', help='print with frequency', action='store_true')
+    parser.add_argument('--min-freq', help='minimum frequency <default: 10>', metavar='FREQ',
+                        type=int, default=10)
+    parser.add_argument('--min-len', help='minimum original text length <default: 4>',
+                        metavar='LEN', type=int, default=4)
+    parser.add_argument('--num-proc', help='number of processes <default: 1>', metavar='NUM',
+                        type=int, default=1)
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/train/bin/pickle_model.py b/train/bin/pickle_model.py
index 06f6620..11bcddc 100755
--- a/train/bin/pickle_model.py
+++ b/train/bin/pickle_model.py
@@ -16,7 +16,6 @@
 from array import array
 import json
 import logging
-import os
 import pickle
 import re
 
@@ -174,25 +173,6 @@ def _load_config(path: str) -> Namespace:
     return cfg
 
 
-def _load_resource(cfg: Namespace, rsc_src: str) -> Resource:
-    """
-    load resources
-    Args:
-        cfg:  config
-        rsc_src:  resource source dir
-    Returns:
-        Resource object
-    """
-    cwd = os.path.realpath(os.getcwd())
-    train_dir = os.path.realpath('{}/../../train'.format(rsc_src))
-    if cwd != train_dir:
-        os.chdir(train_dir)
-    rsc = Resource(cfg)
-    if cwd != train_dir:
-        os.chdir(cwd)
-    return rsc
-
-
 def run(args: Namespace):
     """
     run function which is the start point of program
@@ -200,7 +180,8 @@ def run(args: Namespace):
         args:  program arguments
     """
     cfg = _load_config('{}/config.json'.format(args.in_dir))
-    rsc = _load_resource(cfg, args.rsc_src)
+    setattr(cfg, 'rsc_src', args.rsc_src)
+    rsc = Resource(cfg)
     state_dict = torch.load('{}/model.state'.format(args.in_dir),
                             map_location=lambda storage, loc: storage)
     _validate_state_dict(cfg, rsc, state_dict)

From 4d6fc457ca588e9f8b5cce39d8221056805d6071 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Sun, 27 Jan 2019 22:34:05 +0900
Subject: [PATCH 32/41] =?UTF-8?q?=EC=98=A4=EB=B6=84=EC=84=9D=20=ED=8C=A8?=
 =?UTF-8?q?=EC=B9=98=20=ED=9B=84=EB=B3=B4=EB=A5=BC=20=EA=B2=80=EC=A6=9D?=
 =?UTF-8?q?=ED=95=98=EB=8A=94=20=EC=8A=A4=ED=81=AC=EB=A6=BD=ED=8A=B8=20?=
 =?UTF-8?q?=EC=B6=94=EA=B0=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 rsc/bin/compile_errpatch.py                   | 106 +------
 rsc/bin/compile_preanal.py                    |  61 +---
 rsc/bin/compile_restore.py                    |  51 +---
 src/main/python/khaiii/resource/char_align.py | 128 ++++++++-
 src/main/python/khaiii/resource/morphs.py     |  27 ++
 src/main/python/khaiii/resource/resource.py   |  53 +++-
 train/bin/extract_errpatch.py                 |   7 +-
 train/bin/hd_validate_errpatch.bash           | 194 +++++++++++++
 train/bin/train.py                            |   4 +-
 train/bin/validate_errpatch.py                | 270 ++++++++++++++++++
 10 files changed, 688 insertions(+), 213 deletions(-)
 create mode 100755 train/bin/hd_validate_errpatch.bash
 create mode 100755 train/bin/validate_errpatch.py

diff --git a/rsc/bin/compile_errpatch.py b/rsc/bin/compile_errpatch.py
index caf7d8a..7e49fd9 100755
--- a/rsc/bin/compile_errpatch.py
+++ b/rsc/bin/compile_errpatch.py
@@ -13,23 +13,19 @@
 # imports #
 ###########
 from argparse import ArgumentParser, Namespace
-from collections import defaultdict
 import glob
-import itertools
 import logging
 import os
 import struct
 import sys
 from typing import Dict, List, Tuple
 
-from khaiii.munjong import sejong_corpus
-from khaiii.resource.char_align import Aligner, AlignError
-from khaiii.resource.morphs import Morph, ParseError
-from khaiii.resource.morphs import WORD_DELIM_STR, SENT_DELIM_STR, WORD_DELIM_NUM, SENT_DELIM_NUM
+from khaiii.resource.char_align import Aligner, align_patch
+from khaiii.resource.resource import load_restore_dic, load_vocab_out
+from khaiii.resource.morphs import Morph, ParseError, mix_char_tag
 from khaiii.resource.trie import Trie
 
-from compile_preanal import align_to_tag, print_errors
-from compile_restore import load_restore_dic, load_vocab_out
+from compile_preanal import print_errors
 
 
 #########
@@ -104,100 +100,6 @@ def _parse(self):
 #############
 # functions #
 #############
-def _split_list(lst: List[str], delim: str) -> List[List[str]]:
-    """
-    리스트를 delimiter로 split하는 함수
-
-    >>> _split_list(['가/JKS', '_', '너/NP'], '_')
-    [['가/JKS'], ['너/NP']]
-
-    Args:
-        lst:  리스트
-        delim:  delimiter
-    Returns:
-        list of sublists
-    """
-    sublists = []
-    while lst:
-        prefix = [x for x in itertools.takewhile(lambda x: x != delim, lst)]
-        sublists.append(prefix)
-        lst = lst[len(prefix):]
-        delims = [x for x in itertools.takewhile(lambda x: x == delim, lst)]
-        lst = lst[len(delims):]
-    return sublists
-
-
-def align_patch(rsc_src: Tuple[Aligner, Dict, Dict[str, int]], raw: str, morph_str: str) \
-        -> List[int]:
-    """
-    패치의 원문과 분석 결과를 음절단위 매핑(정렬)을 수행한다.
-    Args:
-        rsc_src:  (Aligner, restore dic, vocab out) resource triple
-        raw:  원문
-        morph_str:  형태소 분석 결과 (패치 기술 형식)
-    Returns:
-        정렬에 기반한 출력 태그 번호
-    """
-    aligner, restore_dic, vocab_out = rsc_src
-    raw_words = raw.strip().split()
-    morphs = morph_str.split(' + ')
-    morphs_strip = morphs
-    if morphs[0] in [WORD_DELIM_STR, SENT_DELIM_STR]:
-        morphs_strip = morphs_strip[1:]
-    if morphs[-1] in [WORD_DELIM_STR, SENT_DELIM_STR]:
-        morphs_strip = morphs_strip[:-1]
-    morph_words = _split_list(morphs_strip, WORD_DELIM_STR)
-    tag_nums = []
-    restore_new = defaultdict(dict)
-    vocab_new = defaultdict(list)
-    for raw_word, morph_word in zip(raw_words, morph_words):
-        word = sejong_corpus.Word.parse('\t'.join(['', raw_word, ' + '.join(morph_word)]), '', 0)
-        try:
-            word_align = aligner.align(word)
-            _, word_tag_nums = \
-                align_to_tag(raw_word, word_align, (restore_dic, restore_new),
-                             (vocab_out, vocab_new))
-            if restore_new or vocab_new:
-                logging.debug('needs dic update: %s', word)
-                return []
-        except AlignError as algn_err:
-            logging.debug('alignment error: %s', word)
-            logging.debug(str(algn_err))
-            return []
-        if tag_nums:
-            tag_nums.append(WORD_DELIM_NUM)
-        tag_nums.extend(word_tag_nums)
-    if morphs[0] in [WORD_DELIM_STR, SENT_DELIM_STR]:
-        tag_nums.insert(0, WORD_DELIM_NUM if morphs[0] == WORD_DELIM_STR else SENT_DELIM_NUM)
-    if morphs[-1] in [WORD_DELIM_STR, SENT_DELIM_STR]:
-        tag_nums.append(WORD_DELIM_NUM if morphs[-1] == WORD_DELIM_STR else SENT_DELIM_NUM)
-    return tag_nums
-
-
-def mix_char_tag(chars: str, tags: List[int]) -> List[int]:
-    """
-    음절과 출력 태그를 비트 연산으로 합쳐서 하나의 (32비트) 숫자로 표현한다.
-    Args:
-        chars:  음절 (유니코드) 리스트 (문자열)
-        tags:  출력 태그 번호의 리스트
-    Returns:
-        합쳐진 숫자의 리스트
-    """
-    char_nums = [ord(c) for c in chars]
-    if tags[0] == SENT_DELIM_NUM:
-        char_nums.insert(0, SENT_DELIM_NUM)
-    if tags[-1] == SENT_DELIM_NUM:
-        char_nums.append(SENT_DELIM_NUM)
-    for idx, char_num in enumerate(char_nums):
-        if char_num == ord(' '):
-            char_nums[idx] = WORD_DELIM_NUM
-            continue
-        elif tags[idx] == SENT_DELIM_NUM:
-            continue
-        char_nums[idx] = char_num << 12 | tags[idx]
-    return char_nums
-
-
 def _load_entries(args: Namespace) -> List[Entry]:
     """
     패치 엔트리를 파일로부터 로드한다.
diff --git a/rsc/bin/compile_preanal.py b/rsc/bin/compile_preanal.py
index 6874d2c..649e94c 100755
--- a/rsc/bin/compile_preanal.py
+++ b/rsc/bin/compile_preanal.py
@@ -19,14 +19,16 @@
 import os
 import struct
 import sys
-from typing import Dict, List, Tuple
+from typing import Dict, List
 
 from khaiii.munjong import sejong_corpus
-from khaiii.resource.char_align import Aligner, AlignError, MrpChr
+from khaiii.resource.char_align import Aligner, AlignError, align_to_tag
 from khaiii.resource.morphs import Morph, ParseError
+from khaiii.resource.resource import load_restore_dic, load_vocab_out
 from khaiii.resource.trie import Trie
 
-from compile_restore import load_restore_dic, load_vocab_out, append_new_entries
+from compile_restore import append_new_entries
+
 
 
 #########
@@ -175,59 +177,6 @@ def _set_align(aligner: Aligner, Word: type, entries: List[Entry]):    # pylint:
     print_errors(bad_entries)
 
 
-def align_to_tag(raw_word: str, alignment: List[List[str]], restore: Tuple[dict, dict],
-                 vocab: Tuple[Dict[str, int], Dict[str, int]]) \
-        -> Tuple[List[str], List[int]]:
-    """
-    어절의 원문과 정렬 정보를 활용해 음절과 매핑된 태그를 생성한다.
-    Args:
-        raw_word:  어절 원문
-        alignment:  정렬 정보
-        restore:  (원형복원 사전, 원형복원 사전에 추가할 엔트리) pair
-        vocab:  (출력 태그 사전, 출력 태그 사전에 추가할 새로운 태그) pair
-    Returns:
-        음절별 출력 태그
-        음절별 출력 태그의 번호
-    """
-    assert len(raw_word) == len(alignment)
-    restore_dic, restore_new = restore
-    vocab_out, vocab_new = vocab
-    tag_outs = []
-    tag_nums = []
-    for char, mrp_chrs in zip(raw_word, alignment):
-        if len(mrp_chrs) == 1 and mrp_chrs[0].char == char:
-            tag_outs.append(mrp_chrs[0].tag)
-        else:
-            tag_str = ':'.join([m.tag for m in mrp_chrs])
-            mrp_chr_str_key = MrpChr.to_str(mrp_chrs)
-            found = -1
-            max_num = -1
-            for num, mrp_chr_str_val in restore_dic[char, tag_str].items():
-                if num > max_num:
-                    max_num = num
-                if mrp_chr_str_key == mrp_chr_str_val:
-                    found = num
-                    break
-            if found >= 0:
-                tag_outs.append('{}:{}'.format(tag_str, found))
-            else:
-                new_num = max_num + 1
-                restore_dic[char, tag_str][new_num] = mrp_chr_str_key
-                restore_new[char, tag_str][new_num] = mrp_chr_str_key
-                tag_outs.append('{}:{}'.format(tag_str, new_num))
-        tag = tag_outs[-1]
-        if tag in vocab_out:
-            tag_nums.append(vocab_out[tag])
-        elif tag in vocab_new:
-            tag_nums.append(vocab_new[tag])
-        else:
-            new_tag_num = len(vocab_out) + len(vocab_new) + 1
-            logging.debug('new output tag: [%d] %s', new_tag_num, tag)
-            vocab_new[tag] = new_tag_num
-            tag_nums.append(new_tag_num)
-    return tag_outs, tag_nums
-
-
 def _set_tag_out(restore_dic: dict, restore_new: dict, vocab_out: Dict[str, int],
                  vocab_new: Dict[str, int], entries: List[Entry]):
     """
diff --git a/rsc/bin/compile_restore.py b/rsc/bin/compile_restore.py
index 132df3a..c5a0172 100755
--- a/rsc/bin/compile_restore.py
+++ b/rsc/bin/compile_restore.py
@@ -13,14 +13,14 @@
 # imports #
 ###########
 from argparse import ArgumentParser, Namespace
-from collections import defaultdict
 import logging
 import os
 import struct
 import sys
-from typing import Dict, Tuple
+from typing import Dict
 
 from khaiii.resource.morphs import TAG_SET
+from khaiii.resource.resource import load_restore_dic, load_vocab_out
 
 
 #############
@@ -32,53 +32,6 @@
 #############
 # functions #
 #############
-def load_restore_dic(file_path: str) -> Dict[Tuple[str, str], Dict[int, str]]:
-    """
-    원형복원 사전을 로드한다.
-    Args:
-        file_path:  파일 경로
-    Returns:
-        사전
-    """
-    file_name = os.path.basename(file_path)
-    restore_dic = defaultdict(dict)
-    for line_num, line in enumerate(open(file_path, 'r', encoding='UTF-8'), start=1):
-        line = line.rstrip()
-        if not line or line[0] == '#':
-            continue
-        char_tag_num, mrp_chr_str = line.split('\t')
-        char, tag_num = char_tag_num.rsplit('/', 1)
-        tag, num = tag_num.rsplit(':', 1)
-        num = int(num)
-        if (char, tag) in restore_dic:
-            num_mrp_chrs_dic = restore_dic[char, tag]
-            if num in num_mrp_chrs_dic:
-                logging.error('%s:%d: duplicated with %s: %s', file_name, line_num,
-                              num_mrp_chrs_dic[num], line)
-                return {}
-        restore_dic[char, tag][num] = mrp_chr_str
-    return restore_dic
-
-
-def load_vocab_out(rsc_src: str) -> Dict[str, int]:
-    """
-    출력 태그 vocabulary를 로드한다.
-    Args:
-        rsc_src:  리소스 디렉토리
-    Returns:
-        출력 태그 vocabulary
-    """
-    file_path = '{}/vocab.out'.format(rsc_src)
-    vocab_out = [line.strip() for line in open(file_path, 'r', encoding='UTF-8')
-                 if line.strip()]
-    vocab_out_more = []
-    file_path = '{}/vocab.out.more'.format(rsc_src)
-    if os.path.exists(file_path):
-        vocab_out_more = [line.strip() for line in open(file_path, 'r', encoding='UTF-8')
-                          if line.strip()]
-    return {tag: idx for idx, tag in enumerate(vocab_out + vocab_out_more, start=1)}
-
-
 def append_new_entries(rsc_src: str, restore_new: dict, vocab_new: Dict[str, int]):
     """
     기분석 사전 빌드 중에 새로 추가가 필요한 사전 엔트리를 해당 사전에 추가한다.
diff --git a/src/main/python/khaiii/resource/char_align.py b/src/main/python/khaiii/resource/char_align.py
index 1682606..a2052d8 100644
--- a/src/main/python/khaiii/resource/char_align.py
+++ b/src/main/python/khaiii/resource/char_align.py
@@ -12,13 +12,15 @@
 # imports #
 ###########
 from collections import Counter, defaultdict
+import itertools
 import logging
 import os
-from typing import List, Tuple
+from typing import Dict, List, Tuple
 
 from khaiii.munjong.sejong_corpus import Word
 from khaiii.resource import jaso
 from khaiii.resource.morphs import Morph
+from khaiii.resource.morphs import WORD_DELIM_NUM, WORD_DELIM_STR, SENT_DELIM_NUM, SENT_DELIM_STR
 
 
 #########
@@ -594,3 +596,127 @@ def add_msg(self, msg: str):
             msg:  에러 메세지
         """
         self._msgs.append(msg)
+
+
+#############
+# functions #
+#############
+def align_to_tag(raw_word: str, alignment: List[List[str]], restore: Tuple[dict, dict],
+                 vocab: Tuple[Dict[str, int], Dict[str, int]]) -> Tuple[List[str], List[int]]:
+    """
+    어절의 원문과 정렬 정보를 활용해 음절과 매핑된 태그를 생성한다.
+    Args:
+        raw_word:  어절 원문
+        alignment:  정렬 정보
+        restore:  (원형복원 사전, 원형복원 사전에 추가할 엔트리) pair
+        vocab:  (출력 태그 사전, 출력 태그 사전에 추가할 새로운 태그) pair
+    Returns:
+        음절별 출력 태그
+        음절별 출력 태그의 번호
+    """
+    assert len(raw_word) == len(alignment)
+    restore_dic, restore_new = restore
+    vocab_out, vocab_new = vocab
+    tag_outs = []
+    tag_nums = []
+    for char, mrp_chrs in zip(raw_word, alignment):
+        if len(mrp_chrs) == 1 and mrp_chrs[0].char == char:
+            tag_outs.append(mrp_chrs[0].tag)
+        else:
+            tag_str = ':'.join([m.tag for m in mrp_chrs])
+            mrp_chr_str_key = MrpChr.to_str(mrp_chrs)
+            found = -1
+            max_num = -1
+            for num, mrp_chr_str_val in restore_dic[char, tag_str].items():
+                if num > max_num:
+                    max_num = num
+                if mrp_chr_str_key == mrp_chr_str_val:
+                    found = num
+                    break
+            if found >= 0:
+                tag_outs.append('{}:{}'.format(tag_str, found))
+            else:
+                new_num = max_num + 1
+                restore_dic[char, tag_str][new_num] = mrp_chr_str_key
+                restore_new[char, tag_str][new_num] = mrp_chr_str_key
+                tag_outs.append('{}:{}'.format(tag_str, new_num))
+        tag = tag_outs[-1]
+        if tag in vocab_out:
+            tag_nums.append(vocab_out[tag])
+        elif tag in vocab_new:
+            tag_nums.append(vocab_new[tag])
+        else:
+            new_tag_num = len(vocab_out) + len(vocab_new) + 1
+            logging.debug('new output tag: [%d] %s', new_tag_num, tag)
+            vocab_new[tag] = new_tag_num
+            tag_nums.append(new_tag_num)
+    return tag_outs, tag_nums
+
+
+def _split_list(lst: List[str], delim: str) -> List[List[str]]:
+    """
+    리스트를 delimiter로 split하는 함수
+
+    >>> _split_list(['가/JKS', '_', '너/NP'], '_')
+    [['가/JKS'], ['너/NP']]
+
+    Args:
+        lst:  리스트
+        delim:  delimiter
+    Returns:
+        list of sublists
+    """
+    sublists = []
+    while lst:
+        prefix = [x for x in itertools.takewhile(lambda x: x != delim, lst)]
+        sublists.append(prefix)
+        lst = lst[len(prefix):]
+        delims = [x for x in itertools.takewhile(lambda x: x == delim, lst)]
+        lst = lst[len(delims):]
+    return sublists
+
+
+def align_patch(rsc_src: Tuple[Aligner, Dict, Dict[str, int]], raw: str, morph_str: str) \
+        -> List[int]:
+    """
+    패치의 원문과 분석 결과를 음절단위 매핑(정렬)을 수행한다.
+    Args:
+        rsc_src:  (Aligner, restore dic, vocab out) resource triple
+        raw:  원문
+        morph_str:  형태소 분석 결과 (패치 기술 형식)
+    Returns:
+        정렬에 기반한 출력 태그 번호
+    """
+    aligner, restore_dic, vocab_out = rsc_src
+    raw_words = raw.strip().split()
+    morphs = morph_str.split(' + ')
+    morphs_strip = morphs
+    if morphs[0] in [WORD_DELIM_STR, SENT_DELIM_STR]:
+        morphs_strip = morphs_strip[1:]
+    if morphs[-1] in [WORD_DELIM_STR, SENT_DELIM_STR]:
+        morphs_strip = morphs_strip[:-1]
+    morph_words = _split_list(morphs_strip, WORD_DELIM_STR)
+    tag_nums = []
+    restore_new = defaultdict(dict)
+    vocab_new = defaultdict(list)
+    for raw_word, morph_word in zip(raw_words, morph_words):
+        word = Word.parse('\t'.join(['', raw_word, ' + '.join(morph_word)]), '', 0)
+        try:
+            word_align = aligner.align(word)
+            _, word_tag_nums = align_to_tag(raw_word, word_align, (restore_dic, restore_new),
+                                            (vocab_out, vocab_new))
+            if restore_new or vocab_new:
+                logging.debug('needs dic update: %s', word)
+                return []
+        except AlignError as algn_err:
+            logging.debug('alignment error: %s', word)
+            logging.debug(str(algn_err))
+            return []
+        if tag_nums:
+            tag_nums.append(WORD_DELIM_NUM)
+        tag_nums.extend(word_tag_nums)
+    if morphs[0] in [WORD_DELIM_STR, SENT_DELIM_STR]:
+        tag_nums.insert(0, WORD_DELIM_NUM if morphs[0] == WORD_DELIM_STR else SENT_DELIM_NUM)
+    if morphs[-1] in [WORD_DELIM_STR, SENT_DELIM_STR]:
+        tag_nums.append(WORD_DELIM_NUM if morphs[-1] == WORD_DELIM_STR else SENT_DELIM_NUM)
+    return tag_nums
diff --git a/src/main/python/khaiii/resource/morphs.py b/src/main/python/khaiii/resource/morphs.py
index be6d232..80c85c9 100644
--- a/src/main/python/khaiii/resource/morphs.py
+++ b/src/main/python/khaiii/resource/morphs.py
@@ -127,3 +127,30 @@ def _parse_one(cls, morph_str: str) -> 'Morph':
         if tag not in TAG_SET:
             raise ParseError('invalid pos tag: {}'.format(tag))
         return Morph(lex, tag)
+
+
+#############
+# functions #
+#############
+def mix_char_tag(chars: str, tags: List[int]) -> List[int]:
+    """
+    음절과 출력 태그를 비트 연산으로 합쳐서 하나의 (32비트) 숫자로 표현한다.
+    Args:
+        chars:  음절 (유니코드) 리스트 (문자열)
+        tags:  출력 태그 번호의 리스트
+    Returns:
+        합쳐진 숫자의 리스트
+    """
+    char_nums = [ord(c) for c in chars]
+    if tags[0] == SENT_DELIM_NUM:
+        char_nums.insert(0, SENT_DELIM_NUM)
+    if tags[-1] == SENT_DELIM_NUM:
+        char_nums.append(SENT_DELIM_NUM)
+    for idx, char_num in enumerate(char_nums):
+        if char_num == ord(' '):
+            char_nums[idx] = WORD_DELIM_NUM
+            continue
+        elif tags[idx] == SENT_DELIM_NUM:
+            continue
+        char_nums[idx] = char_num << 12 | tags[idx]
+    return char_nums
diff --git a/src/main/python/khaiii/resource/resource.py b/src/main/python/khaiii/resource/resource.py
index 3514e81..110ee70 100644
--- a/src/main/python/khaiii/resource/resource.py
+++ b/src/main/python/khaiii/resource/resource.py
@@ -12,9 +12,10 @@
 # imports #
 ###########
 from argparse import Namespace
+from collections import defaultdict
 import logging
 import os
-from typing import Dict
+from typing import Dict, Tuple
 
 from khaiii.resource.vocabulary import Vocabulary
 
@@ -68,3 +69,53 @@ def _load_restore_dic(cls, path: str) -> Dict[str, str]:
             dic[key] = val
         logging.info('%s: %d entries', os.path.basename(path), len(dic))
         return dic
+
+
+#############
+# functions #
+#############
+def load_restore_dic(file_path: str) -> Dict[Tuple[str, str], Dict[int, str]]:
+    """
+    원형복원 사전을 로드한다.
+    Args:
+        file_path:  파일 경로
+    Returns:
+        사전
+    """
+    file_name = os.path.basename(file_path)
+    restore_dic = defaultdict(dict)
+    for line_num, line in enumerate(open(file_path, 'r', encoding='UTF-8'), start=1):
+        line = line.rstrip()
+        if not line or line[0] == '#':
+            continue
+        char_tag_num, mrp_chr_str = line.split('\t')
+        char, tag_num = char_tag_num.rsplit('/', 1)
+        tag, num = tag_num.rsplit(':', 1)
+        num = int(num)
+        if (char, tag) in restore_dic:
+            num_mrp_chrs_dic = restore_dic[char, tag]
+            if num in num_mrp_chrs_dic:
+                logging.error('%s:%d: duplicated with %s: %s', file_name, line_num,
+                              num_mrp_chrs_dic[num], line)
+                return {}
+        restore_dic[char, tag][num] = mrp_chr_str
+    return restore_dic
+
+
+def load_vocab_out(rsc_src: str) -> Dict[str, int]:
+    """
+    출력 태그 vocabulary를 로드한다.
+    Args:
+        rsc_src:  리소스 디렉토리
+    Returns:
+        출력 태그 vocabulary
+    """
+    file_path = '{}/vocab.out'.format(rsc_src)
+    vocab_out = [line.strip() for line in open(file_path, 'r', encoding='UTF-8')
+                 if line.strip()]
+    vocab_out_more = []
+    file_path = '{}/vocab.out.more'.format(rsc_src)
+    if os.path.exists(file_path):
+        vocab_out_more = [line.strip() for line in open(file_path, 'r', encoding='UTF-8')
+                          if line.strip()]
+    return {tag: idx for idx, tag in enumerate(vocab_out + vocab_out_more, start=1)}
diff --git a/train/bin/extract_errpatch.py b/train/bin/extract_errpatch.py
index ff9dd8c..7153eee 100755
--- a/train/bin/extract_errpatch.py
+++ b/train/bin/extract_errpatch.py
@@ -325,8 +325,11 @@ def main():
     """
     parser = ArgumentParser(description='오분석 패치 후보를 추출하는 스크립트')
     parser.add_argument('-c', '--corpus-dir', help='corpus dir', metavar='DIR', required=True)
-    parser.add_argument('--lib-path', help='khaiii shared library path', metavar='FILE', default='')
-    parser.add_argument('--rsc-dir', help='resource dir', metavar='DIR', default='')
+    parser.add_argument('--lib-path', help='khaiii shared library path' \
+                                           ' <default ../build/lib/libkhaiii.so>',
+                        metavar='FILE', default='../build/lib/libkhaiii.so')
+    parser.add_argument('--rsc-dir', help='resource dir <default: ../build/share/khaiii>',
+                        metavar='DIR', default='../build/share/khaiii')
     parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
     parser.add_argument('--with-freq', help='print with frequency', action='store_true')
     parser.add_argument('--min-freq', help='minimum frequency <default: 10>', metavar='FREQ',
diff --git a/train/bin/hd_validate_errpatch.bash b/train/bin/hd_validate_errpatch.bash
new file mode 100755
index 0000000..3042ef8
--- /dev/null
+++ b/train/bin/hd_validate_errpatch.bash
@@ -0,0 +1,194 @@
+#!/usr/bin/env bash
+set -e -u
+
+
+#############
+# functions #
+#############
+function print_usage() {
+    local msg=$1
+    echo "Usage: $(basename $0) [options]"
+    echo "Options:"
+    echo "  -h, --help        show this help message and exit"
+    echo "  -i FILE           input file"
+    echo "  -c DIR            corpus dir"
+    echo "  --rsc-src=DIR     <default: ../../rsc/src>"
+    echo "  --lib-path=FILE   <default: ../../build/lib/libkhaiii.so>"
+    echo "  --rsc-dir=DIR     <default: ../../build/share/khaiii>"
+    echo "  --num-mapper=NUM  <default: 1000>"
+    if [ -z "${msg}" ]; then
+        exit 0
+    else
+        echo
+        echo "${msg}"
+        exit 1
+    fi
+}
+
+
+function abspath() {
+    python3 -c "import os, sys; print(os.path.abspath(sys.argv[1]))" $1
+}
+
+
+function parse_args() {
+    INPUT_FILE=""
+    CORPUS_DIR=""
+    LIB_PATH=""
+    RSC_DIR=""
+    RSC_SRC=""
+    NUM_MAPPER=""
+
+    while [[ $# -ge 1 ]]; do
+        case $1 in
+            -h|--help)
+                print_usage ""
+                ;;
+            -i)
+                INPUT_FILE="$2"
+                shift
+                ;;
+            -c)
+                CORPUS_DIR="$2"
+                shift
+                ;;
+            --rsc-src)
+                RSC_SRC="$2"
+                shift
+                ;;
+            --lib-path)
+                LIB_PATH="$2"
+                shift
+                ;;
+            --rsc-dir)
+                RSC_DIR="$2"
+                shift
+                ;;
+            --num-mapper)
+                NUM_MAPPER="$2"
+                shift
+                ;;
+            --) break ;;
+        esac
+        shift
+    done
+
+    # input file 검사
+    if [ -z "${INPUT_FILE}" ]; then
+        print_usage "no input file"
+    fi
+
+    # corpus dir 검사
+    if [ -z "${CORPUS_DIR}" ]; then
+        print_usage "no corpus dir"
+    fi
+
+    if [ -z "${RSC_SRC}" ]; then
+        RSC_SRC=../../rsc/src
+    fi
+    if [ -z "${LIB_PATH}" ]; then
+        LIB_PATH=../../build/lib/libkhaiii.so
+    fi
+    if [ -z "${RSC_DIR}" ]; then
+        RSC_DIR=../../build/share/khaiii
+    fi
+    if [ -z "${NUM_MAPPER}" ]; then
+        NUM_MAPPER=1000
+    fi
+
+    INPUT_FILE=$(abspath ${INPUT_FILE})
+    LIB_PATH=$(abspath ${LIB_PATH})
+    RSC_DIR=$(abspath ${RSC_DIR})
+    RSC_SRC=$(abspath ${RSC_SRC})
+    CORPUS_DIR=$(abspath ${CORPUS_DIR})
+}
+
+
+function init_envs() {
+    # global variables
+    INPUT_DIR=errpatch.in
+    OUTPUT_DIR=errpatch.out
+    CACHE_DIR=errpatch.cache
+}
+
+
+function split_input() {
+    >&2 echo "{{{{{{{{{{ ${FUNCNAME[0]} {{{{{{{{{{"
+
+    local total_line
+    total_line=$(wc -l < ${INPUT_FILE})
+    local line_per_split=$((total_line / NUM_MAPPER))
+    rm -rf ${INPUT_DIR}
+    mkdir -p ${INPUT_DIR}
+    shuf ${INPUT_FILE} | split -d -a 5 -l ${line_per_split} - ${INPUT_DIR}/part-
+
+    hadoop fs -test -e ${INPUT_DIR} && hadoop fs -rm -skipTrash -r ${INPUT_DIR}
+    hadoop fs -put ${INPUT_DIR}
+
+    >&2 echo "}}}}}}}}}} ${FUNCNAME[0]} }}}}}}}}}}"
+}
+
+
+function cache_files() {
+    >&2 echo "{{{{{{{{{{ ${FUNCNAME[0]} {{{{{{{{{{"
+
+    hadoop fs -test -e ${CACHE_DIR} && hadoop fs -rm -skipTrash -r ${CACHE_DIR}
+    hadoop fs -mkdir -p ${CACHE_DIR}
+
+    hadoop fs -put ../../src/main/python/khaiii ${CACHE_DIR}
+    hadoop fs -mkdir -p ${CACHE_DIR}/khaiii/lib
+    hadoop fs -put ${LIB_PATH} ${CACHE_DIR}/khaiii/lib
+
+    hadoop fs -mkdir -p ${CACHE_DIR}/khaiii/share
+    hadoop fs -put ${RSC_DIR} ${CACHE_DIR}/khaiii/share/khaiii
+
+    hadoop fs -put ${RSC_SRC} ${CACHE_DIR}/rsc_src
+
+    hadoop fs -mkdir -p ${CACHE_DIR}/corpus
+    hadoop fs -put ${CORPUS_DIR}/*.txt ${CACHE_DIR}/corpus
+
+    >&2 echo "}}}}}}}}}} ${FUNCNAME[0]} }}}}}}}}}}"
+}
+
+
+function run_hadoop() {
+    >&2 echo "{{{{{{{{{{ ${FUNCNAME[0]} {{{{{{{{{{"
+
+    hadoop fs -test -e ${OUTPUT_DIR} && hadoop fs -rm -skipTrash -r ${OUTPUT_DIR}
+    yarn jar ${HADOOP_HOME}/share/hadoop/tools/lib/hadoop-streaming-*.jar \
+        -D mapred.job.name=validate_errpatch \
+        -D mapred.reduce.tasks=0 \
+        -cmdenv PYTHONPATH="./${CACHE_DIR}" \
+        -file ./validate_errpatch.py \
+        -input "${INPUT_DIR}" \
+        -output "${OUTPUT_DIR}" \
+        -cacheFile "${CACHE_DIR}#${CACHE_DIR}" \
+        -mapper "./validate_errpatch.py -c ./${CACHE_DIR}/corpus --rsc-src ./${CACHE_DIR}/rsc_src"
+
+    hadoop fs -text ${OUTPUT_DIR}/part-* > "$(dirname ${INPUT_FILE})/errpatch.valid"
+
+    >&2 echo "}}}}}}}}}} ${FUNCNAME[0]} }}}}}}}}}}"
+}
+
+
+function del_temp() {
+    >&2 echo "{{{{{{{{{{ ${FUNCNAME[0]} {{{{{{{{{{"
+
+    hadoop fs -rm -skipTrash -r ${INPUT_DIR} ${OUTPUT_DIR} ${CACHE_DIR}
+    rm -rf ${INPUT_DIR}
+
+    >&2 echo "}}}}}}}}}} ${FUNCNAME[0]} }}}}}}}}}}"
+}
+
+
+########
+# main #
+########
+parse_args $@
+cd "$(dirname $0)"
+init_envs
+
+# split_input
+# cache_files
+run_hadoop
+del_temp
diff --git a/train/bin/train.py b/train/bin/train.py
index b696d96..7684be4 100755
--- a/train/bin/train.py
+++ b/train/bin/train.py
@@ -42,8 +42,8 @@ def main():
     parser = ArgumentParser(description='train model from data')
     parser.add_argument('-i', '--in-pfx', help='input data path prefix', metavar='NAME',
                         required=True)
-    parser.add_argument('--rsc-src', help='train resource dir <default: ../rsc/src>', metavar='DIR',
-                        default='../rsc/src')
+    parser.add_argument('--rsc-src', help='resource source dir <default: ../rsc/src>',
+                        metavar='DIR', default='../rsc/src')
     parser.add_argument('--logdir', help='tensorboard log dir <default: ./logdir>', metavar='DIR',
                         default='./logdir')
     parser.add_argument('--window', help='left/right character window length <default: 3>',
diff --git a/train/bin/validate_errpatch.py b/train/bin/validate_errpatch.py
new file mode 100755
index 0000000..ed73271
--- /dev/null
+++ b/train/bin/validate_errpatch.py
@@ -0,0 +1,270 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+추출된 오분석 패치 후보를 검증하는 스크립트
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+from collections import defaultdict
+import io
+import logging
+import os
+import sys
+from typing import Dict, Iterator, List, Tuple
+
+from khaiii.khaiii import KhaiiiApi
+
+from khaiii.munjong.sejong_corpus import Sentence, sents
+from khaiii.resource.char_align import Aligner, AlignError, align_patch, align_to_tag
+from khaiii.resource.morphs import mix_char_tag, WORD_DELIM_NUM, SENT_DELIM_NUM
+from khaiii.resource.resource import load_restore_dic, load_vocab_out
+
+
+#########
+# types #
+#########
+class StrFile(io.StringIO):
+    """
+    StringIO 객체에 name 멤버를 추가해 파일 객체인 것 처럼 동작하도록 하기 위해
+    """
+    def __init__(self, name, buf):
+        """
+        Args:
+            name:  file name
+            buf:  buffer
+        """
+        super().__init__(buf)
+        self.name = name
+
+
+#############
+# variables #
+#############
+_CORPUS = []    # corpus cache
+
+
+##########
+# mapper #
+##########
+def _sent_iter(args: Namespace) -> Iterator[Sentence]:
+    """
+    sentence generator
+    Args:
+        args:  arguments
+    Yields:
+        sentence
+    """
+    if not _CORPUS:
+        for name in sorted(os.listdir(args.corpus_dir)):
+            if not name.endswith('.txt'):
+                continue
+            path = '{}/{}'.format(args.corpus_dir, name)
+            _CORPUS.append((name, open(path, 'r', encoding='UTF-8').read()))
+        _CORPUS.sort()
+
+    for name, lines in _CORPUS:
+        logging.info(name)
+        for sent in sents(StrFile(name, lines)):
+            yield sent
+
+
+def _find_list(haystack: list, needle: list) -> int:
+    """
+    Return the index at which the sequence needle appears in the sequence haystack,
+    or -1 if it is not found, using the Boyer-Moore-Horspool algorithm.
+    The elements of needle and haystack must be hashable.
+
+    >>> _find_list([10, 10, 20], [10, 20])
+    1
+
+    Args:
+        haystack:  list to find
+        needle:  pattern list
+    Returns:
+        start index. -1 if not found
+    """
+    h_len = len(haystack)
+    n_len = len(needle)
+    skip = {needle[i]: n_len - i - 1 for i in range(n_len - 1)}
+    idx = n_len - 1
+    while idx < h_len:
+        for jdx in range(n_len):
+            if haystack[idx - jdx] != needle[-jdx - 1]:
+                idx += skip.get(haystack[idx], n_len)
+                break
+        else:
+            return idx - n_len + 1
+    return -1
+
+
+def _align_sent(rsc_src: Tuple[Aligner, dict, Dict[str, str]], sent: Sentence) -> List[int]:
+    """
+    세종 문장 객체를 정렬하여 음절 별 출력 태그의 벡터로 표현한다.
+    Args:
+        rsc_src:  (Aligner, restore dic, vocab out) resource triple
+        sent:  sejong_corpus.Sentence 객체
+    Returns:
+        list of output tag numbers. empty list for alignment error
+    """
+    aligner, restore_dic, vocab_out = rsc_src
+    tag_nums = []
+    restore_new = defaultdict(dict)
+    vocab_new = defaultdict(list)
+    for word in sent.words:
+        try:
+            word_align = aligner.align(word)
+            _, word_tag_nums = align_to_tag(word.raw, word_align, (restore_dic, restore_new),
+                                            (vocab_out, vocab_new))
+        except AlignError as algn_err:
+            logging.debug('alignment error: %s', word)
+            logging.debug(str(algn_err))
+            return []
+        if tag_nums:
+            tag_nums.append(WORD_DELIM_NUM)
+        tag_nums.extend(word_tag_nums)
+    tag_nums.insert(0, SENT_DELIM_NUM)
+    tag_nums.append(SENT_DELIM_NUM)
+    return tag_nums
+
+
+def _analyze_sent(khaiii_api: KhaiiiApi, raw_sent: str) -> List[int]:
+    """
+    원시 문장에 대해 패치를 적용하지 않은 음절별 태깅 결과를 얻는다.
+    Args:
+        khaiii_api:  khaiii API 객체
+        raw_sent:  원시 문장
+    Returns:
+        list of output tag numbers
+    """
+    tag_nums = khaiii_api.analyze_bfr_errpatch(raw_sent, '')
+    logging.debug(tag_nums)
+    return tag_nums
+
+
+def _cnt_pos_neg(khaiii_api: KhaiiiApi, patch_raw: str, alignment: Tuple[list, list],
+                 rsc_src: Tuple[Aligner, dict, Dict[str, str]], sent: Sentence) -> Tuple[int, int]:
+    """
+    오분석을 정분석으로 바꾼 횟수와, 오분석을 다른 오분석으로 바꾼 횟수를 센다.
+    Args:
+        khaiii_api:  khaiii API object
+        patch_raw:  raw part of patch
+        alignment:  (left, right) alignment pair
+        rsc_src:  (Aligner, restore dic, vocab out) resource triple
+        sent:  Sentence object
+    Returns:
+        오분석 -> 정분석 횟수
+        오분석 -> 오분석 횟수
+    """
+    raw_sent = sent.raw_str()
+    if patch_raw not in raw_sent:
+        # 원문이 문장에서 발견되지 않으면 스킵
+        return 0, 0
+    aligner, restore_dic, vocab_out = rsc_src
+    sent_align = _align_sent((aligner, restore_dic, vocab_out), sent)
+    if not sent_align:
+        # 코퍼스 정답이 원문과 정렬이 되지 않고 오류가 발생하면 스킵
+        return 0, 0
+    left_align, right_align = alignment
+    left_needle = mix_char_tag(patch_raw, left_align)
+    sent_anal = khaiii_api.analyze_bfr_errpatch(raw_sent, '')
+    sent_haystack = mix_char_tag(raw_sent, sent_anal)
+    pos_cnt = 0
+    neg_cnt = 0
+    found = _find_list(sent_haystack, left_needle)
+    while found >= 0:
+        # 패치의 좌측 오분석 열이 분석 결과에서 나타난 경우 우측 정답 열과 코퍼스를 비교
+        right_corpus = sent_align[found:found + len(left_needle)]
+        if right_align == right_corpus:
+            pos_cnt += 1
+        else:
+            neg_cnt += 1
+        del sent_haystack[:found + len(left_needle)]
+        found = _find_list(sent_haystack, left_needle)
+    return pos_cnt, neg_cnt
+
+
+def run(args: Namespace):
+    """
+    actual function which is doing some task
+    Args:
+        args:  program arguments
+    """
+    aligner = Aligner(args.rsc_src)
+    restore_dic = load_restore_dic('{}/restore.dic'.format(args.rsc_src))
+    if not restore_dic:
+        sys.exit(1)
+    vocab_out = load_vocab_out(args.rsc_src)
+
+    khaiii_api = KhaiiiApi(args.lib_path, args.rsc_dir, '{"errpatch": false}')
+
+    for line_num, line in enumerate(sys.stdin, start=1):
+        line = line.rstrip('\r\n')
+        if not line or line[0] == '#':
+            continue
+        raw, left, right = line.split('\t')
+        left_align = align_patch((aligner, restore_dic, vocab_out), raw, left)
+        if not left_align:
+            logging.info('invalid %d-th line: left align: %s', line_num, line)
+            continue
+        right_align = align_patch((aligner, restore_dic, vocab_out), raw, right)
+        if not right_align:
+            logging.info('invalid %d-th line: right align: %s', line_num, line)
+            continue
+        if len(left_align) != len(right_align):
+            logging.info('invalid %d-th line: left/right diff: %s', line_num, line)
+            continue
+        pos_cnt = 0
+        neg_cnt = 0
+        for sent in _sent_iter(args):
+            pos_cnt_sent, neg_cnt_sent = _cnt_pos_neg(khaiii_api, raw, (left_align, right_align),
+                                                      (aligner, restore_dic, vocab_out), sent)
+            pos_cnt += pos_cnt_sent
+            neg_cnt += neg_cnt_sent
+            if neg_cnt > 0:
+                break
+        if neg_cnt > 0 or pos_cnt == 0:
+            logging.info('invalid %d-th line: +%d, -%d: %s', line_num, pos_cnt, neg_cnt, line)
+            continue
+        print('{}\t{}\t{}'.format(raw, left, right))
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='추출된 오분석 패치 후보를 검증하는 스크립트')
+    parser.add_argument('-c', '--corpus-dir', help='corpus dir', metavar='DIR', required=True)
+    parser.add_argument('--rsc-src', help='resource source dir <default: ../rsc/src>',
+                        metavar='DIR', default='../rsc/src')
+    parser.add_argument('--lib-path', help='khaiii shared library path', metavar='FILE', default='')
+    parser.add_argument('--rsc-dir', help='resource dir', metavar='DIR', default='')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()

From a5c512ddee78c6d5786a2ee035d51315891ecbf9 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 28 Jan 2019 00:20:19 +0900
Subject: [PATCH 33/41] =?UTF-8?q?=EC=BD=94=ED=8D=BC=EC=8A=A4=EC=9D=98=20?=
 =?UTF-8?q?=ED=8C=A8=EC=B9=98=EB=A5=BC=20=EC=83=9D=EC=84=B1=ED=95=98?=
 =?UTF-8?q?=EA=B3=A0=20=EC=A0=81=EC=9A=A9=ED=95=98=EB=8A=94=20=EC=8A=A4?=
 =?UTF-8?q?=ED=81=AC=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#3?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 munjong/bin/apply_patch.py                 |  79 ++++
 munjong/bin/make_patch.py                  |  82 ++++
 src/main/python/khaiii/munjong/libpatch.py | 426 +++++++++++++++++++++
 3 files changed, 587 insertions(+)
 create mode 100755 munjong/bin/apply_patch.py
 create mode 100755 munjong/bin/make_patch.py
 create mode 100644 src/main/python/khaiii/munjong/libpatch.py

diff --git a/munjong/bin/apply_patch.py b/munjong/bin/apply_patch.py
new file mode 100755
index 0000000..44562ec
--- /dev/null
+++ b/munjong/bin/apply_patch.py
@@ -0,0 +1,79 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+apply patch to original Sejong corpus
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+import logging
+import os
+import shutil
+
+from khaiii.munjong import libpatch
+
+
+#############
+# functions #
+#############
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    if not os.path.exists(args.modified):
+        logging.info('creating modified corpus dir: %s', args.modified)
+        os.mkdir(args.modified)
+
+    for name in sorted(os.listdir(args.original)):
+        if not name.endswith('.txt'):
+            continue
+        org_path = '%s/%s' % (args.original, name)
+        mod_path = '%s/%s' % (args.modified, name)
+        patch_path = '%s/%s.patch' % (args.patch, name[:-len('.txt')])
+        if os.path.exists(patch_path):
+            logging.info('[%s] + [%s] = [%s]', org_path, patch_path, mod_path)
+            libpatch.apply(org_path, args.org_enc, patch_path, mod_path, args.mod_enc)
+        else:
+            logging.info('[%s] = [%s]', org_path, mod_path)
+            shutil.copyfile(org_path, mod_path)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='apply patch to original Sejong corpus')
+    parser.add_argument('-o', '--original', help='original corpus dir', metavar='DIR',
+                        required=True)
+    parser.add_argument('-p', '--patch', help='patch dir', metavar='DIR', required=True)
+    parser.add_argument('-m', '--modified', help='modified corpus output dir', metavar='DIR',
+                        required=True)
+    parser.add_argument('--org-enc', help='original corpus encoding <default: UTF-16>',
+                        metavar='ENCODING', default='UTF-16')
+    parser.add_argument('--mod-enc', help='modified corpus encoding <default: UTF-8>',
+                        metavar='ENCODING', default='UTF-8')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/make_patch.py b/munjong/bin/make_patch.py
new file mode 100755
index 0000000..25e3bd1
--- /dev/null
+++ b/munjong/bin/make_patch.py
@@ -0,0 +1,82 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+make patch from two Sejong corpora
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser, Namespace
+import logging
+import os
+
+from khaiii.munjong import libpatch
+
+
+#############
+# functions #
+#############
+def run(args: Namespace):
+    """
+    run function which is the start point of program
+    Args:
+        args:  program arguments
+    """
+    if not os.path.exists(args.patch):
+        logging.info('creating patch dir: %s', args.patch)
+        os.mkdir(args.patch)
+
+    for name in sorted(os.listdir(args.original)):
+        if not name.endswith('.txt'):
+            continue
+        org_path = '%s/%s' % (args.original, name)
+        mod_path = '%s/%s' % (args.modified, name)
+        patch_path = '%s/%s.patch' % (args.patch, name[:-len('.txt')])
+        logging.info('[%s] - [%s] = [%s]', org_path, mod_path, patch_path)
+        patches = libpatch.make(org_path, args.org_enc, mod_path, args.mod_enc)
+        if patches:
+            logging.info('creating patch file: %s', patch_path)
+            with open(patch_path, 'w', encoding='UTF-8') as fout:
+                for patch in patches:
+                    print(patch, file=fout)
+        elif os.path.exists(patch_path):
+            logging.info('removing existing patch file: %s', patch_path)
+            os.remove(patch_path)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='make patch from two Sejong corpora')
+    parser.add_argument('-o', '--original', help='original corpus dir', metavar='DIR',
+                        required=True)
+    parser.add_argument('-m', '--modified', help='modified corpus dir', metavar='DIR',
+                        required=True)
+    parser.add_argument('-p', '--patch', help='patch output dir', metavar='DIR', required=True)
+    parser.add_argument('--org-enc', help='original corpus encoding <default: UTF-16>',
+                        metavar='ENCODING', default='UTF-16')
+    parser.add_argument('--mod-enc', help='modified corpus encoding <default: UTF-8>',
+                        metavar='ENCODING', default='UTF-8')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run(args)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/src/main/python/khaiii/munjong/libpatch.py b/src/main/python/khaiii/munjong/libpatch.py
new file mode 100644
index 0000000..13e3ecf
--- /dev/null
+++ b/src/main/python/khaiii/munjong/libpatch.py
@@ -0,0 +1,426 @@
+# -*- coding: utf-8 -*-
+
+
+"""
+patch library for Sejong corpus
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from collections import namedtuple
+import logging
+import os
+from typing import Dict, List, Tuple
+
+from khaiii.munjong.sejong_corpus import SENT_OPEN_TAGS, SENT_CLOSE_TAGS, WORD_ID_PTN
+
+
+#############
+# constatns #
+#############
+# line types
+WORD_TYPE = 1    # word (eojeol)
+BOS_TYPE = 2    # begin of sentence markup
+EOS_TYPE = 3    # end of sentence markup
+MISC_TYPE = 0    # other miscellaneous lines
+
+
+#########
+# types #
+#########
+Line = namedtuple('Line', ['type_', 'wid', 'content'])
+
+
+class Patch:    # pylint: disable=too-few-public-methods
+    """
+    patch line
+    """
+    # patch types (actions, operations)
+    SENT_SPLIT = 1
+    SENT_MERGE = 2
+    WORD_REPLACE = 3
+    WORD_INSERT = 4
+    WORD_DELETE = 5
+    # patch categories
+    SENT_CATE = 1
+    WORD_CATE = 2
+
+    def __init__(self, type_: int, wid: str, content: str):
+        """
+        Args:
+            type_:  patch type
+            wid:  word ID
+            content:  word content (tab separated second column)
+        """
+        self.type_ = type_
+        self.wid = wid
+        self.content = content
+
+    def __eq__(self, other: 'Patch') -> bool:
+        return self.type_ == other.type_ and self.wid == other.wid and self.content == other.content
+
+    def __cmp__(self, other: 'Patch') -> int:
+        if self.cate() == other.cate():
+            if self.wid < other.wid:
+                return -1
+            if self.wid > other.wid:
+                return 1
+            return 0
+        return self.cate() - other.cate()
+
+    def __lt__(self, other: 'Patch') -> bool:
+        return self.__cmp__(other) < 0
+
+    def __str__(self) -> str:
+        if self.type_ == self.WORD_REPLACE:
+            return '=\t%s\t%s' % (self.wid, self.content)
+        if self.type_ == self.WORD_INSERT:
+            return '+\t%s\t%s' % (self.wid, self.content)
+        if self.type_ == self.WORD_DELETE:
+            return '-\t%s' % self.wid
+        if self.type_ == self.SENT_SPLIT:
+            return 'S\t%s\t%s' % (self.wid, self.content)
+        if self.type_ == self.SENT_MERGE:
+            return 'M\t%s\t%s' % (self.wid, self.content)
+        raise RuntimeError('unknown patch type: %d' % self.type_)
+
+    def cate(self) -> int:
+        """
+        get patch category. EOS/BOS patches are 0. word patches are 1
+        Returns:
+            category number
+        """
+        return self.SENT_CATE if self.type_ in [self.SENT_SPLIT, self.SENT_MERGE] \
+                              else self.WORD_CATE
+
+    @classmethod
+    def parse(cls, line: str) -> 'Patch':
+        """
+        parse patch line
+        Args:
+            line:  patch line
+        Returns:
+            Patch object
+        """
+        cols = line.split('\t', 2)
+        if len(cols) == 3:
+            if cols[0] == '=':
+                return Patch(cls.WORD_REPLACE, cols[1], cols[2])
+            if cols[0] == '+':
+                return Patch(cls.WORD_INSERT, cols[1], cols[2])
+            if cols[0].upper() == 'M':
+                return Patch(cls.SENT_MERGE, cols[1], cols[2])
+            if cols[0].upper() == 'S':
+                return Patch(cls.SENT_SPLIT, cols[1], cols[2])
+        elif len(cols) == 2:
+            if cols[0] == '-':
+                return Patch(cls.WORD_DELETE, cols[1], '')
+        raise RuntimeError('invalid patch format: %s' % line)
+
+
+#############
+# functions #
+#############
+def _load_corpus(path: str, enc: str) -> Tuple[List[Line], Dict[str, int]]:
+    """
+    load corpus
+    Args:
+        path:  file path
+        enc:  file encoding
+    Returns:
+        list of lines
+        word ID dic
+    """
+    lines = []
+    wid_dic = {}
+    for line in open(path, 'r', encoding=enc):
+        line = line.rstrip('\r\n')
+        if WORD_ID_PTN.match(line):
+            wid, content = line.split('\t', 1)
+            if wid in wid_dic:
+                raise RuntimeError('duplicated word ID: %s' % line)
+            else:
+                wid_dic[wid] = len(lines)
+                lines.append(Line(WORD_TYPE, wid, content))
+        elif line in SENT_OPEN_TAGS:
+            lines.append(Line(BOS_TYPE, None, line))
+        elif line in SENT_CLOSE_TAGS:
+            lines.append(Line(EOS_TYPE, None, line))
+        else:
+            lines.append(Line(MISC_TYPE, None, line))
+            if line == '</tei.2>' and 'BTJO0443.txt' in path:
+                break
+    return lines, wid_dic
+
+
+def _make_sent_patch(org_lines: List[Line], mod_lines: List[Line]) -> List[Patch]:
+    """
+    make EOS/BOS patch
+    Args:
+        org_lines:  original lines
+        mod_lines:  modified lines
+    Returns:
+        EOS/BOS patches
+    """
+    def _get_eos_bos(lines: List[Line]) -> Tuple[str, str]:
+        """
+        get all EOS/BOS from lines
+        Args:
+            lines:  lines of corpus
+        Returns:
+            prev. word ID
+            next word ID
+        """
+        eos_bos = []
+        for idx, line in enumerate(lines):
+            if line.type_ == EOS_TYPE and (idx+2) < len(lines) and lines[idx+1].type_ == BOS_TYPE \
+                    and lines[idx-1].type_ == WORD_TYPE and lines[idx+2].type_ == WORD_TYPE:
+                prev_wid = lines[idx-1].wid
+                next_wid = lines[idx+2].wid
+                eos_bos.append((prev_wid[:17], next_wid[:17]))
+        return eos_bos
+
+    org_eos_bos = set(_get_eos_bos(org_lines))
+    mod_eos_bos = set(_get_eos_bos(mod_lines))
+    patches = []
+    for prev_wid, next_wid in org_eos_bos - mod_eos_bos:
+        patches.append(Patch(Patch.SENT_MERGE, prev_wid, next_wid))
+    for prev_wid, next_wid in mod_eos_bos - org_eos_bos:
+        patches.append(Patch(Patch.SENT_SPLIT, prev_wid, next_wid))
+    return sorted(patches)
+
+
+def _make_word_patch(org_lines: List[Line], org_wid_dic: Dict[str, int], mod_lines: List[Line],
+                     mod_wid_dic: Dict[str, int]) -> List[Patch]:
+    """
+    make word patch
+    Args:
+        org_lines:  original lines
+        org_wid_dic:  original word ID dictionary
+        mod_lines:  modified lines
+        mod_wid_dic:  modified word ID dictionary
+    Returns:
+        word patches
+    """
+    patches = []
+    for org_line in org_lines:
+        if org_line.type_ != WORD_TYPE:
+            continue
+        if org_line.wid in mod_wid_dic:
+            mod_line = mod_lines[mod_wid_dic[org_line.wid]]
+            if org_line.content != mod_line.content:
+                patches.append(Patch(Patch.WORD_REPLACE, mod_line.wid, mod_line.content))
+        else:
+            patches.append(Patch(Patch.WORD_DELETE, org_line.wid, ''))
+    for mod_line in mod_lines:
+        if mod_line.type_ != WORD_TYPE or mod_line.wid in org_wid_dic:
+            continue
+        patches.append(Patch(Patch.WORD_INSERT, mod_line.wid, mod_line.content))
+    return sorted(patches)
+
+
+def make(org_path: str, org_enc: str, mod_path: str, mod_enc: str) -> List[Patch]:
+    """
+    make patch from two file
+    Args:
+        org_path:  original file path
+        org_enc:  original file encoding
+        mod_path:  modified file path
+        mod_enc:  modified file encoding
+    Returns:
+        patch contents (list of patch lines)
+    """
+    org_lines, org_wid_dic = _load_corpus(org_path, org_enc)
+    mod_lines, mod_wid_dic = _load_corpus(mod_path, mod_enc)
+
+    sent_patches = _make_sent_patch(org_lines, mod_lines)
+    word_patches = _make_word_patch(org_lines, org_wid_dic, mod_lines, mod_wid_dic)
+    return sent_patches + word_patches
+
+
+def _load_patch(patch_path: str) -> List[Patch]:
+    """
+    load patch from file
+    Args:
+        patch_path:  patch path
+    Returns:
+        patches
+    """
+    patch_name = os.path.basename(patch_path)
+    wid_dic = set()
+    patches = []
+    for line_num, line in enumerate(open(patch_path, 'r', encoding='UTF-8'), start=1):
+        line = line.rstrip('\r\n')
+        if not line:
+            continue
+        patch = Patch.parse(line)
+        if patch.cate() == Patch.SENT_CATE:
+            if (patch.wid, patch.content) in wid_dic:
+                logging.error('%s(%d): patch conflict: %s', patch_name, line_num, line)
+            else:
+                patches.append(patch)
+                wid_dic.add((patch.wid, patch.content))
+        else:
+            if patch.wid in wid_dic:
+                logging.error('%s(%d): patch conflict: %s', patch_name, line_num, line)
+            else:
+                patches.append(patch)
+                wid_dic.add(patch.wid)
+    return patches
+
+
+def _apply_sent_merge_patch(org_lines: List[Line], patches: List[Patch]) -> List[Line]:
+    """
+    apply merge EOS/BOS patches
+    Args:
+        org_lines:  original lines
+        patches:  patches
+    Returns:
+        modified lines
+    """
+    merge_patches = {patch.wid: patch.content for patch in patches \
+                     if patch.type_ == Patch.SENT_MERGE}
+    mod_lines = []
+    idx = 0
+    while idx < len(org_lines):
+        org_line = org_lines[idx]
+        if org_line.type_ == EOS_TYPE and org_lines[idx+1].type_ == BOS_TYPE and \
+                        org_lines[idx-1].type_ == WORD_TYPE and org_lines[idx+2].type_ == WORD_TYPE:
+            prev_wid = org_lines[idx-1].wid[:17]
+            next_wid = org_lines[idx+2].wid[:17]
+            if prev_wid in merge_patches and merge_patches[prev_wid] == next_wid:
+                del merge_patches[prev_wid]
+                idx += 2
+                continue
+        mod_lines.append(org_line)
+        idx += 1
+    if merge_patches:
+        for prev_wid, next_wid in merge_patches.items():
+            logging.error('remaining merge sentence patches: %s\t%s', prev_wid, next_wid)
+    return mod_lines
+
+
+def _build_wid_dic(lines: List[Line]) -> Dict[str, int]:
+    """
+    build word ID dictionary
+    Args:
+        lines:  lines of corpus
+    Returns:
+        word ID dictionary
+    """
+    return {line.wid: idx for idx, line in enumerate(lines) if line.type_ == WORD_TYPE}
+
+
+def _apply_sent_split_patch(mod_lines: List[Line], patches: List[Patch]) -> List[Line]:
+    """
+    apply split EOS/BOS patches
+    Args:
+        mod_lines:  modified lines
+        patches:  patches
+    Returns:
+        modified lines
+    """
+    mod_wid_dic = _build_wid_dic(mod_lines)
+    split_patches = {patch.wid: patch.content for patch in patches \
+                     if patch.type_ == Patch.SENT_SPLIT}
+    for prev_wid, next_wid in sorted(split_patches.items(), key=lambda x: x[0], reverse=True):
+        idx = mod_wid_dic[next_wid]
+        assert mod_lines[idx].wid == next_wid
+        if mod_lines[idx-1].type_ == WORD_TYPE and mod_lines[idx-1].wid[:17] == prev_wid:
+            mod_lines.insert(idx, Line(BOS_TYPE, None, '<p>'))
+            mod_lines.insert(idx, Line(EOS_TYPE, None, '</p>'))
+            del split_patches[prev_wid]
+    if split_patches:
+        for prev_wid, next_wid in split_patches.items():
+            logging.error('remaining split sentence patches: %s\t%s', prev_wid, next_wid)
+    return mod_lines
+
+
+def _apply_word_del_rep_patch(mod_lines: List[Line], patches: List[Patch]) -> List[Line]:
+    """
+    apply word delete/replace patches
+    Args:
+        mod_lines:  modified lines
+        patches:  patches
+    Returns:
+        modified lines
+    """
+    delete_patches = {patch.wid for patch in patches if patch.type_ == Patch.WORD_DELETE}
+    replace_patches = {patch.wid: patch.content for patch in patches \
+                       if patch.type_ == Patch.WORD_REPLACE}
+    new_lines = []
+    for line in mod_lines:
+        if line.type_ == WORD_TYPE:
+            if line.wid in delete_patches:
+                delete_patches.remove(line.wid)
+                continue
+            elif line.wid in replace_patches:
+                new_lines.append(Line(WORD_TYPE, line.wid, replace_patches[line.wid]))
+                del replace_patches[line.wid]
+                continue
+        new_lines.append(line)
+    if delete_patches:
+        for wid in delete_patches:
+            logging.error('remaining delete word patches: %s', wid)
+    if replace_patches:
+        for wid, content in replace_patches.items():
+            logging.error('remaining replace word patches: %s\t%s', wid, content)
+    return new_lines
+
+
+def _apply_word_insert_patch(mod_lines: List[Line], patches: List[Patch]) -> List[Line]:
+    """
+    apply word insert patches
+    Args:
+        mod_lines:  modified lines
+        patches:  patches
+    Returns:
+        modified lines
+    """
+    insert_patches = sorted([(patch.wid, patch.content) for patch in patches \
+                             if patch.type_ == Patch.WORD_INSERT])
+    prev_idx = -1
+    curr_idx = 0
+    while curr_idx < len(mod_lines) and insert_patches:
+        curr_word = mod_lines[curr_idx]
+        if curr_word.type_ != WORD_TYPE:
+            curr_idx += 1
+            continue
+        wid, content = insert_patches[0]
+        if curr_word.wid < wid:
+            prev_idx = curr_idx
+            curr_idx += 1
+            continue
+        mod_lines.insert(prev_idx+1, Line(WORD_TYPE, wid, content))
+        del insert_patches[0]
+        prev_idx += 1
+        curr_idx += 1
+    return mod_lines
+
+
+def apply(org_path: str, org_enc: str, patch_path: str, mod_path: str, mod_enc: str):
+    """
+    apply path to original corpus then get modified corpus
+    Args:
+        org_path:  original file path
+        org_enc:  original file encoding
+        patch_path:  patch file path
+        mod_path:  modified file path
+        mod_enc:  modified file encoding
+    """
+    patches = _load_patch(patch_path)
+    org_lines, _ = _load_corpus(org_path, org_enc)
+    mod_lines = _apply_sent_merge_patch(org_lines, patches)
+    mod_lines = _apply_word_del_rep_patch(mod_lines, patches)
+    mod_lines = _apply_sent_split_patch(mod_lines, patches)
+    mod_lines = _apply_word_insert_patch(mod_lines, patches)
+    with open(mod_path, 'w', encoding=mod_enc) as fout:
+        for mod_line in mod_lines:
+            if mod_line.type_ == WORD_TYPE:
+                print('%s\t%s' % (mod_line.wid, mod_line.content), file=fout)
+            else:
+                print(mod_line.content, file=fout)

From 090517dbc977ac7ac2d44a90e52b86683d2adb74 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 28 Jan 2019 20:21:51 +0900
Subject: [PATCH 34/41] =?UTF-8?q?=EC=84=B8=EC=A2=85=20=EC=BD=94=ED=8D=BC?=
 =?UTF-8?q?=EC=8A=A4=EC=9D=98=20=EC=98=A4=EB=A5=98=EC=99=80=20=EA=B4=80?=
 =?UTF-8?q?=EB=A0=A8=EB=90=9C=20=EA=B0=81=EC=A2=85=20=EC=8A=A4=ED=81=AC?=
 =?UTF-8?q?=EB=A6=BD=ED=8A=B8=20=EC=B6=94=EA=B0=80=20#3?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 munjong/bin/convert_jamo_to_compat.py     |  79 ++++++++++++++
 munjong/bin/detect_sejong_period_error.py | 106 +++++++++++++++++++
 munjong/bin/fix_final_symbol_error.py     |  86 ++++++++++++++++
 munjong/bin/recover_english_case.py       |  96 +++++++++++++++++
 munjong/bin/recover_raw_morph_mismatch.py | 101 ++++++++++++++++++
 munjong/bin/recover_wide_quotation.py     | 108 +++++++++++++++++++
 munjong/bin/remove_sejong_period_error.py | 120 ++++++++++++++++++++++
 src/main/python/khaiii/resource/jaso.py   | 103 +++++++++++++++----
 8 files changed, 781 insertions(+), 18 deletions(-)
 create mode 100755 munjong/bin/convert_jamo_to_compat.py
 create mode 100755 munjong/bin/detect_sejong_period_error.py
 create mode 100755 munjong/bin/fix_final_symbol_error.py
 create mode 100755 munjong/bin/recover_english_case.py
 create mode 100755 munjong/bin/recover_raw_morph_mismatch.py
 create mode 100755 munjong/bin/recover_wide_quotation.py
 create mode 100755 munjong/bin/remove_sejong_period_error.py

diff --git a/munjong/bin/convert_jamo_to_compat.py b/munjong/bin/convert_jamo_to_compat.py
new file mode 100755
index 0000000..6ea1426
--- /dev/null
+++ b/munjong/bin/convert_jamo_to_compat.py
@@ -0,0 +1,79 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+한글 자모 영역의 코드를 호환 영역으로 변환
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import logging
+import sys
+
+from khaiii.munjong.sejong_corpus import WORD_ID_PTN
+from khaiii.resource.jaso import norm_compat
+
+
+#############
+# functions #
+#############
+def _norm(text: str) -> str:
+    """
+    정규화를 수행하는 함수
+    Args:
+        text:  입력 텍스트
+    Returns:
+        정규화된 텍스트
+    """
+    normalized = norm_compat(text)
+    normalized = normalized.replace('ᆞ', 'ㆍ')    # 0x119e -> 0x318d
+    normalized = normalized.replace('ᄝ', 'ㅱ')    # 0x111d -> 0x3171
+    return normalized
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    for line in sys.stdin:
+        line = line.rstrip('\r\n')
+        if not WORD_ID_PTN.match(line):
+            print(line)
+            continue
+        wid, word, morph = line.split('\t')
+        print('{}\t{}\t{}'.format(wid, _norm(word), _norm(morph)))
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='한글 자모 영역의 코드를 호환 영역으로 변환')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/detect_sejong_period_error.py b/munjong/bin/detect_sejong_period_error.py
new file mode 100755
index 0000000..7dfd4fb
--- /dev/null
+++ b/munjong/bin/detect_sejong_period_error.py
@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+detect period error of Sejong corpus
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import logging
+import os
+import re
+import sys
+from typing import Iterator, TextIO, Tuple
+
+from khaiii.munjong.sejong_corpus import Morph, WORD_ID_PTN
+
+
+#############
+# functions #
+#############
+def _get_two_lines(fin: TextIO) -> Iterator[Tuple[str, str]]:
+    """
+    get two lines tuple from file (generator)
+    Args:
+        fin:  input file
+    Yields:
+        current line
+        next line
+    """
+    curr_line = fin.readline().rstrip('\r\n')
+    for next_line in fin:
+        next_line = next_line.rstrip('\r\n')
+        yield curr_line, next_line
+        curr_line = next_line
+
+
+def _is_correct_eos(line: str) -> bool:
+    """
+    whether correct end of sentence or not
+    Args:
+        line:  line (word)
+    Returns:
+        whether correct or not
+    """
+    _, _, morphs_str = line.split('\t')
+    if re.match(r'.+/EF \+ ./SF$', morphs_str):
+        return True
+    if re.match(r'.+/SF \+ [\'"’”」\]]/SS$', morphs_str):
+        return True
+    morphs = [Morph.parse(_) for _ in morphs_str.split(' + ')]
+    tags_str = '+'.join([_.tag for _ in morphs])
+    if tags_str.endswith('+SF+SS+JKQ') or tags_str.endswith('+SF+SS+VCP+ETM'):
+        return True
+    return False
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    file_name = os.path.basename(sys.stdin.name)
+    for line_num, (curr_line, next_line) in enumerate(_get_two_lines(sys.stdin), start=1):
+        cols = curr_line.split('\t')
+        if len(cols) != 3 or not WORD_ID_PTN.match(cols[0]):
+            continue
+        if '/SF + ' not in cols[2] or not next_line.startswith('</'):
+            continue
+        if _is_correct_eos(curr_line):
+            continue
+        print('{}:{}\t{}'.format(file_name, line_num, curr_line))
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='detect period error of Sejong corpus')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'rt')
+    if args.output:
+        sys.stdout = open(args.output, 'wt')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/fix_final_symbol_error.py b/munjong/bin/fix_final_symbol_error.py
new file mode 100755
index 0000000..5ba1644
--- /dev/null
+++ b/munjong/bin/fix_final_symbol_error.py
@@ -0,0 +1,86 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+fix final symbol errors on Sejong corpus
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import logging
+import os
+import sys
+
+from khaiii.munjong.sejong_corpus import Morph, Word, WORD_ID_PTN
+
+
+#############
+# functions #
+#############
+def _attach_missing_symbol(word: Word):
+    """
+    attach missing symbol
+    Args:
+        word:  Word object
+    """
+    raw_word = word.raw
+    raw_morph = ''.join([_.lex for _ in word.morphs])
+    if not raw_word.startswith(raw_morph) or len(raw_word) != len(raw_morph)+1:
+        return
+    last_symbol = raw_word[-1]
+    if last_symbol == '.' and word.morphs[-1].tag == 'EC':
+        word.morphs.append(Morph('.', 'SF'))
+    elif last_symbol == ',':
+        word.morphs.append(Morph(',', 'SP'))
+    elif last_symbol == '"':
+        word.morphs.append(Morph('"', 'SS'))
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    file_name = os.path.basename(sys.stdin.name)
+    for line_num, line in enumerate(sys.stdin, start=1):
+        line = line.rstrip('\r\n')
+        if not WORD_ID_PTN.match(line):
+            print(line)
+            continue
+        word = Word.parse(line, file_name, line_num)
+        _attach_missing_symbol(word)
+        print(word)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='fix final symbol errors on Sejong corpus')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'rt')
+    if args.output:
+        sys.stdout = open(args.output, 'wt')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/recover_english_case.py b/munjong/bin/recover_english_case.py
new file mode 100755
index 0000000..5a0dd9d
--- /dev/null
+++ b/munjong/bin/recover_english_case.py
@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+recover cases of English letters in Sejong corpus
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import copy
+import logging
+import os
+import re
+import sys
+
+from khaiii.munjong.sejong_corpus import Word, WORD_ID_PTN
+
+
+#############
+# functions #
+#############
+def _recover(word: Word):
+    """
+    recover cases
+    Args:
+        word:  Word object
+    """
+    word_letters = [_ for _ in word.raw if re.match(r'[a-zA-Z]', _)]
+    letter_idx = -1
+    is_recovered = False
+    word_copy = copy.deepcopy(word)
+    for morph in word_copy.morphs:
+        for idx, char in enumerate(morph.lex):
+            if not re.match(r'[a-zA-Z]', char):
+                continue
+            letter_idx += 1
+            if word_letters[letter_idx] == char:
+                continue
+            morph.lex = morph.lex[:idx] + word_letters[letter_idx] + morph.lex[idx+1:]
+            is_recovered = True
+    if is_recovered:
+        logging.info('%s  =>  %s', str(word), word_copy.morph_str())
+        word.morphs = word_copy.morphs
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    file_name = os.path.basename(sys.stdin.name)
+    for line_num, line in enumerate(sys.stdin, start=1):
+        line = line.rstrip('\r\n')
+        if not WORD_ID_PTN.match(line):
+            print(line)
+            continue
+        word = Word.parse(line, file_name, line_num)
+        try:
+            _recover(word)
+        except IndexError as idx_err:
+            logging.error('%s(%d): %s: %s', file_name, line_num, idx_err, word)
+        print(word)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='recover cases of English letters in Sejong corpus')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/recover_raw_morph_mismatch.py b/munjong/bin/recover_raw_morph_mismatch.py
new file mode 100755
index 0000000..2bcb136
--- /dev/null
+++ b/munjong/bin/recover_raw_morph_mismatch.py
@@ -0,0 +1,101 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+어절의 원문과 형태소 분석 결과의 문자가 정규화하면 같지만 코드가 다른 경우 원문의 문자로 복원
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import logging
+import os
+import sys
+
+from khaiii.munjong.sejong_corpus import Morph, ParseError, Word, WORD_ID_PTN
+
+
+#############
+# functions #
+#############
+def _recover(line: str) -> str:
+    """
+    문자를 복원한다.
+    Args:
+        line:  어절 라인
+    Returns:
+        복원된 라인
+    """
+    wid, raw, morphs_str = line.split('\t')
+    raw_idx = 0
+    morphs = []
+    for token_str in morphs_str.split(' + '):
+        morph = Morph.parse(token_str)
+        lex = []
+        for _ in range(len(morph.lex)):
+            try:
+                lex.append(raw[raw_idx])
+                raw_idx += 1
+            except IndexError as idx_err:
+                logging.error(line)
+                raise idx_err
+        morph.lex = ''.join(lex)
+        morphs.append(morph)
+    morphs_new = ' + '.join([str(m) for m in morphs])
+    logging.debug('%s\t%s\t%s  =>  %s', wid, raw, morphs_str, morphs_new)
+    return '{}\t{}\t{}'.format(wid, raw, morphs_new)
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    file_name = os.path.basename(sys.stdin.name)
+    for line_num, line in enumerate(sys.stdin, start=1):
+        line = line.rstrip('\r\n')
+        if not WORD_ID_PTN.match(line):
+            print(line)
+            continue
+        try:
+            Word.parse(line, file_name, line_num)
+        except ParseError as par_err:
+            if 'raw-morph mismatch' in str(par_err):
+                line = _recover(line)
+            else:
+                raise par_err
+        print(line)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='어절의 원문과 형태소 분석 결과의 문자가 정규화하면 같지만 코드가 다른 경우'
+                                        ' 원문의 문자로 복원')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'r', encoding='UTF-8')
+    if args.output:
+        sys.stdout = open(args.output, 'w', encoding='UTF-8')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/recover_wide_quotation.py b/munjong/bin/recover_wide_quotation.py
new file mode 100755
index 0000000..0ee0792
--- /dev/null
+++ b/munjong/bin/recover_wide_quotation.py
@@ -0,0 +1,108 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+recover wide char quotations in Sejong corpus
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2019-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import logging
+import os
+import sys
+
+from khaiii.munjong.sejong_corpus import Word, WORD_ID_PTN
+
+
+#############
+# constants #
+#############
+_QUOT_NORM = {
+    '"': '"',
+    '“': '"',
+    '”': '"',
+    "'": "'",
+    "‘": "'",
+    "’": "'",
+    "`": "'",
+}
+
+
+#############
+# functions #
+#############
+def _recover(word: Word):
+    """
+    recover wide char quotations
+    Args:
+        word:  Word object
+    """
+    word_quots = [_ for _ in word.raw if _ in _QUOT_NORM]
+    morph_quots = []
+    for idx, morph in enumerate(word.morphs):
+        if morph.tag != 'SS' or morph.lex not in _QUOT_NORM:
+            continue
+        morph_quots.append((idx, morph))
+        quot_idx = len(morph_quots)-1
+        if len(word_quots) <= quot_idx or _QUOT_NORM[word_quots[quot_idx]] != _QUOT_NORM[morph.lex]:
+            logging.error('%d-th quots are different: %s', quot_idx+1, word)
+            return
+    if len(word_quots) != len(morph_quots):
+        morph_quots = [_ for _ in word.morph_str() if _ in _QUOT_NORM]
+        if word_quots != morph_quots:
+            logging.error('number of quots are different: %s', word)
+        return
+    for word_char, (idx, morph) in zip(word_quots, morph_quots):
+        if word_char == morph.lex:
+            continue
+        morph.lex = word_char
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    file_name = os.path.basename(sys.stdin.name)
+    for line_num, line in enumerate(sys.stdin, start=1):
+        line = line.rstrip('\r\n')
+        if not WORD_ID_PTN.match(line):
+            print(line)
+            continue
+        word = Word.parse(line, file_name, line_num)
+        _recover(word)
+        print(word)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='recover wide char quotations in Sejong corpus')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'rt')
+    if args.output:
+        sys.stdout = open(args.output, 'wt')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/munjong/bin/remove_sejong_period_error.py b/munjong/bin/remove_sejong_period_error.py
new file mode 100755
index 0000000..59ffd86
--- /dev/null
+++ b/munjong/bin/remove_sejong_period_error.py
@@ -0,0 +1,120 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+
+"""
+remove wrong sentence breaking marks after period error eojeol
+__author__ = 'Jamie (jamie.lim@kakaocorp.com)'
+__copyright__ = 'Copyright (C) 2017-, Kakao Corp. All rights reserved.'
+"""
+
+
+###########
+# imports #
+###########
+from argparse import ArgumentParser
+import logging
+import os
+import re
+import sys
+from typing import TextIO, Tuple
+
+from khaiii.munjong.sejong_corpus import Morph, WORD_ID_PTN
+
+
+#############
+# functions #
+#############
+def _get_three_lines(fin: TextIO) -> Tuple[str, str, str]:
+    """
+    get three lines tuple from file (generator)
+    Args:
+        fin:  input file
+    Yields:
+        prev. prev. line
+        prev. line
+        curr. line
+    """
+    prev_prev_line = fin.readline().rstrip('\r\n')
+    prev_line = fin.readline().rstrip('\r\n')
+    # print first two lines
+    print(prev_prev_line)
+    print(prev_line)
+    for curr_line in fin:
+        curr_line = curr_line.rstrip('\r\n')
+        yield prev_prev_line, prev_line, curr_line
+        prev_prev_line = prev_line
+        prev_line = curr_line
+
+
+def _is_known_period_error_eojeol(line: str) -> bool:
+    """
+    알려진 특정 문장분리 오류를 포함하는 어절인 지 여부
+    Args:
+        line:  line (eojeol)
+    Returns:
+        whether has error or not
+    """
+    cols = line.split('\t')
+    if len(cols) != 3 or not WORD_ID_PTN.match(cols[0]):
+        return False
+    if '/SF + ' not in cols[2] or re.match(r'.+/EF \+ ./SF$', cols[2]):
+        return False
+    if re.match(r'.+/SF \+ [\'"’”]/SS$', cols[2]):
+        return False
+    morphs = [Morph.parse(_) for _ in cols[2].split(' + ')]
+    tags_str = '+'.join([_.tag for _ in morphs])
+    if 'SN+SF+SN' in tags_str and not tags_str.endswith('+SF'):
+        # 4.6판: 4/SN + ./SF + 6/SN + 판/NNB
+        if 'XSN+SF+SN' not in tags_str:
+            return True
+    elif 'SL+SF+SL' in tags_str and not tags_str.endswith('+SF'):
+        # S.M.오너: S/SL + ./SF + M/SL + ./SF + 오너/NNG
+        return True
+    return False
+
+
+def run():
+    """
+    run function which is the start point of program
+    """
+    file_name = os.path.basename(sys.stdin.name)
+    for line_num, (prev_prev_line, prev_line, curr_line) in enumerate(_get_three_lines(sys.stdin),
+                                                                      start=1):
+        if curr_line == '</p>' and _is_known_period_error_eojeol(prev_line):
+            continue
+        elif prev_line == '</p>' and curr_line == '<p>' and \
+                _is_known_period_error_eojeol(prev_prev_line):
+            logging.info('%s:%d\t%s', file_name, line_num, prev_prev_line)
+            continue
+        print(curr_line)
+
+
+########
+# main #
+########
+def main():
+    """
+    main function processes only argument parsing
+    """
+    parser = ArgumentParser(description='remove wrong sentence breaking marks after'
+                                        ' period error eojeol')
+    parser.add_argument('--input', help='input file <default: stdin>', metavar='FILE')
+    parser.add_argument('--output', help='output file <default: stdout>', metavar='FILE')
+    parser.add_argument('--debug', help='enable debug', action='store_true')
+    args = parser.parse_args()
+
+    if args.input:
+        sys.stdin = open(args.input, 'rt')
+    if args.output:
+        sys.stdout = open(args.output, 'wt')
+    if args.debug:
+        logging.basicConfig(level=logging.DEBUG)
+    else:
+        logging.basicConfig(level=logging.INFO)
+
+    run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/src/main/python/khaiii/resource/jaso.py b/src/main/python/khaiii/resource/jaso.py
index 1f5e184..7b667b8 100644
--- a/src/main/python/khaiii/resource/jaso.py
+++ b/src/main/python/khaiii/resource/jaso.py
@@ -17,21 +17,63 @@
 #############
 # constants #
 #############
-_FIRST = ['\u3131', '\u3132', '\u3134', '\u3137', '\u3138',    # 초성
-          '\u3139', '\u3141', '\u3142', '\u3143', '\u3145',
-          '\u3146', '\u3147', '\u3148', '\u3149', '\u314a',
-          '\u314b', '\u314c', '\u314d', '\u314e']
-_MIDDLE = ['\u314f', '\u3150', '\u3151', '\u3152', '\u3153',    # 중성
-           '\u3154', '\u3155', '\u3156', '\u3157', '\u3158',
-           '\u3159', '\u315a', '\u315b', '\u315c', '\u315d',
-           '\u315e', '\u315f', '\u3160', '\u3161', '\u3162',
-           '\u3163']
-_LAST = ['\u3131', '\u3132', '\u3133', '\u3134', '\u3135',    # 종성
-         '\u3136', '\u3137', '\u3139', '\u313a', '\u313b',
-         '\u313c', '\u313d', '\u313e', '\u313f', '\u3140',
-         '\u3141', '\u3142', '\u3144', '\u3145', '\u3146',
-         '\u3147', '\u3148', '\u314a', '\u314b', '\u314c',
-         '\u314d', '\u314e']
+# 한글 자모 호환 영역 (초성과 종성이 같음. 두벌식 키보드로 입력할 때 들어가는 코드)
+_FIRST_COMPAT = ['\u3131', '\u3132', '\u3134', '\u3137', '\u3138',    # 초성
+                 '\u3139', '\u3141', '\u3142', '\u3143', '\u3145',
+                 '\u3146', '\u3147', '\u3148', '\u3149', '\u314a',
+                 '\u314b', '\u314c', '\u314d', '\u314e']
+_MIDDLE_COMPAT = ['\u314f', '\u3150', '\u3151', '\u3152', '\u3153',    # 중성
+                  '\u3154', '\u3155', '\u3156', '\u3157', '\u3158',
+                  '\u3159', '\u315a', '\u315b', '\u315c', '\u315d',
+                  '\u315e', '\u315f', '\u3160', '\u3161', '\u3162',
+                  '\u3163']
+_LAST_COMPAT = ['\u3131', '\u3132', '\u3133', '\u3134', '\u3135',    # 종성
+                '\u3136', '\u3137', '\u3139', '\u313a', '\u313b',
+                '\u313c', '\u313d', '\u313e', '\u313f', '\u3140',
+                '\u3141', '\u3142', '\u3144', '\u3145', '\u3146',
+                '\u3147', '\u3148', '\u314a', '\u314b', '\u314c',
+                '\u314d', '\u314e']
+_ALL_COMPAT = _FIRST_COMPAT + _MIDDLE_COMPAT + _LAST_COMPAT
+
+# 한글 자모 영역 (초성과 종성이 다름. 세종 코퍼스에서 사용한 코드)
+_FIRST_JAMO = ['\u1100', '\u1101', '\u1102', '\u1103', '\u1104',    # 초성
+               '\u1105', '\u1106', '\u1107', '\u1108', '\u1109',
+               '\u110a', '\u110b', '\u110c', '\u110d', '\u110e',
+               '\u110f', '\u1110', '\u1111', '\u1112']
+_MIDDLE_JAMO = ['\u1161', '\u1162', '\u1163', '\u1164', '\u1165',    # 중성
+                '\u1166', '\u1167', '\u1168', '\u1169', '\u116a',
+                '\u116b', '\u116c', '\u116d', '\u116e', '\u116f',
+                '\u1170', '\u1171', '\u1172', '\u1173', '\u1174',
+                '\u1175']
+_LAST_JAMO = ['\u11a8', '\u11a9', '\u11aa', '\u11ab', '\u11ac',    # 종성
+              '\u11ad', '\u11ae', '\u11af', '\u11b0', '\u11b1',
+              '\u11b2', '\u11b3', '\u11b4', '\u11b5', '\u11b6',
+              '\u11b7', '\u11b8', '\u11b9', '\u11ba', '\u11bb',
+              '\u11bc', '\u11bd', '\u11be', '\u11bf', '\u11c0',
+              '\u11c1', '\u11c2']
+_ALL_JAMO = _FIRST_JAMO + _MIDDLE_JAMO + _LAST_JAMO
+_ALL_JAMO_SET = set(_ALL_JAMO)
+_JAMO_TO_COMPAT = dict(zip(_ALL_JAMO, _ALL_COMPAT))
+
+# 반각 자모 영역 (호환 영역과 비슷하게 초성과 종성이 같으나 글자 폭이 절반인 코드)
+_FIRST_HALFWIDTH = ['\uffa1', '\uffa2', '\uffa4', '\uffa7', '\uffa8',    # 초성
+                    '\uffa9', '\uffb1', '\uffb2', '\uffb3', '\uffb5',
+                    '\uffb6', '\uffb7', '\uffb8', '\uffb9', '\uffba',
+                    '\uffbb', '\uffbc', '\uffbd', '\uffbe']
+_MIDDLE_HALFWIDTH = ['\uffc2', '\uffc3', '\uffc4', '\uffc5', '\uffc6',    # 중성
+                     '\uffc7', '\uffca', '\uffcb', '\uffcc', '\uffcd',
+                     '\uffce', '\uffcf', '\uffd2', '\uffd3', '\uffd4',
+                     '\uffd5', '\uffd6', '\uffd7', '\uffda', '\uffdb',
+                     '\uffdc']
+_LAST_HALFWIDTH = ['\uffa1', '\uffa2', '\uffa3', '\uffa4', '\uffa5',    # 종성
+                   '\uffa6', '\uffa7', '\uffa9', '\uffaa', '\uffab',
+                   '\uffac', '\uffad', '\uffae', '\uffaf', '\uffb0',
+                   '\uffb1', '\uffb2', '\uffb4', '\uffb5', '\uffb6',
+                   '\uffb7', '\uffb8', '\uffba', '\uffbb', '\uffbc',
+                   '\uffbd', '\uffbe']
+_ALL_HALFWIDTH = _FIRST_HALFWIDTH + _MIDDLE_HALFWIDTH + _LAST_HALFWIDTH
+_ALL_HALFWIDTH_SET = set(_ALL_HALFWIDTH)
+_HALFWIDTH_TO_COMPAT = dict(zip(_ALL_HALFWIDTH, _ALL_COMPAT))
 
 
 #############
@@ -54,12 +96,12 @@ def _decomp_char(char: str) -> Tuple[str, str, str]:
     first_idx = first_start // 21
     middle_idx = first_start % 21
 
-    first = _FIRST[first_idx]
-    middle = _MIDDLE[middle_idx]
+    first = _FIRST_COMPAT[first_idx]
+    middle = _MIDDLE_COMPAT[middle_idx]
     if not last_idx:
         return first, middle
 
-    last = _LAST[last_idx-1]
+    last = _LAST_COMPAT[last_idx-1]
     return first, middle, last
 
 
@@ -71,6 +113,9 @@ def decompose(text: str) -> str:
     Returns:
         자소 분해된 텍스트
     """
+    if not text:
+        return text
+
     decomposed = []
     for char in text:
         code = ord(char)
@@ -79,3 +124,25 @@ def decompose(text: str) -> str:
         else:
             decomposed.extend(_decomp_char(char))
     return ''.join(decomposed)
+
+
+def norm_compat(text: str) -> str:
+    """
+    유니코드 내 한글 자소를 호환 영역으로 정규화한다.
+    Args:
+        text:  한글 텍스트
+    Returns:
+        자소가 호환 영역으로 정규화된 텍스트
+    """
+    if not text:
+        return text
+
+    normalized = []
+    for char in text:
+        if char in _ALL_JAMO_SET:
+            normalized.append(_JAMO_TO_COMPAT[char])
+        elif char in _ALL_HALFWIDTH_SET:
+            normalized.append(_HALFWIDTH_TO_COMPAT[char])
+        else:
+            normalized.append(char)
+    return ''.join(normalized)

From 762600555835cb954a327d8bfabc063611c2a1f3 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 28 Jan 2019 20:57:56 +0900
Subject: [PATCH 35/41] =?UTF-8?q?train/bin=20=EC=95=84=EB=9E=98=20?=
 =?UTF-8?q?=EC=8A=A4=ED=81=AC=EB=A6=BD=ED=8A=B8=EB=A5=BC=20train=EC=9C=BC?=
 =?UTF-8?q?=EB=A1=9C=20=EC=9D=B4=EB=8F=99=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 train/{bin => }/extract_errpatch.py       |  0
 train/{bin => }/extract_preanal.py        |  0
 train/{bin => }/hd_validate_errpatch.bash | 18 +++++++++---------
 train/{bin => }/make_vocab.py             |  0
 train/{bin => }/map_char_to_tag.py        |  0
 train/{bin => }/pickle_model.py           |  0
 train/{bin => }/split_corpus.py           |  0
 train/{bin => }/tag.py                    |  0
 train/{bin => }/train.py                  |  0
 train/{bin => }/validate_errpatch.py      |  0
 10 files changed, 9 insertions(+), 9 deletions(-)
 rename train/{bin => }/extract_errpatch.py (100%)
 rename train/{bin => }/extract_preanal.py (100%)
 rename train/{bin => }/hd_validate_errpatch.bash (91%)
 rename train/{bin => }/make_vocab.py (100%)
 rename train/{bin => }/map_char_to_tag.py (100%)
 rename train/{bin => }/pickle_model.py (100%)
 rename train/{bin => }/split_corpus.py (100%)
 rename train/{bin => }/tag.py (100%)
 rename train/{bin => }/train.py (100%)
 rename train/{bin => }/validate_errpatch.py (100%)

diff --git a/train/bin/extract_errpatch.py b/train/extract_errpatch.py
similarity index 100%
rename from train/bin/extract_errpatch.py
rename to train/extract_errpatch.py
diff --git a/train/bin/extract_preanal.py b/train/extract_preanal.py
similarity index 100%
rename from train/bin/extract_preanal.py
rename to train/extract_preanal.py
diff --git a/train/bin/hd_validate_errpatch.bash b/train/hd_validate_errpatch.bash
similarity index 91%
rename from train/bin/hd_validate_errpatch.bash
rename to train/hd_validate_errpatch.bash
index 3042ef8..c4b4bb2 100755
--- a/train/bin/hd_validate_errpatch.bash
+++ b/train/hd_validate_errpatch.bash
@@ -12,9 +12,9 @@ function print_usage() {
     echo "  -h, --help        show this help message and exit"
     echo "  -i FILE           input file"
     echo "  -c DIR            corpus dir"
-    echo "  --rsc-src=DIR     <default: ../../rsc/src>"
-    echo "  --lib-path=FILE   <default: ../../build/lib/libkhaiii.so>"
-    echo "  --rsc-dir=DIR     <default: ../../build/share/khaiii>"
+    echo "  --rsc-src=DIR     <default: ../rsc/src>"
+    echo "  --lib-path=FILE   <default: ../build/lib/libkhaiii.so>"
+    echo "  --rsc-dir=DIR     <default: ../build/share/khaiii>"
     echo "  --num-mapper=NUM  <default: 1000>"
     if [ -z "${msg}" ]; then
         exit 0
@@ -84,13 +84,13 @@ function parse_args() {
     fi
 
     if [ -z "${RSC_SRC}" ]; then
-        RSC_SRC=../../rsc/src
+        RSC_SRC=../rsc/src
     fi
     if [ -z "${LIB_PATH}" ]; then
-        LIB_PATH=../../build/lib/libkhaiii.so
+        LIB_PATH=../build/lib/libkhaiii.so
     fi
     if [ -z "${RSC_DIR}" ]; then
-        RSC_DIR=../../build/share/khaiii
+        RSC_DIR=../build/share/khaiii
     fi
     if [ -z "${NUM_MAPPER}" ]; then
         NUM_MAPPER=1000
@@ -135,7 +135,7 @@ function cache_files() {
     hadoop fs -test -e ${CACHE_DIR} && hadoop fs -rm -skipTrash -r ${CACHE_DIR}
     hadoop fs -mkdir -p ${CACHE_DIR}
 
-    hadoop fs -put ../../src/main/python/khaiii ${CACHE_DIR}
+    hadoop fs -put ../src/main/python/khaiii ${CACHE_DIR}
     hadoop fs -mkdir -p ${CACHE_DIR}/khaiii/lib
     hadoop fs -put ${LIB_PATH} ${CACHE_DIR}/khaiii/lib
 
@@ -188,7 +188,7 @@ parse_args $@
 cd "$(dirname $0)"
 init_envs
 
-# split_input
-# cache_files
+split_input
+cache_files
 run_hadoop
 del_temp
diff --git a/train/bin/make_vocab.py b/train/make_vocab.py
similarity index 100%
rename from train/bin/make_vocab.py
rename to train/make_vocab.py
diff --git a/train/bin/map_char_to_tag.py b/train/map_char_to_tag.py
similarity index 100%
rename from train/bin/map_char_to_tag.py
rename to train/map_char_to_tag.py
diff --git a/train/bin/pickle_model.py b/train/pickle_model.py
similarity index 100%
rename from train/bin/pickle_model.py
rename to train/pickle_model.py
diff --git a/train/bin/split_corpus.py b/train/split_corpus.py
similarity index 100%
rename from train/bin/split_corpus.py
rename to train/split_corpus.py
diff --git a/train/bin/tag.py b/train/tag.py
similarity index 100%
rename from train/bin/tag.py
rename to train/tag.py
diff --git a/train/bin/train.py b/train/train.py
similarity index 100%
rename from train/bin/train.py
rename to train/train.py
diff --git a/train/bin/validate_errpatch.py b/train/validate_errpatch.py
similarity index 100%
rename from train/bin/validate_errpatch.py
rename to train/validate_errpatch.py

From 5e572b28b8fde2a570e510bf131a7e17195ae0b9 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Mon, 28 Jan 2019 20:59:15 +0900
Subject: [PATCH 36/41] =?UTF-8?q?munjong/bin=20=EC=95=84=EB=9E=98=20?=
 =?UTF-8?q?=EC=8A=A4=ED=81=AC=EB=A6=BD=ED=8A=B8=EB=A5=BC=20munjong?=
 =?UTF-8?q?=EC=9C=BC=EB=A1=9C=20=EC=9D=B4=EB=8F=99=20#3?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 munjong/{bin => }/apply_patch.py                | 0
 munjong/{bin => }/convert_jamo_to_compat.py     | 0
 munjong/{bin => }/detect_sejong_period_error.py | 0
 munjong/{bin => }/fix_final_symbol_error.py     | 0
 munjong/{bin => }/make_patch.py                 | 0
 munjong/{bin => }/recover_english_case.py       | 0
 munjong/{bin => }/recover_raw_morph_mismatch.py | 0
 munjong/{bin => }/recover_wide_quotation.py     | 0
 munjong/{bin => }/remove_sejong_period_error.py | 0
 9 files changed, 0 insertions(+), 0 deletions(-)
 rename munjong/{bin => }/apply_patch.py (100%)
 rename munjong/{bin => }/convert_jamo_to_compat.py (100%)
 rename munjong/{bin => }/detect_sejong_period_error.py (100%)
 rename munjong/{bin => }/fix_final_symbol_error.py (100%)
 rename munjong/{bin => }/make_patch.py (100%)
 rename munjong/{bin => }/recover_english_case.py (100%)
 rename munjong/{bin => }/recover_raw_morph_mismatch.py (100%)
 rename munjong/{bin => }/recover_wide_quotation.py (100%)
 rename munjong/{bin => }/remove_sejong_period_error.py (100%)

diff --git a/munjong/bin/apply_patch.py b/munjong/apply_patch.py
similarity index 100%
rename from munjong/bin/apply_patch.py
rename to munjong/apply_patch.py
diff --git a/munjong/bin/convert_jamo_to_compat.py b/munjong/convert_jamo_to_compat.py
similarity index 100%
rename from munjong/bin/convert_jamo_to_compat.py
rename to munjong/convert_jamo_to_compat.py
diff --git a/munjong/bin/detect_sejong_period_error.py b/munjong/detect_sejong_period_error.py
similarity index 100%
rename from munjong/bin/detect_sejong_period_error.py
rename to munjong/detect_sejong_period_error.py
diff --git a/munjong/bin/fix_final_symbol_error.py b/munjong/fix_final_symbol_error.py
similarity index 100%
rename from munjong/bin/fix_final_symbol_error.py
rename to munjong/fix_final_symbol_error.py
diff --git a/munjong/bin/make_patch.py b/munjong/make_patch.py
similarity index 100%
rename from munjong/bin/make_patch.py
rename to munjong/make_patch.py
diff --git a/munjong/bin/recover_english_case.py b/munjong/recover_english_case.py
similarity index 100%
rename from munjong/bin/recover_english_case.py
rename to munjong/recover_english_case.py
diff --git a/munjong/bin/recover_raw_morph_mismatch.py b/munjong/recover_raw_morph_mismatch.py
similarity index 100%
rename from munjong/bin/recover_raw_morph_mismatch.py
rename to munjong/recover_raw_morph_mismatch.py
diff --git a/munjong/bin/recover_wide_quotation.py b/munjong/recover_wide_quotation.py
similarity index 100%
rename from munjong/bin/recover_wide_quotation.py
rename to munjong/recover_wide_quotation.py
diff --git a/munjong/bin/remove_sejong_period_error.py b/munjong/remove_sejong_period_error.py
similarity index 100%
rename from munjong/bin/remove_sejong_period_error.py
rename to munjong/remove_sejong_period_error.py

From b9f00010f628b67307fdeb28ec24fd12013c78b6 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Tue, 29 Jan 2019 10:40:55 +0900
Subject: [PATCH 37/41] =?UTF-8?q?=ED=95=99=EC=8A=B5=20=EC=BD=94=EB=93=9C?=
 =?UTF-8?q?=EB=A5=BC=20=EA=B8=B0=EC=A1=B4=20=EB=B0=A9=EC=8B=9D=EC=9C=BC?=
 =?UTF-8?q?=EB=A1=9C=20=ED=9A=8C=EA=B7=80=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/main/python/khaiii/train/trainer.py | 111 ++++++++++++------------
 train/train.py                          |   6 +-
 2 files changed, 58 insertions(+), 59 deletions(-)

diff --git a/src/main/python/khaiii/train/trainer.py b/src/main/python/khaiii/train/trainer.py
index e299fc1..038baee 100644
--- a/src/main/python/khaiii/train/trainer.py
+++ b/src/main/python/khaiii/train/trainer.py
@@ -19,7 +19,7 @@
 import os
 import pathlib
 import pprint
-from typing import Iterator, List, Tuple
+from typing import List, Tuple
 
 from tensorboardX import SummaryWriter
 import torch
@@ -27,7 +27,7 @@
 import torch.nn.functional as F
 from tqdm import tqdm
 
-from khaiii.train.dataset import PosDataset, PosSentTensor
+from khaiii.train.dataset import PosDataset
 from khaiii.train.evaluator import Evaluator
 from khaiii.train.models import CnnModel
 from khaiii.resource.resource import Resource
@@ -56,6 +56,7 @@ def __init__(self, cfg: Namespace):
         self.evaler = Evaluator()
         self._load_dataset()
         if 'step' not in cfg.__dict__:
+            setattr(cfg, 'epoch', 0)
             setattr(cfg, 'step', 0)
             setattr(cfg, 'best_step', 0)
         self.log_file = None    # tab separated log file
@@ -66,6 +67,7 @@ def __init__(self, cfg: Namespace):
         self.acc_words = []
         self.f_scores = []
         self.learning_rates = []
+        self.batch_sizes = []
 
     @classmethod
     def model_id(cls, cfg: Namespace) -> str:
@@ -86,6 +88,7 @@ def model_id(cls, cfg: Namespace) -> str:
             'lrd{}'.format(cfg.lr_decay),
             'bs{}'.format(cfg.batch_size),
             'cs{}'.format(cfg.check_step),
+            'bg{}'.format(cfg.batch_grow),
         ]
         return '.'.join(model_cfgs)
 
@@ -149,35 +152,25 @@ def _restore_prev_train(self):
             line = line.rstrip('\r\n')
             if not line:
                 continue
-            (step, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate) \
-                    = line.split('\t')
-            self.cfg.step = self.cfg.best_step = int(step) * self.cfg.check_step
+            (epoch, step, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate,
+             batch_size) = line.split('\t')
+            self.cfg.epoch = int(epoch) + 1
+            self.cfg.step = self.cfg.best_step = (int(step)+1) * self.cfg.check_step
             self.loss_trains.append(float(loss_train))
             self.loss_devs.append(float(loss_dev))
             self.acc_chars.append(float(acc_char))
             self.acc_words.append(float(acc_word))
             self.f_scores.append(float(f_score))
             self.learning_rates.append(float(learning_rate))
+            self.batch_sizes.append(int(batch_size))
             if float(f_score) > f_score_best:
                 f_score_best = float(f_score)
                 best_idx = idx
-        logging.info('---- [%d] los(trn/dev): %.4f / %.4f, acc(chr/wrd): %.4f / %.4f, ' \
-                     'f-score: %.4f, lr: %.8f ----', self.cfg.step // self.cfg.check_step,
-                     self.loss_trains[best_idx], self.loss_devs[best_idx], self.acc_chars[best_idx],
-                     self.acc_words[best_idx], self.f_scores[best_idx], self.learning_rates[-1])
-
-    @classmethod
-    def _inf_data_iterator(cls, dataset: PosDataset) -> Iterator[PosSentTensor]:
-        """
-        데이터셋을 무한히 반복하여 문장을 출력하는 제너레이터
-        Args:
-            dataset:  데이터셋
-        Yields:
-            PosSentTensor 객체
-        """
-        for _ in range(1000000):
-            for sent in dataset:
-                yield sent
+        logging.info('---- [%d|%d] los(trn/dev): %.4f / %.4f, acc(chr/wrd): %.4f / %.4f, ' \
+                     'f-score: %.4f, lr: %.8f, bs: %d ----', self.cfg.epoch,
+                     self.cfg.step // self.cfg.check_step, self.loss_trains[best_idx],
+                     self.loss_devs[best_idx], self.acc_chars[best_idx], self.acc_words[best_idx],
+                     self.f_scores[best_idx], self.learning_rates[-1], self.batch_sizes[-1])
 
     def train(self):
         """
@@ -193,12 +186,10 @@ def train(self):
         pathlib.Path(self.cfg.out_dir).mkdir(parents=True, exist_ok=True)
         self.log_file = open('{}/log.tsv'.format(self.cfg.out_dir), 'at')
         self.sum_wrt = SummaryWriter(self.cfg.out_dir)
-        check_start = (1 if self.cfg.step == 0 else (self.cfg.step // self.cfg.check_step + 1))
         patience = self.cfg.patience
-        train_iter = self._inf_data_iterator(self.dataset_train)
-        for check_id in range(check_start, 1000000):
-            is_best = self._train_and_check(check_id, train_iter)
-            if is_best:
+        for _ in range(1000000):
+            has_best = self._train_epoch()
+            if has_best:
                 patience = self.cfg.patience
                 continue
             if patience <= 0:
@@ -210,8 +201,8 @@ def train(self):
 
         train_end = datetime.now()
         train_elapsed = self._elapsed(train_end - train_begin)
-        logging.info('}}}} training end: %s, elapsed: %s, step: %dk }}}}',
-                     self._dt_str(train_end), train_elapsed, self.cfg.step // 1000)
+        logging.info('}}}} training end: %s, elapsed: %s, epoch: %d, step: %dk }}}}',
+                     self._dt_str(train_end), train_elapsed, self.cfg.epoch, self.cfg.step // 1000)
 
     def _revert_to_best(self, is_decay_lr: bool):
         """
@@ -224,21 +215,21 @@ def _revert_to_best(self, is_decay_lr: bool):
             self.cfg.learning_rate *= self.cfg.lr_decay
         self._load_optim('{}/optim.state'.format(self.cfg.out_dir), self.cfg.learning_rate)
 
-    def _train_and_check(self, check_id: int, train_iter: Iterator[PosSentTensor]) -> bool:
+    def _train_epoch(self) -> bool:
         """
-        cfg.check_step 만큼의 step을 수행하고 evaluation을 수행한다.
-        Args:
-            check_id:  check ID
-            train_iter:  학습 데이터 iterator
+        한 epoch을 학습한다. 배치 단위는 글자 단위
         Returns:
             best f-score를 기록한 step 여부
         """
-        start_step = self.cfg.step
+        has_best = False
         loss_batch = torch.tensor(0.0)    # pylint: disable=not-callable
-        batch_size = 0
+        num_in_batch = 0
+        batch_size = self.cfg.batch_size
+        if self.cfg.batch_grow > 0:
+            batch_size = self.cfg.batch_size + self.cfg.step // self.cfg.batch_grow
         loss_trains = []
-        train_sents = tqdm(train_iter, '[{}]'.format(check_id), mininterval=1, ncols=100)
-        for train_sent in train_sents:
+        for train_sent in tqdm(self.dataset_train, 'EPOCH[{}]'.format(self.cfg.epoch),
+                               mininterval=1, ncols=100):
             train_labels, train_contexts = train_sent.to_tensor(self.cfg, self.rsc, True)
             if torch.cuda.is_available():
                 train_labels = train_labels.cuda()
@@ -252,8 +243,8 @@ def _train_and_check(self, check_id: int, train_iter: Iterator[PosSentTensor]) -
             loss_train.backward()
             loss_trains.append(loss_train.item())
             loss_batch += loss_train
-            batch_size += len(train_labels)
-            if batch_size < self.cfg.batch_size:
+            num_in_batch += len(train_labels)
+            if num_in_batch < batch_size:
                 continue
 
             self.optimizer.step()
@@ -261,30 +252,32 @@ def _train_and_check(self, check_id: int, train_iter: Iterator[PosSentTensor]) -
             self.sum_wrt.add_scalar('loss-batch', loss_batch.item(), self.cfg.step)
             self.cfg.step += 1
             loss_batch = torch.tensor(0.0)    # pylint: disable=not-callable
-            batch_size = 0
+            num_in_batch = 0
 
-            if (self.cfg.step - start_step) >= self.cfg.check_step:
-                train_sents.close()
-                break
+            if self.cfg.step % self.cfg.check_step == 0:
+                avg_loss_dev, acc_char, acc_word, f_score = self.evaluate()
+                has_best |= self._check(loss_trains, avg_loss_dev, acc_char, acc_word, f_score,
+                                        batch_size)
+            if self.cfg.batch_grow > 0 and self.cfg.step % self.cfg.batch_grow == 0:
+                batch_size = self.cfg.batch_size + self.cfg.step // self.cfg.batch_grow
 
-        avg_loss_dev, acc_char, acc_word, f_score = self.evaluate()
-        return self._check(check_id, loss_trains, avg_loss_dev, acc_char, acc_word, f_score)
+        self.cfg.epoch += 1
+        return has_best
 
-    def _check(self, check_id: int, loss_trains: List[float], avg_loss_dev: float, acc_char: float,
-               acc_word: float, f_score: float) -> bool:
+    def _check(self, loss_trains: List[float], avg_loss_dev: float, acc_char: float,
+               acc_word: float, f_score: float, batch_size: int) -> bool:
         """
         cfg.check_step번의 step마다 수행하는 체크
         Args:
-            check_id:  check ID
             loss_trains:  train 코퍼스에서 각 배치별 loss 리스트
             avg_loss_dev:  dev 코퍼스 문장 별 평균 loss
             acc_char:  음절 정확도
             acc_word:  어절 정확도
             f_score:  f-score
+            batch_size:  배치 크기
         Returns:
             현재 step이 best 성능을 나타냈는 지 여부
         """
-        assert check_id == self.cfg.step // self.cfg.check_step
         avg_loss_train = sum(loss_trains) / len(loss_trains)
         loss_trains.clear()
         self.loss_trains.append(avg_loss_train)
@@ -293,15 +286,18 @@ def _check(self, check_id: int, loss_trains: List[float], avg_loss_dev: float, a
         self.acc_words.append(acc_word)
         self.f_scores.append(f_score)
         self.learning_rates.append(self.cfg.learning_rate)
+        self.batch_sizes.append(batch_size)
+        check_id = self.cfg.step // self.cfg.check_step - 1
         is_best = self._is_best()
         is_best_str = 'BEST' if is_best else '< {:.4f}'.format(max(self.f_scores))
-        tqdm.write('    [Los trn]  [Los dev]  [Acc chr]  [Acc wrd]  [F-score]           [LR]')
-        tqdm.write('    {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f} {:8}  {:.8f}'.format( \
-                avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score, is_best_str,
-                self.cfg.learning_rate))
-        print('{}\t{}\t{}\t{}\t{}\t{}\t{}'.format(check_id, avg_loss_train, avg_loss_dev,
-                                                  acc_char, acc_word, f_score,
-                                                  self.cfg.learning_rate), file=self.log_file)
+        tqdm.write('    [{:3d}|{:5d}]  [Los trn]  [Los dev]  [Acc chr]  [Acc wrd]  [F-score]'
+                   '           [LR]        [BS]'.format(self.cfg.epoch, check_id))
+        tqdm.write('                 {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f} {:8}  {:.8f}  {}' \
+                .format(avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score, is_best_str,
+                        self.cfg.learning_rate, batch_size))
+        print('{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}'.format( \
+                self.cfg.epoch, check_id, avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score,
+                self.cfg.learning_rate, batch_size), file=self.log_file)
         self.log_file.flush()
         self.sum_wrt.add_scalar('loss-train', avg_loss_train, check_id)
         self.sum_wrt.add_scalar('loss-dev', avg_loss_dev, check_id)
@@ -309,6 +305,7 @@ def _check(self, check_id: int, loss_trains: List[float], avg_loss_dev: float, a
         self.sum_wrt.add_scalar('acc-word', acc_word, check_id)
         self.sum_wrt.add_scalar('f-score', f_score, check_id)
         self.sum_wrt.add_scalar('learning-rate', self.cfg.learning_rate, check_id)
+        self.sum_wrt.add_scalar('batch-size', batch_size, check_id)
         return is_best
 
     def _is_best(self) -> bool:
diff --git a/train/train.py b/train/train.py
index 7684be4..ce4b4e5 100755
--- a/train/train.py
+++ b/train/train.py
@@ -57,10 +57,12 @@ def main():
                         type=float, default=0.001)
     parser.add_argument('--lr-decay', help='learning rate decay <default: 0.9>', metavar='REAL',
                         type=float, default=0.9)
-    parser.add_argument('--batch-size', help='batch size <default: 1000>', metavar='INT', type=int,
-                        default=1000)
+    parser.add_argument('--batch-size', help='batch size <default: 500>', metavar='INT', type=int,
+                        default=500)
     parser.add_argument('--check-step', help='check every N step <default: 10000>', metavar='INT',
                         type=int, default=10000)
+    parser.add_argument('--batch-grow', help='grow batch size by 1 per N step <default: 10000>',
+                        metavar='INT', type=int, default=10000)
     parser.add_argument('--patience', help='maximum patience count to revert model <default: 10>',
                         metavar='INT', type=int, default=10)
     parser.add_argument('--gpu-num', help='GPU number to use <default: 0>', metavar='INT', type=int,

From ef8dba1b32f7208ad6b188ce4ea7b528f3f9370a Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Thu, 31 Jan 2019 17:58:21 +0900
Subject: [PATCH 38/41] =?UTF-8?q?=EC=98=A4=ED=94=88=EC=86=8C=EC=8A=A4=20?=
 =?UTF-8?q?=EA=B3=A0=EC=A7=80=EB=AC=B8=20=EC=97=85=EB=8D=B0=EC=9D=B4?=
 =?UTF-8?q?=ED=8A=B8=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 NOTICE.md | 20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/NOTICE.md b/NOTICE.md
index e384491..d6d5505 100644
--- a/NOTICE.md
+++ b/NOTICE.md
@@ -97,16 +97,6 @@ Copyright 2013-2018 Niels Lohmann
 MIT License 
 
 
- **NumPy**
-
-https://github.com/numpy/numpy
-
-Copyright 2005-2018, NumPy Developers.
-
-
-BSD 3-Clause "New" or "Revised" License 
-
-
  **PyTorch**
 
 https://github.com/pytorch/pytorch
@@ -124,6 +114,16 @@ https://github.com/gabime/spdlog
 Copyright 2016 Gabi Melman
 
 
+MIT License 
+
+
+ **tensorboardX**
+
+https://github.com/lanpa/tensorboardX
+
+Copyright 2017 Tzu-Wei Huang
+
+
 MIT License 
 
 

From 2b24c8962bcec8e8b1508ca135cfc9ae8fc4f424 Mon Sep 17 00:00:00 2001
From: Jamie <jamie.lim@kakaocorp.com>
Date: Sun, 3 Feb 2019 13:08:44 +0900
Subject: [PATCH 39/41] =?UTF-8?q?=EC=84=B1=EB=8A=A5=EC=A0=80=ED=95=98?=
 =?UTF-8?q?=EC=9D=98=20=EC=9B=90=EC=9D=B8=EC=9D=84=20=ED=95=B4=EA=B2=B0?=
 =?UTF-8?q?=ED=95=98=EA=B3=A0=20epoch=20=EB=8B=A8=EC=9C=84=EB=A1=9C=20?=
 =?UTF-8?q?=EB=8B=A8=EC=88=9C=ED=99=94=ED=95=9C=20=EC=BD=94=EB=93=9C=20?=
 =?UTF-8?q?=EB=8B=A4=EC=8B=9C=20=EC=A0=81=EC=9A=A9=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/main/python/khaiii/train/trainer.py | 166 +++++++++++-------------
 1 file changed, 76 insertions(+), 90 deletions(-)

diff --git a/src/main/python/khaiii/train/trainer.py b/src/main/python/khaiii/train/trainer.py
index 038baee..6bc2f37 100644
--- a/src/main/python/khaiii/train/trainer.py
+++ b/src/main/python/khaiii/train/trainer.py
@@ -55,10 +55,9 @@ def __init__(self, cfg: Namespace):
         self.criterion = nn.CrossEntropyLoss()
         self.evaler = Evaluator()
         self._load_dataset()
-        if 'step' not in cfg.__dict__:
+        if 'epoch' not in cfg.__dict__:
             setattr(cfg, 'epoch', 0)
-            setattr(cfg, 'step', 0)
-            setattr(cfg, 'best_step', 0)
+            setattr(cfg, 'best_epoch', 0)
         self.log_file = None    # tab separated log file
         self.sum_wrt = None    # tensorboard summary writer
         self.loss_trains = []
@@ -67,7 +66,6 @@ def __init__(self, cfg: Namespace):
         self.acc_words = []
         self.f_scores = []
         self.learning_rates = []
-        self.batch_sizes = []
 
     @classmethod
     def model_id(cls, cfg: Namespace) -> str:
@@ -87,8 +85,6 @@ def model_id(cls, cfg: Namespace) -> str:
             'lr{}'.format(cfg.learning_rate),
             'lrd{}'.format(cfg.lr_decay),
             'bs{}'.format(cfg.batch_size),
-            'cs{}'.format(cfg.check_step),
-            'bg{}'.format(cfg.batch_grow),
         ]
         return '.'.join(model_cfgs)
 
@@ -99,6 +95,9 @@ def _load_dataset(self):
         dataset_dev_path = '{}.dev'.format(self.cfg.in_pfx)
         self.dataset_dev = PosDataset(self.cfg, self.rsc.restore_dic,
                                       open(dataset_dev_path, 'r', encoding='UTF-8'))
+        dataset_test_path = '{}.test'.format(self.cfg.in_pfx)
+        self.dataset_test = PosDataset(self.cfg, self.rsc.restore_dic,
+                                       open(dataset_test_path, 'r', encoding='UTF-8'))
         dataset_train_path = '{}.train'.format(self.cfg.in_pfx)
         self.dataset_train = PosDataset(self.cfg, self.rsc.restore_dic,
                                         open(dataset_train_path, 'r', encoding='UTF-8'))
@@ -152,25 +151,23 @@ def _restore_prev_train(self):
             line = line.rstrip('\r\n')
             if not line:
                 continue
-            (epoch, step, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate,
-             batch_size) = line.split('\t')
+            (epoch, loss_train, loss_dev, acc_char, acc_word, f_score, learning_rate) = \
+                    line.split('\t')
             self.cfg.epoch = int(epoch) + 1
-            self.cfg.step = self.cfg.best_step = (int(step)+1) * self.cfg.check_step
+            self.cfg.best_epoch = self.cfg.epoch
             self.loss_trains.append(float(loss_train))
             self.loss_devs.append(float(loss_dev))
             self.acc_chars.append(float(acc_char))
             self.acc_words.append(float(acc_word))
             self.f_scores.append(float(f_score))
             self.learning_rates.append(float(learning_rate))
-            self.batch_sizes.append(int(batch_size))
             if float(f_score) > f_score_best:
                 f_score_best = float(f_score)
                 best_idx = idx
-        logging.info('---- [%d|%d] los(trn/dev): %.4f / %.4f, acc(chr/wrd): %.4f / %.4f, ' \
-                     'f-score: %.4f, lr: %.8f, bs: %d ----', self.cfg.epoch,
-                     self.cfg.step // self.cfg.check_step, self.loss_trains[best_idx],
-                     self.loss_devs[best_idx], self.acc_chars[best_idx], self.acc_words[best_idx],
-                     self.f_scores[best_idx], self.learning_rates[-1], self.batch_sizes[-1])
+        logging.info('---- [%d] los(trn/dev): %.4f / %.4f, acc(chr/wrd): %.4f / %.4f, ' \
+                     'f-score: %.4f, lr: %.8f ----', self.cfg.epoch,
+                     self.loss_trains[best_idx], self.loss_devs[best_idx], self.acc_chars[best_idx],
+                     self.acc_words[best_idx], self.f_scores[best_idx], self.learning_rates[-1])
 
     def train(self):
         """
@@ -188,21 +185,25 @@ def train(self):
         self.sum_wrt = SummaryWriter(self.cfg.out_dir)
         patience = self.cfg.patience
         for _ in range(1000000):
-            has_best = self._train_epoch()
-            if has_best:
+            is_best = self._train_epoch()
+            if is_best:
                 patience = self.cfg.patience
                 continue
             if patience <= 0:
                 break
             self._revert_to_best(True)
             patience -= 1
-            tqdm.write('==== revert to check: {}, f-score: {:.4f}, patience: {} ===='.format( \
-                    self.cfg.best_step // self.cfg.check_step, max(self.f_scores), patience))
+            logging.info('==== revert to EPOCH[%d], f-score: %.4f, patience: %d ====',
+                         self.cfg.best_epoch, max(self.f_scores), patience)
 
         train_end = datetime.now()
         train_elapsed = self._elapsed(train_end - train_begin)
-        logging.info('}}}} training end: %s, elapsed: %s, epoch: %d, step: %dk }}}}',
-                     self._dt_str(train_end), train_elapsed, self.cfg.epoch, self.cfg.step // 1000)
+        logging.info('}}}} training end: %s, elapsed: %s, epoch: %s }}}}',
+                     self._dt_str(train_end), train_elapsed, self.cfg.epoch)
+
+        avg_loss, acc_char, acc_word, f_score = self.evaluate(False)
+        logging.info('==== test loss: %.4f, char acc: %.4f, word acc: %.4f, f-score: %.4f ====',
+                     avg_loss, acc_char, acc_word, f_score)
 
     def _revert_to_best(self, is_decay_lr: bool):
         """
@@ -219,64 +220,51 @@ def _train_epoch(self) -> bool:
         """
         한 epoch을 학습한다. 배치 단위는 글자 단위
         Returns:
-            best f-score를 기록한 step 여부
+            현재 epoch이 best 성능을 나타냈는 지 여부
         """
-        has_best = False
-        loss_batch = torch.tensor(0.0)    # pylint: disable=not-callable
-        num_in_batch = 0
-        batch_size = self.cfg.batch_size
-        if self.cfg.batch_grow > 0:
-            batch_size = self.cfg.batch_size + self.cfg.step // self.cfg.batch_grow
+        batches = []
         loss_trains = []
         for train_sent in tqdm(self.dataset_train, 'EPOCH[{}]'.format(self.cfg.epoch),
-                               mininterval=1, ncols=100):
+                               len(self.dataset_train), mininterval=1, ncols=100):
             train_labels, train_contexts = train_sent.to_tensor(self.cfg, self.rsc, True)
             if torch.cuda.is_available():
                 train_labels = train_labels.cuda()
                 train_contexts = train_contexts.cuda()
-                loss_batch = loss_batch.cuda()
 
             self.model.train()
             train_outputs = self.model(train_contexts)
-            train_outputs.requires_grad_()
-            loss_train = self.criterion(train_outputs, train_labels)
-            loss_train.backward()
-            loss_trains.append(loss_train.item())
-            loss_batch += loss_train
-            num_in_batch += len(train_labels)
-            if num_in_batch < batch_size:
+            batches.append((train_labels, train_outputs))
+            if sum([batch[0].size(0) for batch in batches]) < self.cfg.batch_size:
                 continue
 
+            batch_label = torch.cat([x[0] for x in batches], 0)    # pylint: disable=no-member
+            batch_output = torch.cat([x[1] for x in batches], 0)    # pylint: disable=no-member
+            batches = []
+
+            batch_output.requires_grad_()
+            loss_train = self.criterion(batch_output, batch_label)
+            loss_trains.append(loss_train.item())
+            loss_train.backward()
             self.optimizer.step()
             self.optimizer.zero_grad()
-            self.sum_wrt.add_scalar('loss-batch', loss_batch.item(), self.cfg.step)
-            self.cfg.step += 1
-            loss_batch = torch.tensor(0.0)    # pylint: disable=not-callable
-            num_in_batch = 0
-
-            if self.cfg.step % self.cfg.check_step == 0:
-                avg_loss_dev, acc_char, acc_word, f_score = self.evaluate()
-                has_best |= self._check(loss_trains, avg_loss_dev, acc_char, acc_word, f_score,
-                                        batch_size)
-            if self.cfg.batch_grow > 0 and self.cfg.step % self.cfg.batch_grow == 0:
-                batch_size = self.cfg.batch_size + self.cfg.step // self.cfg.batch_grow
 
+        avg_loss_dev, acc_char, acc_word, f_score = self.evaluate(True)
+        is_best = self._check_epoch(loss_trains, avg_loss_dev, acc_char, acc_word, f_score)
         self.cfg.epoch += 1
-        return has_best
+        return is_best
 
-    def _check(self, loss_trains: List[float], avg_loss_dev: float, acc_char: float,
-               acc_word: float, f_score: float, batch_size: int) -> bool:
+    def _check_epoch(self, loss_trains: List[float], avg_loss_dev: float, acc_char: float,
+                     acc_word: float, f_score: float) -> bool:
         """
-        cfg.check_step번의 step마다 수행하는 체크
+        매 epoch마다 수행하는 체크
         Args:
-            loss_trains:  train 코퍼스에서 각 배치별 loss 리스트
+            loss_trains:   train 코퍼스에서 각 배치별 loss 리스트
             avg_loss_dev:  dev 코퍼스 문장 별 평균 loss
             acc_char:  음절 정확도
             acc_word:  어절 정확도
             f_score:  f-score
-            batch_size:  배치 크기
         Returns:
-            현재 step이 best 성능을 나타냈는 지 여부
+            현재 epoch이 best 성능을 나타냈는 지 여부
         """
         avg_loss_train = sum(loss_trains) / len(loss_trains)
         loss_trains.clear()
@@ -286,38 +274,34 @@ def _check(self, loss_trains: List[float], avg_loss_dev: float, acc_char: float,
         self.acc_words.append(acc_word)
         self.f_scores.append(f_score)
         self.learning_rates.append(self.cfg.learning_rate)
-        self.batch_sizes.append(batch_size)
-        check_id = self.cfg.step // self.cfg.check_step - 1
         is_best = self._is_best()
         is_best_str = 'BEST' if is_best else '< {:.4f}'.format(max(self.f_scores))
-        tqdm.write('    [{:3d}|{:5d}]  [Los trn]  [Los dev]  [Acc chr]  [Acc wrd]  [F-score]'
-                   '           [LR]        [BS]'.format(self.cfg.epoch, check_id))
-        tqdm.write('                 {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f} {:8}  {:.8f}  {}' \
+        logging.info('[Los trn]  [Los dev]  [Acc chr]  [Acc wrd]  [F-score]           [LR]')
+        logging.info('{:9.4f}  {:9.4f}  {:9.4f}  {:9.4f}  {:9.4f} {:8}  {:.8f}' \
                 .format(avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score, is_best_str,
-                        self.cfg.learning_rate, batch_size))
-        print('{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}'.format( \
-                self.cfg.epoch, check_id, avg_loss_train, avg_loss_dev, acc_char, acc_word, f_score,
-                self.cfg.learning_rate, batch_size), file=self.log_file)
+                        self.cfg.learning_rate))
+        print('{}\t{}\t{}\t{}\t{}\t{}\t{}'.format(self.cfg.epoch, avg_loss_train, avg_loss_dev,
+                                                  acc_char, acc_word, f_score,
+                                                  self.cfg.learning_rate), file=self.log_file)
         self.log_file.flush()
-        self.sum_wrt.add_scalar('loss-train', avg_loss_train, check_id)
-        self.sum_wrt.add_scalar('loss-dev', avg_loss_dev, check_id)
-        self.sum_wrt.add_scalar('acc-char', acc_char, check_id)
-        self.sum_wrt.add_scalar('acc-word', acc_word, check_id)
-        self.sum_wrt.add_scalar('f-score', f_score, check_id)
-        self.sum_wrt.add_scalar('learning-rate', self.cfg.learning_rate, check_id)
-        self.sum_wrt.add_scalar('batch-size', batch_size, check_id)
+        self.sum_wrt.add_scalar('loss-train', avg_loss_train, self.cfg.epoch)
+        self.sum_wrt.add_scalar('loss-dev', avg_loss_dev, self.cfg.epoch)
+        self.sum_wrt.add_scalar('acc-char', acc_char, self.cfg.epoch)
+        self.sum_wrt.add_scalar('acc-word', acc_word, self.cfg.epoch)
+        self.sum_wrt.add_scalar('f-score', f_score, self.cfg.epoch)
+        self.sum_wrt.add_scalar('learning-rate', self.cfg.learning_rate, self.cfg.epoch)
         return is_best
 
     def _is_best(self) -> bool:
         """
-        이번 step에 가장 좋은 성능을 냈는 지 확인하고 그럴 경우 현재 상태를 저장한다.
+        이번 epoch에 가장 좋은 성능을 냈는 지 확인하고 그럴 경우 현재 상태를 저장한다.
         Returns:
-            best 여부
+            마지막 f-score의 best 여부
         """
         if len(self.f_scores) > 1 and max(self.f_scores[:-1]) >= self.f_scores[-1]:
             return False
-        # this step hits new max value
-        self.cfg.best_step = self.cfg.step
+        # this epoch hits new max value
+        self.cfg.best_epoch = self.cfg.epoch
         self.model.save('{}/model.state'.format(self.cfg.out_dir))
         self._save_optim('{}/optim.state'.format(self.cfg.out_dir))
         with open('{}/config.json'.format(self.cfg.out_dir), 'w', encoding='UTF-8') as fout:
@@ -347,31 +331,33 @@ def _load_optim(self, path: str, learning_rate: float):
         self.optimizer.load_state_dict(state_dict)
         self.optimizer.param_groups[0]['lr'] = learning_rate
 
-    def evaluate(self) -> Tuple[float, float, float, float]:
+    def evaluate(self, is_dev: bool) -> Tuple[float, float, float, float]:
         """
         evaluate f-score
+        Args:
+            is_dev:  whether evaluate on dev set or not
         Returns:
             average dev loss
             character accuracy
             word accuracy
             f-score
         """
+        dataset = self.dataset_dev if is_dev else self.dataset_test
         self.model.eval()
-        loss_devs = []
-        for dev_sent in self.dataset_dev:
+        losses = []
+        for sent in dataset:
             # 만약 spc_dropout이 1.0 이상이면 공백을 전혀 쓰지 않는 것이므로 평가 시에도 적용한다.
-            dev_labels, dev_contexts = dev_sent.to_tensor(self.cfg, self.rsc,
-                                                          self.cfg.spc_dropout >= 1.0)
+            labels, contexts = sent.to_tensor(self.cfg, self.rsc, self.cfg.spc_dropout >= 1.0)
             if torch.cuda.is_available():
-                dev_labels = dev_labels.cuda()
-                dev_contexts = dev_contexts.cuda()
-            dev_outputs = self.model(dev_contexts)
-            loss_dev = self.criterion(dev_outputs, dev_labels)
-            loss_devs.append(loss_dev.item())
-            _, predicts = F.softmax(dev_outputs, dim=1).max(1)
+                labels = labels.cuda()
+                contexts = contexts.cuda()
+            outputs = self.model(contexts)
+            loss = self.criterion(outputs, labels)
+            losses.append(loss.item())
+            _, predicts = F.softmax(outputs, dim=1).max(1)
             pred_tags = [self.rsc.vocab_out[t.item()] for t in predicts]
-            pred_sent = copy.deepcopy(dev_sent)
+            pred_sent = copy.deepcopy(sent)
             pred_sent.set_pos_result(pred_tags, self.rsc.restore_dic)
-            self.evaler.count(dev_sent, pred_sent)
-        avg_loss_dev = sum(loss_devs) / len(loss_devs)
-        return (avg_loss_dev, ) + self.evaler.evaluate()
+            self.evaler.count(sent, pred_sent)
+        avg_loss = sum(losses) / len(losses)
+        return (avg_loss, ) + self.evaler.evaluate()

From 90bd9107f35a5a17fabf5b1d4bd4d46bbf2e80e8 Mon Sep 17 00:00:00 2001
From: krikit <krikit@naver.com>
Date: Tue, 5 Feb 2019 22:28:49 +0900
Subject: [PATCH 40/41] =?UTF-8?q?--check-step,=20--batch-grow=20=EC=98=B5?=
 =?UTF-8?q?=EC=85=98=20=EC=82=AD=EC=A0=9C=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 train/train.py | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/train/train.py b/train/train.py
index ce4b4e5..8943041 100755
--- a/train/train.py
+++ b/train/train.py
@@ -28,8 +28,7 @@ def run(args: Namespace):
     Args:
         args:  program arguments (config)
     """
-    trner = Trainer(args)
-    trner.train()
+    Trainer(args).train()
 
 
 ########
@@ -59,10 +58,6 @@ def main():
                         type=float, default=0.9)
     parser.add_argument('--batch-size', help='batch size <default: 500>', metavar='INT', type=int,
                         default=500)
-    parser.add_argument('--check-step', help='check every N step <default: 10000>', metavar='INT',
-                        type=int, default=10000)
-    parser.add_argument('--batch-grow', help='grow batch size by 1 per N step <default: 10000>',
-                        metavar='INT', type=int, default=10000)
     parser.add_argument('--patience', help='maximum patience count to revert model <default: 10>',
                         metavar='INT', type=int, default=10)
     parser.add_argument('--gpu-num', help='GPU number to use <default: 0>', metavar='INT', type=int,

From ec1ef731f16795bc4acfe2900b6835130742b8f3 Mon Sep 17 00:00:00 2001
From: krikit <krikit@naver.com>
Date: Mon, 11 Feb 2019 08:05:48 +0900
Subject: [PATCH 41/41] =?UTF-8?q?=EC=88=98=EC=A0=95=ED=95=9C=20=ED=8C=A8?=
 =?UTF-8?q?=EB=94=A9=20=EA=B4=80=EB=A0=A8=20=EC=98=A4=EB=A5=98=EB=A5=BC=20?=
 =?UTF-8?q?=EC=9D=B4=EB=AF=B8=EC=A7=80=EC=97=90=20=EB=B0=98=EC=98=81=20#30?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .github/img/network.png  | Bin 85450 -> 85343 bytes
 .github/img/network.pptx | Bin 59222 -> 59223 bytes
 2 files changed, 0 insertions(+), 0 deletions(-)

diff --git a/.github/img/network.png b/.github/img/network.png
index bfdaf7c2d584b033c122e68978133dc7885af0a0..59886ce43fac28310f7212f13d08bcfdf2eced14 100644
GIT binary patch
literal 85343
zcmb?@cRbbo|Nc?-J|cS^BO$ZQ>|><NL}VnABxGlA$0pGtp+ba|8QCkTl$mTwWF<TN
zp0DoD=li{XzrTMT?ngbkIp=*|@8`Il*L6MLF^2luG?Z+V7z~Eyln&k)gCQfwU=B!=
zlfu6+9@aB}KPcRE%sns|szc~MBFy~^77T_5a|*9!;+yev<fiZ0%U?*B75%?>j$No(
z@;~}9%4&$xL{&BVGKX3#Qx9FHk#=LjY=aPwf|wAc^Z`Bz@?S9o!*$K8M;RT5P8X2W
zd{ocA@a6c`U%QPTU6adncVs$uUyGG=J<Q1XuJ3S0|GmfPM%vwO2Wi5;zd8hUG|6~y
z|N6>VZzko%{_87+#x^|p|N5P%QZ*g?zrO4>26_1Z^%A%;UT1oSf4u~Dh+c*~=HIX3
zOwJzhuUEk1Wg=+Q|NRm;_NM<2m*L(+q*GhJzq>?EP9Bn(nVGzO^X5%Le0+q(^NP;x
zXj)En6O-HB{VKl8JZ&m_5@EH?yeT>n_WwGNG69B{L(Ydeo4#1pcv4)bbZOn&-MR5=
z`qb^)x5xi%t>zaLs09Wp#cT0BHONYgkEiC{I?hIO>C&Y~J&qB*AKt&$Qdd9lb73K}
z_ru$_>X$C%eB3^D>QocktDO0~cxsKg5aQnytRyz=7kOF#bw17%>=D<Uzr4D7{>3Gd
zR~E-6CnxFI*eHX8gMDw_Y;0{MNlZ*M6i!R-YNh!`O+nFE<26su%zWf=kf9;HjPqc~
z-McKjYX=V;D80GrfY;QdC|FurO1eH=!HdJG>jbK(sQmm=PGxIrt3J0s5v+Bo<?+t=
zcgCpEM!x^&vu9j-mC$|v1eI1UR_fjvPsSYDqu1=XxH3Fd-a6@P446>)Kgpf_cPyCd
z2`hvgyXcWT+cqaQKQ6ox&gzg+tRSjD`@i9k(q{(&6|rg(fx*N+`(-QjfH)5>?7v@S
zE*Y0@6Cylr`8OEN#9HhM-0shas`se+&->U<?X6eV^GZq@4Hl{HY%FQR`QrHau$-L7
z_z$8tF&Bqe%hw?HINj1%#wk>LxBWkXG{ThIL&#M0nas-yYI<SU?)L2+>+L&ta%*bX
zyu7?5*b?q42P#NONx9rCk{$IuXl0zrdne{zKtaw=oS=RN;p<mRViFSD<?+Vl)zw3{
zBGK(CqJuE*<t2XImQ%BGKz-y8Ui+U<LiH+mKj_ciM!mMZJ^!uLR7v{|dSX)2rmik3
z7Z(>dPtOS5Ct;RB{&&>vrC8I8Ey%TxM9b)p)?BP)`8_{CH#9UPBqr9zY!eXL(t@XE
zk-)rv|6W}ReY}*li*R}h#7`^}ZigK2k^CDq9qP|7RMJkhB{avgO5?SwkGyk`uJQbF
z;JIlWd|;fSB0I#roZoLDzu&W=qjmloPmE&{PoV@nto!aH>whzGBYDU<3Xj8bPCYzZ
z*r#*)bo1?_XCbi2;NYxYduNxG8G9Z__o?h6$N!Xo-p1;hnn|Df9ZvM(pYsfw$Hr_t
z|JH-Di~m~(%*i2oWynw!n3Y8VpzwJ6DVq+z@t_FDe^0H9T;<R5LCB@;jU_s5ZS9^L
z=+osNoMMYOod5VSv33#{3Fht@!Ji|qnMp`UFd=pR8#xsfruHXu#)`(9)c&{yU=z)z
z35n`|Kao9lZhrp8cmolKoEwHa@He)?bu97eQ^B4We_uBIQuxQPSx<#`cO?%7Z*|#>
zPfi}d_)e#2mtOh61<6N&xp?JD^r=)q+li(~iF;>oQw$>~4D^m*R9BwZN!+In`%I5+
z2-YirErIld(gp`){Y43u(!d?>fIlnk9o^kCYeeX@YG~Ob4(C9ab#`{bJ2&KL#*Kgf
zPKI&w@X$DQiXv#^^%xO`fq|jaeMXm&iHY~vF+F?xoY&tjeW&p8w_!Qsmr)&{?#Y2(
z3WtjuUtCOt=$%|?=hn{C&osPnA+G$|z@^485(t^YoJr`jDL}nF7V+|>bj9`IsKA}|
ziR$kUHLhMgiW&cGprTnHv@790t%Few3R30@{B1NiJZuM9j^jJDD`>kIg9&MhyBU`H
z&xbQlfAomw<kobm?`(GT_|Kp5tKUC($TOfTsfn`m#MA@^1qI!eb>(`WsW{+%N#({Q
zKEKbYrLFA%hK-$l<wsstQ)lNXdy||o(FS=2veT7Q#?3b^KV?Q2)>ZzuavavcrSS0b
z5=lx*c6|7t9S~6a&;RKSEmd<2{j<A!VJ~oHvNgY?Bw8z%2=n;yV?wN^r$OgqcSyX6
z=}1dP)0ylCyF5K9=(7_DqKTGb+V1Y|g-NUlTF~-@7-PNlHBm_9+=w+2aVwYCvA<rd
zFKrR<Lyo%Jukq?XzgBeG)bz0O-liRgf+xAdOY?tD+({dSI%j(BTc)0eN8KtnLeIQa
zwD_l)Ej~G$r=R`wDe1NTA~pDBQaY~q@$rlQi7!q!VqUD5W6Ybno=$lVV+xhxf8J;D
zNz9=mm!>;X+}z#O&COZgY|n{^+PvXl7PF-1Oh7;Lg+=>+`sWZ&3_c7FmrqQLR%QQ>
zbkO#XBN=ix4nX|R_7yUdQnT<V;7}NrV`)48TSm)lDI*VTjQDU{2drp%{`f@V@RC(H
zP1jQt-8UzV|4m-Jo10q$0n7KX&fll1rY1H~10B<K0v~xy$ED*Z=5G{at9w$T-@@Sv
zzs<~$M5_19*<HSTc_Aa}@0A{a0%(sFlQo&#$+fv?utF#CpgU*kpV%47;L}dL2!VON
z+Fjt;vuF5rbbUEvA!p+mQZ(I6{Z$1jHLN~gHtEd_t+Y%2Z#c+2dGm$~t{<|!6;3Mo
z{{7fDfjdp_-|PQVd}{t)|6w%<n%-kX7`M4g566$A%f{m?j+{8bBwo9GK*48`I<+%G
z$$x!3H%Ynap_2c{uU|zKxX?k35J7|JxD-*<s&;@v_Znr-#r#8@(SPqv$2U27d0V(O
zzS}?YunY`gQBkDMU#qYX^wB+zjQ=E;%b`3L0cad{o68f11t#iqchb`*L&>=4P3voG
zIW(eanN0Tq8L&v%li${TlDy5t!~~`Ekm-I?B$bf3xVDZ?$FYs_vvLg|j#3Mp$TKwY
zB7QOY#al-QZ*QJq$uTke6H}eDYq07^t>2RF`ZC*4PF>9Gm46_n%5a_PPp#D>!;Xkv
zapemnHU_dCmoLp_9V=;hKcbQi;6#1y5H@9Dd2Ow=CFT%9DfSKyI3Xc&cX@GX^wG^J
zA68?(e*L<<_<9K^Dtbm|*S0yzcKzp2Wgm|d)sG)P?n+;w4GRyC?B!PQ6#Di)Gp?aQ
zm3J+gnl<d^>bFKH>^i%zULCKht{z$6n^=qyDv*!-d~S=dq8o|FU2XQn<I`s>y@=Y*
zaK+Q_lW&r4zq)q4I-J!>R{B*qgOLpz1>MrP_w(RBn}@8S@~+Aes{6zHF4cEPlXl6M
z^fSqW#eI0!xbOaw;o#(q>=ne}h8bB|V@gXU61h-6DU*P9a$8pbp&>v!uY%5;i7(w=
zSy>^(#xhu_{KYnr>Y<yPzP2=%BGqqb=Q7uE9GH0IsRluIyYF!0hoi4(KEo+Q#n!y+
z5E)K*7UiyUSt<fg$}5L^WF>w4L41CZO0<5qwFJrW#%5(w<5N3H@6Cj#2}$c3_<Wz&
z1EiQp%fzZu)rhT1*YNfI`L2j=N>vEE(}JmL-r+r~TU%QYM<Eu^UA~*3c=VKpD^Af&
z(cveoE)wMyP&{Ol%;kz9D8!zoVd@YcPWV+bnSPz=yBK9BF7@q1!Ag5$?#?(T#)MtA
z$z1uoPAUT43Q@OiF%((Xv8}JKSMBfZ^gy|1mtjC9C_nHNg6}nF)rxz_Jq#k$+pr9H
z``2vLYn#=+BPY7mFmmGgYUGwz8TNMj(j-{>9A(V1+wM}Nnv09lK|I6=-08wHG7@L%
zN3YEfOuhV|+zc6qZ%65l`v>LQr~G7Od#rgfjOWhmwTdhiGXbF<rW5DZKAN>@P-?Su
zGaK1`B~GnA<p{S6$Iv`H%$}sOC)bw9NhK>QYgy$+)OAO^-OMI1vaC$1X9Z=MGaKX&
z{BW5e<vq$T4cW^#sMTCL4+9jKbz;Z2A2=f{CldAO%Q9c_*Cat~Cj7>1;t5$<yTM&M
zUZ&#$PXVLo=Y{9y<|c1%7@s*q4d^c=H8q@5(R=)kO%RkDs^y<AKQy&#zpJVHTND!i
z87-)or`c%e61XYFzr`eeQWxxLFqdRB7UE42`dgT4?2HSjrN0MK`X^}V>>6DSd}LY}
z+3P|jb&q=FwU4&3F@yKw*O)kF@iKWGJip=P%L2l}!jFs#(6?z}ws8SenE1+3KPvcX
z=38Bk$8oZ<%-jLXp~>6$3N+o6ISP763(ub(U!WIKpu`%zmVD5itW(UHmS?2E+f!Nj
z2>(3wREwn{oQ>q$uCA4Z%CVfPDi*+|&ljaVzMZM6s(R&7_L?sRb0a2mVs@5aG~C1~
zP^5WqXb7QFRCsggu|vo2oryYEXlk`_v&gjYh360B3<WP;onpW}&LcHz=`>BO(hlP#
z!Mh@ilq4{!WMpPeOihJDp0~8LJTfb>o*k|fpLVZMCAdWFd(o$~N<HY_Txq$~D|^n)
zvDePcq&hYckM9*rH2XH5Q%rqy@2X+v!rudIf*Vih$<@w%@nYraj@|>A?n`VDy%j-$
zFF&7?mzQ_DaYGGmhvE^q^Tz-L!x#od>WD=}r@a(4w=xHUV5OH@DS4}*RnRjrd3Egq
z7?-&I<@oHHUvGbZ&a-EfP)g-IzESIEDn!yLZ(>r97_<%81Sb?*y{^!3$918})m$_N
ze?UAM!+7po?(!urg`JTj3Jh4I2Zf)~rrfr;?%FoZL{a)GlEx%@YQJ+h#CB$@G__3)
zH&GyjO)2+XX=Uyi93<S6zJfb;jE0MgD|y#xLAk($5@seficSU-Gc!AY*)S}SQ&DMH
zSg_UB)oDVylv!9jU%yW}u{_b-Q*ic3Xv|`~rC2n4YvqO}CDC27vn#oIl8^7;eE)O6
zzO$r7uEmtNvZE6&fVOPJIMrTu3dzx6jfSP}Q+Ft=^1YMbZQ>i6ydRD+9Lu1wS|#db
zscdlm_FV>>a^m9;Pa#p!mJp(Y%WdqgkD%(4)6#0Gsy3Y1KmUnDn?WPD=TY$fo?&4h
z`myg>!HG~W5K2bfGp-+kv^kniqh!G5rPA@^jMmoHm=Kr<ay`Dk$4!>A#^9TdK96w;
z`x~#t>!WVvt6FDYa$PAjK-i%7DQ)_qDpXUs!jY8O&oBDfYptBjMfpp^8!e;QrcCd4
zcjxe7u|2N#uC8?4+}yu^4xRO6EOT?O=73h*5-0wK{JAL=02&yk0)F>s555Uhynp{b
zUYj3RCZv!_``UY<X>5Pbr^uq5l8lz4%7q&ynAurJn1EDO)o~0l=(2om7tOS=hppco
z%7`H3FWfd^p7cDY{Fs5YhU}?^VEMZr@-qxq<U~j{4xFI=W-LPyLnvxFNOv+Ml8V_{
zT9377q^z>C{bRPq%AeH_d<qI|q~S!vm9E-BL2qUqVIZ1--hKYH_pxu?52BA<sCWeM
z4&netwpy1l7{=ncl(wU;H8Z5FmAw-=EmI6*Nnc+dAt52M)4mCC%J|$|EP+6dMijua
z&w~aJnHQ?w5#kUY35~0fvUJIhh|A1FP-nv=Jey1YyA%;7M!9aHXq`S|wld|k<+vz!
zywemZ%R7O5Ntf`{T5f)Zq32m=@OaKo+T4K=fJFEN1xYQQABJxId`~R!SLcz5@o^p=
zUUQs}k54?iENyFRYe#42i<*KZ0%5WI^c^wVg)iko!op5Nhp|{9IMC(gWz(!&L!mDs
zTv#vYsTTB5|K&Ztjla7m?fX5`2ZKpE`HQ^9XNeZNDSQ@GU_L!Qh33IS1}BLMZinML
zw>zIu5?8lP_O^z@h#g(Lm%$K9C^~&rkN5t4_K5}h2=XNRo*Ro4)xAvbfAQ{LWq7FZ
z{DvaO%q~yL^4Ca02SZv~T6Rv(0RX1v=H?x5-#U5KM{@_=rBSc3Q&UqTr!1>2zY|0F
z@}RCxN%!>WY!l(0YL6?9<rY=1U&keC;yD@rGgB256m)cSsA_3xb$LU_L9|SonKcoH
zpP&EY%-giPGS}HYeform5TlfsFXyWK8BE+2BWhaXU`X`*5Ltbs$@&K|b9wYRy&d-_
zr+hE-XXA4YTvRGaAawNCE;mtiyno-;5PC2rJv~ayszz0;e(zRz-jORikNoGy8^h&%
z7CE-p=E)MY_<-U_0ZxW;sy!(hQoh%zd{z$M+xSAwkr!(0i|fNO?=lscQG`EmeM15D
zBmgkxCQNaXFpK^9{n-F!pZ1?aFM-KOIB~6?kf_(|lF{ozIP1gA8BWOW-c%rsOZ+G}
zZ7f?Q6s18)R_GMOfE~Kk)~4k=TFXj7K{5F1YU-?<y~Y{ALlrI~vG?z@O@IHM>$5a^
z?4>Ax9=?}D@l8$Y_E)Z;TqnfEy>`hOA0J<*)50HyyrP=?{yiE%-o)2xkNoG)Nib(J
z<giZVY%gBCz!d+H*ntb2n4adB5(e1Nb3OG`mA4c&MF405sOydSM!7J<v-73tuFU2p
z;iRPx!FV!QcA08@VK|!TASq0Dp|afAls-o%r>1O;=;X}I1S)1R42GIiqY!dBXSkI#
zo8;HVhU?ZwYaDa#n>TMRvnAZV&1m#@R3{KRw3If-!%<pV`!Ae;fi*5sL#E~P=R+t`
z^|O={-n}~mkbG*`wO;d|LF`>QRZK^Y)tdlT|BXdW7Z))YM~N`{`uaMjPN@wPTao1P
z;ZC=cW;2a*>_A)Ue^R!{@z_}Y>p~!28{1~zpWEX}?fr>=bk_iHPt=ygodoxU2MQ+L
z$z==#l*R6I1-Eh4;u!pO9Q><+%I<MmPK5}dVRBx*VmzXsfmj+*Q#(68<*n&MED2f|
zRaI3CP<cQS@T7JU!@FN=K<(<u+ZqqIiq#O~wm3Q%0DiZT{*-^o<(Esl+Z04br+1_$
zB?1aTXpd#6g9j&vecd2GKW=XP;mvRN(vN=FTS*EY=icAcF5Jn40(;;524PX4z$A8h
z+P<Z=l^1xQks7Z<oH2yi;sTh;+uNJ%rLy5X5Q7E(T11VeH?Q@~;fT-YqI-rml<|Jv
zSeS)o6MrgJvL>|o`8p~zj*NN-W1Q@*jIMtruKU4=q2RDr8bVjmFnaSlCv?OyPj1w6
z#Y?<(@fBwZOmYmelw|{ItE=zjx)So6fg%kBGW5ZNVXveF4tZl$XqQX~7x%5)pl<x}
ztXZh&?G<!dZ_(ci<Vhk;D6TiqKX4%&eLU)xPWQ(kwG_jk^C)Gl^v4ENd_HzX@1j|9
zeC3n;eDblt4Ng*84g!FFV2e7UfbOCN_Ds%wnjCPUq{~-1U<Eqb5SBG<|8SJishYHW
zTNozJ@H+?dByc&2C8oiwKua9N)OgO(C44&N_%IPBp3u|)9;f&RJ&p<#st3q!(xH)l
zmQ+ew`h`=UFe5W_^F@hQH6MUkcJ=l1tMHJ@*{=k`&k~}o(Db?bf2e?v5DffYht_df
z*{Ru{TqtttK=*`*@co=|mxnBUY+Nu7=YJZp=Y@P}I@r0xNx4izLbbjG$cee-GY*H!
zJ`_dw*sDG|%@}5Bhu5ByxNU4B*GcxMt%@Y!>7fuKO0sHGG}UjXK?E|MeR`()hcd8H
zdU|>bw+Z=Px!OruIy&Tt1%jNn?a9%Eyl;kbHZeCR4AZdrQ78()?C~BsLRMa0j=?A@
zDk8c7imIEp_u-6Wf-@F{NSgWxgIm!haa||<?4JB^3m0*~R6R7Dg@bX=qE<`I<Z@2u
z5#>yI%IOwxXH4(HNI-e7MGX-sPOW2m>$0eWz?-{yd9_00O}TfE!p6o%)Vz#1D{!6l
z&7WVVJ?DB)d76~nFfL$(%h3!7IEi=;sD@<Pi5!h9Q|;M(PLhrvdGxR9<2N$!I5i_9
ztkbn?6F+`XKtp@{m|uvIvZn{`{|j`ZJFS0^+vFME?q|%&pOd&$YpkgabX7Lb082)R
zVmR^AL&LWwaJ4$RSC3Q)YoTB1CF&a($c0{RaB_thvVBZSDhHw~oRp>sE@A>GH^kYr
z5U{0{-@E<!AL>3)-{0Hqd456Gd8mv8goLIir|%4M@aM>Ut`81Ae&Pf#OxK=_I2;CZ
z<b)uukm=h)a*Ujn7Yp!8Zfv+5b`L{^+~7eLUw7T<2bVmO3Hf%$g?)xXjL@4g7&!%n
zCK#juG#}gfy*R=Sw{x&QxIX_b{yETEO;Bx@-${Sq0iF<!xa=Yg4bA4(_bea<6HU44
z$jO0zCnzu`qNtjgqK|O=z(`z!UllV%LtTyxF13^X@L`lF70##i^ji1#mi8yZ>6Aml
z!w<?_|3Zv`f{tZMoqZY*80flrUQdrEBO~LBBn{U5*4wwb2-lF&aYZy<3{lTkb#fB^
z`ST}Un-5oWREC@aMwRpBPE0|2TUTt?>bJf<c<?a3(rY(O#H^{kU0Y{&d)x2N&e(ob
zWo2dZHo!d*)1r_&ckYyK!!KTUcLrRD9~*_gKw3UC%4_dmdp*85BltKdyqIe^_>WE1
z%~u~bVl#7D#YR<!3PfbXdgp&{ZhQ$`UZ?lK8VuyO%ieoE@ncRDe&J0)_Zzbqf{O#-
z7T=jPGc;HrY=pF+mEDc8<_{mx{0khdrl;p|;9l|C0yTWh;?^yI{;=%qanQ6%JN57w
z3?vAUe5jNH+3GaJ1u_gk*m8?w5}eqc1(3!*_Vz|(WN_FXsZK~p0N@q+^5tEstA2G{
zaQm7TNqgtl24DM>EN_Ekqdvz@O$yTtZ+Q#>FL_*6jrVLt_lpPAa^CqE+Qj=Z@_WyL
zu*Rt{uhU-Nufv(?NUqO_eUv+DWMm!x_4@Zmu2PqlLT$#Ug}#y9Ol>3tYU?*PT3ZGN
zl`y1`L+&?1Zb7KhVpwEdIDY;5wY;~p#0Nttz}LHqKE>m=@jeKyGQ<!APa)_1lLa7N
zB4Gd3NtKj@3!pUVXUGv@AP9BNoQW_kWb!_!jO`hLMmGNKThzv=Ki!QRH`M3IKm-BC
zh!7KlHCEx_!FYUorv@ro^7h`|p6~wd1{^Ez<Gudp7bIOrPuhAsaviO$p1~c!rW}I_
zb#S+9U!C_XZ>tK8@K#->cUw;@(TDmy;xN7gi$89-4s<Mj7{6yWrlqR06HxNwL7d@m
z?V%A;(wwDnvKs;84z*b;5;*nN_;+JvQzvRSKG~74GD}N0_SYFK`==2iVCEUveVTFN
z+Tqclo$Ds1rVw_~pZ8KVb<do^U%bd8<2s6lAG?jk5p?H)2mkowG;@c&R1Y>rCj^i#
zKnOko0TL)dGoCqk`;>olO(*Skk&W!KH^ffWb+baIY64WDtPZ^=mLiI`FxyE#jzTR=
z(bLt_(}cG$eQx2^?Okmt<}6Q<;}1}rgihs;1UhIyC>in|q)6c?d-+n*wu!jdK*zrI
zSz+OY%r#MYc~%e@LrsL4l4A+8qvJ<dutV|p+k|0YU<CN*JSZ@7lS8b&?3=4o!??e)
z-$&De%h_(pR^}ewgO5YyN({;M@p$_Sp!-l#QQb_rUG%WNUge?hBn1dUq%dv6U2RIA
z=gFzClVC}H@F4Dk@(#64(3U{F)|A%{oIubLHl=ln9L6h(D`Huy{rA?Ey~&#bvvW0p
zv}LzH&YJACu6@^5J;MLWrW^oK`*XczZWxDNHlJLR=S`<FiBi^`85ywUWRS!ve<tY8
zS2(l6l|&x1xhV7C=8N<%n$FImIXO9oW@haL9x+8aP*<Rql|t792?m3TkC=5WxvQ(>
zX%>2Hxd5VQ1vll0)dW)}0<0wGe&3mR<c<Z*;`}cAQ)YgjO<3w%&X_lW`1TjNPUggi
zyKmA=S+mg|kr(d=Dm)Kph$(1u<aBh}Fr))O7@SjBJxE3i!ta|xHjpzRKcV5}KYOOG
zsY&uaLp~gmd;)$$?&dH4^kWy+FWBf(LF)H)-BEvjSax#kq6B%Np6rp-PK!)$dd&NW
zN>r|{u0m;Vk+9%{8i)V-&qjDxTd8{(3`im$oywurhFiWrJ&U)`ESmhh(8_Fs2eyAS
zU{#M_P!LVioXN(;S|O(yg}3Mz|LX+^s`g}zJh!DY*?JS_r2XVXp3(U<vI2b6VyFgh
zszvM3^U=5KR(M8l`e6cU|LSICM-7zL7%3^idvNdGy^eS9oW1H_Uir{8HnyLbwBLWh
z)wV63^||u5>FJ20XP=J4>`^`c&F&fmO5%)jh&t~#|6ymD-0Es`o$EFJ&`HpZ15FZ!
znrJ|giOvFZ<M85co6EL3PmBRnsulL(BqR(0bmYifVKU1@s26(4rM0;}7L}lrDDiC1
z-lsQehn7bGYNJ*BoHqT^$RVkMGaw3<Vb9gLV^qTo#H$8xQtI37GCI5jD2JDUv2#?-
z%WF+<1IW_lU%zS!r|Q@GFMuSwt*1>dC?H@vRCamxOZmK2-{y&XCb5xOXQES6?>^8s
zE&AeX!2YmT*;9e@>&VEB#jo-KP=#P@(}FwS)ZDD8&6mSI+y9&$bVXxhV;4E8!SvqV
z-Uhg%J6}Q7{y4v;?sRB~t<aJFKuBe!EGZ3J#O1E^2{@_7&6O!huQ?O`BS+#AweXw~
zuj|AudmmPPJs;!IM?&tNa2N$7zSh71a-w-;)C_g}92rUPpMx?bymz|bJG;km9O$*d
z!NF@>A38c3K&f9e2L{Lm=GdvuBUvpI&CxGlI5jab8MoQrW`+w!IH%Z7i(dz#C6Plu
zEc76m^~T-saE-#@>uHcDz}uk?%ey9S^CljY4ax#N`)fr{)n@1C;{b(R0^7h{NjuWL
zwUS`l&RYVrpY!Mh4Iah=)8E(_dR7lRR3ov0|0zBc*7WQolFF@aI{v|HmtKA1$M@So
zTD_OD#|V^dV@Oaa4tg6#1tuRIk6%v*y`|xu(R8mGj1h=70hB#h?QsNgaWGb*=4xbQ
zlrx1LA_n0FhG0HunNa^QA<4<~?@#(r0kmU?!4K}3oIBTAa4+UdOy6<xlGwz=wqG+{
zl7ZWv@4#nJXjak&WFAne5^xCdY$qspcQ&-$+)8Kj@%(^-kJF|tv_x(*^d2{DJhdPQ
zg%?}?qQ>D+P;pN)vFZVyyLvE7`ip@G1QS@*OwhxJ4+-!NS^IykJkHCj)R4m6n0A-<
zTbql7*mfR!!+FHuAznv^Fnb<vT9TbjR9#&S&Dyo=_W%s?pw<E!qt8er<WH5>kWQR~
z_P4n`U$O!-`_z0%eH-Ywr7j~1MIgOnkTiPei2nKMUfQvsonvKXWiFeWzkhs&nYlMH
zh9DZP{uF;@Z0BN$XISprbAORB8LiaO$>WsM1MH=h6%-6Q^p7jPe7=ao@p1)i+6Dlx
zgS;hUZvwGgT(ls-hlL$Lco>8+yS8|Qf1!)4G#sQGUmLJ#FEW3L*FNIp0TO|ZfkFEt
zweUxV%US|C6}Pjpl7Rnfx+sCEl3H*Dz=ou;G+hBXPOMS;r8P4xBnLH7iH}DzTQ{I!
z1o`^s*XlzmyT4Os9eoz&=$?m@(dB>`hY11v2XV?HCYBg?_t1!XzJD0xuKkR@$zpW$
z(?fqp=j?sL3$rF?Qrday!3%}a#Hu_GB7p0P&&uM41cd>r<U;$^4-b{#Dz(9^k}hTy
z3zh*p(BSVX`OAQ8fWW+3LClwdg!|!|nwpq~o=E&g2oNfezD-R{kNTVdU=YRNaSDgA
zM$z_XQ@7gM+wI=oK8i#}U0q#K_uo7iD6ly|d~8lcb4vtlx*>WQ3>3hDH32~jT#_ab
zjv^+YoP)mCBdHLNI0qP7L2782>nmL860b+HB?!*0L*G-hxf*wt>cu4lV%*-WQ=<r6
zZ%T7KK8J75-e)K@!o9a9k`DQ4q4a*v^^GY|Y$I>L26X2B@gFtccrd?bdk!wW*%2G{
zUsCEF7~li003-x04-e^}KP%cqS7Q_szkPE+Aq4;2R(83|(A*pVMT%yXg8K5nNhKvz
ze@<@C<^o-A54e6-5Hj8I8uvO(d4QUa1rPN1yIUCHYQD$D#xB1VDM5rVAox(NTs4p^
z0MPR*+&4t8ZI`4J_2*AOL{yY|$p9a4(EtHN_jZ?p4`QGhK+h8lXxRp^5&~l@HHRF%
zVCs*Wb`fD=dXV6td?~tMg4g=Yz)00_TD;D)WYXef-3zS$2Nw*c#rJA<`Y@F~zW#8Y
z5pG>Yuc?SlNf!|n+ug{4G0{W<&Y2OnEJ*Dp$Kw^Ecd+0oF);WrY*CrqasOVo1HXWP
zT7k*rbZ6=&NH@#PbSjn}Q@q&S`e)CCod-*aV0dlaTK#@uq<v_pI`b^BS%QUk&8}|!
znn~_@KKbNpI^tbGD*;mRe2wQ(%*Ek~>ySN-pPrmX&KsD~DfLg``*}n}sDO5A0ZBMk
z(yr~|&icY&jh9Fi4SQ5p&^E<iuEdA{Z=@ldmh<S*qnXwI_I5JCR0@~_YbI9Vm=OEa
zG4sp&g6k0_B~D&mUPN;3ZG8y71cS!F$-O+4T?(MSHh~2L^N7!IdtJbmly%_RIRK-R
zldW+Z3KgwiY|2<^+We}9cX+>uXnJzgg!(2hIAAmrwB$OAz6HtSaJHb!{Q2{z)N9`K
z54^UER~C3Uh%mqX0=s1V{r!PYs64Helc+0pC)aiC&FrTqRA|BnPJMcMnnzrmE^zyI
zsA=I$O(Y(JxUru<oik3}WB{raV`yl2@r85?m{y2Qd(N#Q8YtV!3RF5!RuEMRR4Wk%
zQGOwxJ{iLp8xyZ>Z^V|MVCj?}Qesy+E5vD>r5xxs*LxKD*Fg;WWT<X`eyZ1eOW?F~
zh1z)&lL&At*my{5p#ix*c>jU&2YV@7fLO>u@<rhiqW?dr>~jEU0R5si6j<M?8JD*U
z!<BYFK0N{4<h!@^eQCCem6bJPZ`LgKS>iZI%G1UIHDt7*CQWT;wbLIN;XdgTrS(=g
zWb@#hM1ToDU**PY?e{|f>d4f6Ru>tUuY>_>e;TObaB?IV==bALJmJN2aEH8vXlZF#
zlmp~|jf4D%fWnFx)U~xWmXp4vv%YwB7HZPznHe<m!Lb0YIujKUK@8mU=JqzFKFIV#
zMLCdzfUEcv4y2rOJl<<JvZ9%+8isvgpuu}KCb4&7b7#k<asip5p6Wg5cXtoU1Gd7*
z!Xmb`v~-wl_qqg&p|Nooh;BMh%-F;Lk3y#eixD{`WzEc&LU4kFG+wM)Y;T6bbYtt+
zWnwZi4Z+lr^=lwJ0f|z#u)7p7<g9h$hUB#S&BCSq;A;zYSM>m4Ti={vLN~;<Ztbx0
z=J-MQBxfF;WCi`jCSaviFKhoA;!|M8T^Y1XnkX!JQ?u~h{H4O~We%YcQq1h90`B9w
z`d95JfXcwt_@Gy?zkGQD^07r_Z-oxvUt?b%Q)Fc1vCo7c(8|C)f^ob2)*Cwh;>hc`
zc)i{UFepwy$yBqo<wZfsrQ}CQ#Pgp4;_J=+o<AleH<u4EP?XC^jaqIGxxN<Op40c*
z>&UJ3Vb>Lihp)e~KroI=)a{)pU}Ezf+{qqt#;-h?%9%PJU&wbSH8pi<_9irUL_Gkl
z4rDnEGc)s<EM-na%$_`X(qjFZi%XQn^ESXh%K)$%VA6*?g~416@RZ=0yO;#6@a1JU
zyp~oIOlPa#x?ex-2ilkhEODTL^8t1MO$-Q94P9MI!1-~|bkY|lCME!sHo=v<Y_VDU
zPLK$t>016{2#qu@h=A|Liwz*3KyOLO${L&V;P#v41tS#3Ou4H>xf1UB<t&g#Sr!);
z$tft*fC^Udnu~z<8|(M?5P=S;?dT|km|_k^Z|bG`edR<>#VD2C9|kar*Uqe#08T||
z0a|^;(90`h>#Id0d$*p0360z>5np<(2&4u)e&Ky%BN5cJW-{)8kdGe?Ig|rf<_Ai&
zL8=CuPzWU4WI^6Try+1k9H2TZs%mU}tDzUqX?FfR%R|5$*R-Y>hVFxZ397Lz^sT!J
zUc%rLQAVb{Mczc7R}yfJ$vQ59m9KSMQw;HyeD9R7Olwz+&U=whaMjA)kVZUJ#dUUs
zHsS40uCyGg-2FN^c?ld_s(GI$X71Q%!EO7sx+>|jXbC(B@Lbj}>#a+)fu0ZAoFF@M
zSAqyg6hsq(NKhKE<&H!Mq`s@E5h0rJ5Afai66-pPr}<JW)DkSzEQ;QuTJbEQS03IZ
z1*NMMj!nYh9SiaujE;_;FTF^FI9rEqHuTbvV*{7BT%nkU#l=y9yM@ZYzyQ4iNc`{(
zR}Zgv>>BT%FTk+TgA8N-(t-EnNe<xEkiF$<A{0|j=_~K?aQmL>XU2k~LDZ^-u{nxH
z13m!Ys({ZM-%NoJlL+1mpk|dKMw1zuk09La?d`wz??;*G!Al?w2rhK9`DeZ78h1!-
znbqu^imE5$c7pyawOjh>iQL*<qgHT_MCv|i>Mb~nAOI!=JP&nsfrUG?tWuAaa<S%V
z@b<{51_iEeBG?X}4y|>vBZZHghUWU(#wL6O;>W<W(}2Er7y5!CSGfjWw!c%JTDV&2
z8N2wTdLYHg<C(xIyE$~<kK7mw=Ds6l6$VuW#-_t!mcHHF{*_X8ZPXyWUxL{S4R;|T
zp9MC44jn?$qjFwpDScM(p8Q`a;=1k=1`upj0|ElNWPsbP{r&0mV5x1h*R?^wVv~TD
zNihJ4C;(Ff4IUO9O$JerZCbc8ooX<d6#Rz|sjcAPxCp-_8^Fn@6^~qL%U~bGNl7s*
z)vmt3UQZUh+02bh7U;G}O0L}hewdN*A~Y8G#K<US@zgqSxsgoLbJh@*KoPUz+8_QN
zipZ1)&Hfd~{OSi4SJpmX#r`7maG$Y2X7TzzQhpnYG@1H7+H246g@0mI3U~Rl&sj!2
zqo>93%r%T!)%D7WTBLQ<h|J8)u*f*GDDV7~Mh3)04SO%tJRo(aKnFyo3FEV8U9WAC
z<|IBYE~Z6k2x>6U9ntVd4hZr&c3A8X<3lC-6unekdvdS1_7|~rf9=V(A~oI%v7qiU
z0)FY}>S_isCIR>SU55Ojo}L~|amHY!O!s9-9l!;qCDsXd@7_)SHhvDQW}09x2Cqno
ze=#1{s6nUX*<AseUV3mzVPSK@**uH;Cg0y@zFuSjBVO&dC9tlei)E3#)N&%>Md7?(
z28c!uu9^?FfGO+&9upXX2FK$FqfH}`9FnkjsJ?vToA<plPhb0u0uL_<22XI5Pk@1B
z-ik9kKX=u%gZ^3ilxHsXbs>9CyE%<GQrs8s>gP4X+Zmlb%cnq@at?5XhJ%CPf0Q~1
z$BW|)p+Jxv;!K!LhyR%Zf*>$+z(cTN#upZ7#KpxM!7J;G{MGiW|4|>_n=(Ioxe}B=
z?VZEtB?MDR8+@1xd)5H_kw6D-kcy&8AF#rhF8%K6>I#RM!^z1B27qPLByM{8k51td
zC~V3pLPfftlTaN-u>&(f&9@hYDx!PCKnn+VkAaX1dgr4(S18q(kf^9AQ|XFwr~pJ5
z>!7X6wSRtf4JZYDdw1{Jp36dAH7CeDP0!91ek`&1>!}A!?cAmF_Cdk(=dF-|7mu^O
z2-y0i$mXSEe~O=hqDDc;mejD<v88xU>$jI_b1KZMINTgc8wnr;85tRQKs$vf*!=du
zH90LU63lkQ#KeR|9nZ|UIVZ41Jez5=1ONq|2*4u|5fSA2+Iag1MP+J(;NV35bVwkW
ztz4^@Pvd*p5<XEVCpHR|?2@$JN#y<XGHzCf^9wY_Qm_S~nd8d)3^gDv0SE!dQ#>py
zAwiEucDLqoNdTLL!x|bICBrg7e|(E9s~Ny^07t9^0cGpznq*z)SMC>YQ_}9|MKi#D
zoq1-C&<K9@7kvd6o*4f)5nQ=G`c)S?4>SX6Gc&WI^Dho8)opO0Jog#BNd{%K_J;=m
z9Rg4V8rg=H((WsBjx9)J6&0m+eX}LV<^S`A;(Ldf|5AV9UqV3lktGv&=P=H^8+<=P
z%*j^>4H=*ofMx0ctTO<q+X|q=ZT+V?$TF=C?>jro<mBWYf%O`cY^XiJIz|>27Q(oX
zm+2N}j{wU0e4XD3(_%{s0C|8_E`p~MK|@e!@$Gy$v_kzNOrc^u&}Hd4TXB|3F(oA>
zWEc>eBBG-!Yh0lbJT|`CB~-odAa&1evL$BtnK?|AF!0d=h(@l4p}nzdtLD$Z!-#1}
z=GBn!oc$<MAbI&MBMbz{vyD*21=qSrCJ|E<j!<R)oS0YGvw+zc`r3^VZ<zdnr$q}N
zgrua(dq-_O^~wi5omfwv$o#0ZOw{qj^rph)&k$S293}DmqW$bIUe;hY6bmQX1koCg
zmjv)GFxY6onuPb#X!pEzuPqp^5M2YZ0QeRDnstj_sa1iniQG-#Py!MW56OyVo!a#;
z*TCs??dsvt;Qif!gtt7n;pyxjEo)>KQtsS|2FuvB3J6!=mS~n&RzL^nJ(dJzJ>~Z8
z2x!}OptA%0WYA>?wiN_JKuL1BnE_FRMo=N4+4~DHnZD^Z<9lc#92p+o?3H^A3MUn2
zb9iho3Sz&*et#^Y%6A0^CZX~9dH(?~v*#9(peN;A`S8$pt{@MCfrCNz^R@)`XxQsu
zd+oFIX2NF@<W@;oN0_ge_f&8^&W7gM)9Iqs^8x8C&N6&EKS^2cuSq;;{Pu_`Rfk{4
z%cO}~+UY^JTk=Xq3Srh;UO{2<MLXAPhh0aU(*+nJ4w6#SbA~(L?X*`1Bm%?>GHJk2
zF~7b7BUExcf#A~ORN^GBsA$)lZ^UxKi4ijnjT;WK1z8_qs;1`-+T=yFF>FCJo&pDG
zF_uq@M@Hr_w7L*LBWP;}a39yXbLY;N*>_$avtJpg-(&jgHu(Hm5?pWqSTDg?yZNh2
z9Em*Pjc#3bP>CCRda&%WE-k=W!~>|d-Q8Xr1SQZl@ZN$JF6H9KM{1h@;+6qkgE4Fz
zEe!yrl=ARlB7o0au=^pWvn}j705uS>K6K^$D@UV3nwsGoUQR~UmWeQ^{ki|cq#bw}
z=#~r@Dx4dUqpUOa$XMYWr*dF^v@h?<=TEQeP_XbT(3Zat(Nkx--yNbZ>@9o=hdVv_
z<3~*J-fC8k9eq#me~?5Kqz9m@V+yq9_rSfvanhF_xWp>Y9}JjB+1c5B;EKR{iJA`c
z)avGawnZ~~)eSzF=mw?ItnheZz*opUF*k^AIZ#`Zu-kBL=rA*LGz?32Q2r_dGk*W}
z6unSEjr@TCJ!r?@dVU10u>nRvBNG$P?mdQ>>6&l3z@?fNoeu>#fhHU@Pyq-B;VL?4
zciZaaexKlANz#qn1+)D(BT;!K!gVeqP5p^IU*_*hU8YRrRzcfdz=?w)j12@9L_Z<7
zJyNw(1@)Zd`ymEXZryVKY&ZH2Su_v>&}U|&sjlASl@tv09zgpRxH`1VG1^m8XEt`Q
z=bBu_9dHKA1?<6=$+KX#-e3I#ef;ydG@XV~9>`|%GMMFWL%y}fGQq6a1pQC`R$S0a
z$f8S2qjhp_;~2;k3`R^$3`(6Q06KJ3Im5PTG)>5hkAedZ6AA%Moi$yWZsgA5YdW;L
z^64}9q8q*1AdqK9e8yT|FTIJ2jip3_72Fb7RC!%!m%y#U1&y#3Za%nO=@}ScTLvQL
zLjU-|;-b?RrS4fMY5Dmnz)K+#o+Y6z{c7Pmht(51b-U}szB>z*CH{L66o;VEMa)-h
zEDUed*nNEmro7_e^pup)`1ttrdI5%0g(bNFme8&ev?QR;55gW0CPXW@5!hCQt2tDH
zH~?%mHZ~4?edqpi<C|8xsJZ*+A1Yu7z7X_o<fx+JYel~?(R)4i`2Ks`uw6#K_>kBp
zkN)p=h(bK@8epVt7~N}={QzDlJD@H$v-Y>L(pR~0PTX)n;4J0?0zVYy78seC(Om+s
zCGfkjK48uh-WVeOX7d2ZM;Lkz4r(wT<mU^CeFyDL4Vti=Q*VCyS|-xnPJ&;FMdBjn
zI9tNFJ>H%oUXV}snE~%dfwvA?S|hP{?l9NZ)&{NKcnr<*?3W^&damyuvf4q1_tkEH
z5%nY6)iXC1Tn3<6DCT8fkT<C~2)-4$g0IQb!)Bfw4Dak^rY>3?QeR+nOIbZYjvXq7
zlbl3UF#P!l+;T6k7@){RmXO7h+mN(vAZWz2bNR>9sONYKrgj38ipn%hh5yff7~oa)
zY(qlgsl)heLqsQn;}n<^sP9OBBP@oUlOcy{W*Y!>ppb}2Gccw-A3ugbDpPYR+*ZL`
z7y-2b-Yr|7O_26H7rvZx+y%NldoA_Wt%!aD9UWMdfDI1;0fA$=oeE(Kks?NYZd{c2
z7=rY(Va$(29|RZ*0Q-9(uzUosfuIW7>Hu|g@6UIYIDkVMz-BS%@TPTqRFdG7y4tQB
zSmXAFGKtRLXz;tAo!y~_TwYnF+uK^4w@BX{yvja0F6rnj9v^&F&b>>_?)|Dw&h+oC
z^mb1gy;+$>0bAL-vutaei=ZJFeZGI<+H&t%1CTq$0TqLRw-HLhUsimyjvcs(??MbY
zv|7(i`M@B2^ypDU0Mc?R$3uFdpJ(PwP{@a$0u%I$KGN|=YL=;)gpN;MKSeBkuEOAP
zSKdQ#eR1S3L_pr9ba5M6$U}h2cCdB=$eeh|WeZtn|IFL$GgKIQK@KDH0#tv>f-$f_
zB5N$L5$bbCU`mIrEpi&B&jei^od!U!V^Zhr{Ns|6q5-8Of03PBVu%5I&fghhLdn?F
zG^+m3j5ykI>~%Wv+fNDjOdw1U7PRhoN~?!Ud2~82i16ztCD@=mGcLGqc(%qTS*xk)
z9N}F%Lc4oNYH!bAlO%TQPll8*$+f)|QT6bX!$nb@)ErT)Ze9(qyPoDqpZWMFkDDcU
zO!2(Z5)JXr?z+){Pj1&>!G8`H%!RyZx<HX64Gx~in7JrtKrKc5DUP3?82SP-DfF@d
zUOf3{xp{W9PLZ1QpuN<+S)JZSa9tru30jc`z%G<xD5LXf&R-6vuX2|(DC|=&?7IX<
zEF>h<0OMVBN&Oz%hm)(+4+GcD!4|rFfU@|^322=`22bhi1N-aG8W+539x>;bWQ%T|
zQOPZ;AYaovakz(mb{1*0h%i9TR4{Zvp8(|=DN??^zJx?=&kLe6FI91A4V|qN#av?Q
zbEhv$$yHZ1_SG%0dCKyoo|;&k!-ywU?2dJPc5XB1^SjM{KSexcDv0?+#<X$AW8<pV
za!p-&@BEaF3m!KIf_uYD0o)j{^f>m3gBSx0uc*Z<azFyUfS5mm9!1SAbLOHhuIAo@
zV<TWUA%0>!lz3I=I6y;~nfQRzhT#-bZ29trw|5jc?JQl&@1%e*ncf#%29{VpFn1w@
z05o@URu&xw)Vx}aSc1iqTf@8!@&{V)Bu#;Tnjye1CefZLEG&%vcks`VD;G^JkCH4)
zD_#EN?-~(85lYYdoBoR~6Fo83|AL97rxn#Ku8dGv7#2&yoOW(Nw}&Jm<Tq_L1u6L{
z9Ra0#XWsta2{OES=LyN(>CQ9nmo9u8z2`ySskgX4`dv%Y7UY_oLVE0oMC6&MJ!`g~
zy{v{V6}E-%L#V}>d&*$}F5a}TZdx-6SQ*GAZP1Huz)?(sK5uy}tfLWj`cRui5WAxz
zBrng-Qb>Ei6~Ga=q*U0z=lREP6v%6)I={8asrv*%#@(~D<fnCYlkeR_93QIrpbii%
z&!-t7cLo9y(8~cQHtNOX*Tt@$)>0ceH9f$geaVo!<8tjIu!e~Jo=^TrqUc3x@<)dM
z_pk31jJP?LTeohNUhX;!tu8L8avhL70()xVq`aW6fC~8hQ99mJ30x4UJi?ut0y%kh
zpoAN4^an-C%=6$pfQ!X)aKr=D^_=fd@S$eNdb@n$zg_^W7e;?8QyE?itVEa)D5;1M
zfb|0nI1mz9KAiT%5*MjK0{2`(pHR;XuIa)%X{ng}1@r5sKSj2&6hyMk9}+%@JQX?n
z)}Cd2CX7?zpyiBI@4#_FF>G)MJ`CEv3Kg3XMzIzUO2og67Mh6%Y|Y$36$0@hpb|b3
zz*;3$Rmq)PAGAZr0iZ8(G6N}W+mXx*2#4&Bw2wOQL1-5-CpTBqj28s(>DvdJKuWZO
zWCSbHSEhlpM{qT_<m&A?#74qu?@w{iGkEc|cll(WzjH7%o3hgL4H;T57jXK1TZ)y`
zuUeH>wd_>RJorFjcr~KNk*ODk2RJK~y*<h>F}4FRjjsoZTh4Qq9`uIqh9X<%i%cV0
zk{H)thUne;`0-=I%&rbOP*G>@Nr56mf`%HjYzs8w&*4hC$`S8jfjcZu*9!{@C?L-r
z9UU8hJQM%&zTN-^%Fd%?5il?iVL0U7i9s7W3SJdJILXP$A!uW+A%rN5ogsyfu>OPc
zfxYId4|)*JM(hQSLQMs4MLT#jmX=lA<aCZbo!%D#Z=oKDfzNheSQ>{7t83vh>v@Jk
z$>lc$PQGHMQyEsr>uw8>mEf-*6&9w1wH6+PEn$5|E$=g!En<r`kC|9oC!#H00QRNl
zMU%T`X2<GPvP4dq*&k=KEw@<ta;=1-6SUp1)=Su1dc8ezGa-hz_))MRD%n|Pm}D5u
z-HJ=Khq}^_dE|>QLzEOSvhfo?jxAQhUvIIOylB(ohiTALySnT_hy4mdC;?CPdm}tA
zy2_SZ+%#uhRWKDT7=i;s1o)<jZ{GxK7_36O9h!f9&U@iHrVLi|tY9T&Wv;7NuX50S
zIydAjqjM_)i~+n`uwQYjCC!}hB@D6<dY%=^T^Po}lP2c!X3GKgPH|A8;tn5cLJRLO
zD*pjhlG!Q^F5*i}(XFhH&MqNY7szW!pT-nx=r>TaCkj7Haa;>6^w=BJ1KyjHU|fJd
ze|}J)vai1%n!vA{xeE<%(?o9k*}Q>)Wjw7UE~Vu3bW-r#H+FXz!TYI$*a$F}>6Xpe
zv*Cc})OBQxw5M)g5mx@QtO554DT6SXEA0D0t3g&6#l@<F`<v}55?9`HSf4;kvT_%8
zsA4_yvTqkAAHG2~Emb|MJ58J@;$)QLsGj*XV(%LdE%0E3j-#h;MIb%^5F_I2+})*?
zf|zEXK5#)^V8}A$atA<)QWC##`*Y?lP*&>R-ZDTi`!%{P*CXG1>5%jHlBcjdo0Hh(
z^m6jMhawT?Xg2Fdr&<zO+!Z*As5^QITK7m^pPP|r5UxRw-bfiaAtTe|@G={a7UHN-
zf}j6#>!-mjelTEIxnBTWApuKezZZt@G^>F-7MI6P4=chTZi*B~@9k88gQ4P?CZ1OR
z7JtU)tjN*xA~lrQp?m7Zs#<wRb*Sg!SjO-@UL7&{*{Uk1c}`vzKHecZ_(XUr8Lcd4
zB$^hs!b}wpjA1jl^no*K2AjtUC;-;p1kHhVdnshLf=QPG8~~Is%l0UPM&(}pIrEnQ
zF2{7-QyVOn31I35V7ymGCv8>f!inLLm#;onh)p@DX#Iv=@t4_z=L~PpIYK^@W1lr#
z5nm0dJnsp{Syu`L75=Z!MWa(R5WD6s1zkkR?>CZXkQfZA3~a<vA~K4Wb8@Li_8HK%
zV4d<^?UsGvy<h>ygPyPfOFcP-aE=E(YKV9-sE8oi9x2D$+uN%pr*B0VC~py$yxA6J
zICKbhzVb{vD~Furyyh5@Y6eW)(AvxArXm&2LkBQmF@nWkGR*Dt^c*NRn9{4C1z$Tb
zVCjF1kboq6hvsD;e-#ZDD}T#oiJv+wT92geYrNH4_w%-gtm5D}I~D=?B45&xhrS2E
zo1NTg4|V`Xt1eq6X6C<E6p)r*_;0v?+y`NF5wz&d&slpBkV-vFF@%oLqngZ)G~fzr
zgIHc^B;y{3B`aAFvq(WZ%TNwv1EU2X{LRTYaTg)P0KqEBAuy-|Mx2|QGc+<n!w@4p
znE-as0q3?fhLbJ)&N~3o$_hZ<DR@C*40^@_(iYDeDE-~dLl2|qmK}>n?hcR-kh=rk
z(4jIQC-GfZE8X6rZN8JJn(fG`maP?9doYU2cf0y9EE=cref>ro>0zCx4WEQ~eBiP+
zA|FtDF1P6M;xL7$7(f%sh5~x`E&<$Ou#;ipc?@ndgT6eENr4H}NHSzXLu7_rz^26g
z2M-_}LSHY|AnD(;O2iq-cZhTt3*J{c&eqMe5eS=Tk~e?ZBgqEbtc}3j4Cf1^c*L#h
zX`Qp9faeI@P}P|p7iNp4&rhJY3RhsEVbqZ&^N>SNKH+Sei3$fRA>8BI*UQ4_85m&b
zfW@~%z^|bmmXdnoEkqCqh9}p0&mz)_=M;nJVJZ{hv=QZCV5>oOUKW7ZkwLUL)aNQN
zP8*q;dN;#1PIoh+Tdsb7IspgW4*hmwX2vThrUyjH)g9<%(VHNVEQ2X&D*b{>^@KBS
z`&^aTN!mLuYMI##eR`LdJ*W;_#Sn@$UWk#srt0TuDwN*VN;qs4t-wP2_gNf4fhXUC
zcH;I~PZm+fusY6Ub?PqMZq)eTUPq{BS6zGnHPp|q=8*#)KmC&<NYzn-<-LmklEqDb
ze+{w9{Znkf6oEpHR<1x4_ue`Kdt?5K)nq{Q9`RoILV|&vn?{chO3|=a3BC#pBqZh|
ztVy2y)xqc8DUh?Dmn=d*^&sMG8d=JPz}pAaab&&rE)9*7378PfhFjxx$N|`c76k4r
z-$WQFcyW|x$i@Xk50vxR;(oy*Q<0#p8D`Ym|GLg#ELRqW-}-X9nT~uP9vNc`Z0o?<
zXig3;B3iM(Q|*cI&Aqq0aODZGZ0&g5y7cEP_@_1d_}u+&ycuzO_KPQWPCxg`AidDc
z`fKTX1dBWTrv`0aaS0Y;dATHKr^l>31#lSEgYPyq&7&FPxJ_!C|88rCK&(D~`t%aq
zJj7z4VGq_YE&&lYJMwx%)8~_X9E{3HR)-1*QG>z2xC8N$+Vvdmsd{V@*@Du;6|meG
zXT1+Hz@veAMArcbJ1=?|g$RH@NLHlI!tyD}@rwFFCD;@Hsa?3D<uqtW?8GO}!S61S
zS4f+klTFSRDdMLYPj3}HTig8Bskb9RLg@MI6^e*ke{16$204s?-v_4}8M}-#dLjs7
zBEb@ICR5>@r_$>#SbvLU61n)|JO)EVL<Bp48sMH@*glVV_jLD7qKW@dI#kvdUO8eR
z+g?C%gG`zZyAF#+NCbgro%jrq2^_*&oyxE&o2A5iCd}OzS@9Us$EcPNux>^*6fXGR
z`^Lxi1`C8Wo_{4My`}s;^S6&EB!e9N6(UD2aQiqib%5$%<z5d~Y5=aiwo77v0-0W-
zhhxAj4(_}#b}ZIN`}^|2rqkwN2;uS_4~>fo$9&8);I7@CEA;*Oq8&Bxzl}K~qtBsn
zw29j`)tLZ*KmSV6tS+9}Eu3MuTMe#5JK-y#4z6S3pX;~_Z;}fZ;p8syJysz=E5I{H
zsb@QGB=t+OP=nv9de)1Mjt)I2#m&cu`6H}=&R){@>Cb-g0t+4zctAnTPh}9>B3dss
zg5!R;log6n3`2fGNOT+Xh{h+t%{BZ!`4SBr#LTBcovN`rx6?|S1Z%)f1~F07>-N#%
zB`-)rm{O1L`Y;JsdObf}@a$RBm-1`KHNm5zvVUs_5GP=Wn9nriBO@bN*zxFoAZz2D
zRa7L33F&ckF0_Dh#ZOQ##(yZKrcTNtee(4vvjNYM30kJ@eHs3uTkNGa<^unq2RHO=
zEWq3t)m!|CaX@P;g*E_XVACN0j0nE)a1r!m9hk&Gv?3I1fm(e||7N-~7;rYhzx7ww
zg>D}ON-yRSG+1aV&C(xE9t86R3`i6%y7<K6is}Perw@fG<-n)UDKvb0E=PpkD($fF
zZ)q~yq-PwGrecu>sKUY;Rtrd?V!qEo$ax@hy$OIiN%;Y2P{<w&z&IQJ9RsG(<cth7
zO2LDcDCgmddX1qa1WYB!mijDNqmi0B_>Uy&dYHjVS8gc5m~qhjR8-_Z-{nf9X&t<^
zaqR24hb(<FL1-rc@Mi~wuWDrWHwmGJDN48#Ue16Dkd+vi6L@#{hQ9z%Dh1ep2$~<~
z&Ywq*Il{nJ3o|NDy}c$-o+y6+VDwbVa95p$MIvytPStIUHiKFb3kCKOL<q=GXp}<q
z@u-pxUiRIX67{06!SVa^M*W5%^rM<^U%i_q5Ce>b=YYk+bQ5Qtb@HYZ2nxgNkwGt^
z^`nPG{axGjUhwO@as)YJec<C0V*p11!)*(*7$yWdlXa^G-dg~6faV8iRPHROyR1pl
zl7k!9PlINivADH+qIu%`VbPj@p4T;G4x3o;AgM6$#8GtyFIInt$C^Ro72?l=!{JhR
zH*<`Gmci&6VMQ;@Qn0;x`{eO^zoF@Q`O0+025PFPU&)6zFzWFl{CuX~A1|uO<eZF$
z5%HfdvhS~a=p7s!ki7t8Z;imfU4e}uFpz|W-@FnE0x}FNd_Rfag3l4r$5?uL9)25;
z1mIaf-v<(1uYqF=<O(b}<szo%*ZgY1FOK@xt+9_E-G+j56B{&tL`i(m7y6W`W!-&)
zpJao{3LXrmXk9^8-p%A@b`vLXT*H%-_x|th&GpaR;|ucw{PjjQNoiSu<OZ&i?P-32
z(ua3fk8Iw&7CAlt;Qq?{`hkY_&m+ZIB4<0?J}VU%UhR{6XcqEv-eAhuqbW)-T!R~S
zj{bh-UxyiNS!I8V4#)+cy?NLz8CCOTs#p1WZ|l0jt;e)k;-?Y=kKi+Gq+p0hv00UN
zq0|kx?6-A!FX&_jNaIG?FN}RM?=g3N6Uy>gU3HygD8aTe7g}Z>beG=k4*Ny)a)>hZ
zk%|g`gY%0)l`-lJi$D%5L4jt)QEsd$yZb<9?R4_To}@xZAoSI{o#CP?bcz9i9|&%w
zj)3x{dj9;~S&N1H5UvmX=8oFVKMjWYe{(f!KOVfAl%4y42$x~=myefMO=k~55yT*Z
z5XpRN9~=Nwh+dnVoQ#+b1{G=f7lX&0xH@P+<Jl{rqvzvdqtuC4%_H#F-`k(K7!32J
z>OjNjvJV~8R~-%?I@zx)<HY!OZSu|!mUU#6WBYQrOD!A&-bLn#_3v0hHwSrmx4P=T
zMCUDXB~>l=n?K1hb!c%Z;0Urhkix;8U{a`JpPV_iH+26PZLlSJo(O#9^ndp}s9tBC
z1Ukh8>^`&xZj2m(SoWennjH?Kn&NN6u>Zr=f52ni_hIAsN%p49GLvMq2`O8sjIuJa
zG7`y3sqC4o_A*O}tfH)tSsEykk*yRJB9!qvKDzGbdH(<Nb-nJ_eO+A_=lMN9<9HwM
z<2c?rcOMR<G#~j;(=N-mgLazJIx@3uR!3l}h{i7d(VBJVC0-cDCYx3=M}#duIQ);<
zpwDREX@7q)81GTHl2V_ZYmD<vzeu7;tc2j(uTAP3%+d^MbvFLmtKVeyE;Vl*2a|4u
zz1CXPgIZ?)-kO*gJqpZbqKzVj1Dd9U0pm21_BK0Xr}^dU$Vx>fi`b_lN}BG~W?pJG
zdquj%*QL7w^8&-r-q=_FC7;xFoJ9%mWw$U<NFxTsJLy9bCF}g`1hneKE|o;(UstE4
zs)677Xnjs6vxQtrnYLg|%1}t!XTujI<&vD76k8&<hT4;8lTd!be94%Yn8?7w5|)>@
z36wB4d8+Y~G2<xI_t_qS46)ju8sBW%e^DISYoZ<Gu?)@QG$U>=LM~Fq0rZ!hB|`zI
zBw)Jnc;&kM@J`m?&}vyCHba0+l8i=z2Z<&P^1#N%#?<#;ah4H}9V&lLsC=J&`!cQj
zoh_z~j<1g8>pTCG;NL-MkV`;Z7u<*XBUv71Ga3=swK&^us|3RJ@ripT@#PV`dAB>-
znyraAfE4bl#3)oAUw*zgv?b9l(uiPrj`sWU*~ce^72G0mIcUad-GBewO`b&I<*IAu
zyawZgW(lWtaFmi{_M7vp_f<kWoWd&4NS0g8Ynq99De>ImuPa@n`L|oVl*y>f_;{n1
z^--z=>hv-A&&i9e9_fDgU6JfDC5GDf-**hF3yYO1-3Qyn8J!M<`!y-3`=-N-?A;NU
z^I34c+<x1CQT9dNE-G+fpET$|ZnWmk()g#pI7NN2g^vHo2<Od>%hxd~L%tjP=!z^}
zZ##ARi<nWD=PXqk1c#)f19yl$v^10;^xN;;yZ1TkryyiXR#sLPW%k*PKyM@<<=SC?
zT_0|-Kl^dscpZ!DcXO974QB~D@K)F@YAnOtk8ZALy~wo<9N#ZVma-X@9keux*Ux>H
zc<f_5eBrRGx9mESIzr$Bd9}V(m=8JuhJ=Jg{h9h23S2{>{CGBV80ge$`q_U=;QRR;
zSKRtU3O!C**<K}9K7wM@ecuHrDo;6#cO+@pecDj_M<O~`b7<AgH>?Wg<(!n}T}L>s
zjjt<Z@o?PyM?0?mM@W%^$!U(QFA~;lRV8KK@(Y=40jdhNRTLsK<J;V~WZ*Ajo=;Iq
z`jPc>dSc?}8`qp%xbV>x)rN?Z#8~0}iUGq?84alC`(AO#RCq8b;RTzxZuMe1>*hnO
z48Nna%LCc*rF0f;&eI=REhlh>=|kU<VX7C7Ubs2_)#2o5?_`V=<|r*GvHQ_o!dw2E
zfG^N?uw;-B1VCLYKR$G<2-+YF1IczH;saPiY615rP`a#L8``A2mP;0;Sls2~TJGM8
zuD2Td2YJLxxe_&YTKh5#vxYZITD9yv8-~MxO|(0E<sZ$s>Yi%f*$jJOE8`y{L+#f8
z_fgQ__S)YFfgustpzE;%1pa0O6A<ABJWxE=-y?uoL^o_;glhNWV`p|`>YzPX;6Lq3
zI9*~CA|mb<xo9QjaI4dcb1t;L>uH~gdbufSa0hJcqNR**Z&|ZZ%I|q>)S5kEe$IEb
z;=<yOiHMae#pEuB#bHH!53A`_JGFuIvtaTdE=wSEM2`)0IH%Ns@{B?8azSt>4>`2Z
zvKIVNat`WDVzul3v`JY}>e?$JaU?=z(`*yNonn7HlIG7;_3m1gmh6dLthha=b9F(*
zClOuw#WP4#OzdG~m^mM5_1RoyX8BCSF*>?jWJ7$6?lR-`&34ojl4e2#7zAf&#qbdS
zImBM-r%o03-E0357#X^>Kh01z&9KmJwOHvIa#9vu8xhcCGZ2Nk#GFoCm&-c*g`Fob
zwa<50uFB8u61^1tGB4*tT6#G(5Su*{Nx4fr=-o-x{}WwzqShx~1(Z=$po1tB1d~C@
z%H92=lZ3Q<gjB&~=FGx$JK~n!(~+FN&WL@j9gzErc`@+B&?1AT0hca`#sF=KCO`OJ
z^xmoPN2K?ORoqo=a5;a(=2w3fxA%{}-)ijYP*A@lxm#jizJPMaBD00m2W9*UL{%~3
z4}rE(A>*`PaYs8=odhf1Y<DD;X!qQ=rg%?pZ;;VsOAke1Cp^Enf81+gkMq(ha_1a3
zM~VT&>k*{9f+zNAw-+<XS>@v+q&ga5s&`scAGpmO<$!W3Axfn9Ow`<hg~X)^>J8)#
zxyq3P8s7tma|#8$NtNTzrcF_`O`YHCt#lp>qPA{gZZCdyw~RdVWv~15Tf#G~Gu>pg
z1KxnNW!YN$cZLHz^hiCdO?0Td>dH_5`VTsKG3d{2zKCE>Q8W!Hq&rGl(IgyX%KurP
zlV+>YjU4UU50sehWE}G0PCvQn=^y7K$&0+MHyiO2yext{8KziPQ&SYIN@<<5;+c*E
zojZu@4tzREo`q(66|RE3@<}GRiAd!@+&83>zqjv32)r@Gl?JJ?)**faB`%|`4x2)%
zNWC3{Zr?XysTLjccaSC9I#f53?vg$Zy!#(s=IuP7zKueO)RSME_s1C<>EA0eN8x*?
zHMkH^<|JE3I<Gv0lvlV_2@^N|Vr_gMk;dX}yzM=TjDnl?XIB~>NQvqG<}fnW`+h_2
zad|!#VqihzC49(Ih6jZq^P5c@HZ1$G&*D~v*x`QTHSyJN&3{)(9qN;1*E)D3?2_a9
zO}A_g9U{GzkFqJ5YtPU8q|$xV|DwcQ!Tb^9)*EZKR79sABVYqB02H#oL9ei_Q3zsP
zEl)%wh#RoUyKvIaFH${?qdN|HcDC_+#aV?eo7)KKpeB@<F8wJV@*8`|p%|E_xG#DX
zbQ&62M>faZwmBHtOi`QTk(P4xJ^4oXbgwN-*&P+v#&4Vd#^=?1JW^^OP^8a(sl%R7
zg<0OJ2095Rrn*nAk$f9=tFdN7+`;89318W1tS}jTSV+}A<k`fdV?v@hDI7vVaqyJz
zkv^HWHXYGl+{2Q<IbQqcWKHzst`|9qpI5CvjCka}4Bwg?a(lyv(83gT|C16v%zHaq
zb*U_c`)Z2Kfw-x^DZE!xEUhBa*xwdd1{A#xN^0~{rLLb_*H_`u^K4+znQ8Bqfq3ul
z04S>Js#4-FBfj0+O5;+a+5TUNlWzJM`x1l@Zbk|PS`I?F6YU+Y%vIeNqLdz)_lnOS
zsNh{!y3Iv?N43c?E_|ah+X&V}jZ&&~K2Q9g8+dV&s^uHkN9K#-igjB*{_o>E;^RZ%
zuRzq>o`dxes+0Rwj1rWs7h~9@aJe-_4O^*u)n#;QjLHl*ZF|nLy4>AW-255d-nV82
zI6RhL-}nIKijTbHYT-tk^T8#Xt$N>@QuBX^(}!tRPF%9u{BaF6B_Xti@!!8rKsG*<
zx>)*eDZ|y#hx^xZ{NUb15YYhJ6+-SXI%PAKy<H4CwIsyw>znw$N|KTawRlEG27Eec
z3G?{$A`%kXI`TL_pPf+EL4N}p6Qm9~?*0`%1V6H9hZu^K+F1$Gp3>j<NEhun&1bo4
zIu&o6H11Rg`CBLDpS3=>&2?0uZg|J`v?qs~0|Rv*jd~X0eM<?%7<QP4EVejseB*b1
z+C&b>(!;=!a5qD*vhrf@$Bzt`UW$v~LJfOZK|~z1ycpbE04O>Y2=Rh>GHlncdYG2b
z=2QRmcY1XF_DZ+|)qm~XyO)!<VP%}A=5Ir5qnJJGRiwxMme#pFU+EXO>bi$^rDvX6
z>z%26PSWpwo?+uCw94(UFBtWdd%KCB^`6r0tk=~)LgYqx<#^JP3r^0t;eDl#atjN`
zBUs9IaehvHGUdz<MYN%pG4}6jDpBG-Alb6^>DS9{LW7VWu;5PwH97&{#)<qKtgGl8
z<RC<aGB8*tqsR|>06Kbx)y=f@^lPyu@Q|a8q_?TB-zM7mS+w;0!AbVtb+=L~E{x1F
zh|UfERSs!aIUF6ux22(|Y@2<-aM^th;nm{eDtEgc&wfo|=HL~Q5m%XHpXpjC?gTi)
zJ@GY5<D+#x-NKy)(b~10x&KR#ZhcbQy_EoK!Hcx=M>b@iFg_GYU2+&}(lt#e?Q7J`
zDAlbab-9mKQuo=2{0D1cmZX4vh@vUbm0vaLW!PaTgOzkkF*j4|&TDG+TTK1(IRHUE
ztOu&MHowgK>!DHov&DARuiLu+Eyhw$D};sXB+TC&d?7W*82p2i?_m<woKg19^~cnM
z)2X}|i#xutho?Fha<)wGWqcDPuG=hWoz`~h|Du=nXsl*2HL+(aPu%rcjn$L7<3qK^
z;M&3_=miU=?-xZyGr3Qk(f+iRW2W+VqV;5qUL(4$oPr&b^Oo8y`8ENb{14^3<c8z<
zdYX>Uh1WUeTRSY~r=9zLbCWJtx|q7fe6py<pF#zj!CdY7wF{|64?MU35;qlmQEq!i
ztwr%}y>?&G`k!U@dNf;~z<Ph6CXy_BP0f@%J(K@w0m!c35w-kAjhR@*&XtN(epWzd
zB5jXLY4dI{At(I2pZ5hnP@imn*c<%W=}!Lq(tNR?_N7DSi|ogqK6fi%dS&y>Nm~3?
zPTS<C($lKL_aaK_!*s4uG}%~+ZU1I1xv3>pJTQ6mAAy0JSDh3tjH|xfx`3yvV0ni6
zQ^;QpQQjGMMO?eCKGhmC`WzQ+Ob@eKp9{^Of9HM0#m$xHs*TkFxRqsOD)Td0OtXtj
zCVx#kDj50-xR!7^l-++WXX4$m+nHI<F6Y71GxD58g1Q$)E_r^nUH(4zAf1o}EpNoH
zYv@!UVy4ZNoAk4BL}-d9<XLi@wF))tvWgS?bac54<4XLbRZ5@!p*2W^*P~c!lMowM
z;^49EU-N&Bu@@_z<(u!es4(8Le2o=tLVpo=EmY#_lecc(Jk>91w7NeiAxbBo=n`al
zqHwFO;kbD4z>DFLukE@%tYWTOiY9wfw;9`=>CwB}&_ieQ)Ru69c#RNE@rz+ELiegq
z9&>fgcWDvBrWK5}@}YW^N3u1bI7mOn(IOW$r5~i1tG)jqzrcPgdkB`+7<j!<)jZ`J
z(cx76r2I;zyA&@PH2+??K`YehffF|p#P-<f4T7XT%(Yi8{I-OCiNQ@<vt5r64MUO_
zeoRuj+mcCpiPL&p$-cP47l{nN#%LbDiE^Kl#v>A4W{$}xx2B^HV{;)%cu=09I}?iL
zodc>i5-5a{4s?mKL*KpMXluRLxX+YXMt-;9?aVuguZB<qF#&&XJ{0@$>QidpH&^}-
zBxqT8LUkml$A46oR9$Fp)RUY95bH9B%gvW76G~NZ*1;DtFV&MCsG;t3b?~}Dy^yZG
zwstL#L^f5|FSBa<Z85o8WKrX_6pD}C3h~Kp1<!%`c8G!D=;8iI9<<r%Ki-l@vm#nt
zBlPaAvW|?Rnf29Cpu_vp!CpR)%k_Vr-6DNcw_+->%n_#&TmW>CN0pW;0X}Mz_w=oA
z>a)|D+_$upsj2wLvPP{FYsI*<M9pkfrxfrt{d<jURS@w}MUDaK9L5RlVgh*zOnx4v
z<4NN-^AM1v_6w!p3tH{vb*n4RU&_~$*DR$M(OWIKe4@3m7IcW8U6^pqlyeoGyD0wL
z?eo?*<Ri8!;RoB2WLWSEV-H*I)_m^vNKM$l@k5ZkXyDI}D(GJbDUH4!P76f$wSE;y
zxMp+K`o_{5*T-oO9}3u|QOP#t<wRDbwQQW@$>ts&mTSM=>$dr$Y-M^nPnzhWY=D8g
z^{#yqG0kVQ+fII(TbHTEI%TNAnCH(s_EIIH;^@ov=|h0qZWmd-_)~mTHRJlN3aF*7
zsGC>dP+#8nFG@UBM{J2z{Ve$Xe2W|;_K?2>ViUv}!LxH)$44DcDC=Za>y*5_V^tq$
z1f{?Ct=hlfi=&1JX#BVPT+oPLc3`(RdE-ux(<^d+mrk1HemYk(hd?3!Pt(JJGbz^U
zzbfbKB9kUZ!MXaq`ML1t@Avlk<l@>B^QT*qg7!t2CC}5`k6V>mkS28X=Ck&B^0uW`
zpZCl^vXV|aKGJkh7O0}b2!2jDH=@z3+P;0el~)QZJyl&@o{QgK4SCH^+?YGj?hgvH
z{}mhwF(qfP_aICV#%}nGPgsX^UBp{bUD2b(Uqu$I;8&2F6xZZ0>gD4Z@KVx4QuX+B
z!Ke@i9mY0%s#)3lUD@%=N!f#*xqR{a_VHe|IeZ|D|57BwD-|=Vpc@DGFud9%`??|J
zUVQb_uE$;CO)k=^p^odVhOEcG{Z&bub=$ajmOXxtrcc3G<BI0h#|q?r?qrF(#gfyx
z7j48xp=eETqJD#ow0Ix@O=!<ds^uvxiQ5K)nx7+Z`}vzU-ye0gxylWBZS2q-dZE6{
zp+DUxvd291e1tuXYYlkDQn6TNCcP`dL&ftG(NPQ(LZHOWn`F95)qUvpNaXa}+1+OU
z<F%;R-_rxp>&d7B+vn65__mO0)=+Yz)zyO=uOXno*pQHi$dt2L!=ZEo+7*2tTCq|!
zuN!L$Ds%}4fosQ0RfkPY^!@o<RZTY43_}Oo*EDWTHqqbid4GuvJMPNf(uPA)4l!tm
zpMvbA4mYXRKTw>`Gmzs1%E0iU#|{XkbQk#fWcDu~WF4`3Y4xes8OMEDweOS&Sx#1H
zU7CB>B)vBMchj3UtWbPIJeM*eMfMXQ!6|xCPNSVe_kOEJIkV`E{yLGnTbL9s(ws%5
z9$Wne#vhdPeDC_`rs}?no=G!KFQeJka7@$fm1S_}XRoo{e^*4*84y}0_dR+q6HS4O
zD@nVWM1lS9jrQSs$V}2>`W9K6S~2d)y2?6yg!;{gm6XdTOg_%g*v`CdDPCY-VOjXI
zB+KH+7w!(Zl&H?D7HMk!W13?d%wvWUt}b+~PyhWDYfby~X+`4lLwkx8kU~Pdn*`_k
zLfO@`haU|sf1>YH0NjuOJv|l#7|mPoF{{~S7lUfZfzR|;x_ghOBykAFMQi+h;?lb<
zJmZwZoK%_P%0xcbaOQUo9xJUwqmDP1cz=ObgG%=zsag@#HgoYWlxNH621xT{2tf~G
zuAI7$R6RP_R6}CQv#ldT@w<nLo!;J##{MP>03t#5|2@W-F*)#_m5QpI5=Zt;<}>mz
zqM|JBg2p#()e!v^Q6^>rZ0F@A|B|U>pJ*u^r*$|eNiYV@Bo(i_*oagk<QIyTj8NmZ
zv@Y_yYi<6hPnA%kk;r!ALOUN=o0r*({I|R+!dl(|_~9qO34^9K3CDu4hrC7v$mKso
znX0!~v#E>q){&Lg;YuR6S})MO(x(KKiE-<d#P?}y4!wRTFRgDPs1Dgf*Q9CHku5|}
zW#&1W^K<~kpz6Vcm;PHgt?K3TDTa`JA@H^e8R3k*FxI_J-nANx)bd?Ej#%KDSFiT!
z1h~7)CIgU@^oW<lP>KHKS{XE;NSTRjU-e1By2=25=DL#M7d<muUu*A;>DlH|`{vYO
z-OH0NH#Kew&n{!0UvbWIKAEb<zf-Z7dofscX?I(o3wzlx@8@gPpFVL1v?uiLSlbaW
zcY584nG>Nls;_?iy}Q0ccIjR~%ZS`|PN_o-Rzo{FhPOSK9`R+J9iJX<_Ah0Anb+Pn
zkm6Cn;+J!MP%kH!u}RjP(SY4dbgnc);`Qmj+fEh7yv*}(@tPgXQqQgG``DGgZl=p7
zGHgydLFmltRwiDuKezrU9%LiUUFc}}cKL6I-;By{CMI43X0Ff}<|Wyc1V0qf0|r7b
z4y`LJA7bK5$U-4u=wwR35Ukh&7-s}43fIL-7|7rp7ezRf&LY%zBpeamMTVC)NR~qy
zZi3sM$+}UM>N=Cq%N3u(hIogrR#}fmY<xMPKG~GKA>@OGRX~6~XJPV>fk^}8ARUyJ
z30(MM**7X2d&@(J;l@6RyDN%r>t$%}Q*V`Hz5M;?--yttxSveIDkxo^spM!MxVD+*
z&l&Rjelu|G)T-{ad>!P|_W7B<r<JeC(L~N2{FYvoYVO{&Plj{m-E}fKtz)A$E@u^(
zm~d122IE!g8-xSTzFa_(3P^;Wt8Gy)bh%fpN=3R?;NPy^upwA%R^wr*lgn}}e1&=l
zRBnQP-NA^37M;s`_FI7PWy_5-5|!eKOe?YF#u)}bkc6Y!h8V>OB+F2D`?gyX^Xo+w
z7+Rof677hqRe8bdX`Sxb^~oV@C`Bnb=52IA^`wE&E4B8ax7Sb2xQNbGu2GD=n{2(=
z?vJGC+<LYo&6ytEV&6_#^~u1ujoNL8*d_f-7+>B$HaPG(gmaK?t<r9Foki!L3-mSh
zh*`LBeXz4Jup;E^(+ttwrmw?Q0?&Bwx)-ar$Ex@`a|+MKzaNgqW#>xC(lzMUUoH1A
z0?ht?S~^i^C#Mwj5}ipT73KbWYfoW(5o75(VqUx9z<{(9(kJyvR=O3hogOEUiuMcv
z)#TN=ZbuP4ep)r|9;5eG5j7c>EjtI;fg)R$;GjB>L$38XX+7a}T_vz0Vg5_<88!jd
zO&0hm4fX%7Kh7e{AFc+tC@(+<ny^M$6^BYGVV4O&fzZEr%v#PSaG@xF&+6rwBQ4>m
z=W*DXCi*Ir)%{JD&I`HK(RtoI>P2En_O30GP9dOpefS~meU9%tpza(Z!w}ilnMsy9
zz%na2gEd>F?y1b$$IRqB%+iXV%`|l*M0buO#sHCM0)+ICNUy}Tdq~i+sHo^H+8fIT
zdNN}O2u{;%xgD9_OO7u5hhjTN%Xoh7bnf9YKlaVCfR$5GI{OV%I6yj`QDht9<}u%)
z8}`cL!ML4J`%qC3&qz$D#DF<FH`k9t=y|c{AW|45inow=8U4m7R9si#%<qQ@XOr~l
zA<H<j(sw&4*YHZzP%?hR@9!W;rx_or{?^7lN|WRPVcrxH0V!2AH7k&>O`Kvv@bFMT
zajYQ1W_XR!Njm=!jI5>^^y-wDx)AV#3`Epub@g^;?D$g`a+-^IDZKeix9Y2xZQ*u1
zv*&@$Dy}k-VVcyi#3CJpxb?o+jD9NHL2T+|7y?Yg@E*`mPMtXulSL*@$qpg26|fpv
zQlfJ}2+KeGg%+>K@=q@{{kH@|wENF*T{twh(R@Yr8>vzrqsS>lUD;v-&VI|zqfnsW
z*TAk-*U(7Bco+!f(m=^VnQ;qTA71TwkeuxF`%ykLnT^R;nR#FC9DGsm@9RLL?j;0>
zp@R}?IVYYgp4$!CL-78AgoMnTU5m4W3IX#9hzww%5P?;4;NodyH<OqJU?k{&7e$|l
z;~BTG{+X#<ip3KffM)EA%}xxE8yFlUu|^=r5ei{!tC4=J=D#(pD6@$-kR@)xSI68U
zC2|+-F=&pk4SX@LX(Q}qn87nX*G`gKw#&<t^k@i%a2^*{MUaoO<2X#TB*YBE5~wH_
z#rxsr)INMzNIPx~BcvuoQwV$>a*L~oWqIx5I&UtwUO~=z5CwYQygjqF0%E4KWgrmi
z`u3ymwDP2{=z%zbqMOl$lwe0}ND3NkiF0|Ch-e|EIQZMmp|~YdFlf~<H^Rh9H~m=E
z|2~hhaJ%Z=hVBTRpEN$d4M+#Q1_c7b(-K~t_D5oT)NM<INFagqZQueLWU5eG17B^U
zq~c40dL3nUO`pLyn-j#2#JejXAV5q=(2L_vrcg+#G^L8PjS$!v^>BGiS?lMspF*0~
zd#zgKm4S;~Fk+uX4K!;7a9<34{(OG96V>WUl0d+qE00&LARbLJuLGSJEG+7(yuAsX
zZt`K_;gBA+Zk>PAp++Y7AjgzX;e-eg-k_W(fpBni5hth6>7m45W}8TTi}zf&KILo(
zY-t$Zw)e!D<%KHdWlDN_a5<294+KP$7!Z6_n>sy!qOju%j$yJzHSsAx@#BE#;FAk(
zlzfYa`4%!Km#oN6Qidmlcw0!I+zJZTa>0rI>R(;>o>(CXgof~ZmX6PU6pkFbckd>u
z&_wJ~<m`}k4Dbg@YCIgc;7d~KNjINji^L^V&@{xm0qJqE;_`7|ZhDK>>qbbE+FI9j
z?7OBomK!qvq9V)x*RQYq+|-Lc0vJUabd8{&Kn04VH$vLNS_CU;9pV|IBZd4&A{N&y
zy>cqDSMOCQ3;0~*>f+M(a~g~2dstc7PGXxx5|{eMYYxy-Vm8v3r<Hhfr~1QrY2?LA
z<xx*CVC!Q53DOsBv4@E+6oTD}cY8PWL%{kRVWs0OdxWVd7_E|jYQkRf%KRwK(!aBd
zy-T${Ijb!SJ-iRDotpNAN-!$Z5uKxx$=0lXSJoeA6MicyY9x(|<o{0+@H@@$h0)0&
zxl{hdOfUVSPv0hL_w<wXs3_1A<07<&f|}Q6Y(>6pPArsEJbkkCgKY9@=Rbkh2=`fw
z_S%(%%ORHwgJnCP-g~sh`BC8UVKAMt2R$;ieFqhdb+n+!)?r;;FDjatwExD|3=@_A
zE9_mUs2#!JizsNX5KHpvc*$0ZYY`U&oCnv>lc{8_0sJg<M1&0z523pK-t;d;p*O}e
zHEUa^$gNSo&$XqZ&e?n7`~9CqUU2hXo$zI@uDU+)F|m^Av|NK|=$Z%p8d~jQUQ+l8
z`BuY<=a3DD^oo;BrWoP}8!%2yBeMo>7HrH+Uvz^>7#9)gfy8X86vC7%l8RhBnp;Nm
zWG(rU{#_#nnOn6%j+`1(C=)1=Yk&%%d7E1lC?kC`Qvd?oR+AZK1_o6a3C1DO^gOze
ztZ4<H0xatWuuPm_wsA?4By04Z$t24k6@QOIE;z0jHne#toq74j<}7+560gG&B}8Ac
zFL+zq9?xQ*Y$!4H9Ux+d1tVmiga%B-kaAKH9VoF+fg%MSCF3^Hnsk7hoq$)TrexVt
z^b<H+<W8gUeqy{&jgV+aimHM(o<tFoL`9PN3(QJ<#LLHqUaMF+lIH<d!XIZ0&mQe<
z;rVWPz=-uyM|UFzKy=WY(YzDbYCa}4%kGuMwpM3#FghzRsI55+r#O1P#83vsjdi{|
zwlT5&gQFiu(_0iFY)^Xwc9I*MSO*C4(+VS6)JP}E3kP9uUVdteu~Zvv7d16CiK!An
zhaUS!)oM*(Q|j;df2A7lNufnr+AmU;g&vPv!XhLzq<4skIj<Vv4-GJKFl;k#gaRpv
zOqD=PK$TXF8imBrkaoT@F0$pkC?r`Azq+;-KU`&TcJ1;7F|^ss+q(sm6CsVQ+>4~=
zTent2QQuX}-!lCVyLE<ty|Zoe&z7UIoiqFFHfYa#9$)^&+~4@oS^5}rSmlr>MNP<v
zl*G;FUu=QXgbe2aZFcPw8Nh~)HIe?3q<3iaktfKr>nVB2na==vHSFyA9iO(hwx+@f
z0pGm)$MYk9Ciug!-mw?@9nHoX3LYg%ei)9*RUa}0JmW-1<DZZfi7(U?WRlLcBeT-B
z$iniOFu^;|?q+0=!8?c^Z(H}Y>BcsCqWDdD@=(j?(=r;werNlBZSve~i5o8kAizgZ
zJ60eZvhM*rculAdX}icx=bUnm2RjGxP}CGeHQqt~J36ys-^g~bzD7caFr$z}3P8QN
zoS%HGJ-3cTZIM_V7`F+O+E_BCWOSx>o%S-Nn^~h|RKdHxapv}tg~!A2w}!L;0>L>^
zYhIDa4}1mo5GBmc4#wj-F~IYSHJNOqjR1ZUgGUL%H&^5Ic;cHs*;tafxw+(Ea(aC|
z7=Pe^>_c9+%I{CQ#^3Jz2^j-}tB#Mmf{5D+)9fp~tMIh3sBcJI5;BR+NKm6=d^gt~
z<LivaQA%dVAi=#BbzD4t5HTe}g%05)`NBvr7xXIQ^NUa~kTU`c#V<IAe41Qq<jaxg
z1kq|Az=))*td_CyDpD6Bm5j*!$q)=cCz(>I786g^kBtO9`TXeSC;e@}9Q@}<w<Q7s
zL?;sgL5``$gH<zTz~3ZL(5@+Pc2}>PJoB5SpI__>*P(rlhn<dOVwV61hdcRajRI%l
zYMb3u9Z$=>JpwsuUf%mVj~)(w9`$9-l*#70Y2&HyR+08VS4ZA@XLiT=>T-Js=0)dt
z9`kF>DI#1?ee#Wyw0`mijmK)Q3+mho23k9~w#%&!L+9L7$Sg1TR_BVZ1uN|ex8E_p
z7Fu0;kcbiNzUvN+ZP0LYdzf$-CmW<q)sy=rR-z@(v2x{myI;u1-0a$ldMr-wPKzxx
zw$~9pPWf+XU+bR1rFQh67Ve~-{H>;?+$1;%0lt`@d>I}lb9sQsxR7j$E_KvRP1$-X
zJLxlss~^8sIZu0Qhle&a7$l+s1p?p-#-a!z9u15vnNvl+Dp5hC*=#xXK?bHK(K+H6
z2C)}54D>JvOu>6|2`PV!uksK^1mhAD{G3ETn})p#*FVq2pOTZ_=ko&_?A)xM61t;^
zt$}SmMD~~;)2k)LQSkta$JROl`|(fNQ)=%C(+r!n^<TSlzhL5TPW9vh^`)>Q?~C5I
zeVzEQS6DrR>z6}vZmxhwQy`C?TEmrZrP|8!fA?n6F3gQi#m>I(y{5j-WtzcSV`Y4H
zyV@NeEeO#D<}bFm9j|?&p)&Bc?B4T+_Psj7F?Ht~Ck1H2wWD6{xMg(kG+Rr5fcoTq
zC+4?TMmhM`h)DF>tC_g(x*2P~(dPuuwbg87b1G&5QP@I6Z1xFk<&PTVVA0?az%|X{
zQMn+GMc0)o+EUaXl+QXlJKOi9>v)a((@zqmjqAe0!ffH8$Ld^lQJ9-MoMfV4SF2Cr
z&d*0vJtlY7MH;6yaoKPP3vUuvL4GUwUaYKcC(Of}KoXc<o)ovfG`QvS@BZs&dlf(g
zC_m;)TIsOmYjT15B(L%UPpO1oeQntu=9ei;t+#Z(b<nQ77gKVnZ}`xX>T(Hx^LqAo
zn?LVIb#(Ngb!>mTqiVHBz@M$@gLN!C^)shw@&e1bIJelx9rJUp9k}n2ta;WgAz?n@
zc$59Y-(UZyLHn!NKh(<PN7Rf#aCSCxMaQCnsNng@z2mc&uZ^GnEUGrAcxvLg<q4-I
z5E;&2Zr8{jyxMoI?I+{QrFBm^R;E};ig)poXc1hSMA+Gaehkq8f$6x1kWn)$**$sf
z4kYCpp%yrh$PG)#Bdw)|o}@^vLq&m<rME0BWg_QgsHu-jo)205BPx-3;AXK&>P06b
z)8pY14@%7LMSQj#x@dXt^UdG~QN?#lQk_IX7$ikvL`FAP%`c8S228Nkr&iTe&3IdT
zZ}_6v<lVc%+sXEa;urCC{qPTRL~Ht9mo;iLCpA8Kr3)At8U`UXdH^<5_4`DJX?%wu
z1N^PZ%qc6YHALb{<nNP1FDWG17D<|eFkpo2=MeQ5;$iq-=PgRd0q!Lj4EZsU<c7!u
z@^NV`qobp~Z9>=<yt8+pHfnqV_lc;20w2nAYA0=D#s8oG93leX-aDzOstL1bHA3qU
zQY(@E2mTp0bVj&4i5&oOdHl#LqIT4pIx9n9mQfDh=31NY6^zmGeMcHt%l_ROuE5aR
za2w{kb!=4|2*fPmW`z8R%X7Bb;3y?2bV!Ye1MZ-r!2b~mhZW;(exu>#r;={Lso;0B
z^W4+#N=y#NKR`KqcLLEW5GK-NwoBSUB23BRhC>R#(BvpkDpjscWgNm7F!)q$&mQ{E
zpFdL{(YgEY(PWykOpo0;+3U@lduRLTQCTO#gZMsw{h9?<U~VN={49SpbUKM68PvSU
z1m@bDkBOm;w!VVA=zf?h_8?mla3iJ51Em)60PfqrAEUJBWM!|~Wd;nay+jrmUal8V
z(4q5*d|C=TiPvx5WJ)4i`+=^d<!Z4+22_Hp!~-B2&OQ80N@pw!=@9IM!yuG8BGA>%
z%$`<Hp!6dp94>Hm7*44uXe7Gyl$UUP<uZw;RmYnpudJ)cqH=&t?$qqKC@H!xAl3%q
z(8Yc~g(BDli89Tx7i5<MB6bpV#HpPydQJ0N^M?-+<}wlY@81`hKmhJ}1e9qfd^Ax6
zS<v~^BOZ!oGk;~cBO&681kD1Kw^c(2r-?9g7^|8tWBxJT=4U%OxTvQ_=`RRM`<UWK
z;?+ta$-6-6k{dPzBfk&f?c?yw3?R6IWXch<Ih=pbKYZZX6JJ-hKmLDO0HHc$lr-Pf
zT`6N#Y&g<N7R_OEq#t-r3}~gj7s=XUpSCiUce}R)qA(vXeHo?-2&nR?q`I|Z#g``L
zn*w;Oc*$9T1XF_t8AL9`-@y&>!VW+%r-)q!9ca}2+`u~|WC*cJ7&&T2ba+_#>i6}-
zkkv~<dR)4m$U+omnB#pP(m3QpJ!+H}%j!l$7+GC0NVp`Xw>G62NItJCc#6R%d&PZC
zpmDppy7~#Y<Gu*_oj5@pX9#6Aj6B5oOyR9atMSpJP0-^lzkEkvo*04SoG=28UC5CV
z{@B%3)7`xWLT#fb$BrGVK|FJCa4_f*?We~-h3VX2sCqn*irl{ov9aU;gNUf8$wYD$
z!tr(>hXFn6+h3k=8V#d_fXp3NcqlU8R4pybEWYePs3D}H>WIU0b9Yx<)ZqY^RRjP0
zVaH0?7OUU9$zj_n=C3Hj;hHrt;QX}fX|R9kQR)94rm~Wf66}3`;qsQ2mgbh0HQ4s{
zo^^HS@oOj)P~<Wk7gI@jU0{9zo9#L#ASJOG-SPS}b(F*c5I+#&Il}ZKN5V<w628pr
zN4>yfheX6<tJZ1DwG|z!O4?a<Rko_|sJioiy4zS%bN}dhmi_LylT!qAU=Ipa{HGWw
z%}A88Xo?-_?XC<I#w-&@&jcL`!bRvQq?K`-u_eZ`GGG^kz{bb%_K6K2j|;erL?-Va
zF^+Dn<`26mxqETx8gJ5OD)T2SE}h07Z!hoDs7xdH+sG2%_;2`tRM(0XE9_icGT#NF
z%-p<V#~vJ9-?r`-<3_%ZOXG||6b<5ls3~hg@3ye8u-Mw#roQW`3=Ew5*((fIW2EEp
zdf4-0@ij0ch#F1jt7aZOveov3OjQ+S^A@VJ=Yky?!rsD7M_vNau#?|GpDMqw@L}`r
zlVj#rD&U3&i4csWcIyl?2fbITW?i0d(aA0`Fx|i5proOdm5k0<#Aot+;UE*a_R14a
z9HCvyv(HI&h6GF=oD<w{OV0h3x3abd1)~o0_5<$wzyX-*sD)Ms&MfIQw;*;W4n(A7
zq~bo5^KOXRt_qe5i{(m_Gd3|tOdWypTL1^i%qGN-8h7kD0X6b0_Ozj~@!dy0zK%pg
zNhZqUG=yH&J=_6zgG*Z$HzGme;e@zZ=B>GO2MyEackik}UExTNt)fo5H!)3HT^;06
zLG@>f>d$Xlo=rST(&()9UHsXr(K`Igb$Y+SIhvB&9!_Lq8*6K8yYec?$jG3;HhRV1
z#>Po`#dEKD!DEYV)T~&dPQpjl*!%a_@7uR;979Q}aGYb_9?m-Q*UKX~%%#3SZ=#CZ
ztgz7dy8^x@$o53Z%n3wFJgzPx;$kdw@Nf6Ef2>Tr*WBLTfWRit;YMzx%iALcF(2Wf
zkkR3Ji(Y#LFY`UGo?rE!e-T6m_dMB+;D@c#GK-y?gjVtJ*dq@{&MPG<*%=F;9EMkU
zUTwU~KdaNW{r&wF)}<7{y0Ov0hG2R3poPETZ{Ud+D14k2={e|i*Wz4^<JMI}U)sFD
z$Y^R7?G@+gU(oGwvCrJ3Gj9;;r)HVj5ySchmO(OS5!e46><Cb2hjMC>-B&}5^Z2Hj
zXo-5d59^B*J%^?;v>RT$U;sZ!wk5ILp_*R};p++tQlB5x4`9}Y10co6^3&S@FM$x#
zl%K%ef?!r+#gXHQ7r2j~2WNm@#sYpQ8CnZ!K!hurq!Uv;TAQ_-he0m({+Gv=C2FBT
zONZz}Q6AshfBFntd41&r)83&lh1U~;hh7nzC?{vzCR0;W(h<O!8i8ih$IedbQ>M3Z
zDb>TPPWv_?)9%;IOapKV2?1~n9>>_l;*;M6h>~Ljbo*`5qoue(Ma0F2N*bVo#QDHh
zh_FDm$724TCY$>9O{c`M^$o62P@?wt4^n-V@{gIpT5?*J2CO2+0SZaa$F&}g87Z&d
zyrF#?mucti>q}a;Sk(wgr|UZuEJPu)Akl+&=rAOP1M{?uOBTMUED`Sw2qcpF=W|oj
zb&#bxLe+pBsM>j{sgX;b_uthlCdPWe!h$JW-N<gP>D$=3-v(fkA~Q19A+IhL;*LIY
zJ|puSC9u)KuUv8pS7M)HHYVO8v7=5POV1t|S`gavf*&B=@N4@oVKy!sQpydF`gwcT
zkg6MpG?^B`hxz6($<Vw=$n;lOrhxt(F&Ps3Z-N=`$Ft<_G)PvSQ*VF&?p+9cQOA4A
zqbnExFq11jDM?Uysl=Lvlj@ZGBBc-APVP?R<((j&kkZc>V~}?Lek6X=ApS3or~Wau
ztZk$}D9-hgxiW&IP44Y-m&#vs=-TErvBeKfdj|%%jvP_R8Ag;ORzlh_oC(Vg288>^
zJUR!GYtz{`*lz@YlRFu5kzfp|QIq-XJQclYHn%$o%LL-qgK?M0y4$g1COr43y?qF3
z7(1}S#Q1<v!p)uowF0khjP$PA<-ZCD6{gFTO59>a`O9?UbahnAAPI`hZhMMZ3UIm=
z-CkaPaz6NPm$O+&ggGwS5S$+5J0W44<8obvM?*;`1pXGBAYh^`Se^E@p5PM@AYTWW
z#CSU?X@am65|``<xAFcxA%U2!+k1P3(WI%xa@>p>7JI7w#OGG?Pvxgqg5KutgzEw+
z(P*kfVuTDu6%Qd?CqA?DAI|kp!{<lp(6#2HWVITtSjQ71J+_U30-l68An*mE@jpvN
z=4~N%G#?uS?DD<p*D8rRcek$0e<Z)$;UT+sc$ijNI=iz(LsQ`H+1v9!`MQyGB%~5>
z0n~p#Vq(H8Ds~eS5kPN0?&QYB;V))#5>zC@7)uM`y~@{kGX#q4YUms}^S2^z6_A8)
z>k%D6C#|W0|5K}orPIOtbLZV}GD&Tl(`!|XF2A|MBUff*!zw06;Y6J<IEk6Qn0zqZ
z2)d548J41(I!QW%v@|p#A|k38pMV3gp)}!WLeezpI=CD=mOaJH#H4O$m@qjSxH=H<
zj0~BCLBD@s;DEPZh`=tih>FX~vI}GC?9I*1Q(>fV0q68-z4lbd|MSkY_XGL^1j~8l
zq^FCWh1r&z;P|Y^hHh%ICSeQ8$jTlYvcU1IfovM!u_S>KAtObki$R*im_%n3YZv^l
z?Th6*vJSB%cd&$D=p%tQ$TMQYz!o11%~&$BfNUhvOaTLrdq@4gCaWPrRmt=(I7!gn
z@U9m*y8ZvDtq#A#Ay<H#0zJlZbU|<s7B_F;lp~{QU3x2aVVO)#Pfy^BGAUINZFz<G
z<4ACt=*Lk}Al@QFiU`C&ro}w&XJCh+Un)mgco-OQ_mZe97FJf;=oHNNC14vP84_mk
z2V_F3OaS}?n1Yyw-LEr`rgJ&G_91nFlS7OMxkL!^CY`hUIe9Mytq$z2FS!)Rz@^fE
znCtlAlF+E;z0`FEwcn4mFf4r!5t;beATjVZmhbl5bh+w;u*jbeyKXC;)W1IKCtQ`U
zdq&Ov;Pmj3U3HVEZ8QvQ%o1ih4)L4+dRf-ovhuU<y0y~m6_1~)*%(Oa%s;%Xyw9?n
z>Q8BPZQ$hgl)P*?a%H`E`L^!wC0-^j8&h$Ox0m+4PExYCJ{wMZ>ULs`+1U?_Tq>2S
z5rq%kSk8ReW;ejgBf6j)CsGit^Eehtq^no262T<O({(;dnz4G2ogzeCI2CpZ(nA9>
z4A+kqNF}8!C_z&EEBTBBfxE<nF=5<}p>c6+1l0f+0q4`-KBOZMOsgJ}mL^JiQ@E*^
zxKv<}LrtUP=H^Bd9XMFeN}gl>lP5gliy&MkkZ~ckLoE-tuyrVKTx`y-!`6X|bN6lQ
zj_!ZgUCwW@voq~yUvfxoZ{>H%hi<aX4R<0%CIWl>X+E4Z9odyxhQDD_ykMG>)Bm|=
z|Ml5k;SlScmQw<Kcwo<Z7ebtzs+IQ`PJjJ6YB%yx&Bd<6Dq*(3?_!^#EA`D=%)d&1
zi~N!5S@V-l+IywifK9Ak_2}AeyN-tsL?-eU1RL7ht{I9q?Kz+?GkH_O*0Wj6=j)i*
zJ|QhFQgq-guXdb=T?-(dfu|DDCV={=`^DaTMsg__U6f<4mw_KAko<~mB6s5N%ujNo
z;n)%7iYBSy<XMn}V=&zXsP}z-_6XvVtpm6U*T0sj2by4NYKmkJqCjWlLU&mhxiaXk
zg^El};}1@a+ZK9kdOdc1a@M`;3AcDfZtg}>)#7@$$e;E1_YbN4AYLz_?eeLIZDCh<
z2g6+VOO@tMqvBc4G~Sl_mc#k8{p&ZT?x*h`NKq7uej>PrCMNLPN5$=1IPZN_Yg^ye
zrv0?X+P2;FI(yH^$no%)F3QJ`<e}x{%G1qwYg*EKs|}3m74^5g=g-NJrHRg4+xB_X
zRqmx8Q{ruJcHNFGI!y)<)J*1l9sTda)zt4VqcqLwD7;UE9x^2A&``#I|BgWU3qrmR
z!N$(c&N|{;!(4YYrv-WKB?vS~*M5+yz`Nf5N&mN+-U$U2MkYxL<wJe-s^Wiy$)W&m
zbC@XZ%1{szS5se4hoga#oO9XC!XisXgsaVOnQg%>ap&$PIrYH5EzLYjOUa31QO%E{
zH%m+70(o`;Eekt9SM>jjn;&q<>5<ed96aIrp{^~kNjBfv!%l`dAp`|SToBebserJ4
z2XmhGP*a{?ZCszVapM*}f<XtG3r-aMt*_m!>b>`9n_zU5tNFHbds%$=A|FrCi*CB~
z{B<*(l7-d1D^XF~4NyanS+`T)+L$qTME$-->j=rGy0*!t9>JsG`r)n^;<gupBsL1b
zj5da$O@72~WbEsZslm<3lqh6WRIq%37umn}Q_21`W2dL&f2*V29@f4l=PZwjhdQU0
zC=X-17?P0cAZLpSVwnCBboDBsmPLJ(1V+)o(9zeg2fV_TK~zEE(VAZV?M9d_Kx7>F
zj5W`ntBv^R)WBc)yt{mI;Rj}A^pB0PV1|>@+w&uJ|NVZkv8i=w2SvoPE~}@h>fTt>
zd2uPF#4XE%%`dV&^}f5LtrJL(W|9{U<{Qa?x0@fhY)4iQCRq@HJC6<87+kjw>!Z0w
zkS=n*nUeILnJi<%U<CT8VMh6v_O1yzfM{Ii+9TVeTv?dI2=3Qxb;0#ub?y|?A=(Fg
zBb#%pz0&tjTQFP)g5GuR&nkA|-C@Y_^zrdAj=OgGG6s@Um6n#`5HcOTiPAVSIavq=
z(_)bg_=t;_F5z%ze=s$-w-+?MC|JRhCr>uFwVgX&iMJ{)A)yW@(08j>xVVwsl-lUc
z%|r{~D2m%49$A6*V;I)AqVtJ+3=j$<*Aj$Axl>!RUbdco?d0evil`d$)81$M>z?#&
z;7;u^H}P|$n5d+*YF@wyCuLOd;Qu17n%11Y_3WQ(3UOSQL?-Y%F+qs1U(ng}B`Nua
z#K#+CG~jo`amzwaPY-uxVccWX76r|W{}6oJm1FD%Yj=<GcsU+=am5jw17@d@$*<rI
z=^Y##e8xU<qFTfSP(C>e-e$A^lnZcMI*yY3f`Z@&4>rKSoXzxjA4W8!fmOzdMu;I)
zBLF@@ILDwAjYcy`@-KGZDj*-kvWNdGE&E1B7{H>Qh2d37Q9?ojYDdBP`g)vPD{$eG
z1bpRle+tzr5oisKQyPk5K#s`RSk{LRAL7ji1G#R9IZs9nBN`qP+rYR_&f|zk2b<gI
zjUe)9qWe;3!SW$_@VL9lTPi701lR+y^UF(x!S(Xu_bryk0b1}qYvW=AF|r3AkGvIN
zr0f&2z6~b6^{$G(^>pL&UoK&Vff^w3HWWfa;{W5Fne5a?<BCWs@Vv-i=Gs~s;JdcP
zo^~@#he*)NAd_&1PNE;@Zy96&K?4Gj#-V!2R9-@KHQ~ufb&`0cu5z2sw(Ym$<^O#(
z3o+!Y**GduE5`VzBUarz^eUi`X6uMMJj94<iwF;=#qMME_Vy+-fe>LxO+lmcc}E97
z;wp)m15-_r0>^@q)&UKDBBj8$dZqo+%Ok{Djak0IpfT^@!Qj}a$1@@rd-&D@$Z7mV
z+>szj$t>hSh_gx2M6wS_4iwyZG!&>$sxj&b)Cf-JdO-UE1P_ft|A<z;0@O}H&fBh{
zega2H!J@#qAd2iqgvFCM98%-rN!f!zb$Ne}mCRfMx&<I#LE(|L2qSGx9MQsO$Eo0m
z9LM$`gSAP$3O1LCLii|Rn=dXdwuKG})dHbz`hNbjaaP&k%9rK;ntExe7)0+bG$Se7
zwr#tOk?&;YB9eU~Fc-tz!h$-pr}%*1!^LfpS@rdmyHY^6k@PPz4Hjn_@?vc5?5NDl
z+Fk2#>@BB5B9J(K_a6p)^jeUcBnA-FAs;_~TSrs;w;r?}NS}cWP(pw)Nth$SjV9Tq
zcKFu8R}kIGW<(379b9*S`G4n8pLO7>xHC75mSoxD!b*&!(++~MnG)K4Hj04X&Q1Y@
z@0)DA^mTMJd+aepq43MwBQJs5=?_*N8M6ganap+U>+4HZz6w`3u?P?tMd>56Qn8#d
zaDoxW(kc*&9Z#_>NGMvZ{T3irQXRnlPLDgygl8-DiWL~S9*rVu-z|Ao$~dMXE>A}U
zFLmO}TS0ONKoY2f3~3~N5ix?K14++oT)KAkD)|yI3O+Yq*$La=L#{P54jP~x$!pSU
zpm!jn^T_B5)Mn`I*MP4jR55}%4wt+hyu9Hci6X$?EoKBrv=$prUk7g7h%V`h_3PJz
zNMVeKX!`A%3~rcIh^RarP@6C?GCnwA?OXrf$(Z@;>?7c!vj~{MYOlpwB_EFREncOs
zgX<E4aS1mep?{l!LIzNPo8Jr*{>{1PhR@p4wU|lV2&VCpt}OBsX+Rm{^sYuCC7FGU
ziA^%gbOhI(e^*HAr&B&W1^n%zJi5(Kuk4VyFczi{gh3(tKgb9%TJWp)O!H1e&f<`f
zDu2nYY)SnSoz)Pljd}U2#3+OnLi?;Sug-o8k{0PHQhRyzG4;rPVySYMJRSu_Rb4&g
zkvl4KOa!YsmiH_I<<@Y{*$5qU(kZsskT}Gu;B?RW<xZiHDjl5~j9gUJ)s5b&4T2S=
zNZtj{p%<Kxt0Csf^y3PBu|$-Ishyq=JhQ)TIfKYk@<+6^8)8C-=NA@Q9y#s>oX{Iq
z{(qn7r1H0VN)RZaR75f~4%n8YX+)-jSZzDM_vR}+9Wt05SGc{s7}-?#g8~H|fk;|)
z0>;E#^WXit=(%H1pjK?lpF#nuR}#O;$)&l(jbUs=g5r_U0L~=g(xnyX!bWL|T^Stq
zsKwcD>LLuW9avzgLSMjKGK|nL?Hneqa98{Q%U@VnczJ>B58Q&}`1$ye5BRmQ3+RWU
z+?BE!Vq44vBqhJsZzJD%oLrP3Y}!n9mX#}4lF>O^ws2A|pQ*UADQ4E-!-o%TaeHI0
zDaChyrJ$zZbKLXVhlq2Kt2%mmkQHU0c$npBqMuZd_2Xmax4EI-%++h*0*28y$7h~u
zO=xT#bbI~ZzsEAJl&Auok0DfS;P9F&&Yc<iSa7Ghrx*bB8eiLH5Gg=}Rp2|xU~RaO
zuHU*fh*8iOHWUozAYf_1uY9smTKzu3hMM9E+cb{(6nRAkKjD0MQ^EKgHTb<7xCJRe
zxw*$P8e(E%mZs+?`SrM>@2a11Y`>(Japmw};9ouQi%i>6AUvQbOKbm`oi+B4MNq%#
zX`8GALSOJKA|dH(oS3{#<b9buN>rJvjz%rU;^R+lnQhuKVPIdt2kJK?$fm8>|FDXx
z@58K}@`s&RKGe>xP@r2K3eRvKv8iCLEHRX$6;}-k9^su0z`*@z=a39rq9MZgjj(@(
ziE9G#MKL+?P|4RRTrVe3*qud^<}CIxp;}2L1EH9fjt&D3AjO;;ct}wK&rZIv4$}{g
za=oD=u72pyjpn0<a#uDr@N#q0f>ORRC~Ns73|I^t!J)5nem}aYD7Z;JSj6G`#8pc!
z1I&R)?#hEq7V+)X2g2G~N_~$}nyzk`e|qk9WABMY%s?>d`L$swfa3wWFu9ElZitMd
zP>@4yxZxx)qUN4GfV2H2w6ix;m%sVD680!l)n#}F5mOOqwK4L^lP8nrFqdHnP!{*K
z?Z}{`dHwn|cwrS=#C0~F#Hya3`;tNwZKM&4SWysY0>OJPxJDspgiKYeMvED=(sbUr
zUbL=ADhSaH;UA7l{_|FeiP;g&D9&d>r}At%jChojB5-@-Bda0)5xm!xZJMnahLYnV
z!&`o+>{~f>&rnieeyby@KS~gBm_m!f9d`vhy$lbD*oFEEC!s<3ApSYgNZTGc5{9I`
zHj6l}Xd3L`I$XP+lS8^!20`E*`0efI*MKj195d0fr=$btqwDN}jIv~6ln|>!A?eA$
znil=$PEHcIq$w0csZ&!(@+SBB0df@Lhs5;kcMSumB@hvvc`~`x5FAtgELg%Qu+$gU
zC_Xa%W$fo>Jbm^+E&v`yYUBQg$5|d2pNB_+bnK;?=!uxZ-~Ug{C*5~^MCA$+J$%4c
z(cRr0XIMYfvtbzdNFl9MXAak>b$hRpL-IPJ?_jl{!;~<-)vHPPBjxfjyA^1USrR4D
zL+dggY;$fYM{-=HE3KE_5Egee8APFxAa2;zi+@HpVJG&HArn~hBuExPeYi=<>{MVQ
zlJ|`moz{vU`F>YZFXsIG7N#SP6NB_gH#4>dBXzc_y`2{a9V+!*92^`JfHZQQkqI84
zoykz>dE~X80`~R%`feQw0K&;dW_pqPmOx_C!D>^+RIFfZVxo{F<AKnAB1xZk>!NaU
zCW<wlXgHAeCn)JKG_47gBx5YGp*}aI#2ekwRukuXj;}tFUAP|k&UPq2&@-|}7nk%3
z0v4vtknYS2sxgS{s86NgkkPFb7pEaHJeaOWDq3g@gvq}Q%s>zhqOhq!A{PBQmm0Ar
zsbLHlVggTyOn)Y`2yD@2eDOkutY_?zHh=wdoZsSDeUl9+wn&2~+FMdDB5x8ha~h<y
zn@Q7A$b>{P&JU<KAHB80W9@7x66)|(+1lGzcZ{)iW4~B93OhT0QF;-U>4XlTD5hEA
z4UrlF*g=TWH!u)Fm}e5GB&2*Q)#_&%VW$+m?<IUiWa^5?-oGr?eEJ)@+zJ{x;=J4n
z%32IK%kQ@>x8_s79N(z?N+m(DEC(!U#H1O%6+3+>BjfC)ryEJw3b}I0)MuPqD=6To
zHskkFf>6RiYjB=$?!d}~QO2%lWKzLc?8b^9vBb=G&)zn6HTLe@urzb}+)~MXM4ztK
zPgafw;I?baGHE=ydi5=)I!@F8RuyOY$w>-GP=Ejj)D^_f+0%2lW0Qyo1rz?j=!clg
z)B{)|`t>|EE0Inh_IT%JxP}}z;``FVLc4rg7_)CY$|p(^F4+H13qS|iI24jh*i^W~
z#XuA*4#wyRep7Y|4Gkr8ov`ht8!}Ps6Yn=3o5?JPhHw3qQU~rJ3x^e%f97aTnF4Si
z`x+_k7+YI|d8y>9+ZlmSKr*8m9<IJi2D_=o-8KxMb!JBriZPk9t@o5zx_=0*I_sT`
z4B{Q3KJ{D77Y7W0^T^1E<X>a?Ivn{Uj<XfJ`FnbLT&J%<Bdw>GMOqwE=L6=~AJ5qR
zSHxQ1h>jp;0hH9t2<#$5)ldBF*%ei1KV0EFyhc#ab8-GOgpmVahVvgi%JTV;swsB*
z$5;{>;0I9Om<Iv!c}OQrJ7)r~9iT~;wYd)A1Y+&P*RLma1{5t)bqZjWtV+CSQO^(!
zD$-3;fuy{0(0(98xbSMEjtLH?f;1r?`L(zsFao+7x4yqWw{!ii8#nsV03_5mSml;9
z4SL6vhzc3mw(=~Y$GBKQr&QrEA%bU&(ZYpgTKY4!bTCWFf!i?94G5N+g2SSL4B)tN
z<AF?A)@qDS20Ka;uW&_%LlZP9fK`?~wHsi-Ool_AvLr*KZEX<{_oQLptKV1CG=6Y>
zD!cf`AGeN7p$~c0AMQS@0=Z%2i=!L6>W9r=)GH*T-}kc7FzEU=i+8QHwRan>Q2$yU
zEdxyp_?m8tVSMdUe>orA$!d6rD;VeD=;pO1|F$v6UOP9p`2$<<*t5s>Xd0(Hcssbv
zz=^%%z+PG_wGLtCpE_6mzNb&^OwbHo#!LU$AAE-!<Ok`FD>Knke2yvdvE56ec@s%4
z{itv3KxkwapybhlGotw`N83&k+(BO3^@Ek!cP#DnZZ688x7gv$qX6+8UReehId=~2
z%jNOIZP{ja$B(lCBNV!{P~W?EZ}ZlzA^1#$>qLWtIccL&lEZ&3i0yF|PelK8#%d<Z
z(7WF>d1k<}+{WVMTu}M1t)R2^5MKwCwz{+0@<*6%AzfYs=K79)NJG;CNCSx0IPV0z
z`tn=B|D@}Y;urmwhmR1Q#7|-ziQt9pk(lcmP(Nf02u2{7b<Cht{!i>ODzfd}Dl9mn
zaRB_R9}qEG_rF%-w`XFfcl#8vhUte!32^S=JySSpPhX&Uy~(t9ZW~+s+ujI7e~C;m
z76&fvM(z|f&Jw8@HZXT*evhdEqoFW9xQCxh&JBedsYeToiZCOtF!1{sps;x)c9Jm1
zoE&#SM;*<(9shND(*0$(?4&chSlwK9g%{Y9@PTayq0_*(PzA1F0{CYwKr}QJ<eq`a
zh$LVk+JXooe4ky<q`|vnkJdc!>W10I-=9~IxGAn%$puA4b*NL(aCLUPrJhz5Ld%?r
zBLQV0!*xNUnlAv4V9g*sk?KN#7)hQZ)2`Uq=AU7lz2m>$qo(Gx>4D$Y^NFXZnIsde
zufL@h<uv(*Q<+@#gscJ|LuQrW+Cl${7L2;xks~yMf`TqbA1uj7k;9Wexp_xxPTlL2
zpVxi{e(`dl#$-PbX9vSE%3v#j1lV9JC8g8rYhZx`isJ6X(djw%aSLW!9!Px%Y|4h7
z5a<o^G$1tx4~{vKkkHsP&ysBE+D*&paSHYh4z#htD_5@7Ig5T3>Nhp}V_fdLws8Ge
z*I)frI;1x6_C+>Whp7ggRri?f$)ElDfIlk!w21BezS=r*23{4}m0A`Ch6hZkWw!+S
zX8t_y=w-Du$Ro5a0}UJrY4z4CzfN72d_C%)xG!!z#*g#NQh~^XW1W|#)w9@-^z5|f
zZQi9GHoN3)@3~>ZOfprqf#>*$OV8es%TbxhWQ-7MVS<wwl9&<?VCduB0~^4oh3SW9
zw!CaYAL9TQ5PWL3^lSQE*qA#jUhP2_xCTtKovq00<c%nQ>M#>W#C*QIT$#AU&~ZqZ
zg|%xIfQB|E1w;v{1W03{4z7G|@h1{rx1}>*K))?ABht{=aOM?D#xA`*+7>)Qwd>kX
zyHRw`T#tJ_<tE;}mrHWdLHXtD=L-n3-Dz-cmaoGgS!}oVb#^bG)k9G{xfaDYxTxB%
z3Yny+hDg2MR6B8QeOmUW*TT=k=Sma`0@wTagrBy4`Q}>x-%p9n?c7fFS^;S+D)W;&
zr+mVqA+c$-J!~fYxx1#0zUh3Zpv9Et`j{S4nND9<PYXg%L5Vdh5{;J$pksT)EzQE{
zZPM8by#gpS>8^HUh`~tM4{19vLJ%r9IMK<aNb<Tc*#NU(lusEd*8GAdEFvi>X;Ody
zK*+_QU813kbeC-V{j)b7=XKljT@qO~Wb82d$Dn(2s*|&@-oRdHqM!!LOi{j-cI@cU
zXlPRj<i~`g2GqDGroWj$f13uZ1qseF1Bu$1ToE|H4<4Su*gZ6r89?=|D6wLbm6atd
zFrezuq#h43vHexlqW=jq!-1GZCM=6w!--B0cU?&B(}IHMfD@#4RU5z=>HrW8^<Iw1
zM9#y9MD9o0MG3RG21%8T@ASsD<J?~7xAq0PCN@7XTsw7le2f04u29ff`P&~wI<7pE
zb}&8A&aIOsredDqB*1#KN~gQT?!(Xc<$qU~=RQ+>LA7&-zQ%HQ$nxajAorBOAoQ>C
zP6W5+sFmmL_7{j`W{=Yd-l}kOTGcs*w1IJfG@>UA78$OcOhXZ6@GL`KiXbwL(gM*E
zz!}%_^ylODWb8S#3n&yI8ZsS_Bwym}BfKjFHqGUyy@yF87RHv=0N9Yfck3)<M8Q%G
znht%G!+-s*XD0J+ci|x8h|#x1Xvk_x0yL0leUL1${|3-f`cQt<0t+E{nnHnDGXk|C
zR?>|d{5&$|LI2H=2VBBqgD*IT!vng36JB1`ei%N6Tf`0znb>@w?PB-yn^Ai96H774
ze1s+{N@rJn>6`2yn#RmGzZI$8U#d_HR8(ZqT1QK6Rn|jC4{gZZ#S%&F=;HB#_3hP#
zWfp4tb+pV7*N>{eS2<vQX>Wos=DP!ZpnYL>d`|SzOEUMCFb!8W>ds;$gafn|6bgWk
z48oqV6vNRJd(?zNl*ILlNJ=tbBX2z3TM>(bzaP9Bqyt&QXZ9K|UFN05;NL6mkc95R
zxD8D)qwwF$18h+=80R`TI)<aMu_5Lx87KlXH0a*zObbwxoq)D(_RpUII7=F!P{lHz
zhMk!EA3QK|?JZwz!S_+s^gxibt{wVZnyl9~WuB`>7M$4nGj`+E6;MCo5DtQ64C4nS
zW7v0i^uU!G;?>kR`5QHMDx}qinon_}Z9`f|lgGd*Q>{pTmb`8q<6}qa#00<_YOL*f
z;P}QTP!83ALnYxvWL~8q8isJWh7}dbfwexY(1Tqrny7Qx<{IxR#l@&7<Nw|ySG4BH
z(VeS0xdiVH^)x(tEeX~(o$j^H{SCKYe<GK4lmG*&PJ&MNWsu(ik_2fjY5>Joz9U--
zbudh8{{80Q*1{4Kjl!Rx?xLbVdulUVs(I55f*687AxtMB1JIR$CvpG-2F*eP`fQ0H
zkN|>BmCocVEzZot-v$`!HP6Xu0-y~Pav%oCru6AqS_;SXh)gn%XY{fVK5NOM&?T6t
z)3I~`5&&<jp}0zg2&um&au8J<P89414QvA2qethe3m}*<=>*|+SmK;m>`arRHl$7@
zNQaD!L{tCZlgP+Ojg8kTeS*gW8&XmIn!kQp6&)p2<Gcr`U#kD5VE4XgO*PM4v`RO0
zTrUi@ejk^=<tY0uca607qUm%&dEP*Vyf7(8V-*b|VTJJfuwBtoNVy4X@QK-RJFCio
zg^a7b>(U!b%F1dWxC7`R$xCb2tRZNk(5-{Zy3|Jv8gjVm0odaqFaXlyOstOv6^xoA
z>zt5EHGYwSmdfHv1IHUKS?$z+4&2Ey5XoBmDQA}`&%`!u&8bQ7m7JbSyb3qUo)>7b
z;|-GU6q`Qf>?9+EJ<K>vhh;MG!|}8S-W#ZdtEXjp{k{mnH`=_o??=Vl)9KWtuGpi>
zMKo173I){Uca_nA*h~hd!9P+9@dG+!fAiX1-`!=TkfH(|x~#VH)!Dd6fs|@scu$yh
zbaaFki%VQSQGy9m>jB*$ki;N`VApdWiAFu$Uvk%yXmaD$wCs%dtMxhI`jMH`JoM3-
z5=VwqYD>a2Buy@hnEloq@qtP#*xSns>g^SHhg%>JfM_1}xC<AEs^I^l>%GIdeBb}^
zrwCagl$q?vNXRUNvPX7CWn`38q?CxP3S~w}Mj4?(k?f*~C?l1SR7gWbnZL)S_xtnt
zevjk(bG-hkSDxd(@9VnG^RZ6iCnZo?QPD)5)6V}FOaOSbQsyh|f(_(L@Nk(~SP1Q!
z$BNEIj|k0xC0t&QzzdTn63?jA8-NV4U_1!T10v!`QXe-CBZ5PqGAAPlku=i)E1CC}
z-+DVuO;=-ILgqG`NG){zfH-KPjdML&1k|Gd16_wQh$4rfBV@!31}GKUe<}liZMpif
z+dCS{tE<ODZAI9Wc~NtxC0$;kXPXJu0+(EDq?%P-aTCHHHgjP);EM!C;XJ>lzHG%~
zH0e?q8n*q5ZiJnOE(6oEL4-8N@-w=)jC}g)Z~ab;J=*L0I|ZN7LGSC+0UXgf^Czue
z!Ze6Ypc6$RfD)VjcbF3b1*f36_)a&b<hYFBSSjqDL~=x;xd5gQfBF;&y^U0`z;oWP
z`I7~vLMNxUSwbj46YXqkE3wN#$hZvBO2V=L6mC^QeFz$pTrx~KuyP7Rv%DKqIN?)B
zIu35s!_$)xNGR1mToAZSn~=mIr1Gi%s3JP2l^qegG~CWxaFs>xl1^Q@vGsjcakOyY
zYo_6E$hEN)CsT^itQ3?JyCXIrbI-NWIz<2HIrIHHl;Q!byu6I~S;wvMhCTakEkzCv
z*k@O_eiJRb8GQHt{hh!>|Jh60MvtOhr)S@@8?)DMz2LqExX?7yvNlZz7ZE6!CHxJ8
zvB-TTlj*0Yrzs)mwfEy_Npcd$)Dcm6aiPLsJcU0XHr2XrX6ez_+pD*0SCsu@Ey@tj
zJ{~bq(Us6O5GybEVmK@quzxPYI^TiR7O6`_?J9d4Jhir!Baf68XUn6}9a?RymZKJT
z3D|&*2A-AdK$#*zPPQL(<`&jWdlIHeZ+uzL8wJc=#54Q1fhmB@bKWtYr-HXDPB10K
zdf+loB6(VgaE>LnsOSQjp$>e}%lQIs!b^}M4zON8&M8IN3oQbn#IbUGCtoXqJvqBY
zBs&F12J6z`sO}!pk@5=tR^Mo}!^B*yL_<C1;r<@GuR1!TJ~uKWAKqH0)I8bw(p_Rk
zVL9~!j+m6gqWh*>cWh!;5&~2F=R<<2<!h*J$v8R?rB)Mx^A{X;Ob*tq$8JE%YCI#%
zXrO3G(x<sO3NE%TLy1eDvMtRSf({AW^7J!iuw1tPVw|z}&|5B1x1*#>ay#FeVvSV0
zm#vJX(v^L)Zy7K^qg#`^z4gL5QM|h=jsh$Q<Qe3(Zbk>5u6~jog=kN$%#viECL~tk
zBt8vC>kvxXW<ZCe5=3rPis2qsPi1P)H78U|BCDIc9mK*E3}%u`HBH~J;#vGN_rr~B
z@_`1DUAidU{`<x~A-(PAF@hzkU&>bGU_K-RB-I<WbrBy}+9sGNkYF;c-S^S?BafK8
zJUc+4R}fl02>2G;*=hFnfj=W2N_JI6p~4OOFRzPG6$9vDDQzL`_^X_I1*J0K_$s{l
z|G*A9-M6E4G;P}W0Q0=3?hrX$LTIDByq>^xgrYiBA1`h$v2k!<B9eu{TE|ga9MQNs
zX8}Fee*bMJ*eyrMUU*wJpZ?=Z34zps43mLFj8YGo@MF83plaxk?SS?A9W#T(_nq%o
zJu$ohDO?myqVv3^+e^QH$Uvdp`z5Yb3GrU{cPa0GM2~4N))VPCbP4hVVe<3IGF;PF
zuU-uyhCgWR=i(S}qIGJ=H`(9!A%X(n<@Jq?D+nTs$ihPZ?cGOazB&?fL}q3trUnI$
zO>jaVPjs|IXtJ4(j*igX0E}cgcTD7U+)x!yG9WM?++eWlXpNpgi?MBK{umdA5AI^k
z*fc97Ez4GQcN2Mh8F%m!5}Ha~y0tO&&$WF%9v3o|c2wqKBm{muHw~IjBkM@#vx24i
z84Jqp$pcHWvh45v9KQU0!c=d^9z%<=9|GI;z=~9K=wl8Pl%0Uf<Y2=XeRMMzQz6OU
z;AD52Rj0s5Ks*E3rs&Y56WtW{0F=;rIBm#yOV9&ETQN~Pe-(%oK<wQrhCh;QoU~4y
zPzJZ^^f{0;qNq0^sa8pH;aBym$1}<TnQlre_ujK+E*Z*4vR@qEvZYM5;~Y@NCH+V2
zmHjMqS(<AG+(~V@;z^Q*)`|HuPYf6Yr8b6Wo6{}>4Hd3{W+B)hGo2x=%I~Uz?Td&K
z5IIUdB7W7<JNxU6>!m?XJ^T|_qE}_fO7A}7xK_7uEw=P<u<XOA%H-ta@_S%mx*7>Z
z3K6ox^nyKbNB9V^6NvqdDyEjULPZL%%uZ|T<h#<C;R;AW53c8m1ApDX^rMof4E$Z#
zx*5O*Qc13(^OVil3abx1hv86$jQGqd$~@i0ti8r@H_m^a0tY*#jz<d@v*S;f%av<6
zSDRmo`kSf)m`vlQNk*aHn6yw~5Zr;psp^YI2)kyXzk7t}?<x21aVs{M1+>Lj@5Co2
z)BgxjidBEB=ckNAWe8lX%4%vO5H?6+%O<>Zu6{yyM~)=UtbcnK$Xuc`d|Mr0I_aI4
zn@fbxNnCVu82y0msoVAl>)x}s8@j+msIC+$yfOct?`_+~#`f-Ai5puBx1p=b7idjz
z6INfeJJR{Ao@e2Up4NrOB!we|ott|V%;1k&S{O4J%yV+gBuF#^ncsvzRwS-nwPwx4
z&dX-k&GteA=P~d=kvMSRtm%Gz62KQ3K8|~EJ<lp2I1YDZEo5<326IHmgmMVTO<k)u
zT%B3KPyfIiU>VCSpuP@8x}1_y$>q~+cQjws6D`-Kk8w%65+1oh#rJC6{jk{NlKOsV
zbgYo;%hEV9lVg?{`H@z-Yw(Oh`J>m`Hxb#9`{O^*5m;y7Eao}nR0o@OG(yBv<PN>>
zox<(9+Cf1<m{q={1+);QL(jXHcOHBis60Xq@z-at0miwG(zpqaU_h~YjgB1^TvqTv
z-DXu)9`MdPA$Mxsj7x)Y_Ydt2%L*?fq0*;H+GucO>1mm1c)8EVO}<Jx6`vXsJY3h_
zV9cOp>xr-C?lTvBu6ry<>eqdjBamZznfmz1V7+#qaWO}<BQ$SrZk#sW$hc{M5CdEF
ztzLilHOUYe6xfn;kf#k}nOPUfUKTT#;2_{5Wbn79RR|*3(P8ab`c)YM*S~#R``RE{
zTchauyLi{<;bAJM%-=$_4+(GakDo8V{iB%%x{=RUeRPxo`)c+>(3?i10byyk5j2ie
z9nHK2l-s~r>d{15#ggknP0N)3>!df4U_xI6p`MqzO$h0Fv_wC>4ZD>~LCf6vw%!1x
zuh-*+TXo-ezF6g8pnmp+{O@;9HVC!9czl$inmVHuJh!<sc^3+2^lDj<#t<_W#0S29
z0EwdY^V$}Z;MBFTS&M3bFmij{+*V@Z_UK(#P%MpTRA6R{!d~2n4ID5tp;EKVBoh1V
z9{6npmWE)B)i9A~f#o4mHWUq{k_C5||L7eiSKTTu?z&~Sg-B{34#V;<y*<;33K>Js
z$?(&{^zDc?1OWt~X9HBx76q174yka^;FTdua+=<wZ+Qr(HXd;u#N!MP4H0S)%HIu$
zOV2y}_m#@>%#4A3Zf;ilN@e|<pKfpYfrh#BhA}S~>+2|a$O6XSSiUb*ySEel7S8>o
zpD?U3gCPPe2-GBom(FqEK#@8>W?%wr0ge8mUqdIx3V1KTrJSQ}&z}c>Rl&UsN5O{i
ztoFq<f4fy=QU2AJ*+nkFnqq~hlf<cylar3#Wj}3-?a9DT%{&eGUr}BVWAQ!5hbP#G
z4>$0^c;q<3IAXyZ^^98-yyZ9-wM+9z>Q*XBv2ith`YQ9uJDWY@KL3(BA=##5*{V``
zk59zUK*zEP;u6)g8DMr0oMb_fi6&$<5cUzsZL|@#a%67$5CL@&=?noQ=WLs+JkES(
zG@C7CDh?nbg+s47upvUcL{bz;!X+9;0MOJ35+##sKx-#%Sp$Fx{{hZt0_A)fu?lhu
z3Xfo)x)Z!~21!`Nz|(q#j3$N#tPAiN;Xbi3DhGV!t3@s|rgk%fzyqCY>2&|D9V0#t
z2xJN&O2JPNbPt>{yaZuT^MT&={-I`USp|ytJxI?`+;@DeRT>%^BoqeKf@DzJ`VlO^
zZA9{ocn2`@E6jq^DbJgmpc(W1_1T=X7FZ~n&|+ERy5pjNELlN}X(3)QlfkiN!IH4g
zImqnv94q6thK1y8PL421kbw7eFT9ArCt!&_S$sM!+G|JZips?&>j0_$nN8^G?2YgK
z{2LG!73LTwO@o)u28NZC=>iuKjtDZo1x#{1TCJY*c^xN-L{&$qpgaWH_5DWIw((O)
zI{^}0zqnG?3-)Rp_RZkvCCdkEPOZgb0r6c6ehFzA`38)I1UdpI0=&*k(Ug!fG*jgg
zjAMVFcCvftSnb}4`m#{be!n!%^p=8Cr+j69(r&u?s_DfOlh5UZG}^U;pr8Lg>Mw)m
z(a@rSu)BboQ}PWK7Rl+sn_FLrhJct3`VGQBKtYEyJ8bF-wC7l&Npp)Nss|71W88vR
z@Z7&9`c{LUBW~-V4<F6}Lfd#@1MF4j%F4Eo5wm2-4~j0Y!2?o7OD^9!MzSSmwuc0F
zq{g^K@b_zJ+Ue9QY?ewnYN&9mEpJotrk6!_sdffy`{t@stFrE$s5#yyQlYS=?X87_
zPg?aynS4QpyH7fp$`%&oK1PJBSP?n;^Gwa@qyw2-23HHWH2lo?bW)w9Y-2Jg-75^b
zCma+uD(k<evtP$U!+^dl8=s!o;;{`;y{+OR;w=DCmM1bXOb^eB^p=xeM>EK9QmFb)
zAfi&b1r^+P2sieFP$EVd%HG%Io;|hqz8%2e4&=&_$U6dy0z$gk`iWhU#(#j1Cd4E|
zbr^z!(b&1!*>^u5LS{my!4r-g52e*;;EbASo&|a@%32u*O?YVV<V(lt>gfr|Pj%45
zjCY@AMKK;`a#Lz%F&On78AL=(y!eX%C*X)R3mHwM5E6;M*%s<x+1R&%hQV&&mauWR
znX>(T%{}+kcy-pB^wSn!_enk?nEk-!mh8EAwu|dG>*yHAazAP7*y)q3P)A9$<0@5o
zgAawT<l@Pa9Ej1otbBY9!z!VIM6ci$3)FZO>H*|O`K69{Rq&9s5X?Nr$ksg{@4%+@
z`~@Ek#hha^QSQ})UZ+q%|C6c6KwPAa<1A3Nhl3>t=5G~KdL@6D*g^4s5kTSN$+99K
z?GqlBTfZIxMoF>zSSr>R4SP$K-FdQMKql)Dtl3_2DEr1US6Y*|x88gdq(Gfi?$uNH
zRyJ>H@r?<AZBgi-K#VAvwk6s3aPTBBm9y4Gy0}OcwaE_z=~vhaZ2F&E0J6^4I-R)w
z)&=H)lG`d654^s2f{vY#j=O~3O&1TBc&-X7n18{gV{CqVqNy@9n3Fra@5qN)7WH>i
z+y%mH=8B%b!(S!*;^)X-;Yb~!`E+$9tN9vt`PA<noe!K}7s&t1z!^<`e$WU0@k=Qw
zHe)Z+(8yJu{OMjbf-b@>d4-Xw=}y!gv`p*FhOc>!S$G6IKbP;(>jaG`d3+se0y*n5
zQq3u_mqi|Rag*mpF-6Rq)+dB7<zJU_EG*6T?tJs6`o+($PRDPn_UVtE*@L?@T4KY?
z%Zuq3oP>A5md-@GBL>%Ryj8;EN-+m--l}`8OVWEBU-SL?QSj#DP}osOoDk@GF~#{z
zl|A#Fxr`aqzNFhH0e|3j$n5IAF@ZBP!smr*|GZ}er$ELhCB?3xp(30N(5#8g{JiIM
zbmHSLFg(!>@7|2sT#)|rho2nx935hKPwH0&&TX<+y?0rai_PkCuh}^becnr@Z4os|
zjL$w!of3!JLNwbbLU()mxr$;1<_C8iYg(T<c$Xj%gNg$AR;i}B-OI_Zc{GLw24C%V
za(vr#S2Q81=5nD$(vqCf*fSG4x=8uwj<bB68utb1-O|=_PM)>y+ALHsFv0WC=4JI=
zZ*qBb`_}t@;^EL`LwFT>j+tl9T@{BEc#t>r?UXTl(s`z%J7OE`#@CI$ec(tV{`k<j
z2cA?4nIHESU@F`yc)F5DL%m1TacxnCyI>LjnHyY+1-LXEEuN`~g-Rq$Uw~FXwzJYK
zSV|EB4}!fRGz#szOL$x`PEF1Zg<IYSmm^JT?wEUZHdNESm?~{-D@p1;5|w}6FJixl
zNKmA@Bi9$pj?V&e9FkSP7feO6-HCP!`Jj(bJQFqO@D=nyc07_gR-+RWjOH9fbpH8w
zpajR#Tu2_$6mbd`{!&Mu20=doEF@1q{N=uH*RKABVU*C)r?`>`14<Eu_n6XI*pqS1
z+Y~Z*XpNG&y_K#KB?;6lh@&LR4hj(HVfm4Adubf^HhHDt`P-7tm@a`7OzuCCXHa^z
zFu%ipzXA_7(#5Yg`X)$n04hKHw)NcWdu3(Vznbt7pszq~TmDp^cKh>`rL7j@^qlrT
zCO3SU_1_vY>&e0pYUY_xC!a0x?q@m}@hmti002wAEH529Ng*INL28M#6ocYJL95sa
z?3Um-0kQKBFgPiDzE;XS9X(149kh?l_#&HHtRy)h!}hkTeOP-3S}-)?K+z=^EKMA%
zkBrdHPEShNEg)r_1cO@H^H#~^-n>Zy`{H}*{3p2J*Vt&c04pOY*HLlAcZ+Awewtuq
za*jMdbN)JAfSNGf?&f)ZhCGF`gt9F@`(79AxNf%VS8dwm$algingy3WCd6~!|H-;=
z-!4ALVvCe`UGJccf5iST`m-!muQ$CaNbl6R)UG8Y=?!HGCY#HZ+fQn=v|rgA#;jea
zgjtK&EQbMm6VU?3R!PSN;k!P0^5nL{z*k)kUE+d5>CwD(@H7H;mjcRcihA}Vcem3+
zfka!irlVbs#D2QRXWPe5pIk5rmRLI=jHaTXqX8kfg-lo?$pp`~9YQXU9v}&#pBOS}
zT0_CU&Iorz)!!<NYC|oV(&ji@0fM2RanvMw+#Y>2g+hkH-!RGW`kN^&AyE%8QsKOM
zPnC%QKcrU(Dkc%LqpzTjM2JlUh$ZU$#fnLk6qHxUNlT?2hQzDQjZ2r!t0Ju}!6MAY
zc?;drgP8+X0}E%bRR5iu%RzbG+aWbNG!z2K?25``OB!$&p*;Hj>({O031#<i1dwRs
z6CYC#+&=V13AtV*>Si?AwZs(53(yG3>~9kv9N**Tr?3s($Wg=9-6=OS0~ydvZkid(
zn|%Qk8^zJ9skn2qz10DSH`(2#L9Y|v{S7*l^bFp296>_a%?2Y`4Y^nDeE)SMC3kU8
z&Zqs2r+yB7+}J+%ny&PB%X_<~p@~ru{RO9{<e`y`vQ5?D3f~TO1z)u)`c@=LE6pCS
zBxrMWRZJ_}0Chy8fzNL5=cYQwO}vL6QWaZYAaT5S;*f#5@xD=x#Cd$TM2QWG$5Hg@
zLHbSrH<HcxAI|}%%CZ2~x?*ItXF3r2|7E`}^7-(<X9F7|Bm6)z5^j@}e7IC@9sO{t
z5a*GNa}X!uhX|tv*b9yBroop4DneK?NnyN>^O@wUkysCu^T;P|fNYak1IbWiDr(xj
z07Qt|1I#fdk;<K#K8WZR7?&%J7NGGQf}{8`JmgUjXTR$2AI2*dhub9Ody%jXu<0cF
zmjv^oXEe98JcoM!?az<%1BJ&j=B=<KiFOYH3KGg{d)IXpBvlyH%G7pc+~E!3R-lsk
z4#<2M&nhJCYoRNXJ|(i4gyRK&I`e0kj%3oKHa{_zmooH0q`gitN^<{VnabzM0p(=Q
zq)@00k#dk5OEuJSu>l3E0HHde)$W4FN&n|<tzBl4N#Lq<G2z$sH?imgW7VIaqchxc
zHLv^(7N=-IDIjhxE-||>TP$&<Q|1hTF00NMF0PX1f8$Hb$=0)~N;@LC_;lCZ3WfFa
z=z+g8GDXDIDSY!b2<JN#s@|!pnYJEtJyo=+PqTiOHV(6vN$DA%8~^>V?MaurX0c!8
zOFm0q*-b){MTk<2j*iBKqC;&Y<Im@njfE14qT2=SK1my+qM(SdEpulk*d@`<x;5L}
zqp{c2q+klmc^jvbD^Y#J1I9<G2h4;|-*kIlQ)|W=DDKdE*2_!j7h8`&p#7-&TkRhA
zx1?DE#t*_{HF3;@sI^5i^C-|0zE2@-Y0|9<al-^~COL?Mi_=>PKnqh%1NIG!ufHL0
zm<N<+2;nG?@HsaF;l#!gfep+iRFH&FG&0IQ48cym56TQ&<`XC^5v`yDq@nuR%)zlG
z8$V|2PpVWQ2Qp96f&oD8)jhYmCREM)5H<&?K*hzPj*0<vlO!^Gyt@F&m7JuV1IbR#
zV5BCE9>u}M20igkbb31Ggk9_x%%v6G+{Td67<AiJ^=ep5ctwSOcv*5@JuL=dXc%;;
z#Fft%$))|mKCuwq<r?zzV5oNQ>1}eF-Ro$SE5iE^hTe3f+~WLwe>*jsqJWZ;65xpz
zw9}IQ9A589A|}aShr;cgLEha>H#J<JT#_GgmG9R$z!7oo93VgHC(oa^05BgO8?*E2
z9dfSld`S!KWg~9m^&T%a4i3wGBRe>BNsCI91%y<zwT*&9jA47_2t;7;XfYE`-`<`F
z=EyznhaXo5fjlLv7uDSTPlEbD&)=h3(=st(A+eG?JYIG}-3OIYxYe*Uu^y@$KNWQN
zuXyXJB!ggPw5N~z%ML)DVK!#b(^GX%96D)47%`*k2!V~lP3il9Gb4ozX+l)`CO8yn
zAj3FnY$R0RfBt&|dk0@G&wiA~`2wQu1N{nWohnn?grn2q)DLlyRWNNWbK=p~*+p?y
z^BM-f*h{J_*c5%yX*Xf_|A8vO@SZ%Q;_7S(ArX;W>rBI(E@CaQ5^(N0q)(%jKgFPd
zt!n0=$zoDXjmDSB$s&d1{1Rnl<t)g1S1+F_L{D7U>R%LisK^pSO)M8aITTng3Oz{4
z_g()`;rl)%x`O6c$O$PFw+r8TIj23uz@1sP+S;OJgO{9+RzzdG`UkJFp^_gv4R?Ab
zr;8E36n!sA%SAy=Oj1y>+9)^ehg-$0NQIa9FY=Yj8HX1;e(nADy3zds?yG0j!nMSv
z5;S!3x+Uq#e}0J-?h6>Q9WV8*_<gQ8{#vY9;bGbY0s9?n-47R|^3zCq1A*)hDvjWw
z1XQ;-d<3jsD;`4pm{T__KaCQbJub(i$YmYkkU7GW9U)#{Ano(fQOoO{8=t}A$^q(Y
z;S!01Z^EO-BN}pBYNZBqG*iO9o_(&Jb2L$volQ}QB#xjz=!B(#0C-R`MLIYw9uGs_
zCCylr*9|C1%_KPR1gEGnn3th~2an(8rXiZ$H-W<U2wVXA<M8=fR<0Le>)ECRgkij)
z=gonA3cSNNro-&5enh<1X&-2MS98h#hnV~YpP~8UCXbO`{qbEo?V$~>^|ex0FPUFq
z@DAr$K|>lN;p`qSQjMOAjKrzJkDE}ah-O<C2>^MqTHTvgg@z!6au<J2zvLlSB+^;O
z&+6q|X4cLQZ2WgYde>@(gjjc24N|V=SD8O8?tFT4HG9dyN2$LCaSy{1n77|CwoJJw
zMm_DkK!vVspXhzvnUL%Af{bid7m|`Vp@|@<c0mrHX>)1f&b26b1kt8w!L^JGW^k(d
zUTYM$I#d_2A(K=aC~c|>*)mN`Vs)aU^5rjQI(BXM+ob+wMQ*HicUG0>JnM%S0@;Gu
zSt2&$$~wlcGd>Qvzm?n-k@S`pXZlf3bD_e^Wd9-t4Um%E<^*;@DX$!Mo2~Wf(*~IP
ze240(D4MX`5n(^N0ul!aa4IB>di!7F?oL~&Zi}6QUL1`B>vv=wqv7S$@qgkWKU80_
z{<4!VbK&-@VaKgaEL!01_X58~h)^u$kTVOjq<FYpQ6asx1T6<T<#@pjY>J9FK;YJl
zFu9qKdj`73nHFLEW*l>)o(I4T{ByDw?2}tZvZzUrTVn_nb#5%u27cAMW~aAnWvd>q
zR_4FY5Pv=1F6UG|>nl4ILpbw!?hu9^*e2yR=9l0b6gD0KJ`U4s$#V~#Zk?yD@+ir0
zKJ2o9I~@r&`W6@)nSbxL)+_o$Z!6iL3}w}2!_;r$ae#`LEG(=<$s>1uH|Cr%ZPMFE
zcS>q)!3e9l@K!XxA255$(^reN6P0Bk{9wk8gqB3~8jvaj3+JL9?L3SB;~g{>e!YBE
z8>9x`8HGs5cj97;kkD`)otS9xna85p_VTEv3WV`c;IUA=gHII$qz=M0btsP*ime%;
z-(=#J*BQ6f@k_z^P#GUYt0glzd~jleKVP=q<m<_wa#svkKl5$y=Mc)aFFP`#N3b#2
zO(KzcLJ@>^0c`@*jb)F<q6!ZHhxmg51cWI6{Ck$6<<_42EC}5BoxYq~6eZb6yH6Nm
z{Rk)!gh2dDCY^69XH_&yDKmMm=fuu;YYz&OUzEESUieD5N$F_O9^f@qwmqhzlRYk%
zYW(D-dqcy6MW0rP`bJO}k<tf6*_TtXH3(-TX%QJFkx;b=%FEx)v{*r+gA8SBT>sX<
zniChdnq*<(@b3BR<O!+QBY<C|1S5hp?ARm?8-tg_fYhrN8(Ur_#x0Dbe|PL6QY#>H
z9*3@k{0ziN_)`i~!SeyxJB!DLwlH60{nno)njAWR)WY@oBlFFV-9551WVf=ByD*E`
z5>w7oL4T2Ky(U+s;h99|+Y?h;cRjDOKJ5SU<x3*f!x6s$I*CjdKu5blH@!zifex_n
z=D7f<f=2KsJiEGJzK2Ff(ca3C99xp`1h{TOF%I{jIc|L%gxPgY6YLuKOLKKgVGyyQ
zT)cxOh7#b;k%<Npu|DDU#m)GhaGi-Bu~_D&B_HEvSlQar7hCW7`Uj%KmS8_NMT&m^
z&3zHvPREP)ir&AhiV(uSjxt^a$(<j@?)pj|szbldvl7|MO&i}_N<T&wDj4Q)<JJrl
z6u_ubtn%U}Fx=enLAk&htAiu}4V(u_e?bqw7q~J3eTki?R0Ro(+x~u7gK*7S^1-8!
zkwS1sEcA6MM@0lqec4Q@M+el2Lg3l8?dy^EL{PYzVET~deMYhcpZz#GZelreJA)=|
z6`&1^ox)zdM(%6k?=Dd&<wr~E++IYb$yeoTieIZ`6?|{;@ks8da}uU9hq>bn_%nOl
zp-p8q{<qDNmhZiS!O@bMhYyiUK;+T&uJTsOAAvgQk#Yo{<*}ax?g!3^W;DMj-wiqt
z0&7D^N=7O`z}KekBe%}?@01%+t<^ra%DB@*67L{67&vBx)Ch2?l(4g9B?Uh4=R*lo
z|9G^6eF&HgLe3#-yY2Q2PWM>UHG$WX$dnZwog|mA+m{1j!J?wzM3tVoFqg!tKcVL&
zR34%2VQJ!P)i?Q0SEoHXA%^$$g8*8}eCAz#)g5;pVb>DF9U|xS_+;6oRY(1PtfyJK
z=Xv0Bh*so$-YOHP3!bP#xlh2qm=MwgK|$Ty5d<bE0`W$02f>Z8191`6a*0FL3JNsf
zI9jF9Q2=T@)OHeG6=5w9<-owe;FUgzoJ~M~@K91H5V=wlb)$W;9_nzcDPrbE5-F>D
zS=F<u9_~{!-un+7ih&@(E|xn+5402+D@G<-4~>mQLtp_C9Km=DeI-ebs?&gUAae*s
zjRYim2yO9|E9-!}KYvlbq0ks?3d-kK)d8iWwlGM8!*H#l4fyhkiUQBpB@yA}6`LfU
zqDtv`hV5Ps*2MxBR`ZqYWhPXV6&b=JCfeP<&Q%_N)IMMIWRub3^<zuAbfp3oJB7S_
zOU6wE%ZAeOR6cu2wNL%Tw<9G7KF%E&9+(8wE1!)Juotv)Bn}pv6N_j<K?D+72JWq4
z^5ozYLbgE&hy{Wtk#ZD&8;cnxdx2O`zJBS+NKe0rq*z+7iU)fA7537QGE*Ui7cMK}
zG{&a~soIQYwG*SVi2Wa8Fp{}MN-UB_Kq*wbclUL`^VyeUegrSUj(#4?cr8F-+uGAA
zEbF&jgc@xG4Kyu<kYC_fb>Iyl)PxSwOtFL7-rh^)r^>IuV;tkOI(zOUqQ&K)$su?W
zIj=BKA(EsDKx;zCb&|P)kN}i8hWv=8!O!9&QoKmZr(Kjwi@}3d?#GmbEh9mgeue5^
z_3)qEk!-{7bWuMr$A%U0A7$FQVU6qO_uE;eMKYg$#oY!G`MyoIZh1@!`qDNRG9d}%
z1R3x~<}r+RWWbpp`0F!O)qKZxJEt>$-OvN<hj5zAoP6qfVzQ#|wtG7dfx_2_8Nmj>
zDtVJ#D2O}~8PU^GgrW?=ZsetK8EFZ%S>+ImLtq0GiO#2gtOO$D!o3}N6y=8p1<2GV
zfV!k%D8AEkIUym8bf6$QC?Uk-jbsfhLtQ}EDA!4-Iv^rpoPX%nX(%MR68A&`m{rk2
zAgfgm1jJqU4?9CTK}{iZXRrkiV}Um#l85*efd=eKkN<ZO&l_H3XX_D4m1s&h+r-L0
zC7&oSg^EB?|DJ)G-}iB<aHEa>9NDzu;=-#p^u~=GSvu=+eHvRcWiNsXDIr*=cCia6
zkFXGAkydfYQ{rtg<T23nA_|@a_n_~@K1CJQu(v<hl$B<F!g(oF;yf8(v^aggP7AFu
zM6Jcv&B5~{>pl8P_>s0iIDZ_aWFk-UGwc=G5ZU6j9Y3lA8X;kqfi>^$KHTr$Y)vv*
zkez^IR^oi}Av%K;?g4NrXfdSQ8!XVw=mNK`@MgC{tPdwt1c-3*QJKcvdTLl{$W0-}
zQxtXjSN`S0eQhR>6#s4OR|foGmeMwQr}Dq;uC-{nP2xFF-+YJRlpK<j5L`n|4&8?f
z({GY81F-S7Z9|m98Dgg*CWxVq+S7U1&e)LR3(e|2kit;gA4wI{wYt*oP9GYAZd!fj
z?J>gOlb8}@KL839#z0EaObF6*o(ZD(LdWP8Q(S`2*wB0L%k#CczQ|Fs09RlS$DX(Y
z^^TPKV9U*Y3Kt7JA=z#O=F`TX(l!d|LjLza!R3S26FBVhhVDNO5aBGQAc7pboS1kH
z;-L4?<B)w)iZ2*x9p#7KZ~;XmGZum2Uci%;c$M1*(y5}xrBae&KZNJ=56qu;@X(XF
zKFJY!`ua~bo^w7{Tgr`p>e|Dz$I;q))JDuqa%pRVs+t-vsFVoKt}(x!5`9Zf>|!XB
zlfS_bf*&uwO3cUv2@$B5K&3l-+QTx|V&{fJoBmKPlaCQ~x&F9hY~0+@Pz`7m^)z}Y
z*e44VKf%psM4g4DF#^P^f^R#JjjB87bC%;X(M6NE%M;j1h}_>pA<h2jR|F{H>LO;M
zu(1WN$w(m?9=)$#f5hQn&T#~(m`CF<HV%@@>%HFWQ~}Ae{KyVL4MXOGAaPIsDDIfj
z#^o@t!n4fU<g<65|F!bf)p9o4-KzrWYAAU|qNuV0k60h)O>sR*zg1bN-c{T`eJ&0A
zB^lp@GK;`{h|4t0b1~#ShzBLv1)61<zIlw+MQxz~pnq}UTkr@LE{RYlnbIWcry_pO
z{{=hmcIV2$SU(&D*SkEvKU!mI>3#4}+1<O^uC5});DaYV>15D$)HbfJuJs)y!@<|g
zMikHnL#e}8Y+s=bFuDs}E0S=NdHVo)G8|~K#$?P8lbbtGC1BFZk%?aE+}zwDd{i=Q
z48ey{;1%c=EL!ERUVK$uDa*m&$rFcLshYv!Rbd}LOP;Ona`KQ_GMe1v4{5Ls-P;$r
zFXvVUH=~ZFqV-YUCps!qes;>qvr+;-CZ%wpR0B+s!wm1L)4SKVzEwVxyB<#*u~VW3
zdv|PFkU|71$WS6<<-Wi6E^WZDdz5eJwJ_4K;^G_`Ndglss^IG&g2>rG;_FG!3q&sf
zC~b;eoy$Gg$bc=_(1!s802t?$-?D{7*rg*<8778^$jF@)Exq@kcF1<FJ429C;=jSH
z5E6{)yfnKO+tBsCiHTg9F4T-({d<mKwV~TN6+QxO#t_sclb^03=Lq%`t6>+&rmi3;
z2BM2MJX_aG3Dei-Ob+ZlB-%b$_G6DspNIZ_JT*eug&uqXYoPp*3)7ElOBK)Bnp8iy
zkxyxF*roh!(V<a0PtWb2dt8cWIspNm;6Ez8i&H^5%#TlQrd#V)#rmwiiFWOYG^uyT
zOboImlJ|_8LV1QNmz=i1IP87qotM;bev%T1SLxs~ZfsFNDy}O=V&sg<mUvGJn&@aP
zUtc+rzT0*Tn=BJH=j8<hA}!xC>A9p#vCV_Qy>1x&Lo^-4vV&-ujql&Tr%-SdlPNFg
zB&~Dq1OyM{*?m1TDXD60zKCce0uDmN@xOp+^>Jw;!G~Pv?av93!UEJDB>Oefg+@yg
zq7|=S^APCifGw`#K8HM4__j$HC($<GrHD05vCy2&ob~&QOM71xB+aaMplfdV8s^aP
z2GMaqq_+NnS|6c?NFs_v^u5UMpFe-z!Ct94C-Qp%+BZU`5(p-jB=gt){m6^TNZ)?2
zr^hOO7BLi<&v+^K_#yp{DoRZOO^1?z@9pEH<0#8DS9!nT`B&xm2Wwg^0vaCouDRFx
zu_`h1-mXZ;J=2X<;RzzYFI-@_GxsUR*RP=8c!<u*Hn(Z(RhFi9tF-M)Ld7f-_a?)N
za)L6{>x_BD^MxJuDWzVXdt*GGXmQ-RYX2equoKQ2XpF;KK<b<aOP8#+4d6Riho-4h
z*s4gp5gDIG)Lg*mZQW_59UpYdo7{YvbgBeKBZfGeTs*TlUafd|F(3=|uHg3V>j^k|
z^X7rP?RD1v2nyIlWB_$*X{f{!3ZAV?HM0I{TY%bQAM^%pu`zrq?9jFm!1TIB0phDx
zQ%D|!=o{YHv|0#+`%W^$#)to4S*onA&cPApCJ+}BLy{4^D*hcP-HBNABwq?e#ihhV
z5&a-+%-H1`-qrf}<PIQgY!%>fNjvh_fd3U`e=+_y7r@}2_(iAE78GDvUAwkS7?<|h
z{1tQBk#xoXz_nb>PxR3Vtcr)JI)<SESXpo~S~+owF_ZqpbZuE^nov-*z}m%1riy0k
z>{(R@^<(V!ygc98_=(c;Y4|fA_woj^_L*7vxC|T5)z_ti4^DZXTj=Lgzx$Qm!t_Qm
zZ_d)OxRM)W76$M!ix!C;IbD`jROT6XD)c#Y&z+k;EcTlW#ryTkPyOe!Ex$9v(<0%J
zS;rJ{D%QNf@^(dW^^BMnYvJ4P;$%S!DbcSANk~|;`1g4*jCn@!vt+asSQ*=4vFYFB
z(Esx}Cl0cXP5KMiS6q+?fbFV6<afa2>u`id*xofw{PyY&HNtJ<yGxk+b;l|l(^>M2
zR<55DBsJKn)u(2(dmk&fa@}};x_RHZdr|SK7Bw%G^EVIIq=5N|6ZsvdA<r^$qa$>g
z`jFjrhnjJ58%Hs9Zfa`Xmf;oKP9I}hQR5xpwnF&XE2+9@^)dl|Bbjx|t`c&&^OcT~
zCQ9QoR)KOv(FbIfIH0y~-;NQF`vCk9a>`QDDmBkG+G{r&9JG+WgGYd;;QVYj^{pln
z&Wyla{}MBkHr8EU3011y{^^YVFlqiK!8LzV8HEZ8KJ-RSjWMSRMq0Rw0@AS;4hv~v
z@`;kKY;4I*R%KB%BbYm~hVAbiT!lv;j3U$eZG0I12zvD1rmY5+Z?Ls7Kmv)932S>8
zD(nU@H9v4!wOMH-k_^baYH)SRuw$Yuiv&kv$PWw*D45;Iz(LGlk=#6f>GYgBMMl=3
z`%<O1eRl9?+x!nNHf`3ov`M_w&~#APpUJC!+n*uohiC(dx)ZGn6Y5Op$_U{rBlA`~
zbvzc$CMvAkNPzhTH{vswx`k<8l-tB~gA<3$FoQF#3Ho_LAnNGp86=X4YEk(oT?mLD
z!C-iFU$I$}Ucp(tQ3-!Op;<;H_@=6mq!WgL5D_Q^z;ozf$^K}SDoyleS)qwGV|ng3
zR65Jku+W<lH<t&-TA<bX52Afpr}B{mn>n)zlt+U@n?HYfV>&K@^Z1b~gZ-6#$*Qz$
zT|toOlTom!Ig(<>NWdX2HM~yr6j<)iReL8*OE25>vL6s|yo;KZ=Ij>qSNYbCoIj6f
zX2!0$dyv4o11W%_VDMbl!l-^kVm*bDVRAsl-Nl7Er_1)!U-E6E^7sD>PDp#`egF9g
zzyg6EF%g3F1d#4lyh9Op>kpy4O6A0#g$m<wX29Lfez@j07f#WY=snw9z*#kwd%n!F
zDCh+2LD7Wl6US{Q_<{nVH2@JsA>oG`Yz3@O`r}uP#4w+D8Cw3Q?ntoS-~Iki%z(sZ
zui^C3Xu*`o+*lpU>#lfbSGy3(Ic`2SkGs`N7~089*mE)^h|sq1NRbW$Bq&Ng4?8_S
zV76vOJ(LiHD6r#Mj8mVAO1y&tFn<MfSddWS$icCK&o=W?tWf+v?|mcD?s`suEMMic
zu*tw%x~&zfzZ5&<#_C$?owp4c309(evyQLxJk$qXI+zeh{2Ulb#zCUq$VQKLuIcI1
zRT_c6h<p-m(f!;Qdg)DyKzOqg<rGmWTj_A<(ua~rL;dkjA3pewuKy2!uZRj&#G#+E
zwL&)3;q?x&sZ#Uo=K8i;&K2peg*MKu<}-SCKC@VFYBrB{<biFJ3qH&aW)_b_cT+%T
zm{$~7N)jL03FNF0<&Q4#`vTuDyG=E$u66yptt+y-sZh5h{0emi46n7ncVkqj5jty2
z=1KMH9mRf{B}!{<zGT_OANl0{4O<#|&8P+et%EG==E6wni@=7s4#Q0@t1{qA0L7#x
zvZbC&SNs_W1=BIg62-$^m0vGh8?-^Q{_!&xK$K$sB0>~Rx-doorAon99|Q^hOJ@xt
z8z+g7tnBhrjv~NswZonwHfVxd(9Mv27OkrLbzA2U-jmU+f{lX(#YEvln||;u+s9t+
z;M}FP2YmQLuQr>4Nnzbk-{vai&qN9{y%CVIByJI8g2?Zg$v2U(U>N=nnEu3}_B1!l
zN&wSggKBVz)1U=-jp4@7B>_;jcUM$VQ!`9Y&iB-BU8HuuGJI%|p_k#^6H%{Q+Iw`g
zSZT_S`iSEzke@=TaBV6nDJcZ87+4G<v<D*v;$P#tRD^>r5E2)sBMDEW3C1isN+l%f
z0ztpb2`GYUK^rpg+i?BQqgQAlm`|4nvA8VOYg;ldLyxA^fR=<rOm^t3#{Rf`V9+Y=
zwhQH&b{GBNFm{cs(Pn5)kqbq@EdBAd92}begX(|p&=LLtgS0UA3!s{31?h>D6Axo3
zRF=frgWaDfI6?R9b#V#XDR7bo+f);(h`d;nZV}*{T6bzGUU}5^`9Zm{7z7hz_Q34o
zQtQhfSLeAd${F%UM&+A$DDa@vrbX$Q_<uYQF*;~XHp5>*e3yDTYy|9ph=ih9+9wP=
z4q9o_T9a{5h)O50HICc&A9RevI<mqo+`~~cw8_`vDq8s)N;H3)+9S6rB+PCV_a7&1
z9peOmheSiy*18q{0^6h9k?1Zx9Sl+kU+Wa&R7(7Yg69KftATLSP~Zn@lJ9-<W>5S@
z9pj|B-Rn;ChpG&Q<wwh>F;!Y!TCrL!xFE-DP2vWQkxS~-VpQL^da?TW#gmnkAtfOZ
z(tj`l<l29}npMaadfp9)fdvlcG`>!!J=Q33hs@M5P95WzF4spLZG(H4de&dEyCMnQ
z!G3ufX_nS4d)c3Pcv=d-{Vah+kR7-gz#)W%R`_l5W4Y?x7@<?$jm27ruK(DHl}DEl
ztaUBR-0X>^MN77r;O@@%m*%;QLb=cADGFUQVXZ8eU{j<bW0b9}do^0#)&}t+M(Ym@
zj?ygz>&~}49k=Wb^rBxTh_c;Fi%%WMT`Mwj{D^97-d8>G!JF7ejWv_A!p-HH`aCY|
zB8S(ilVlnb*f7bMg$e6*;HBQOz=Dnk>?|6Zs3-wjFy%OGU}?{dwGbvkI-Jnt7K8QZ
zef<-Ta89z_=bPv3_4eN5{oJs^l*9GJqx-V=oacG^j&?LOb*!&tX~1{VlX>-lqE^HH
z)!m!Cvp%-s+Y0nKIXM7|9DSZ4p8|P~VX#wBLL|<Uw1d9%f7eQdaI2D+Z?-po+gO)X
zVR-6~O2jVyn+^S71_%R?u6$$#S`P}v)YO!qlEhwJU0wZXRnb>y>q)NzNr+PsGyKT5
zry)rOt9t^6u=nK20v`Yt9-sP|ckT`t{`}j~_s)-VF&4u9o4h`#KT`s=C{*ZAD9$BF
zicU_s&V!RDh7z1QSzYtrf{vkDAhaOScdSYv!^GIk-E<15Sur689LQzI^w^7y90fTF
z63Eyjl^CRR7r<MQTziCU(UKe$6etKd{9AkKg@PiI+gtGAW6-+Gw;-e=2A%J9(9rCn
zIpfd=gxqu#{3XWEk~hDitbY?8D|pNcfYK=k%_@){4Sn<m*Z~kmnGCfk`LSXnn>42!
zp83EzCW>I^M(y_s%x%QJ03qOY0E%R63?>DIfv?yLq>4OG1Xe<wQsB~CL<A@(B8i;v
z*gR5OA0aLWaJ;a&d|`C#yPCa4<&kkaP;0blTlBo`<u&(E(B}_}%6CF)wNZvDC$^O^
z%#Q;2dKTuloVU-H@25`pci(7pp=<_`PY#;V=5*rTTa{l|>f?rzsJ!o`!3&IVG!Q@Z
zhpDN<!YFw-Grhr!AiM-ff5V$MbF~#9O9=A?n0rTo<u<dg<Gm%a$U#EJ!A^YMeTNPO
z4gRS3n`Sm@dDQzCvjDUCC4K&Ip6EA$!qXqTuCvm-QYp&A)@J#Q1h52rb237ND!I|X
z+fg89S#SMXx2D~F=uy?C{6icsy6N`H|7a2?pWbrovaUgM|MW(RX4bLXNBTU;c&u=A
zqXwyL{J2-w=WXR#V7)_7nLut%O~NIG#KdS(ASJs+^rGw}Q#B!&peATOX#Bl$>#{5*
ze?Un69-$dD6qKDpn>VvzCBjd%Y4-#WG7^^)0=`rjDXgs(Qd7^iJSXYm@Fuv$E+Jlz
zjHp0D##w;Y+34LMo27(ck~Tf0L{*<Mo5)l%qEmt!fpC6o0s=lKW0&9@BPfEGmm~!c
zC?>+LBn<?rs7)h<vuS3NF2^nvO1A9P)-K@++|sGNA$#cf_3q12Q6kszX^~BtgLfiu
z3BZ4lTjKt===Lk5{+Q~fz7+jpQ?08KHHN?fQGxuEoXt04fVpK?f8aPESS`WKV7_vF
zB$^`y3FNSZfPsWV*D3D@_&4l4n#e!{%pQqOfw1O8Cry+9KR+6i>`&xGI<3a|@C`d6
zi1&qjX52Ic{)XLu7s$AbpJ!*2q-4aw?#KFs6ZTe~HC_@4<`<!lf@dJhzCtSd<=Cc}
zx;puFOt@NL0<hmaM#*{`GeC$Cnb=zJ<ieP~nnI>UIhD$m#W|wIA%;vWS)5%YKN)CJ
z6K49NyHO=kj|8#p_14yT@kH^;ZSkAyMBVFmZ`^S}jnhc%>mGfb_BWtduqw~OsLPMJ
zBo-|v1?9~j)rODkJK!R!AJkr72T!lwavh=5{fnS+ekqS`sg=>~QH{`+F(^wn+mQ&{
z8xpzE(-gISB04Cb$b|SJf!J+7ApaxcA0h?>r|azgTpJ}e2`J8%*pCyDv>lMDhC&K0
zgyWdV`M^dZQ_lH@7Q!>^i98*^lVXXZBS0eJU?LqAxgLmg(Ze<`#UG3TfP_tjTAtX2
z-TI5!g!OEFavx*F+~Cq;K<dx0GihcEpxRki`V1Nc%vY#<{&C=#*@w&|9UV#t^3M@9
zZuFm1&~G)WZZc^Gad%Pd(pMpX`O&#!wZ3&7@*A)R#P`A|lu~{M0>j+jyLay%{`O5)
zf8ji;=Tt%HJ<fHT^CY6gKx`5pU*YKvc!5a58XoZ!l@jaav_%(;$^xfxI8|%_%}(3t
z?Z@G!qL8hG5-_byNF74%5U>qjm=c1h2nG`q6HIRMi1F^LKn=GlutorK&e7=)+^|%`
z!V(6^AxZ7qnbK7~HKaW<nhpNpI^=7+@nBx;G1Jts{&i~>XlcyT+^wQZXi8bW1U?r%
zJ#nv=YuL0S!Zy;mF7lexyE#9;FByu;zXyM=oBL`b_c?|VvcFhhGO})Ce_v@&l0xN-
z)pC^yjQRw&8(=qtB!XHXzWXq^m0rY1WeIvFAB7^K9xJ8o%be9Uq)tIsC<bGzfkhqm
z?LXKTM5`~qf^Dzu=y_({Q>Ef;Ul)}*<nzc?!~PzX9fhT3<9QmG0o(#8P(m0EC#HOT
z8h~>^z6IxU+k7p)EI1(S|9AOLK`KGq`$UgHx;+9a1KuNJPXX!{9+`Y_9M1+t6Ui{G
z8e*8lh3Fa3!JL4YC6ijGY^|Ku6siGow(+^9rKK-(rmmLXae+>se=IZIk$<!ISoL@2
z@A0{d7fg*Co{2g)wv{fXeG69Hx0mjqy`sR`Gt%R?nx(h+HZv(VUeu3UInut=QKrt-
z>(8YmlX*5d)8A&v>9vMhRYOhb@;Qu?xT%SDgO*rA0|uMQ$jIE1H4g%d{B-<6w#3Px
z0(aSWj90@(?J!3X@k%(TcKkCDeN(B;&Ha!Wm$F#Z=6YAYrTA9s*B?R&Pg{fHz}T)s
zIz?1mTnm}<MV=MPrnU>c<<kQgul&sYSgM$d?i{<dm1jw7?I&IHbAt4vX=zp`jq|?(
z_b`EmY4W`ns_Vyj&(?TMzAzcLO=8!fIzG*#tlV#C;mqOI`AV(+U6)<_ti1WTgfMA`
z^d7*aJ7g#h1{Zy;qb?2j*PPF$->IhP!1rbR)?!Umc75G?QmWzSbM1*4$QlX{B0U&`
zy)jIrIjAKNqcji#kB+18`w!3rD>P%=Wl1y_h)psxi&(RvUF89Wr3|voyZ9c4j`^uN
zR(--gwX@=9vSdcVX{cwxthBq6JPkBYW|I9`J|4F?z0|)9tfBahe{_OEP{c&X!uxt+
zlK1kgWKwu2a*=-8Jn3<BGl@c9TAZmPbmzL5=A$k!?`C0^^m!oy_+w*vk8)(=7QO#V
zh3}O(nwR3Vf<i>x;DySb3DB5HRV>WUQz$5JWR9JNN^-RMb^k!>M!z~yLjZQnTC7Bu
z55Y^y$%7bSMo7yLD(bI)`9$~+&-H!}U<&Ix_~`nFV!5=|;3c3C|9%u{m$8mQBh1!@
za&{e(gNTolsGUvKf6ySE<Wf>nI1;)>pmTTxE}RWJ4CqplyakIY2ML_{??Hu8e&F%%
zocQsc=UY+_&<--H5JwOR>>{L|`eIfEQr@CLIxdgn1kBZi38KKs_e`_R7yZBW;E<?c
zl_8FbZ`<6LW!Eo61fl4K2vdYfvM9ral$7r2mD2UZ9nT!NY<_}Y<u%j%<>M6S@$zDI
z{BL!7Xy19twDxgLiU?Yc<sUC2V%Vau&wDoVD!^^c-Sa*M{CN!%?`qTtih)#Rc0|>+
zz&F0kq7yUV1Ai&R%`pW81W+{pBbg?RBWk9{h=C&fgsYp|K6sn1sGqz$deU%apvdFf
ztrs<azgnwaATuC0-8Xgi_e$xV_nw3qMi@rkWjP#qFDMDIsARV)!DotiU?CE53pn~c
zx~-C5Ct);;MFn$>+)yHZz`;){)pZD01N^cgM(?L4_G0O9qiB=bE>=N7MOhGf?K8Q#
z5v5N&8Z-H%*^ib#={?r`FM<M(CVx5(R+^!sz4{*}e1z+PE|~MIgkjS#XOp5LCw%C>
z$Bu=8S*lw23Bd?aAQDqBE(S4B_<npxN;pWy7$yd4_zn~R`QHzHb<)RMUX#%9F&y}c
zn4(o=5+70Pz)ONu$`zGm&O-yW^HYYT=Y?(ndbgx)rQpw%4)MGTnsF2n6}^r*EZE^W
zzpB|L^qm{eEQ?VS@cygda41M{EugmRsur#a`vPLQ!o{|ve-KiTKT{IByz<kw0l~=W
ziUM=C%L3<SWux?=sQhVCJhZ<n4&}giY_j)ZTwP5;xuptU6Jl~6d{d6m3q#S;y(9Z|
zaj``K5q##|!RCemDWxxLwOCo#t+UNB645YKTkKbKdmP-sw@D#&SYmBg&pQT&%Ug~w
z^Qie_TT+ccV4(jf7y11=FerTeGqLQ~EkGa@Pet3^q)--3hFMp3ux=3Y>Np=G_NAjI
zd+7Pcy<1O7N%Zsm!z<m%m<m)USm^QHvhc&ec_Uh2-2L{u$%!>(2=avL#2Dp5Oa#1d
z6ICf2G&MtxRRRx++NEuuz<ugyg@jz%aQ#QY|E}kK2U(RY=vesbL35M2U+)1lCy$Wf
z9Vj$Jjr6)J9i9?&XK3GA3U9bQ;IHmo+{UyHPm~!F`bGRX0x6Hh)|88F`Qo<XYS_Pi
zMW;KLs2+(H6KhW!&0rU-I0Un<TUxB{z7xg653N-0Ewzb)`$tP&3m}(_Q^j|3XZpjc
zwFh3VM&}_AlH(8|&UdS8pn*pc&F}J^i^^X$B5*E-=rl=m5((0-&+t!8f%*hFNxRZ7
z<{#<ezHr@IWko13!D|EqV`Drr83l=@a=--ZTg5ZJcYZ_4+;b@`{o1vEiMA-T;{r5!
zt`qMFO1tF2YruOlGas_)TRMWo^<!LjahsN`Cad!9b5UL;;jei&l&o%I+Co~&nmanc
z)BgLD^qvW_SDfwZFi!#LvoNoa?i_f}F8&MoU2rQx5o*4svv<Kfms3Re<{z%u!B;U7
z*S1;g%swTtIrWI);dA%TZ)f?>Mr;1FNlhr`5>RqN4r0^SF0d3*P_SHn$M#H#C*Pw?
zoxfC;)7l)q79QEWejAGx5Cew9E6;X4>3M&z;hYIi{qytN*~!Z(iV;fyYFKzf#q}2m
zaV8P9ePRup`ABp1h8SB{*W7M4awf`txs}d3*3a2u5&ZaacZRRksU`y-H21TNHBT?r
z06smWOI5;9QfZ1W-Tf-xQWC@kYAr*HLF7mp<XQYLp!X5!)D6%@VWH74w;!vq=y1;=
z-7@5l16b1UJ<1n-F?8Y}i;`wud&3tfkD)qk&?8C|tSuLq6Bv<T@o;HzF1{O~rr`#8
z@wrO)KZo`~pa)e36X;l43^)a>i2w2|R%sy+5=!Sw${Wm^?k|nI=YQz+nCRv1WvCzz
z!m9BveWv5|ers#`F9ucFyK)=dW?SiZ;Mji$B&KavqUYrXu=(nh<u+ce0!wX<lVun3
z<ySu7d;7%joYAEXFLExBRg^EVZt)<CJCD(A7=Jw4{&yX@U>$8944TQnwd5&(^kp}N
zvOaRJgqt<{O=R26?v<lfP2lbgmXRuS|KE2dNKJk}fg7L!ra<z$@d%HQmyf|k`!jM{
zLe|+KSOfdyuKs87<0x8DuO_0?ebC`~%|-ScE*9$&-+M<=R9V;*g`cXd{r<8l81zGY
zw|Sbhl=<HOd7ICs4-K_-+V#e!vd?y~v5RlW{&>UInGnE{yYKDrmWam-{C7JzhC0N&
z>{3y?j{82z583RsEE~XEWV?$!$<qM7nYD#$#JBZ0wnk6+swktCHeLTHK_Q_BYsq)9
zlh{1I_weY7?JcYlt_DP7y#>Ow?9-W{7B?E(=cOvsPg;!cGLG***mz}irtbM9P7}^V
z0O-iQy(~zNz0sm&S-|6^@aKn@kQbN3p_)$d&nwp%(}IpM-{O{ImJ?=UI-fkh#^jxm
zCL`E}Tj{y6J+4>U#Xj&(3x2u0_sKK&KlVRb9E7qz<_2TX)7Pp`zDN{S|JAvJOm!xc
zcrb(p`5jt0UBG~~WJ+h}!n#}=v+SI;WV!_Iq2dPk1R~^r8u0)V&}x^X=e^^iaz{p3
z&t3g%<NJHsj?ZY3Tf;u)oKd1ae^FQUOFkpbXcg}8laD=Qw=`W}dH4oA6XL}B1|9kB
znWg^fpPwb#<`zc!o{1!69261RXG*`}+&Nakrdsh(>qmtlq5FG!W$YhUWOLbc_8Rou
z>rJ*V2(6HuMKgXW<F=$4{;8MtrAt8>qcf4)cfX{eH=J^fpP7LFQ2p5RpH*VwWv<(f
z4bIH|2G^Yc#`*?!4M~s&pf)6>rLByOV=_-q>9@nOWB%zhu1Ytga(0`V)osAV?trE<
zqMMdN(VS>eL(nayJ>R+8BSU7+K(WKisyp$?pwyGkyL{S5j}|_<<J42KM*eBxR`&&~
z!_RB_KigvAK|tz(BZ&uj6r3u95Z#^8?KoBdDdt{}AiP|;j{^O51c20#(PfN`TZlU0
zI>`yII6V075)mfM0^2WzA8yAuN&s&NUl>+xc2BK}k7uue-k5=fMJe};r)L4cCZ+;r
zir)8fTg8WOME;(b-TE{$e^hjIwr_CBRBOXz$)76ToLjlJ>t#+jPKKrppKKS&yU1vM
z)7H7iQl(;aP(}6j#FdW=Zb{mq^J%qbs0~`Bezi7=GO)<G)4F(`6n$QVk4vEpg?g}R
z$A9os9raT!I7~@t!vsZY%{?y@Ppy6+Qk5>FlF_VQ``|<T%hC9kJ-?nyxhH(^R4lmh
zx&UofU;pl&qxwsAwMU-I%7jf&FTbDzJ@X&E{b&BG1<;m|lQSa{avT**1$8iMw5O}=
zhVwh3I~b#0f7es5hH$&%((DrA9)1mlt(H_$>b*2eKrw;9&l`Xqtg~1G{B`hdkYXXf
zm3?ZJnPcet?wd`lsYW-%3aBa4(as^IH{Pn~5A0XtwFvHi-~M21`dZPknU5!(j$zzb
z<$3=m>A{Y3&UGK8HBJPi?axtDFkD#rJ)^tFf9cl*$IFe4mu4B0Z@Je=E`NSsEd_e2
za(nR(0ZTjo%N>cIki!sS-RgSxNt5d{`Tpg9L;2cz(2qj#khD86y(@mdc07~G3d*NX
zzOR4leho~rE(_rM_s`M<n6)XzFUIbcS~~day%VS%U+3{`s({jIJ7;pJclf1R(vhQ{
zWnTt9#X88ONb=E*9_^2bU9tSTL(?9tD?E~FuWG!~$c=j@cA>$OGBp3b$zpV5taQt0
zmFktc_Rq@F7OBEJ++vwH-)!Ha%CP*4A#Yq(<4sMHYP6lX+zjHgcd#>?P*c{r|Nfp_
z{PRT6@h{Z;$`w{@RF0hEyq<oyxN^EPuN!>YpwIQ!_I*q&HN`_1{)7@}?VcP|`dHQj
z@YKN6bnC5cl&nK#Z({6f9_K&M+@ll5s#%l}bi6KICQ^v{LEH1!9iNpcof5-~8Juf-
zog+=Zb@wJ1CJCrHW^xsE-C?rv8oXifNlV|o;^#W)qN<#(%M^-q;($*dI0Jb3#E=JM
z1`5N{$te?0z9mKtRu7bjXP(eTck0vq7nqgoH;QShkul#qpX7OOKX|OdBO`g&P$->Z
zoSd9bhkZZad*g#^5F(~WKHA8>>BP9mYx}hmF5-f&%`=U>u4;AEB0TA4#oR=N!Q|<r
zuFFhT?t}Q|R15c2UOoN1ixR@+h2HJKX!4$JXy}lLYl`m91|<cDfItbf$ry$VS*D?Y
zKuj?2BiH1T2V2cE-!&h!``++na_aYQ#atyy)+7THQzGO_kX14lIGdSe;X1Cf;Pvn(
zchf)c+_aV7u52Uu&pEXe-{?*57ij%cvyxI7{pY_>`L8xHyAu57;Q6HffWpQq(|CO<
zfgD%G03}s+O4iCbBsnTT{zY6ldON8o^^<u1FU*>?Cw&+m-jHDaa^|d3MEIA1SCIEi
z%%A?;xgiMM8EqbwJy+S&XBSpcd<rrG^5b*#r1D0tD81pEPu@MKNLTahbRY-C_l!)+
zg+ACCDD}EoTA!PV78mm4ow1*{Fa3?v^2FaKv9C$X7q}i26T44hd)!%usVfq)$82(P
z{ypdAstWW2-6+tLu3jxd*|m;;C54iuf8t75OzgjZ(J9#_=(igLTp~tafRb?j^antt
z|A2qr7|QWx8fvEZN2@OW=O0rjA$#ySBDP^D;KI*qOWgp&%lAKK(3tyDR?!X?f}=e{
z%?2<a?M<3A8=t3}rPbzFXKYeGJ={PUifApoFzWj{CbmmNl0S0PH;J47-^P^Hj@NW2
z(n}2-k+X2NbFNvFgKE(W;r#XwPuvmNsH|)YCx3EEN=k$N*^mCeo(XW;WSHw37j&6Z
zhR)ADP&Sl`eON4ew2zm6^n8);7n=f?3+KI(5o5ZJpKke!W*L8%Q>DmsAjVDsI`i;|
z2!r@^lbeKx^9RX?7(YZ)Zc*OQIggK-XmixTJwbOM^b{|Cy+nEabBc<?aej@T;aE~q
zcO3(AgdN#9W2q&N9ZV7VYtQZ`Z^vXpv;01tN}Ym!ThJ{|&CK*c@cj!y<3a$uXgCw#
ziOIqu3OvQW;!wk#(4hF-j<Xu0qr)z~Je10*(4fYin@whuPv2x6axUq#(m0%SKeDvz
zj*MaWNLDM~8xj4*<c6txk9aWEj{Iazl3O7uCFS4e7~l!};W)ftlRuE(p$~Yd5_^Tu
zUe&CVDvq(QY4;g<#ABbz8haG}+r~%#o=Hkj!K8|m;=?(Ay`2wSIb(N79&_ClKffvL
z!=dOc%4~S~xBBkxH&FVQV4?F^1j7{1qY+q+K#FYWwx7Ov^X3xvQU?~Eb3RYN*@Ct1
z+POX`7w0D;ea4ZYYmcza{VsohX}nIBy%1a2>Qtin&8%YPf;=Sb3fu`#MfQC9VW0W;
z-ciHYjT6rV?$7Vl$s4)d%C|>U{|wjF)0=9RuV_f}IHbAu6<$eRUS8v;e~-P~iGs!q
zYoDo&i<LDCgKWH^H#9IYNp5#98eqqz3D6AdG%mI)(<2asB6Q0AlgBa6Vl2(Oq{}g2
zz|Bt~%g!yMUs(VAs<BgNFH?pVPZv{-4<5O7FLU1VT5yfYzB3u!9cqt0+pc{o^i*Vf
zn<m+RS@5ieaNPm}RCw!FH*lZ{zH4&(LJulk)hZCV`LY0uQbJgG2R3Ecn6JV~a|0Yx
z|3M|3lQ2PYC$FZ2G~<5Vz%<K3h+n-Rf5vXXG;*}6yF*Gx)apQw;uUTQ%IgRjts||{
z9lcveo!f@*^74PmKDj6GNfV#g>fKp**s-lP0@}ml>WLz<q6Jc@q_i}zy@iL50YNjw
zK&}^nH$XXDN06z+mMyz60FFoAH43B2ncL|p_2Hd@G(-1{#1du$c#6)Hbje?Rrq@Ch
z;-2v0$VHK-FZU}dKh3+Yf~y$PJl<b+6Lz^9*lXKxhw+d(UP#O61MqK*@O>=Jm*}M`
zAj}dK5n0F2NC{zl&+kw1iTfxRa<BR2{^CA9KV^Q^Bh#0(qVcgq+HA1tA~CpIZE@h>
zL*fTJdsg~qlRH=y9?Wk4TDNq@9~@i5W^m-~=?s*RV=lBC@OxLK3e|Ip#kX5k<h%eg
zZ>bX3wrTsxv?g+VkUR45p#zlkFVQ}K@Kx>x3aYzvXP=V*Cubh7Po)-w@3c1#@|9H$
zuigyF#b=PEHh{Zb%lZ?qT#;D4`du2b$0HBmjVN1_|F(a6olxz>lCmS-_4ky5a@YOB
z^cJiA&f*E`93NElxe~6C(-C)|yStm*Y^Z(BtgMo7H4#PEgFx%R|F60?aqD^Q-u^#r
z<3`w~$QZFzL@6pFQ&gsA(tr#VktT@-WUdg2h%{-CL@6bqO-NK4h?JeEY%)~F@_SwJ
z{r;Zgc>aQ?<38?T-}D*Y*R|HU&ULPH*;bs<^q9(=6ln+Ryi=_WKU){l@G6GaRKqZe
zOx1sI4@p7CLLP&=rLy_{@(mJ+|Dsh%i*z5T)&vhqa+z|~cn>QAbN}<w9od`{WchL8
z=a{w!q6HM8?;8oz5E~miH>eLm5=$56KKq|MZ%?mirtsb5U5KZT?QNn>wl&Rv6WsFA
zxt(V-L8R>;lJcIiev)So&L@?|M%ADDrW>&6LXUB$q*CUe>Nn19olNA-eUIW(LQ018
zXz1gxOi^LuBL8!8l@mL8-QWB1Q|%hNA$<OnDa`$SLkTFaN6kjqG@dul|3+NsDeU3-
zB7$*{L_8<HAe<D6*>4wo?AY8Lc<-De*E%Vw2Rej}o-#V~lJ4rmx4<UL`d_+x?pT4m
zoAkTH`y1Tr7PWRS(MTvhb7`aW&3&Op6<_~JRXuL`QRVZY({p@sF7gJ#Ylxyj`KyiL
z4H%t!aAkwd+EQd|B9Jp*QhsU80Tg`!4CvU|0-_h3H5L`fXtbBta$YO~_b+b>Tkv<2
z^!Wx?t%2IDD__Vx2~&4yC^Ovl{Pu$&Dc_^lbN)N?<-N%*Kl!?<I>++z)~><*LgLad
zL@jZ7Y#Qer7QFlH?3xZ|E_&w1E?#p&eMjTZ(`EbeOr0+OcXP*ycyGPqUG{~3{`|A$
zo6M)HfWKu&=q~pkd}cNyoJH8z%3c+x7Nun={s~Lp76JIH3x@e!YyGx1V2F14g@tQP
z&CIaGGu%~*MBndB({Aa3tL<8#QTtm=n=~n{&|vS%Uja<d;of3NmHW9N&+7+j&e$+z
zg0*^Z*Hx;M8mtplyfSJF%^Ql#e?E<w{b_1&w{oRYYpdwxhpmk!RQ|9qHFXcT=5loT
zmKrnDveDzlCCwjG2sW^|93QU;ZD!bg_e^@by15$-aIq~CH=Q|m<O=gAo|w;gv!nCv
zk&>!=k<&Z7)fekwp|Q8EjHu)(Kse27D<c}155~YE8p?ZS{U-hx?vd*26!y|lWS#Mi
z3zK@mfN^7Ubpy1GPbJ*E_vqurFR{kM3^qt~X5QPEC!JtrVHGuC=8#8A$I14S;=y-x
z(HK;I#pSHv)q_Jq^43jUOZTKp=g#A`LZ8-EWTr8*n?iRGUmi<LABi{zp%8<Ef`((7
z8ad(*$+Q08{`J+FPj*zy=(SSi($8U%6t&h;oyZaV_Uk(0yvCa*xW2kSK+lLA!`=^F
zm-oGP`uW+Ru&1>2cW-!}NUJU`JY0u$s0Fmq-QFao#Y3v==C+rlD~QWABQ?uNbLZ}n
zdMa{zJDwHq8wa7U-~d(^sMfrEIX8&bjXL~$j@f<($E#PbsxiVJ9TTHSxMm|D2vr>H
z!a|8;y+0u^ZBg|(qggtWG_4kV-O^#eZ#?tEhbY6y{UMg`uuvfmjF-1!dCh61{B;wR
zl$96oj3~WB5T<9T@GSY4Swe}gAL((%j-I|t-(TwJV3)Sn`HH2$tZnZQ``Mv<sY;rf
zOW>S!bM1S>R1BPQjXuA4q_l609US7BtvaDOKyuwmBoZ@1U!edClw5~N&?{AszkFf%
zdF*V``-%M3ixjuoU^M_lNo5n3V?~B3@9N@{b@k7)GR4Om+l{9oetdtvaMn+d$ylj6
zFlFCf12XK9?Cflnrfm%ywZg=8Ml29HeMd)o4El8E+|;z(@;0@;5-otLDkFSTK?HLN
z>PEi<I|dI)H@a@MnLT^wf9{#WcW-VBr`puqe4qgQ9uwPMQ}V@T)Zy)(2aY~@bgZVA
z!NpDQM@_ZxV@+Y=;^Y}L;DY3FWG))E?geA^u4wd?#*Uqjr&j@qMie>1p$eDW>f{cu
zS4})IQV-C(t2ik(c3B6N@Z;~-Ny2BlWSnfvuHx04hOiz}Kv{-8&(KZlm^%eeF)>B^
zYGp`H)b+=57d_OkJ~TI;94F(^d+09Cz#|5#n4TM{N8wUTcaTp1N+(sdqkJ*w#iN;|
zwTISLR8~I92`Ef1xXCNrU6EXiH)FC~6E6QvW!bGI2g`Mq<|yTxy-F>!ew3K@yngjh
zS1D<iE0z}CFNbz%{?lAPSLLF`RS_$p9DEq!HktP0MGw;@OEy2>D7uV46aH8)wqKOc
z0;+Mo^78H5L|*#t>PvSSWv{hMZ;zn~z_USf+{3$f3u&eWA3Qj0|E(f_#9g`(08}yX
z!UR<OgXDzs8Tvw0^nB~5`9O$6J{bjyFM|h}>y<3wY;O(?SW+sm`t`3ZA49TDR;`LY
z<<q)3`s~>R;B(XSWfM>26ZTa{%lYoFK~-?+$vMO`4m3tTHmR>3H8xaiQVeF39&Hsn
zjA>-YFOy7=mI{{^D48MQDCS%EWo1L&fW)`$^^G}XdSOmbpTR+Q%By)*ovxI)=Xv}%
zC72<*0jByYt;BSv{?XY3G?n}Vx3BIwbZ8P3QqAkv0&`kqduR3P-FwxgC*7Xz?k*{4
zwCkA>hq%t3!DiiDN$9-IZhM{gUs^Tsc!{?B4H6wcvQ;ehM$;)hK=q&6(Gun9_<-Y3
zRX;e*B2lf=Q^q;^V5MD4Gxy$|I|a;XG$Px1!#Y-Tmb9p0T+lqLgRj}1Iqs6NMakQ<
zobN_2c)OzJ?G2+*;v*Qi<*S$3vBWBLqGKq7L;iv7!vA9RuPD)oJ9Z3??mKW`l2$JZ
zUo_z0qy)kk8j(kgiq_QC>ACcjSSSpskV?^55MrF$rCARK<zd#|w);{oh%+s>`XdbJ
z@sV8-<td@dd7hs6F2jIwr=xuP_hA4K7*10yWG3?L6<gSScTSawojqTl8h*&+HVY^L
z=|C>E>WTdPliegvBLPAM|8BL~Ixo0cwS&jb(a%n|s1uNUkB*r%$s8z1geuQ)OF(*G
zUX_lO*rASc%MP-;jrr;YT*rft>O@sfqw^fbFkG9wS`WK3#MB30KYwWO`MDG(m58%1
zAb>@a4zFCnp|dFu*m=5;hVeBj!1#m&>=0rXPW=kvzuxzS{bj}CNk(Bi7M~JjmR;Su
ztk2e~R<5)|aRS|9PKMc$R|v`&aIm`$5ypDUTXnMay_Kp8BS+4T)D18qU-}1joMdWj
zEM}?VA|lvmyo9|jrbJKR5jzxo2cF?fqP>si3!^&V+TL+URiwyeqC$60yYqXxOIa1m
zh3^OPwPAkA$vIpHkxQ}2Pqe&J#A%YPs;-Xw^2Gz=YRJ|3Oq1!(*?!_vV`K3+pRWd-
zPXqB}VSZDvwG!pf?*QI*mj9HbOQ_sN4<9}Q2q#YKpgh_Igya)xX=$|3MZ_C8xWo0$
zKdL%SJILm8t&SF`j)qDP)-u6^t&}fYz)Wx+EsjL3%^#z@KOR2afAC)tC#nDBpWb7o
zMe_9V;x~prvD{SQW^r-lUY-slT8ce^s+|2>J3<Nv%`#RwkzY^Z_HK+ET8D>-nX~gn
z9_BjNF%q2-<9A#L)+}57Ui=z~+D=3VOBfYDM9m`(`e<!!E0l&!$Y~}mz$lCpXpzSC
z6rYu5@1eXJqtEXu`VJUi?>Z)BI`KJ_1+<#urg=8TGLhV;fB(n+aT3XE(xJ}&7fxC$
zRX)Y(@;cYUo_T$)_(mU6UXL;T7@}2v$6T&YEZWQb0n&CQBA?-iBJ!Dgrd7ZFvmsWZ
zzHaXq7Zal^_B8RReqEB{Xaku@TPd9rg8ACN$mDB4>FO2(?HPwl{$8=CaAVu2D!TM=
zY62<dQ=rQ6d0l$+*bp;I=gli8jTBbKaK_KO+gJMewHF@|hw6@J+sT@lnnrVb2WvR#
zAU-W4?{lX{p2&~kk8SpNKT%zxHfhLasg%x8pz-B~(TZDk{yIaJehS7h{Rd0g88dF0
z%AwFRLZ=aY>eP8sn(WCg8q;WZRd`efkW)&D_5$#EaSQ~~PjRl%59qHlgH}5@$9LjR
za7hl7NSMYJ9SV`ql`t6m$gE|8;~QMDD`U+i#l>nGud+TrH)fy3gumNaoMtE9Xf6qr
zsY#M`7SYM_2*!{0RDH6*T#OhK(8;kNNx!6|glphCakqGxzQtSPWY?t5@EXzgWzIl4
z1QOrLU#o}yjNL*D;i@g7dXG;s>Dy7}+PSu`x+qpk@*;>a?_o|qYnC51TX??8z`+v9
z$F+e5E0}X7(*@!TrNnt2YmpM`{v=1_40Q0GnB$ULl5kME)PA_rU-$0aYdrn)#~hA0
z3}(SgG>^(dho1J#wNegY+gmt~rSx4Im&y6q^v>58h1pJyeYRwD!YzZac3WiIfZFxh
z>?3%DC1LdBcj+@Ru)IV7k@w1+!_whU?@uQxN7GWJWl;djfxGEirxoF=65`?t>7Jx;
z#=tCuGy_c2gW$dj3w>8(bfL0p%^Ic2lb7<YjE#(fpde6^1q61KB>pj>vs8bF+8*EZ
zf-*Y%eP_k135@>nb$w`VK1!5C8#zu7w7dwQltNN1S3&AxjPo2LyY?pY!f&~Nwz9UQ
zS$j2^id)2EAaIs2zD{azG;?)LL9Z9XQ;g-LW3*Fzy0!7|79K?b+$L>`fWXd@^#Plz
z&#*HLWsU*SY*+C#mskGYO$jyw%(BTgs6As@{4SmF^UjC^v1rwH6^|P;rWljHa|qmM
zy`(R2_{U>l-hM&0ufMY*dGx4Jx{UoKX&(&@jR%k{hKM>024YE#g5Iu{hn-Sfp}yGN
zlXv9w>2Z4d{QXTlJu`Sv>z!1uFUFY(DWLceeD#r>E-&VK@C-V}k&i*DZBLOUfP32P
z9X&<TBo}vET7QdRuA-7WzigqtzJpU@)2kvHRXOz=Dio!=WG;}hzOk!V02Qar@2-AY
zp#9^0&No&%2w6$}(WUDPvwya1px#uNIMIw$V8ls>YOr%P37@Tn>VnC$zJ2;cq4pZB
zqGE{Xu6D7`{-SGMTdZ92)n!q>awDSgB2ZG;+v7yL>owuiJGay=^ifS|edyghq273?
zc%UaTaOL}78WU&Ic<xl~rtI%w;%{?z-yp?O)fideEZu*+&gx4`i2a9mciK=Q^|~4W
zM<Hnkl72l1d)!D$yGgc+5~qb%>-p9Egs=o{DSiZQf%(TL%y~HFlRp~~69Fx&XG#ST
z`&Ok|fOy{wN?SZeR>LWLdYo@*4Jw~fA*m-X0u8x)dQ=;<<qC!$A2;O}F&e&{;F0m+
z_P{rT*+#X9!TJ2@DVwQFkb{;eC0&g`M|87RLnpz^$S8ud5H{7zLe!PWchBMp6o27D
zL9Lyn;NS5u3F61HS4dtoc#WDgzjn0F{uInS!I}EFg9Xcr#S(nx2^xJ1JV5|)DD$<P
zs(h&z8d<pDz!n8E&?Ta@x<m13(XSZXTVm0k;!XTi1u?ii|MXNVhNX>u0km#+UpQZ>
zAN#a(|2jcOCa>7ZetRW7BV?|mpR-iBI}S@Fp1!rhaK(xUKw%TFq6h&ZA~Y!VLFOho
zNs2oCEAjiUkNp0khfgensR_I<Kt&k?_kED`idB)-AZ#x^bqbgG_;?e(3hoksQ;GFs
zP<i5Zip?3w*{t5pk2~A=Qbu))wZ$u?8!w#tn+*(khf9P=2_yA@WU}mIV2ICdMZ1o*
zU|zsm6!W#Bp+k&S)iT=un&zVO6<5}CxAi<>%Ln}Kh={l8YtYQqUd*coMTz)GXqOJm
zyL<Oa)TGs?J?j-j!$aB{$BM}nSjKqb*E#OWSC0kCXRmJ3R@%Q!JFLFmp?Y6+km$~^
z$_Fe|g6?c36P|(XQ#K<^)So}!@%=?SphT>JR^Xwnd$n@xLYl06{PowquAW&xIcx!W
zq^(giPr3YqcwdsLUL=(*3^-^t^gnzz3Mq!Ab7a(+0e<wUiiJbg{b2~8e~t{xmg~1)
z&*0Tu?~e0_4Vqv~UOObpN-3wi{N<HtPy6=lxy-eXB;W%0*zx(*B|*hlFAL5#H#d(}
zbtmul=RA{>%OKUdB}`+$)YR0}s`<((OyjjGrHXHL`&GU=-9@5Z^iqAwlqsU;!zeGL
zTr3khZpx=|`DEbbuu}Q3_78IFT<d}M^}NTdyscZecAMM2%E)WdB`5MQrMw}vGK9ie
z4op0{0nxeAkdvrM_nWp`KFsAxzi`d>pGDkBPhu2{Aae3^0!w`LCU9Dtm1fXPx?fDP
zT9ftZ!DKD(AL#%L%b4)+VH7=Ywsjg(pk8JfNoYt&PhTvkCmeWSp{Br!=o}`<=oRr0
z|IVyf@7gV8I<V%U_UDs@D&p(5j3_QPp<i<9b6#bo4LDOI>#6vbMUH3}A`ePu(Rl#|
z2f?RLC-HKcgSvxx(r27AOX%2rP^`_5e`P|adSyV!@pXp_7BW&w=kT(<hI7BUUpdMd
z0kr(R?bWSH611f1`TDCLbX&L{i`WQ`gU%VU&7JjWN>C#>Qye0RvEbkw-n&5ZTEXbs
z9?*`#;GjQCpO5@LN&@^fzU>vkpuIV>5<egQk%()M5hHxXtP2q-b=~=CxiJ|T7BGY0
zMn{vb5_3t?ieY;6Y|I&-U+WP0B=`uDgUZU<{^Bc8#>)2y@HLf#Le4+`U=?qsu*BRZ
zk;P&4jKGw_VyQ81)beUlb|Iq=LfNpxs9gZ_+fkyPzWDU*#l>SNM4-AgzIXe>(XYh8
z@%<OFP$R=43Q#VbfA(C8nn-!5f&Yx4K73hxo=DqpX~i*POxlVH5k0DUp`Tc~>G|e0
zS+Bz7*<G}F+|~Enh1m&Yfr6Hn1yAE(&Wi$kM5O0-wLFQDN)yyDJh;diM^2qu;L_8V
zBxIVoPkG<}XaTIO1gTXgI07D|8}<{;IUTCZ05y7oR3Wn_5#k~v!vm=5){Nc6`?x-z
z*Q*1P8*+FbYQ}J2`H-SiW5HqqSd^<%M;j>rSf6xPbRg@U_D+{LmWlQowkV>YVoLws
z{re_#nFKBYz|vPrZm;6o?|}KFgYxtO44Ki5E>tP#C=6k_IFY~4;~(OJ&7N%!#CSbO
ztX9-B;$FU0!PHQ5P~fPCozzM(@uNH`tXiekNlGn>&=bmLupu5*Qlh%2HhLuB0Q2K{
zb*{4T@ydG+sOSWV<}%P~;NW(S1w2pjKO=3$<4>OjfJ>!0nnXEOk9W`cOFOSKxY`6V
z=2RHqnsrdVuZ)Zc(y$b<cXSdV)5J;~EiB%~^xH2Ef3IFEC~DUKQ&wiq!7L*bx-te$
z>3GVeVTQy`52Cs#&EVQrAgFGx&OnjQDx7kpb#ui#aB53~)C2XOvwvm@0)?i_XbMJ3
z%hbF*tu^C(L}-(@$LYXXD@<0+?%<*1!rRn__0$)SIHd?lNZji9_m*cHOVSrpws?JO
z)O=k-d>7&ztg`bC{9*T(hx~SV(M<4sd1wv|m6geh{R4MQy!paO@cF_j>X#0psu#ZH
zYe8O~R0yfKj{{j2(q-%N=bwdyICZRfrH4+v>b`Z|x^vGxs->IWJ}tXmY?nZ_-Npr$
zr|5w$VF3h<M_YToOHZA{-`F|Q$gQ%kxfcw5<169^NPuu+mm|b5eR}nZpoV#07i{qU
zvF!nS44>G-aug`P)^$+YC;kzQ;>W+Q><$34Xew3s{A>w|`uNfL>$h(cHQbplp@@sE
zM%^tsl=2F2rS!6>QI6e+lccyShZs=K9eGq|vVQ$}#x|0;?~3fu<9@bf&EdWi_pI;f
zAfOQ&HdVBCE#0UwNVLk{rB`<sd5K$a|GI!Y1+T&qG&}bfrd$GXiWU)$ZjsgKIDU1E
z{_jRfDqO2*Vo(qqBoApR`cCxgDCabQBLvD9;wk=V=nior%2Np5EfnVkV6XF}4C;;8
zTnO=^C_T<h!7lN8z}Kd<J1%di<OH}Yj2^v*=G^Ms$nH+Zq3avjd~tG9?-k<L*>x4+
zjf#(BOx^Fen7<~v{9II_8F;|ooy#n6sHRH{^rh?|+q9=AH2eBq_8Z`Lebv6ez@ruG
z&8ZK5RhkmJOC;Eo<vAVs>vvFqpd=7<TtbV)u8_QN&TQ?DkSv6iKpfG=!NL1dHz`X)
zwa(k}8Y@QK1PG{U{3+g+#PI-96j6+Nw|JVGn9L0VT&Xj8&6n^@6xxDMWrwaH{HZoT
z6sG$6`<M_62<#Nzm$YUX!-N;PTKq1ZPs~vJvJd49fmKhC1Kf`ox)ETv>owF!rRZ-{
zX&(>^=fkcGm<PN`c=>WAU%SwW)gEi>AXKUH|CW61-u4DfkvFf;ZgKyGL+wV0*b1LS
zq7y|sxLggHh@}dYy11319RWv4B_bdy4SVzH)3>e9YBD1aJK2gafWW`#^<?p@Q+RgE
zzp-~i<YX{=H^XJi!kI_Sdu}PQ7z_VNOa|O4Oq*sy+n|lv>nE8sXGhQUd6kaCx1P`U
z6CXQ#?A?3!irA}0d6@u^QB+eauAOVwf!Y?4ym>=<iGwhv2a8B<W42aC64>$}kM-e@
z*`qsj>LiMTcY8R6$ALF$-0RIAPMbt^R`5-fK3ZFv&$Y=$Joq4@9`eGvGBuYID~i6l
zxU;AJ^%-RuM4lH23*X>U=}FJRMK&2BrL&(N>3K@Rxs(C_M{6KilCtQkR;Tm6Y%sW2
z+opCX7$aKhaGIXE4G?nTMfCOb3elgP3k{XcYARjoNo7|9CC1RlGv5>^b8<U#C#3)?
zQ$$O8AR-OA4H`X=FOi)@xFWL^!DAKE(dPM%95w2mptSc$EJTY`{8U{O&L5s^UZBSw
zFM8MOZ#W^awa4!&bi$fI_^PRd!GS0HysHw52UH4_N@MjzuA<>yLNkn`rIRNxvXxIu
z6RlFwrxw(XU^#FeETr3{#vCAJTKEGuimMtY4Oz?E19aX%8MVC!%cIb3_iCK$u=-27
zG%XXA&MA+7{}eopmTT>ztVtucHP99FWFOFxu&}Xu=oCu{z$G+ucx~S;ACRFi>8-|6
zYL?52gYu0XBBFo`s7PHK!bCyT)AgAFHP=xoZJE;clIdgMU4x>~_LOFhq5KXK6@sjv
z#<=~rBB(F#=H|{p?I2p(9+zY#$_!rY=2VKaRnPD?75F)qE05}md0-~Lh}BtW@Kr}E
zNJhL@?7IF1!NB6$)+_X7o;#e)9IEYU!q|?VxC5dD4l2L>R&YF3-q^vMV^UENBurU=
zk2POEhHnHmKABK8gYQ=z9na=1NOU?#EP|196;F15dFL-|SM-C#!T>JpsU=T!214+y
zUryuv8=Vn3=FKZM<_@O2)r%fa5q3I%pU5|RI&59nw1K>;?-spz^GrC&Y+@3RiI7ul
zjs1nmh978oXm81$wAyp|E`SvN1!kd1xjH9VQm_?z+T!(V0Vr~FbHmllPvzA(I4v=i
zR7OnioHiblDb272Xf_56jt(R$)H7UDEJ|LSe3ZOfk9uZ`i<)BP^Cf3!{SV6}OXASn
z)HAfw)H+UN^C=z|(Aia1-QRiie%+CG$8C>3S*aRzEw5$~1T<?TrQ<652ywvMr)PN!
z>{XbH@I4K!wtrv;iK7sNxRUlfbD2vY3&+>;-<oM_TNee<uro0>eiWMw+y_9hBhHXd
zPW*D?c@O5faDU8zXZ@5dAOEvg>N6b4`sKeQw*G}G^-Fk&VwNAtSDM~~N}~;v#Oc|Z
z3wr?B+SIHbBu|XOgXFQVNyx)>LGOlEDGlvmE9*DS$`C7&aWfqz?F7g}XB*l0w4}-_
zOqpWoo>@_4A}Mg^vA#K<`3};{{A~yUiq9zuZq_y2ImjQfVi{(3RXT>=77wR>KZiYz
zVC{gqqVA#>ODrrSmaDUSxA|g+2RpWOvIyX%zJB?lR{v->>V^2YIF-iU=ScRni9=|h
z!Jv(+8$`?iP}Q@}g<tt*(Em`yg9dGl9^Q@*r*mR?Q1RsAnck;vt92fkx^Gn>-&QIX
z;e$1uXSG3gN;?{;S4^Kxl{lFrX{LF1ZqqX`U((f&`=s`^zA0r8vFaFovb7w^{DU3s
zvOGq1OnZ+_L7qdZ0~}<K{O%k%+ST(<O<Cge^Lr)iuHbu7RVk{gn=*VIoAvH<uJV$L
zv+%50L|q~jMX)ntmt2f;{|`0?nAen9*>c``7g&kD!ab_a{o}43EO96iC7xb=S<-2e
zFgdHa>+^n)D9DjS?|}tm^#x~k01RB&fwVuPBKvFphCZ^g@dN@zo{gw8qBcy^?v3b@
z_)Sy1cGtl5i$Dd{KEKqrmQqmr0uZ`!50@mG1q_&=3nEGqBt&;oMt}bSeokG%eSH1U
z^YPk(%ZA%=U^q<R%|3$$otxJB;XwH{uT1aeOQ-9ELFoi{7{718-bBHm0<{-$B&|JO
zbYN&p4`#58ogWTm>F!!#O{ylQrX}Zn2KVb1gBH!?v{&<SR>Q_V-H>f1gerP^SE?L0
z*Phwkh<$VdPHq|B-(`xFK~r)jJqq*44WB8S1k?~RMx<P!(|oRpHVwxO_tbA80)e5v
z%0E5LMiBj^%|^T+EsGo4WW~ODyviAn76scpBIG*L?+SeC^vSY*PHQQLJdsta)5Fs%
zdg#;YLPae;H$W+V!u(eBP(?PaA7xq)Tc6i1WyY|=_0fE>oGyjkZ@Ogx`}VbU3jRL5
z)d*2fF)fO-Odndl_vFBOVu*VL)eUZOA~*-J<Wdc%IV0><986k@ij2luUtP>|3W^fU
z9{^Egi^ivwt!6#68xP9Y{XOmLULa~L^y)DfcYea3RJcENyx$KTYq!l~XUDWxfCqr)
zwdOV4X&S(KMq#UTN5*m2npW3Yc{P=MY;HDV=jNl`!6__rg|o4vXP=rsU5~m~RGzY4
zr?{U1vg+~8ujy-+w&tj>x_6|!nnY|)jX;-DdD^e^BQ1QfVWgk5Rc}aAMp;MSNEDxk
zZjW1_U^Ij1)mp#yK-n}eCgVgN3kO9Xm@=k>0xfjnxhUr3^^JBI9CAl?S@?Q@mTN2*
zdaY0QE?taJ5)&lHxgo%N@EExwA>MLD8wQrhHdB1m(KXBz6MkcDu2~YZb?5L5bw(ur
zGx}L*TH2;8Tf8uDB=h>xLpEKpq=JD^kJWCju-;KwR*<}(Und{u{mJN?`-U609cnKh
zn-}9UM6foJDt%3`l39{<BGkLyw}#ytva^i@hP0R{>t{EC5y~YEi<yxSeGslO;UFH+
z9u&nlR5rkzqyPZ1bl=73s$P(wUjQtnSN=mm6yk-bpPx|u&6}@GT=!V46GL}>Cnskn
zje{cYhwKl@IPpTtNBhq{A;#xLXX!yQZ4M>IGsw<Cjc=#!3`*hTYlb~2-^gP%WyCYf
z#k*|L{rmUVnRGmnKT1zDg6dCZv(=OvnYm}Fh=_n{q0c}TIvf$CfB^LSOQ12^E_}M&
z>Yfqvue=IftHvm?5GabXD>5I34B8B?2Fa6;gq#p8WJoW=8_8xue7wG3ogv8T-_~p@
zF*UC{G-pU5AuJXy6j|_sN!$(1H$O%3!^aP%4=Ln?6itix0<1mjMnX-f*E#<B)}6a?
zU<mmFeY5CQ2<d48>yXRXR#Ns=*()>c`DdTj%H*G!pPwV8J&Qe)HXekj-&;V1V=dDW
zj4e;C7JXjP_^0Xe?bbMNImiZ~RxZbeTrAwudOcrXzsNDIBF3lXjZFs0%az-HI&He0
z>hZP~5>^)pLcWh+RRntS@ozrH`t=W)Et%t9wdn&SGC4do%qFthxDKZd$H?n>wLNu=
zzpc0N;{sa$ZZHZdF<F5+B<zzXpN{Iy7@Opmb#<Gc%R-K5(RC3H9Z4#il?gqf0EO*-
zaz~JC66GA>G4}d<aMkjO_H_Z%KYm*G?5Br$`GwhQOf8Lu`}q7^op=JyhtRA@lY*0w
zZz0f8{_X*Ob}m#`D5{)qY$VGESgVd?jD$od6Fg${=vYHd@%trq4p0H2yhbiHmJCBz
zwhdAz3i*}z@tsKNr!T4j=zTinBZw-j-m(+zm^nNko!$Vawm(_W@$IuK#<+hOE?IJ1
zUSaA~Yf-J+*~Q*KU?M3fzOUO!5*qqsT`9TkEJ#OIDBO90e$}kK0#MVm4Rg&Myh(^3
z)H)T0<IA#$e^$bpws9KGU{O!plpF>=eC-13MTu|v6HiIm;A$MI3J%bf)(vp%Lc8Em
zgZJmjIrZ%l2YenE{<Uz}tb{c@SR8TJGI04@LKS|CEe277pA?!16wxYL-7PL)YiUWz
za*Cm&aT@Ldhtkf8CoMK?w_O;f70rKJ$cVYUNg!cqExYRiWVb|Lu~eorIhO#g8a1^b
zbmPxI{HOme2*UfKL~?we4ruYDguQQHP|hN~nGjCqv=k}8=a#2$j69#!Y(a&2Zb?1A
zO8euT5fCM2mx|jt)uF*``(jgO)YQ~q|9G4yP!=04w@pc4JCUPGV8Iz6IGD2atFsq=
z8|cV39Q6bnWw*TDFi?9!f8ULEJtOi_afp$OP`CMS_AeQceI}W9NfCX^(BW;p;PBvN
z_&^51L^o0N%GuG<QSN*!+6|%C;)=?+X>&HfW!$St4qNr~_5%6GWL9|JNG9$EJTNgc
zi-9Uj$jHbr)KvZl(5<9o09|GN@}!K*QVa=YQyZ0@(PiAmNPjHapw$UWU@~F*)Ys<U
zu4?U|Hdxkgd~uR@lM*UywKwe&^J<D%1@`pZS__-|vR{8ro(PH;4;RKZRPmI2UCfE~
zMMM<9a7mO~+qX17htiACQ?NEED@L@TE?zWZErmc1>XcNG)o1#$M(W)8jCw_CWhIU$
zjSWwXWZ(F2rXU52!Rb6Nw*_e;9XEAyx`2A2go+dXu9linx>V5iSS9!E-#-@d5Ijrm
z$-KU&rffD6S`rF=uB`yKSjfgPD(;Sv(3BAgCGpS=3dlYnJs8&r!8EuLqM$X%XDGoS
z<L%|A)b?jZr<QgYGP5Gse#^<D;F-yxC}CqawR9S~brGqoSv!9Jm9aviOqGZOX*87X
z*PspUERGVf3@B5LxN96VQ8<vj_32n8PIyOvGsfXP>rDf7v{`~Fj*2P5=?&1Qjk|hK
zl$a54NUd*vx*?>oGCLZ+13Y`>LCZJ3Vk1LC!?&9^y?dsRP*PsyxBvf#so=?j=uhPB
zxds<468RS_MMiDqt)1^|hI)TF18g0_7Cg@~Z91u_6WvG-B}1xPI~2cO?n(hADtd8I
z1?ML{^~{U5u>{dKC09<g55r;4hK~26tD8SHCfKyZ>#c3$h0dQt9VjS-d-?gn@@};S
z)4@tKX?EYZTM+sT4px8e+ydaAIjYs^Y~>8}Cd97=BZ~%G)P^;3pdA*2lvp>M**!#j
z)93i~b&vx`4;(lk$WMl8tePAImmnIZLR27z>p=6E5k1=A^L;4?P608M`#L1!fMroO
zbKcBEl|8xPY9!r0WJD3o?H;SEtAm{j1%-kVNZ?TNk>KprG15R(Hj5ssOv`s+FVx5z
zaKKz!<UKsNg&NSb$J8(3I3FoOwiEx};5Spocap@N$ZXD_tP&p%*TD#2qGhAwS6F2Q
zG!i1u(HY6E{m*W*;2yM*E+Tz=MAD@zJ6?P%V$`A}+s8-TY|=u|HLv?gyp>JeVO!v5
z=n08-023l5crV3>5}1==jyQu)E8p^YLmiEC(cBTqTM*08*g{4kz6dT|=>Yc2yeawN
zeN<#XG2?Jj(?#@h<{wv`O^Y9dlwDhy%OPPM#6cL6eq$}7WNfw?uKwindPWv^H9j`B
z0Iw){s~@%<K95W3tJZn>u}(ugbkNjzvO&~8uqa~5n0}(s?(k7dbSg<6>R4x%%XDiC
z1E1N$Gm;x1&<93SKoz5~Ne0`L<e){#O^0r+D8yS4iDncpwvZfs-mzzE65Z*n=JK@+
zr<5X`c!>*}o)8%z*Ilw<=UGlv8R%a`qy!R|evrr<0-z%`cDwxm{RD5IP;}J=V-XYu
zv~@u3EKry$4i`;`_T0B}IYE?y45SSj3)Sq&a9I=e#it-yKk-fRr{IE4m)r7K1ba(>
z9>wDqHBokJQ&Otai<?@~H`x-d2#tsw?a1Ag8MRMos6^v>cdxAMbA(G|ji&R<J#AtY
zh;Z8ViH`arQ-TZHa+#P0^!fg@Qr7DMl%bdf^CW8si4?msz1&*OvTc(Wr@$>kE9re}
zzIL2ah3x|NByt(`c61@?oHYgsJ`F8@eg61s{R6J}R5<626_sQY?J;}uQfjr(+t6MV
zWl@lFYx4H*GUZ*Y6>f^9;g6`=n&=%Ujvs&Fd5z4aevPzujKBYRw_!3c2r7q0hO$Q>
zdu(*y<`z7*l)0qaddpus<o5Y6pEDNd4kG9`GYOV#@SrU>j=xQl-;JcLsHu7U)v(Zi
zI1Q_k7d!a0o>w?__^``^H>^*c#a$Fh)zM|qx629}>6aD{{D~!8Jn=p?|3(8#j;!G{
z-2%n9-@ZM23IZlQkLB){Cp8@H`-2Nw@1KqN>zikjI1dVfpZv0Uebdji4bOK(C{}9t
z`fM^>_LpBzses9&!?#(^nEQRzeZRWx{>9<;&wF0=`?h=a>TPHIVgiE;J>>EashF+W
zb=A*fsMN+8V^se%ocC$_bOry;H<hElf8X)s@NTPRWnE8vY`L%TVDn~;hApLDOQt;B
zd^*jRB}6kmMkFWyMV6RtY~0)Ta_~W&r9rybbVDM!HYi}`#@wB*rRB2`Rt9)cAC`G|
zi1tW$YHD9Z9u`Waum$h=EdK(e7DhL4jJn*NE;*<l-G7?*6iF4!W)?;-^s1eZt(I`G
zZ<n3#y@Y);G^h##lP}hEe=T!p{SU`O{xevNR9b45aKuk8HT=`ZkK<{zKYseu78$QF
zy(V6W6-#F-^76XkcycE{zeF>TnmmX9&Aai#S_cndvCTD6!|Qg3{$?X=Di^0(NFSfK
zi@^)Ktd8GGA7p!c7%uFd>!D``a2Fy*wqh}L3^zO>nYpx5Lha7afDrxYqVoOvIT*z9
z$=|*D%Wz>Y^mSUzI7Y|CDTJOlG5N=k{O_j5rSF&HI~GyReR#no)3SY<Tgm!lX=&-$
zxVUX><AN=GzXs8SB)n;Sx}~zr;7;w^heN^5D7vwz$B-%42?<r~X93i-@7QrTM%I#9
zJ<k0tFRznmW=oMO@y*E4f?)3<jD+|ZQfD#5dhG0n*JyW0bN~{iWmx#5C*Ds@g^hsr
zhih($ipA$jy^3FcJske<2WfxCvSq`lto_PH@gODRslg-%nlEktwBNJK(f%WO#<ADw
zgUwzTHo&5*C$>>y9Rs2D0?el{@+T0bvtx}HKZu7N*=QL(TVa3$$1y72*v80cJLL2V
zEY`$Ip7*4ZYPyX%oCU+hi${++UAD<`SMi-m^8F-7^rP)cPRpf!O0von+l18A)&KMY
z0t)*ft?wk4T*%9fc)T&iKV`4a`d^+$p7Ur!%(1sufMvtNFqt+Zem(uSEH2JTu`wKR
zXl{mOqJ@o34~f=cr?-a(huyk-pzPV-A3lBRt!>hB0hVE3V19o7?A%eKMjc2kcZc|3
zwUVx#-DsLf?E8@*l27!man@uhy#%XRN`szIp1KAGT~|Lm)pDoo=9Y;MW#*e+2z-9P
zD6UVhz3qGLon@El^K)vkLu%~&vTJ@;DibC=a(MKgTzRXDZrRtN;;c^=1NFik=4DNd
zoR*d>{ETq?#z#f+U|ht5@zdw(ZhUdG!}7z6rDv|4{++E)xC=>;mLGs0!3!&gtx>C1
ztr|aR(i~h8#d?mdtWsxdLQm~FsId7g80nntFXfw-Hi@=!G8i3uzWd=D2N;d0JoxxQ
zt?^_Z4xOEyo$rh@*J-!U@=En_?(X|u*4)J8ea<hpk=kkD5fOjTy`AakI977~$+%ws
zq)Q~-j_b{uZt~<x-q=xlTE*5ci3C3Hru3Eoj4byDT3K2S3)bI={xIe8<$Y%VO!as#
zVfR6wKo(jpb8zUwZDWt#*-cV|dGYH|0m9Oe57hN7OqxY9$>=8WAQ^+8BOxKe>6qw)
z=Zj?7kF|2uT>3vGC)tpBY=SIa-(p$o?%kc{&7UtNfi1y%<t8lUco4T-GwwhgKYL*%
zTBs=rPf6(`VS|`LV<!pO@~>M8!TKksY%Z#rnz`c6kI*TbRr`&z?uI=JZ;loeYuWy^
zdHrK}eSZUgqcn(y8Md|~1R!#ll+rIHrr+gVtLrw=#;o0YL2MWHb=TgC7sT?oE4wU}
z1#KLA?eev1)hOL-a>-$ZXWhLq+69H~d>MY5mWv9yhC~;KRx6Tv<B(2m;pC8to_w$6
zsX3osQU?A*N8|$JOL1wb*nM(r{PI5l_sijW|9kOQ-AFaHJ`!POS?E40ARyoi9Odlf
zn3$N(ojYGgDwf)7FxB8qh#NoK-o1MZ6E13Ot;54PD+@oWrcf-gB@%~A)8g|A8y;J4
z-8x0S*Pq}Nv5*J9vaX+9{k_Jn+zK<cYu4}^s25W8o8RGSifz~&K5{|K=E^XQjWhZr
zC~b?;dq?%Ex2BQ_u!x5RyLNRH+aKAx#$WDO;6AjkgkpB~r!SN(JnEid3l)QJ78PC0
zX>DA7=f}n5<Qt%i045OEJ-#5t)9YK|QX{@|^PWa&kG5)blU18*_o#ULdX8vnYU+cA
zwKiQXF8y_R*DD3J{7xadu&S7V<hY?wj2QVvXK9?jSl>&DJe|WzAhE74I;-pL=B6%b
zsjpkRBD2W!?T_+ui)UF|WIz;;zP7N{fe>6+&wP(|>28g-2VuC0`tT<0DMJg3eOxr&
zL}&gS=2^S(*GERRZ1w70)>)0&(jRrC2VINbetd7#o2s+d^D72-+4+ag)1RjFDBRk2
z>a_oI{V>mdcoxga%eMzuaWZl_)`i8zew1V{n0B=I0P<yh-AzD*Fy^K88W|ogq4S2n
z`(Gpf7CW7ZO)C)lElzimLu~KF?w|iEGG#y%8k9xGnaZ3Aq#Z9P5ErviM{GbNny_?5
zXU)}+5E<rwcAh*r=-9DiolWCgyOzzkEi-74zgRH`F((oxy{ZDxIOUw2w)SB1++01q
zzy1}Qka>^GVVw~r;}YJ!PnVq+LSEzuJ&9Nd50ag3Wi?n>9y7fB;tjo#Tt+UoiH$$J
zbxuP%yJ8SBs4&VZ@r{liNii$ZuGn3N0j^?j7d|>#tBCLCiI#lv!V;L4cOl(K;lE0b
zZz#(t_{xlsFWaEIbEQ0hSt<AAT)h9PyWL*SZF0VGV2{6JhD7&s{r>H7=t`S8Z{MaS
zO-?#vmLM0bzkJoIo!G5r-Ci{}VC&3ZkDe?kDc{oGwbb0pOPlokRZX$&iEB?_RL#)Z
zI*?{ROQ^btt=?Ei7#bVz`1Z|HZQ{fTKn&oqcg3=|{#iTQPb>S;E^gz;9qF6CD}p9e
zn8~{;MqWPe8D6g(x<x6YdCsH%0-E1{$ni0d?=?unNdk>@3*qy|FYlKh3k?;1XvgRm
zTyWMOp{m-8*_7+}%YLBdk+R!zOp_qb3BZ{4uN65rZ;A~+_fKl}Mu^w1JS;41woApq
z51X2_>>ez)x8DoxyBzV*;%skiHWtu!KX~HAegp+!UbOp>7Y+r#%=ouP&$_GjQ}|v9
zyBx(r2MC4^1!kuvVqFX*@FTtU9LQoNG@o0AA0XD-1fX4Xyn$Q8=<7DNz9;cg799)=
z3!o>?2bg(en91lqEWtUx$m@PlQ7_@TkIVP0>hCm-Kj0P<iOnmAY{$->9f+)8YgG=e
zx!M6?W{ru&FxI=0Dlv&M3;FYA?r|GPjO_$+)=5T~s&ua+4D6ZJGNrUC{AbG1qy4xz
zo%mb-(4?12ag1NPeEBx?K8BkYT2ZF#Vz~JSFUy;hDW5Q50+Q)&^hD>Fc&qUhphNsZ
z(@gATe3tDql<YEN(W0K-tzXv^UcoT)#lQdl!QZjW@-%JOn3|Hpj@}!5p0o{>rSXRv
zYHK?O>+j?k>{zkQr}t3^9kFP5vX`{)<@WX-9vXDT+eaJirNOd{zJf*5b+@Jtl4-49
z=sxwAE5G#HvMb}XMPljXjnf;)q`SQCE;eRD3g20Iae?j7DIUJSv}4DQ->9g_%5%Z0
zFfu;A3jv1}m9?c;j7?2<`m}zXv$i3<86o>#6eg@Q7|sUU91UCRt^@eRx2b!b8lG#j
z1g!J-?)4EyaZY%(K7R7Veq`9wgruZ71jA<)6>^8>ro0<RP287l!+^1OYSYcEtOBUH
zuQnDLM0^j;uw+XryCEZ9#++$9ZUbYugT*-vNToFEtvm2(f33~DoyKVs)YW&gX$2N}
z=S17$?XAd{Q;tS&sL1F`Oq#LXu<|#&$X&I5eIV_hGn5BP)1mow!@TS#cK+Pd<h7#u
z*WSC~IUiz1YHwf@5?j5BS2_2zTY96KCY!l)PYFQF7xag}w)^*I8=al&RW<Q&OWyox
zb$Vu|E$1)nw(KX1J736L^lR`AuCy0_e>97dD)~(kmPGR03tw!3sga&Kb*k0*Qv=Zz
zF;I{>ulpa4OXq1XaPC>K!P~L&+F%0dbmaUYud9Cy|0ZMLA7m(&{&IjK#3u~&_x}f_
zNH^#WqT~yxjGJA@UG<HMlHZa0E4U1WPL-u$AQzIi^BOXvZ4AMNleX;~9sAxwTQ-xV
zhq3wI0QMEO{lxz5U~fMi@F*?@j1nQjcgx;h+HzyZjwP%Q63gpxBE>pl;Fh_w9`+g4
z`R-_80SrqIWa&k#_orlv_t@IN@~36j%Ko|i&p+4kt0th{wBCv|^v=C|!pCsuVz1wC
zP`%o5nzvP1y#1Chbuz4s0>Q!w8QLHfy~%}w-0|t04F0`y_wIDmZPp7vwZ%pL%`g`s
z%>@c_-1CGXweyu{k7o||R2E0&`Sa&Ek6pNM;W}3^ebcZ*+gKMkVD{RzN(|a8<AUO_
ztoLU3rcI^W%@l(*d|C-lQ&>vz$o|~aHfVIFsd4D;^VbNWFWB`-jFhrl`h>jU3>375
zG+XF4<$oLtA7(=SLYfSlII$m}@b-}Ihr;pHgeEvPd6U?;8K6CA2km0ML1`Lp9td&p
ziFf=@?k{bZW5eVlHCaA{+H}9zrYc4@o1HvQqKd<geFZ!_D|>ZJP5aWW`0!wILlu}r
zz4`s`XC?llR$=+}gZ_#@bfPM6@T2X9#ReHL)qRob$e`%{gKJBnBFAz_M727k7m!`m
zyLW*QmMm10A;3O-X-ppd-=C<VJk(f{(8kvf25V?&T+nSa{)k@mU(5)lB*#vin2Ft6
zqM3mBMCe&8oRL&f_)ssilR*8I({Go_U+>O#`@z^wcznykwQXb~eX)fF2W}wG@4}@^
z6PgSDkB<_2Lgm+Pu>W}{#XBy<{91(j`<WZ|xY=I#8vgIk|4aYiqt7m%{?dzVVk!k0
Pr#E}Stn)LLZU6rO3iHJT

literal 85450
zcmb@ucRba9A3uC3d+)tRA|pw%$1#$Gglrkfo>63vBRdIEa->LBvNs8pq=*z*m8@(s
z@7KGo-+h1Y`|sPM>-t`oZ|9uPd4FE-*Xz07pEv`3Z7OnhatsDTbw&qkgu#%IVlad!
zNQvP$jLhfG!#~KpbT0W~FckFYKLj5}Lv}G3Ud$P+nsH#x{AfU+vEvBwQqZ<w+8D!!
zAeuaf8667RTVcmK2}MOE@I3W*FM0{omy_d4i8Sp@ZHbuGl<8>IJ#W*JrbN-J;|Pp&
zcvBwq=s7j3n60`C8mBd$`LS=sGW0#?wC{r;;w+yBtIcIE9i+2VvH$&XQWKTO(EazP
zid&X64)>p*x1=0V|9uJ8DT+#+|Gz&c+KKu8`xdEec9ALnc?Vo~Lb;mGf8WDFYnWH?
zzi%Ne;6_XT-}lI(`~TrGQsb#r12#8X4!EqXt;@^HX^m8rl(^d3+6;;Zoh}_r_oR`M
zksbfCeMCQ}b^6Jxj;@om^l?}n;{RNki!yy$j{`lI`ir@tGW*TNQMqgLtncnC#P03}
zTh{pyegFQw@afZtHrrNM=#<P%5y4!od^IyORvGrhl<lKOkK%6KiYO>3NZIb`>A@u>
zF`KIh3u~E0n)vN4+7r^A5+T32_o{t##Pt93sZQT`bCiN)Kb2Zhl$gZhaHLy1I~^ZC
zGDt{Blvh?pR#lz6qboG+;~go!x4T<@ebot;m$c~j)KsMVmukLa$Mke=N=i!3f30KS
z<Kt7G4Z3mTa^C|bQgU*7qi-D@he)V7RaI5PQnr&1`dAYleB$AjC778~F#rEvg-Ykl
z?w?gVDOTRJ9;`N>OWN(4CsPwwY4XZ<37^i1+7`WF=K627gpd@?VB86^Be6B6*Ynvo
zd}uOl+oM0|^?U!P9%ceALwzj1zBECC=@#z|>vv$xiTTgJ522~dy?h6^_)cr_uxq-K
zsNayGK?uQv>TtZk-x5j9+WGQrKW1je*{+h1OW_)&p`jruHMN$rb72bQ-!h}Q1tnU4
zgn+tV&ajr|^w5XT|H=(da#@PCJoF%hQ{IpC%DbGl&Uh9>W8<jj&!vq`OxkUwY~SUa
z*3r{zoU0hSF(>X97awgY?Bw%kvfVXP-e2^?hY#bh!g(68?HwH&I?Sj}@H?GQ{i6&v
zJ6?ANt)pA)@63q&y@mm5bZN=T+aeK>k#rm!@pc_CZC}1{z#a7R@(P1HXe`P!y%MBw
zkuR;M++|PN?2B2I*UIBR@=<oJVFWs7&Zt>ib2o3#iea_2?cYCCE3c|zy8U;tXRIq=
zKW$=<n&vl{HqNmA`#gR3u|~Eu;dAX@>wGn}uQCg#ky+LIg_{(2?(glS-o72Nw&s_k
ze3M-%Xak$8eDk4E5w`2^CUm)-d0p;*>)-t>Q%_h&kiP%;z5C`RTEW4JPVaKyLrGaF
zi8<xGFhW8?`24?jVU$_>|K75D`WqV?TOB<6VGopoCflPame$r7{}ske_5ZWgU4$Pz
zc;H5jHo9SU7c84n%F}uu9YG^A(PvOyf47|rOSRi)!pOF;@Ixf+BO~U1oaoom_Lh<5
zS%X7E_E5!`FxR2dyho1?$I|d<UA%a)Ls=nUk&!96@REa#Qn>h)j43zor~kVP*ah{A
z7nzcHRFbB;lBd$nE4Pg`1sR%}9+788-y-X%ENQl(krDl?WS->NSr<&}kKXLml$0=y
zSZcfBayt=GQ9?|PLI8Ezk@H$wS~1<VSemz&K9XTl0`@2=wVK9d{;oaN9d>;i&%u^<
zRCKgj*^vG3v8JuHkLOeG-X+fO7m?@w`&n%8Sv}Au$^(}#KeK9J7SKv?9^N0ePo5Yw
zwdt(vQ;vtdOMr2Dn;Cvy`OjfgI<Wu6M5X%=RaSb%<QFKcX>m!6b<=aj{=Hj;R$hL7
zFJHznEDu~bpLgMjeqmu@>*I6T<$k}6roVrWOVUOc2kpAuNjhNBQtyEx8+uPz-w)+Y
z&CT2x0|NsL4-XHf^+Ung<#q~5+llt*Oqr{<QyQ@?reXFn7~Y@e_2Yv7zT!fBax%?>
z>p$2OSGonD7CY0$ObrYT`Tn<%bagWru4p~sj)9^J%ji43&(om#vbpdj_U&7}ne`iu
z+q1=(u+~;py<BB0V`IG2gd;}+15uE7(+~1eQ=9Ong5sI~&4`5<EWetuF=KmsdrE3*
zcz3r>e=7RIXp9WMo%T7sz4`l_*UoYmPge(x3T3+PV7WW&4}Lt|NwF+rYMUL>oQ8Wb
zIUoG|DmnS(|BV&H27!z>unWIT-z1JCfUQ9ujZeRTZi}II*HJBtl!5}GW5<rMpCpNs
zzisn`t+yeBdHcrw|8sx2&ieRBEAOpc$WgpbFPwJfUscYk%0-kD?M={x(sG3tYF;Xy
zr<?uCm2}r5<;-<WG4$qihQ9Kl`--pMys7Y>(h@v!#D3;|-qQLy6BqP!98MmS#4CvZ
zCN$DSBmbFLU3B6L+5cX358BPE1hzvQ92|_y%#+!k&29Hix_Yh8UhMBjuPnSjMXLUH
z!$1p?)!|Ql==<{>{JynM<sboymrqQL;`8UvGouZsh^g2G<@wS3R$bs+(fD^au>&zj
z8Ohgvf8&vCTBEP``$d@6b2wdB=-=I}vgiE22lfE9*`-T%Ycu^b{rM#G3k#ZEXrtn<
zB?yql6@DHVobzgWZ01M#_ge7Oy`7yzu(Ulp<P;PZr>Xvasy57a!^&F23lTi$L!&#c
zTi8lxVgJ4hY;1)m`uE$R{wh31<?r0Rdn@Vh1-i37VHLtV*%e=l(Dev^UC$v>+DK~Z
z%JZ)faJaX+u&Wu#iS%=o6Q4ah_Oi1x&!!~=jd-p^v^*oMXknfBQgz0E(-L9iLdC7T
z*#9lpDgz5kEQ?Jr-F=0Cn11J%6CJV3Jx6lKmnJ$2%gR`|lK;M*7T`zW=dIINRYEfI
z*S-e%WsLvZb=~vB!^5_3Gmi%@*8Lp+`7`$3J$AQm&0J^i9t&@?rTyRQN%fD9P%|<y
z*$osL8Wb6;&$2Kw+P=imwa>D^FSnK^+GlM5K6J!TMH&?`8nx8cvMw(#!!Mg|FllOP
zF0HOM%r!HUiv<oJ=55(k^f`N;@9V`c<YaNU5f%;wSF<nXYNgJ5Wp+xjTJ}SmM-4=t
z@nwjaHz{K*#dnFgORpJ-IH=G5v=WVOs5~?+eSBS#Dc$y<O=IR$Np{LM7VDt=b#(N&
z|Ew|UZD-D$Sz2BWw|w{UFF5pLXkn{9L`_W{c(AugM@!55XumsEpu)B-BFKfdDk>`K
z@G;9;+UDkF^;sGoWzKkJ@rat5nv`w$E*+Dd@e{v(9eor6gM*WkbMg9tYM1kLMs8R1
z2g<0rmumc}FKp%c@f+WLF3AeNWhGbX<E;&9sB&?yODA(7VR)f|H@rwodBpUaNy$NR
ziylv6w5JN8XJ62P+f{vyBOwu+SI$L0iO)~j7I12&aCLR%tBS*QD-w~A2=MZT<&&W4
zSP01R3qgSq8d^~6q$NDe%)-w?92^AI)zzau(I5z9lV`vyEp22LX_ja3b!bMDMa<Le
z=z_tefJ4ztbxY;To^5_u?1F~(ccDit+}Nw<=pIS4?MB=j=a?RR$~K!we8<>I>qsaI
zYt!tw0~>MCqjl|{VoW~Rm>ay^->L>(4~3~qXk)`p5h_~LsO=oAJGMI!@$&Pp3Z?b%
zHuTQ?Nl2hPOhlwsVp8KHii(mx4&aK534PpQx>6-OsZVW}1e^v0=nO5qn;x^WMUxmB
zXAPB?Y^|K+yX0VWI_KmAIqz)|%cd~tj%&lOW67`5Lyt53l9HR7OvNGF`r!jTjnW!z
z{k4S~U$p-2d;eob{RFtespSe9O=#<k(n%@2{Hqu9%i8;{XLFn+AsD~)keJV_^32xe
z&`_JRLH?NWEguc~1mXMTim-%&cP%Aiq4_^8F0SyOGYMYqWa)nWn*Lu4Gbi{*55!y5
z^VEXh+b>m6xWCwY{oP0{&O=nPB%)N1-pH)Owm!zEgPQf6zxoMi38}25rKN$hMNg!>
zC$VyJa^tgu#b%XGRLy^Wpl)~T-{2C*T}VL9yNWhikEmqRtM<|kpLi&XApOy&VlLP&
z!UWM_(U?5F0|pOkNnyJCaO-E95ki>OVBlkt5qjL?pziN~@>xkq^nj<not+(mh%PRo
z8!(b=LjDk)&5-C^?NgySb?W%s#$QPC2EkG&py)rtvF&H3!gE}I;qzru&d8Ksq>N?M
z)e$8B3u70vjYT^eJ99gX^6891-GXfj2VMB8EF$!THIliNXc-tH_jWd?Mygy9(jhZ8
zGK!d-yj*B3ivA3`z<&=64Nct-7*m749W^b@N~v@9Y=<__O~FUXbSHUG-_c|{w=Poj
zI<`2y2Emw;R}Ru@{8tr4&*#5TdlAH=7I#lL4R<<E7mK|NEoNfpcZ)VaMNu)a?5|hG
z8m|LEaQ*H;qs33ak5}jz8PPH$E_yLl^7XQvT{yz74$^T+>R5+Q4$^F&%4|a4>{OVS
zyn1@E*lU0HI$z!F>}+|^=3;{^>Cbof<6HLE%M3&q-#g2|$dAs>=4fncqGM)`0Th)F
z9XI98XFM-`^uac+uoHme?wu;ai~f$R+$)05E^Z*WUUjQdsHoP&3b*)R*`P=KZ`*G_
zud)~AJ@qQDx_TpNr}u$U!>9~tk>lRR<tH`;j~}<cdq=INrdE;z6TcCvgI7+D_0pwF
zRA1nFwG!E*TXw%onl)T2NwEZ81q7~cE+j54E)&?2pFEeSsHlOt`7Ky+%f>*upMAMq
zu(DBXC!EH=c(!N`m0DTwou=DkhMFC?EJ#mg`K&R;!aN4<^vry}jh`Stp4rsgD(Wr1
zL#hqOakA5^YJ-mJS97MG)&T1XTfW3FY<t-)ES9VDtYlLZpwqY6r)Z#Wd-?c8%%5`I
zJX2s47#K*ZtFxeX=@MJD=QnBqvSDGEmt9>QgM%0Sj`8EE&lV5Ly3i)EpCrF8?@ymO
zv)}#Q2mUzAu-Der@hK`c_V2;h8E(9O0xmMbudz-FS9p%h4@ZpI_1R;D=F@f+hcmz1
zR;azwnQI$Cr>0-AP<omab8-D>VW~=zPlrUNUhp?z{#jDkuc>Kik!q2|6D2ic>NOst
z<8yNql9FXwp(7v3wCOeS`)$Wt!+{KU_w=-7*o4HL%a+Z<<JJ9SV7`|CLd?p_djHEW
zv3du{Ck%GX`zPml{1(4n!ZtDsr0U2OoJ>_kvpxQKjbC(AwAk#df35K~eE>&zW`Ddb
zkLa97tH}b+xu&{mY`n?nO*>yW(b{A6I~><K9G<cueh_lx?Af#Mz4J`Vtnv8<3P&>D
z^cR~-Hm-d*``)E0MMn_rjU&0i5lumx_|j5+KP+5XL47?t5ZKml2m1lfY+AUxl6hJj
zd4MLre{Vp^E~BHW&Oh4?tOL)?G9Kv5X!owluk8%1nSRJ6YKc2es-mfRu>ph*A2v=a
zJ+Zt@VbN)z^Dt)2xK)^s&XDBz<(H(#EUKUN%Q#|bUPCYKh^6VGkZcTtHwPFyKia^;
z#>R#?P!)w~sbw3&+kofiN2(}53qXZXRkbd@qXnGpYpr*U6~|rNj9OLS2~&5=kPzk9
zuV2%0b5rN{<MOqW+sDSZq0r@grbz+n-V3fgGF#ZVpzihDmZ$V0%cRyhviV~<n%ev@
z8cuGDKC;#)#9$lN!%Ifac;UZkOH@Rb1dF_;{POVB_x!KrB(f_6DLM`Tp;B;VXL5sA
z$Vx5iBFf99oU6<Q1qEgPG^}M^c@UnSo*ou8z<pX?{_=-MXS(0K(Qt8biOt?l!WGt8
zwS;hG%et$3czASFKYYjwuRKglEc=HwTD{+V8Q1+5;CBKV0$vm9Xm|Z?+T_NJS9Ecs
zzW2~TgqD%<Fa-sLUbZY1?1{kbUr+ebeQ1yP2m2HeDsN24g)SWSDJ{q^KJo<1i%%+e
zs7T_!UgcG_qLy;u*^FGTx<plMN$_y!y{HUByI~5~KNuas-9>fm05=meGadjEpmq>Z
zboBK-d|H;Lx+>LlHha6eNR*WqJQp1K@t>ToUX6>0AQVa0g&JiuFJVqfO2XUIrDk0Q
z?sz9Xolio78nzbjd$-}m@$oQ_a2hv8TrpvVg(A-^YN#u(^dZ7~ZSm_-QE9_sa=(Od
zT@QlLSe<@D9PXYKOwMEJvgH>P{0pzzc*v4@CJElrF~-Qh^A+HiJ|)?_#cyO}Bo(-1
z3%flZrc25jPbH!8`L~~y3$Ndo^CrxYw0RXqaOh#bGwOo$x&=mdE7Lt*zP^!gsQ@nD
za&koVzWnrwIcQ_P4X+vRHQq|lxVQ0TX=$nTXKrXxZpb#7{i}pFXyJ$!5m7Mn%E&Ns
zobq7g$}SunF65%kTjY}`rZbH9uGwJX%G2yFpf10V5W#4|_4Lmftb@Lbt80btPkpSW
zrd?MO7k(;*?Kpela8uAJpeL#Dj><Q;W!^r=Vh8B5^n|CUr+I~hhyWyOVX-g8ceb{-
zt3#yt@xrJG6crUw(OWhLvoSF-VZw?oJW($(5l~d*Z0+T7$_?E=?dIl&)e-`1Q-4L8
zY)D_2-=%NB<>dVQywtUMvwp+ww>kwzipwt<baZvquU(V9wlES6JK{3XI1tPxJGHR`
zb_0Wh6Clxn5YjeSY)XK+(bU`nnt`E38?hY{7~HmgGl`izeERgM9ed*9XhZxh&41j+
zX?7x(=M`+srB@uD0EdkHIHd*Ny?YlpU(w|?(M}{NB$QWJNct`GK<Rc`+Ht=bgHkIU
z?7~PQ$0@JrZn{SW1u&VIF}C*hk?^+pAT~sNq@x>U<A^E=^POj)qw_x|l%Bzt{(8ds
zm$3XTDXlc>s#hn<Q(i62hC+L?g}XXj>n*Umw}+YuRKux17XC})ZI%pi=20yzDi4Ad
zPa_`L5b#?|O^x9B^XI5pT`PtU&Aqw7MoD}K)B5HO1ylu~mXscczn{(p(?Ae==DXvt
zxt!Y3sU8x37aPAAcFyKnCV7pM#D{kY^zZ581;`9NhKVAJSl1uxn)TeM1TA{t)wx{d
zCcD>_p}SK&TU{#q{nH^C5>_rUV$_Ddw{;6}iHWdohoN@gmsp%a1dBTplpny$30eX-
z)@BH3XlT%H7b(!toK3PG3#TZ^nwcNtX?tgdD!E=#Tk<!yB##^}!0Uy~H=Z!|OQv_&
zuBl#8q){alyCf-#SWEflSP+e)qhr8VZT{5O&d!2#`qa}GMAK)!)ckh;T~=21bB2T8
z<ph)>sEy;}<6duWLb1~UAS=9%)Z>zOy(2U%x+hXE+-rj>s1if}+a`pYyolbs%+6_)
z_JWR;ofyw<uCSe~V2Y~1+|^HIWKb-!05SXNaEoU=JS0N4CP*l`6<ML!@l&*^Yz@JG
zq~Uve%?=zrOhRH04G}?>lVS*i4PGDK(hwhdsBU^|^7R|9ib+N-?7+sw8IN&nn>9u?
zL;p~o;RtE(M`n>n9_6VtT6K0uToQj`bZmuwA4Md3R0~ukfLF<ljfykt+rpcFK!>Kf
za^<IqzSi175|=^@>|EQu?RDs9uRtrA`2HP{@2t4(?Hfp?gHG4o*GJ;z8`T8pyWcqq
z)WVz;T;ZR;i=i)?+Ww7?{z?3W6yF6wbYp(7ozr?``IS7<kjArtc565`E@{9?q+k)Q
zO});J#U=4lm>RiQoe5wN{_R9i2{=z&dDl3l%$`W>@9$r72UmD?>Q!Q7ZZ6Np{O8Cw
z8ItyYR=xuoZXIunwDa_o1P#mI06H*IjsZo(4+g``%slg<fMRlT($>{gOb+(usy#sH
zsHFg0d)A~tx1>t4xyH!>pR=fbJm&P4>F#{83@>PBx00u#=X%m>sH<cvz=QMY@aPsa
zfjdHH$gsar)6&-7P6!te+tj2KL&KBw=#gM;LxcVI*Xdi6@sdmHLpJ%Kb>Z`GXQk;D
z%#2it@e2re&%Ed3R1RiEV#4RjD|v%1#~r$9^{?t<d-eHKc~6}J*3IN~?OKdUv8J9L
zB_nzCj58K19s|6KzTN05RRBjQ<DH+pH8Ko#65gl?M(-xyWGywe#@=z26mb;o_Z|@}
zc)>8wWRH|&mq=(jh;BW}4Pggm(1VpNH9I?=Pa_6MG7%ExKm<E}W6jyjd}UIJ{ycPg
zz*)e(IZ&U27ff!i&jC9yjAxVc;swA>3oKinoX)U#{80~9P!CCO<LXz1^@P)q09A9P
z;6-O^7@60PJKX^<fA9j=dt_3aXCN{@J$)EcUR@niTr6gE;R46Tua8e?;O)HThmRW`
zkGtmp9jwB>>+<&6FI;jm3(T023Q#L{cUSv{YdmezHiNETuQ;uPb&3WpM~C&47*CXZ
zKx~bbNEsn&-TW?m(pi)Q1O&q^p`jhswY74)zYq2$mA7XEkR${%1cOmfP=HTb>>eH6
ztKLk{$mn?W>M_$tTEigNR}li*N#h_-zMoqM3W1A{kBpA~yYA{0YQxkZxL=0y-}>`2
zH+t}3vEldPmSbbVTp0{v03hnKrWY?x0Kz+EX=N2XP*z&{3K-SI<Rt&I5a=TDoEC6f
zN1D#Zb7tcTOP-c|vfKQ0$ji^0)~MxiJ$X>*wvC)z)w`d!qN~WRm8wqs$!3W#zBsq^
z`<(W~tl$qpYyT4JZ$eyAl}#ln+b8SmnRhIO`0LZIetbMpHDuiZA9eid2SLT{nfxt~
z87C$t!T}G#VBmv^3vL5GpMEq0=7L?%Z6Q(`ny!axk;Bu%{FtyFhxR|KKatqJ`m^tp
z+vetGtVwY+bn;99Snjrw<y?iu#Q?dUmP^)ERaNPnI~QG2lF4vY{yue4QBiv-@xU)X
zz=OD@?N6URsn5=>-<a`}M<NHde`rXJ??Uz4in%p8Uuh1LPxqKl?~l&SrKG14pDn(7
zV3{iRD054f#o6tSh{S7+kq2+y53+epZ%LL{Ft0{7>U*~KU-)U~c(hbm#WYSQD(l3R
z%b-F*tschc-IF2v_3PKxQU?tZtdATya?JdNj+ZWeCnGnPCJq|`dIiYV#7Z~TRKYk1
zT*>WzTLStukTMJ=T0I<$7=Vg)EaMDxRhn+Jf3>5n@kkn4Iy$~b%^(~wA2SOBB+9qm
z)Yw=K;1M(=@Tz|Oilb54r3D16J_~v|M?8xpvn<Irm}?yR{IqmghthhYB3pJP8Ks%(
zgI;L}d5&GZ@@(PRQhWCKj#J&mmy(0ZUae&&{VHprkZk$%$MWOgPdq+ErJia-t0j+S
zYJU9wmXx<*#BOK8+iz9Bse+eQ(nyyWPn>956ga=?n~sYD;#0HxI7cN^dF<w<)tx(c
zT6=pFKki>>VC3SW^7QmX?iny3xxgh}aDJq0TQPpWJ`hP~=>ZF&=M@nl1)*+ctVscs
zme;?qakzV}Fqr7*`a@4!_%I4x{A2J&TvPiOt+BL9rmec;UWoA!Pgg5h6BZD?^jU9j
z@1>26SI`JJ+%w*MZ@-(y>mqUO+BH^{5Ji}};n4LE#<r+-Q`6NYf2^0)zD^b=Uk+eF
z_=-|I%xsw9yy?0`k6@efY*C-Vx@%o`Kk6}X2bZLVb(+7i&g8&6YPovqovU+YGPjJW
zOP0zrhf_m2T&LZoN<|nW0LEmozzXN&a6S+$pzZ^25oo!LuwEvHgoH$dk=*W$Kds@L
z{<(ET!9Xt!FKgLnCyt!BdlPuL$|eAU2&4d?s|^&O=wYd8AAC)+x7QU1`mIy9?SR0a
z59fhG^LD-Sssu1vq+$4mkRa2*P7X&{MgH^YI%%~-K1{QdU^OZ3Z%vss-8kvv;}a!r
z1KM!l^SkDnfq`<M*BM;8)KlaeC!eyi;`8Y7<H_L)`@+&vVhr@33D~4w{!Eoo1ayW$
zuMi|4^Dt3PIpa`klohKxN@Z3wk5>AG?{ixfun_I?eXuWrZ2G<B1*?vczF}z&KO>nO
zk(hI24w&J`TJI^Z-7RmRJN(MZ+>s<S9zOT~v_+C=l$fX(P(`+3?(4_ix<w8GH6kJ`
zlGZO_n|s46kbYWhTDISGz91Ynf6U3^hbHkH%Vk4z(Swf_z&bvsuRw}{dldma`i`!n
z*b{y{Nh*vEpp{c=zdlN_Ct@(oFV(02PZj{s7HlQ}vZ@k4k9xpg?ic<20BQ6fs4s7P
z@f47kuRp0vZ>W%er+C7VCBmjOS4;mY@#R3D2RL&>wLe7DcOGjXhP&z2{b)D*vyoAC
zZfj5<RYZ1tIuetdzJI=ws|k=B!+Z278A!k976X2L5DL6;V{CKz+x4GZZ|^JA^y~nB
zm-79g7rZ$dK=U=&?G}8utO0iek(a?9*?PjDU{MsEqbYs-m>9l%47ff(Y-l++-lUdb
z-7VqGrB9sTOV@?XKu}^r0b8lR0yOz1lRwNiSEkR*e{134kacSXaK8+FpFI0R6mFyx
z)8z~ThfK&(rD&ZY>y$H&FosjJL`o!p_13M&+>fptm*L872QNMx*FE0|a8g7}tYJzS
zR8z22Y5aBr#;%k2{&-tAMcSU3$@<K?iQTH<+RcaTEkbF!74wy+%PZ)7`+G*#zhqZU
zoGHl%5drvNEkMAK-DT}Zj~?;DoevKW2WsE%?l<>|3Vfo`hJcoZZd+d?YisHzYsvYW
zZ=7Y+pL`@MI*08o9+ar`_4n^g;nOI;vF?IaDpT>gv{UcBXfOyq-BQK6=YVB5Kk%;B
z&?4fE<L9Lc8Ya*@-eMnFNP?Ex-v7bMj}e>{=HuRzFPY2uGTy+Syng5LA3tJ%^J#4!
z5f;9@`lGk^6R+S6TDaE0A6ZvYva;fUlc6d+*dA)3Wo5+y?C|0`!LJ4d;tfa^0Pfb^
zooCOU-TUfM3jY9Nz63faMN#O~E2_2G!LvH<{4UldpLqQO0=hu;lM32!gL}?`-mcu|
zyYAY;7nT!FZ(Bk0w{z?)dW?IEm~Ybe-|G5T<q^_+nj^!F7s#w4yiXvXMi$}{jmnD4
z==$~3l*Aoen}$WkpcabX1+BmJrQ!5|7FJ7(AI2tRApl3HyvtQdhMt3F4kQ!sQ*>EG
zpoReT<_ETU{rYtbMny&ConinB*dh?d7@jjWH-BA}6(^4~$t_fY_i_F7tPQ4-ROo)N
z+i*Dn;L$MMf_AtIh*?r{${}Z>1YDpIjAw&E&atn_#TWmGl!-?tj;MjY$M0g7zOJ=v
zSsF<fb~sEjI|_^Kh4P_>Sr#zb_xUqhUS8hTYVJV-%qSBbEx{y932hio$iD{(Y7(q<
zux1#z;tM;9i;C3D&DoPUPk)fwtI{n{cXK<Ilaq7!38(=)mrYiPVf=-Aw(KzjZ$H|+
zw_$P2rkOMQl;<mOm<ce|)zzS(yJTEY2nLr0I6(8l9u#kQV&XY5Aj>L@(n9hLo3RBS
zh(C!2A=$Z*$*kuQ?!8EgVL?~wEqNW@@MQDhA!85}ukLOwZ_QWs8CY8<h3swfNlDQ{
z>4Nzu7rZ3}9vQUsu&-Y&p~8^?6}Z}a>Nu=je}9~L4w#<w^z>2yl|>jE4dM!g9332N
zKR!MOhz;YRzrz7e8H{6%0wXegZLEWJRPJ?eFR#{k$z~if&|o$WHv~v-{Qi~<j$>TX
zMr+7kp8w=avbFio3>brd!Z&m>FFId*!%{#3+PU(z&u-g$eejFOFUXz|S>v(0tDC8(
z>Qu$8E$WwoS#nB)q3{_r)Q7M=p*ht&oATRQvI7CD6_lQ0lM=`lut?dpA3kPQj$t^W
zkJZylcsPJ9nY-582Hz1gKBJCzf(=CcPj^KM78ou3skkIfti#=r0f#o=ewn9zg+G;8
zK!j)a+UXm#7<_(yKM0#{pPv)uYbCZ98s-C?iHEmK0Ff4C8W{JpVMtiMa6v#`E4$)y
zCpp?A09+s;a46)@ca~#T^^6cipxRN~7`8_^is-Q*U8%9<%bxbV;p|H4su-7k@0pe;
z*E_g!<JQBx1?sV9>lMn<fp~9996wG6OWsfb8ZJaLUV@sROryMof%jY5-0Zr=BxYb^
zlMF?4M%5f&oe!!kDkAtb>N)^ga7ow!iCua1rJ9`k_m8`NI)-Ew@H(7)Q*b&DO<Za3
zeaUWuOcJ2xjs=)u`eM3J^1z77nl{+${pW}Q!h`ty^5sigqzdJeRl)aD{4<@_-Q67q
zY6Xbu^b8Ekz!|0x9LKaeNZYN?4z_DQ^MnP1U=^!kpj3j^l=vnK3yT0~^?;T)M=|?9
zvOUMYe~$)Pxvp-qI!InaJjrr|7(?pYuP$E0n1c%uQ_sh4f!f#eU49w~6>AOfl_$Uo
z0yDC1K^dBNbepZSnS=m8KOv|$n=d^047Hu4v%q0F1oj^I-4jrWt#`#O*sTI}um@!?
zUNC;CcJDya8IVQc&5Eij_wUA_XWcQ$s=3;G?_|oG-Yd5Xp{WZ<O42}*#jMJO31bV1
z5)nzszpc=1SV~SofkOWbmP-u`e|YPdBM$-z8HDv(!SJXE+Vn`^JWU@<&4t8ttTy9i
z!2c<_1>Nu7c?7ojgOUe+GIS!39XUW&r^W&g`P4NIe)@7oM<)!3H6nkZ`>VN`vhL>j
zJ2p4hFFIVgf-))L(b1hsJe#i)IRf|Ahaip;0o>wb!2Gp2vwi&m)~DmnzrqbN7V%T>
z-LWorqzJl8PPS@&dMw<?%kPpQ!%l_dIJmamKYgJdKDnTy5n}x0tvX~uTZc(qqpO9l
z)|JRWehs4Qh~K|>Ap>xVAJT9)cQ&F;ioe%KV=)qzwPAPeFd|3;L?4rK|Nf}Q6&$oP
zl9CTqt4mAHiPG+`K#DW8v`he!o&WG527}aWjEIPcs+HCK?m+B<Zb=^dkn61pj|DM#
z-RuEinSigkVR&J}hHOH)fE9Jkf3B>6IaBFQ`{TzCJlw`aiC3(^gNdm*1G~D|619-M
zoUFK_b5`4pH&tipdxj1ATY+;WHBuqFTQS8)A%+)!vDiedcyKbB#k#g^N?2IfXFLB}
zx42R?4BC{l=`#~bltyG6=iM*e(TcOBATl;uy%!yUD=dIcV*B~IW8k0f8OH+_ziOTM
zioAQz#${sAR{)HYU{Ea-oXvPR=%T49>~aAnsFcfq2y~M9;fh1hRfp?*j{yCb0@<*C
z85aeuQs?}6ZNTWa;^PT0K$w9af_LE+5WoU@4Q}yeaLjrYo{T{lm!vwtH!xdV{%77t
zhqrBJ>F2|gBCNKHucDHY=jV-Bm0bcDgC1KxAuXuc7#M(a6c)t3S28pQfRP4Jlup4Z
z4hiV(FeU*vqHQua0{F{1+PbXedsj!5>yQ`Bcu;E9h0{h?eWh3_%WG;{e(wE_1uqmF
z!fBuDX=MkYuCrBFbpc;n2Q8d@o2$a}yK$8ok$SK<Al3ji%L%nP$i;JDf#j6iLT8tz
z)6Sy^O5M*ce~0vX4ms;`63qL8j~M4WIUgP?*+Qo&{1d#^uZ^^<E}qZ?BsGF|6b1?o
zpnCSlbbGVKWwwrvMB_j1ENJD^0h6R<%P#B%64Yj&tP}%)0IGrn+8?CGbTUDh-`$!@
z#$dqVM@Ve6-aisvm6HI=NhEpS(+Nqp5YQ`W*st@RikT_Ix<4?_3r{N`nLS55ZKWb6
zuivmd*Zm!E*v<`06BFit3KjB)hChF{gNz^R@ymx{ShH^wmzBkW`x!;Xumrfc4Wf}D
zAt6A^HSO&9fOHWiYDGe33Dm(3a3z7EdKk}*27HS)DtZYMgpP%U1ezWiSKye|`u(Cu
z-?o~Bn7AF>GHBDybH0J#mggCZdhc!f&jZQTEdb%4{zelw06R#205;M}OTIass{PQU
znBZ%$<iUEy>C>mFc7nk@0dgJ%ZKJWV@qN*SBW|Ph>kq868-Te}gS5O1mpkz`>jauM
z-ap>5p?nqOOPlAWLLFt;QJM=)ObX^>wQ$q?x>f=@YzI2t$g(zXOdzF=u1a1^OpH_M
z1|!eSH9B~|Q~;JG7BzT?)kP9hwL@fZ88C(HJP#vd6l@26NUhqs;H$-Wc`<#8t9tqU
zwb&o8hK5*<A3x4|%7YWZUU*F<s7V7!UWe^Ytgfy`VtqD?M=y#dMe*0O5G3`Bl-5V(
zd`Khpo9u~GC@L=QfUuyH{VPg{wod(elH2^V>&N%+d9ac|@m`6QHAaBq<N$!AAFl~6
zB<NfoFIx1UXaMMf2wTL_3y&XyVh+JWx$7%}pPpTc1Q!CYoy-m0_!ab;M3uc?=OO<T
z(7|{S)*Jo+UU&&1rsFra{hpauk?(9QbijQ_nX20Q`pdsQ6o47g4o&#D<7-C9&yC*x
z-a^GDeHci&*$ZbnG+iMu)9oM)f54l7nBS>q(MuNgT}hhG&LThha+6WIEk}-yKJHQ8
zY56PfIFUOH12XfZD>H!e3YS4qG(N6gy*l%?Zk5X>yP=}IycPDB6wDpPo%u@iMF<Za
zO3lm+JRcg{au4n_HK$w{wCLs7&!N9&Pm}(NCV4;Tn*zv5!SUjpomfWux$D34Z^hwe
zgwu7SFI76(!9rldAjxQEZZ2E*yWFnB`~<sUtdW@+4s=I6jDVWGQ(6Po5a-O1_b1=q
z-#@p{&7*MbSr{6WoWs0sP8Agun@TRRIFGV4cZ{*mS%|fE3_cB(yDATo@(<K7ctW=L
zN^M%GVaRQ51^#tQw}S#sK*rG6*w_}p0U|<S*&l^AX9p?B$jHtq3m-W`0-+0`$LDj$
z%Fh)UPi`$w;Sv(uhJP$BCILiXgTdY4NR<szQt-|~4JHhfI6hffW>CSwQdQT|A_n&e
zg%5Xsoeyn?lu?D}H)UANmFhlwh^DbB-H=f}*!6{dI@Po}*6?#^9ArD~n>UrvZJC*w
z(Ms-xssQjf@%wi&wDg!FQ}_~gl#pLSBMD&8;?~I5c(B$kw?~mtb#1S_k>t=iITjC|
z(B$6cw=PtU;Ew?;-EN)@a4^xeSX-&vJ5U%<W(1k-`BQ3X%uTmJFGs+yGvO%rHQk+x
z<m&fLK>+~Y<#QkmJPw~P=RJ8C#!PH&ZEcWwb9kY6(X(eS`}4KE*Jcbs5k)r$X5qsb
zGq@WGTMIypWs0!G;6dgbRgQp|f*Y-D9)N?u$+#m3V6UzDA?LZ$skgoH`YJ6uJNftz
zmYv6s>k>e1g-B`)6$dRd^QZTmyMvH#qvPaEz~k{2++r|Ry9JVEMt;oAeOdYjrm!4P
z0*$oUmG5tAIPGByBzn*Ql|sRI*fAU@Uxio=O+iX8;Kot&x5+B|xcYkew>hU7U>Ffc
z=J90!k<fpj_}Tg3v2}J3^dFuX?<;)|Iza}fqo<ERArcxtN__RZi+qtS=1CM)#GJ>9
zHEk?TS`G^e^4EI~4i2JRJ(xe5+I+JqlT%ZvckV<{ut*?{0)$ZnF+q>UY75jCgTsr+
z`iT<^>+9<njH#(9tR<;FxRkw#C`Prq%Z2xPyhSRJ?wDUW!B;rVj2Ean3IwGSTow)w
zq%Xrxcva{7ll{{d1#pMbOt%Z3Jn4Ye4+cK#Nf+a^uFaJdm*;A%TwJ5h?&{-CzOK+5
z%xb;(e0+U5MmOhZ;tw|7YtWpm=VD@lkw!Cn@MC6%6jNAKbXD5^tb*?kO3Zk=sdXC^
zjnA7sP>bCB8j6N~Sfj8493$ugu>_gBiU+%^R02nTKf>b)p^E_8ZH0{IqU!*kVGw_+
z&ig@KGKR3P=7*^74B~T3*x$MR&3SQ5f<6Gvk5@r~0~nIX(W7kv-)_cEPdlKL$(uK4
z{jhioaJ%h5T&H6Bp({uq6%(UgJlNjXciu}L2r!vCdlU@dFc4<7k~rgl=*BR#V6o@=
zMcl`o7UJ=@Y*4s4;(h}zY2fj>Oeq3HyMVNMFHdR)1_fyp5ArK0?A*QC)YO!kkwJoi
zjDyCLk0<L!zkNHn>yFkIwiCdo<IviX5r6mYUDSrty0#W>JZmYPxY4kH-<7|-RFGlq
z;xlVfWIQr+?Lt}aC~z5Dh`a)1=>lnQxs8O!c6qWh5K;zGo@0#)%GUt5$pvj3V<gwi
zPmOO$(t>=203jja5~yNRkl<|m?Y3Z8ut@yO_!{2cfC9oJGfo9jJ%@MkJLl8h>r)*e
zLTn#9Pz{uj55^eu;o-_FG!j<zQNT5_dINl1Y@jt5n3xb(nI{epn?b6yZEVaMIc6?u
zuxJooM}!XsKS`Ee3=FT!5blJ6K|y6K1XE-sz6w!~UCUc~4(?Wg)z<piU`VNAR3U6A
zB~{*Y6Vz`|Y&o=g_D+H10Us6g^pc@rI5Z*x;7ov>Jcjqbxvo~3g5`n<gX01yCn|1a
z|Kn{Iy34Tr!2gIqG8c4!PG}~_udhsN&B^cIMYl4U`{tW#KzmWP4a@`Rt}@w2?ck6`
z`P8n9oU1%2f~4IAIZ}D~sjxC0H+==|9=r?NbF2#NHlZRx9xbjgF%=v&u&JO)crA^q
zJ=ny{o)u2B17j2gTR@RP8Cdi1loL)+#@Y~Zy7qf45tM3h)F%>8jkR)kHiiRejd7aL
zpi%OBQ%K1#eI8t4%g}OUFaSL9cu@iwU1;`^5G4bBcku=#tTo7WoQi?8$Xfi?e3N6*
z;<*EvuC6W^@V-M+h>n0gqO7c(SY2jTG0B~q7zkHlVq#L0P(43X#)BF64Bg>?p23&4
z7s@aRwF4eYWJANwp&>}LC?TeA-E?CD@DdTm%)-JRiV8yR^vH#WPhG0cjRWJe0~BD7
z&(>B}RDk$E_(gUjO-`B8@hm-^x))U2H@NhK%TEu7pAn>gc}dJJf{bZh*sN#kekyLp
z`jUkOsz0|EAu@DT$6yXa;wmml&FQVP4D1vR5WYyMs5C7tzl!ewR3z1bR)r^3aCHKa
z0l9O)D==X&g#@72eG6WtQr_RP58hqwLU<CWfu<%Hj;nBc6PK2}61bJv0gnt<yK^Ab
z67=ZHaG{&Qshu_l+olPu#|dHFq9`1kIjOkP#~loO62V9y2y@3_y=~py6BoYJOhI>2
z0}5LXan5-#;5U_fO(4*PgrHr(lYyXRf)KfcjupT_($Udrg;nft2=G+g{OSuSYgHJ?
z&0CWcAZa!B5RW!(`MVg_Std!g>^|@Jdx~%t0${MKCf-Q4yi8y_ac{2Zn%O0xE(0Tf
zHqknv`XDMgu2}QoUg13QbzxOLCIzPBpyE%?&nF;)jJK!l5B(pMnEBSCqOPH#*_B8A
z7~mdEBvs&A2z~tRx6e^d{fFk3<2uX`q((vLE(yNjDLqF>T%$Nr+b@uwtIwnf+kwim
zHN|r<TrLiZf|$uOBys-!BL^#n><4u0;?%cbhSuI+4<RQQkMcFAPSy19)5lHU{gI!Z
z&IE@s;z9PGno3rQfm8-)BeI@j2?U3TWkwE$goizoy&D!<#+rv>o*0WZIgYkJPU;_6
zMPk#t2TJiTUt%HhlE85ag4NV!H6C2nH+~5{v#g_st8$h51my=%VN-#7Reyi&i3&R1
z0)C4;Y($T`12`I@;pZna7QA{6@a1yzr-i-sleA4S_O5l-7ZToz#bPkt5U+fbZEcz+
z&s>Bh+#p7~S?MgCAp#zQ4w1nz4@xfiI|6kx0S(s{kZKtL2V(L|OQ|uoSFQjYkppi_
z^Y471h>{^ae_uWt*uZ7L$jJTz4G7WW&EGKkmy@9$R1Ii{*aZ0Nvp2tm^89W&PzLgg
z<^wvV*48!=o^<m1iUY#JkRgrKEr<XbgBb^>*V8PLrnXKZyx-Yvg%uVamDcw+84&qk
ze_R62TH;0!)E;sL##Xjq$l1QWbF{|o_fXk4@vho5;JRba#X&EF6c<&E5dE37K6^0M
z<si58ys~lvWIrH`m@vRSfMY5gdg$Q@9W&4^P=pgDoa#Qh`YMQ-mNrhU<$}sfrSfAj
z1B92GO&$wbZO*DFO)0}hH^X`;fJ5c)IfIb|z#2dp+diypu=FAQ6a=!Z+=D;4F~3!?
z1M<X_tcan)2BHB*z6BJQmWwM9Hh<fvPbPkfF0pGjl+gJtuy8S9Wo42t%*9~~*5-iq
zc9Ql3FQ>vaw~lRn<hpSqEXpoJ%ATl&<>Ys{LaB9V<;}Z}FA^`^w6sPO<KrnTEiFSf
zy&oA}%|Gg1?J;@_zQ;JEoSMs~{Gxv4c?RU&0tH%6PfzSgUc>22wKPoiM#F`iv{Pf}
z{W$60I!jMr9l3VJ8OzgF50lbml{vn-JJEaU+e=V$fSjVpC8|u62>BkghK`Wbd<)@#
zLr5rEQUWmqkXnEN67OvOMoPzrj~~6a*RCR&7eGd&dLAuT+6Wyu!5*Xac_7?^-whT%
z_&0Qh9D<M3<U)1>gF$u{Id<$N08)@qBOy8q*gHoBdk@7ZAZ10-(EbEo2`z)kTt!ya
z#CZ>fLf(2l9v?MW_VlSbu<{P)=he2J@8nU2ME!A8Y+}od99R2u$2Xw%F(DHX1Q9hz
ze4|Z?<Z01#a@0&TwFT?N;cyGtqZD0-nTh+DEaNT8*#N(F-J=iK|J*L$Tpt;cS3FhS
zQodb3;_{L*gysWtW2=4l%dyPe6_H0wW3BBh&071>VpN+KhX(S#Jj#f+vNf_QcJ<9I
z#ts~J|5{7er@V6v!o=J+X81r!MKm8sY$r^S3a4K7Bz`^@dtfgxp!$IfAh?c9W&c@5
z^!NMw3lr<Tr-&h&m+gk3W@|tB8OYqbs2sIy@TCED#i9^JEKf))c;#UyT|cS*3zM*(
zaH52n&|7S=-_hGkh0L2MapTc~vn<cRnTI2;>a!qkv7s?-Yx^)|8)*8!Oj@T5^c4IO
z^e|by!Qm0e&P*&WN<7+F2pl2GfANDH5}K82!+OYV%xKvqLY{<qNsRuJNa%QJX$&SM
z=A8_m2Oh&5^!p`);d^HVF(_P;uW8b|jm(`M*QJ#?!AJW86J`9%zYn5fYce<cvg{k|
zr&J!qwj}$;4@%t&1BW?Ctg+<=vwR0>Z|HV->fX1Gm`=~Op{dQ}`Vd+x;dd_omR8*S
z1u0B=h}%NOg2ZcevIDe_$=rk8BcR+b?eF|XLhZQ<z>B~pfkV-NfCb#VKeaHi2ZF6o
zsG#&5f@83Xdut#6R-=O8Jbh3cmJm!yhzmn67l8S!vb%H$`7>}ZjSEU&ab<P&Iw8@i
zC--1*y50SC=u+<b?+jy6cVAQHozL35l`~fuO{VcrOB=0Z4v*BMjdJ0!9+Hb8f@A0C
z7@d+rO9`@pzP>)VTcc?kE7rWJfMu>63Ob^|w_YI4<~Gl1`+L}js$L<r=aJWonuz!d
zZ}xAV&eoGNeyow&%*{rBn)~^M4lRjjJ<;g=`zI|qUq1UOkghxO^`|JEzaBcb+8Ta{
z<6b1j780+}KJ!8O<_TURfSgZcs(?1R-WhCw&yxZUhypz{7PUA|73U<yikw3r3BXMW
zYOlY01vyh7Z$8(*JU9~#1Co0pmlOCDC|{j$K%&C^i!3NFnAU@efZTAy^y@6GeQSaz
zel?|Pk!BK}345<jKO&Z=!-Qq*tS+pbslSDff6ZT7N`B2t?UG*!qm?RYQ!KqaG3gl{
zA*HN=o`KCsiQg+92=!JRZm;gVdfS`Bz$j675R^;eQgtsthpzu*(ewtXAGNl1@1<$u
z`|2<HRQuEAqabnvP~8qj0KS-(ZuEE-HX{6;==|WUfs2&=6O&lRD+>KpaP0|N`E@_e
z#efkET@h<($%!@=q`@+aHEBPy0wjj1#R?rClBD5jY3^S|v}-(J!pr&10AXgw>|Ps{
z0qghX_S(d?b6Bk3mncfo_*=J9?%s`oJiOi9rxG~fO}O;sz9GDi+^Zz6co50q7zf=t
zuA4Zju8X(q$1$H`eSr@RJdfzd4B!lv>9)I$b8M42*Xs}-dikIwLUKBPYLJB?y+Cp=
zJG*TA^9VSN-$Hi8iD|eK;N1}DiZG||U+C^LFIpV_KUsi2q+p<^;KwKD?cq!?Qps~g
z=35~}{{H=YbaWHsY+iZ!d*><X>H4`Ce1w4zs)62#;RWpk#N=1-i6BxQ!Bf+H=B24p
z$uV9mbT7rG%5sEDT;Avw#8+RD&h{!FXTb(;#55e*6no5(c2<ye+AUVeoNK;-e!yff
zz7!~v@Pp9(>wq5^(R)J2ExpllK391QEFLKk(~ucP9<8o=mcO10607K7Pr2h8W(Oyy
z%dpH~BejAj@*ZZ4#Fa4uuwC-twO~*&TS4K;)AAi^ha-6QNR(B$HcvHl;Rt_ys?;Vc
zr`T9{&eM;awDkV`PtI=U8DtZcloAF6WD~?;`RnPce0|~$$Gy9jLRg+Dj-R8YtBM07
z9Cb=naH8S9f}mO^?*tqXwc7oajFxdwG5}NGK36hrFOxOSp=>nVLiCoIVs>t>u91ff
z;KN^94b~7aLpb=`j^-dlK`u|ej2p7NoH<hV>J`Ld9Ug)O2Ti$g_U8VkAjHmj9aK?B
z><|W>G6o1t4_-330w_BP@|%?NdtSOKT=%QP4I5#fFRPc&SwS@$avUv|Y^__N)Ztb8
z7IrY<m13^<?!XjR$Z${Q>W6d50FBzfn;&_YxZdL+?cjFlkaq82M_)B9Yqno+o0y>m
zIXQWIezNi{R2fjAU^sy3j(0VrNWXh`5+a*ON`(CGBcr2Wt90KK`V|9(XD6M7$TCHb
zr<+?;zcb_QbJ$Np4yw)-W+()JBxkAYUem_S?YXikXs=y>qTFZSY4W=Sk!O+9Jy^!O
z!iD#9_{*7<lVQ9v0tmkjHb-!v16|J6+>2g5jHf6kdUXf5A=CqB*1?D377-VZj*N_)
z(e*pzRLO>p<G@setYpxibzn14y<T!UryOOo?JCRw2uKRrJo(smdE)gkELV66!2~+G
zgZ3^AuS}SRH#Qfkf`Wn^93Aa|!)?J;A~mmNZ^4s)7B0e2;f~=8s&g?nY8AHdv8E9-
z@&g2{wy(%83GtsREoWKny9^Q%e{t%Y@8GMp{0IWV0BiQu^92_=dj=;-sW1kRM;<I@
z0Z)7!hGH$-XL9&FIGlsL4fU1cr%zG(l%hwNe~#Wq$N@wlQn*imUMA+}ga7!SKYyN*
zRSE+DmS=x|v&9~cVl}#=2OD@*Y@0GV_i0F6<mFrP&=E4~PcJ+qCtj?T`i$<RkrFX`
z<7tq=z{>BrTrm#*3D(MrD~6iu7C<27J<<180i)4)JU`lIpv}VRE(R-r_wq%uE%l#<
zhjDPeTJOI6`mKra&>*-@JK1c>Z|Be{+WQ$jZtX>9g9tDcAxGaAS{^=yPvZYDO?ah}
zo{C;zLrga>@{Hf#pl7kGvvb&IH!XK}aYS+e9a>eq5W1Axd09<Uv-9M)IY_xV^*u-g
zh2K{W2$y=*pD+W73=PrrF)yXHgGN&$qYoyi$S)mZwtU)vtN-EBe-YRH+QEp;s9s;~
zQ9x)^t>&##4K5tLG^Nyyww`CIO)qgbZ_!i#l^r_zon-=n@gDd1DkpdP^aLm!W_3P7
za2zmuj=ExI1Lisk_wt199tTU7_`ZI?;Bf#3Mn_Gta29!aWY_Hxv}11WB<R;x0Vh6!
zXiJE>eEIUob9TDV<sS=nDvkt*zsulZf{tSzuJtQH=XVI+)qvzt6y>4Bv5eTLo$m$C
zKT^Q0!^yKk-Vx*iJ8(%^r5#A&DI4f%ALu;3cJj*48HJ00Uco&@`6DC-fm4V0)9cr|
z{e$rG(zD$Qq5uclns4qP*9;Fw+*an8<B4p7mx<6LCBX2(gL+eP+fz2%d-3b(;Em5$
zAO~1Idj(=u$YI17n3~4InFWlhxA$>&4h{m0MU6+5i!dF5m1X@Q%!-S&ozyR)8u;E8
zD__#0#=K<2433<%&23~9&eR-275-cvkT`6!DLC(=2IpA-nZo&Q{`$MUq5!~=Rw#7#
zuI<789+)bml$0+Yok?4QV*zyF1>nWtAf#1>C3XgQdjhQH1W?ZokZ&bnKO+YWcCa=a
zSpd(ud4GFO;WaWtfsT+jSto-cJTNfu@Yyp;Nckd17#@Ox^hIcuX1^K;AwjtWCeLKQ
zHV+X7-Ay>P)#cl#6bs;)5_u?)vhjukh$tlkO;j2}86|hm4`iMGQrDrekbEnKe|jiK
zGgBc~RhS=t_rl)3;DoZ~QNp(_-7mjtqX!)f2*Xu@ei#Y!6i|4-znclHAPS<T_0Y1h
z#X&Tt4J1%MApdw~I%!^G5k@5cDQ}G*LT(UHJB<q$7{Gp6ZaqYUJSW+J*j^<2BLEIT
z?+D$3`iTm}iG<HTNPt;m3;P?M3gW_1B{)0W{`Im!zd_Xe>ogcxpKG6Q69!%-)Es)K
zuSLXmn^Kt*J@mrxiFeb4Eie?oWm|p|=OBzlaqiqX6pYT#&!3TpFHwW-4x`1uBMy){
z^;c{$I|f!S#HUSmb^*P?69jBll`20cz@c|kEg%_aojKFP^|J|X2V$U7?jtftn1>)s
zL}lgbcOfk>;v{(*;E@n+U#h8))w|fdv#9V+82|0CT<nzDGWEdBYs1fGgaITx{pq+F
zSw@%2aurV{!ut(*q3=cd{R`7zIzegA)GT<zkdACvtfxT7*0Zy-XLQ@)xGwT?SWo*1
zI!J$J`h7HA7sNm^7tpx;u9)9cE&TXP;&)R(G4NO3bn^rLL}O71y&Q%zm4c{Baqwfv
z4uU@91tBPV(Ec!Cpu3_p8UPI~aIB$rF%SRecRag&7JHyAoX+=Boc!nOWZs6A!z58y
z*}{H<WPxZ44CcWsrp+BL6gKe@NKrzBbr(j2T)-kH!o{Fc4uAO)U-}7B^$a0<>!wI?
zgi6JP0nGGZg0<(;Ev6wwMCvj;2BT)Ziin7al(f7gKGLUdcjdkYG#nJh_4Qq4dXq%W
zC2wr3gr}B#E6xtr^6=pg#mU`RRf(!Gcuvkh{V`?=Y9_acu>ne(E?U3!nWOdebeM<S
z#Ji$pYR~$)$d&D&iDK^~B@1X_#m${-IdDJ>BvND#AxsGE!+KrNO1uC4`!Gmfqx}M>
zG%y$#N1#f3*Inf2pY?*}W#!>XhAFxBbg1QEKL}h&J`fQBgF($eysRCbp#sOvq`<ZW
zkL1O*CRy)FC9=5g8#1CSFL!97GaggKMq*`u1a}(V8Dvovs?^SHr2{NwbxqGYPW0*t
zKOY8S)+XMc|2iW91jCU^0F=K2h$%Yp(3^b<MQPwT29s6aEbO%k$Rx2QYGJ~jJUIeW
z^nK-(^Ti(VXm?t_*bak;?K0$>cbUq>ts7aP^oQ&H#bPMgd^&^RyQ45XsJ8y6bcO#R
zz<ws&4D+B&`jl}GoVbBYD^F7(^}w!$P9|~>7mKJ>y<j7mcAV+sL#b<j2<|Fp9=9h(
z0Yh<%>W2U_hZH5m#i_vhsrvkc;-K=@C6GK*;NdQiMQ=loi7+>XJ6`4pdp!^w_)>Oi
zLe-X#y`@hUo)bk+9w0|e4$bA+aXlx2538Z!pD$}=22t+jsB4<?*GC+hW_4bAWvzMi
z>7!nqjEp><*eLWxdr)Elg_n$Ptha<iUK>4csN3QBwU(+RpwSSjxmczkFFy%-G&&gY
z{uU^83srwECdR}NqeF@!A~W?}$a(<vSV^ggDjVS|5$k)Za!<d$|9I+*9~0iyNF!X0
z*mL-jL*X6SsfUf6|EKl4To_`fzSrvnClv@V8<3I#XQH}vkRJ}epkak#?$bYhSg*2D
z!=n-kPh4Fk$OCc@LMEuDA*1%v($W}AZ*Q;rMf7x%!>MP!<5thNEiGM5IU9Y<Lv?`P
zMMP?8lg}>NI96kY9J&SkL@C@@69MCn00UDFb}G17v|ndW!T$jmj5^X1bo>c0A0dqb
z$4wcmVBvM-_)I5|q7A_Uv<{u%8xUw~YhzT;ouiWV_=;n(4kQ5i2EaG7VKPoWW%jYO
z3<t-=6zj2ia=Otd%crzf;WLkA3~a;NI7k=^NRN9T4ThF7`xQVgNa{&q*2Q3}f@Xc(
zXIe+C1Owv<J^lx9apALPhcVC^AnwM?$43C}e0x1Ca21O>an$eT9SM-OD$wJ3%qj@`
z`};9ra0%_sGK3g-_RGgNksAE?FNW4OIIhzwswH0PE}L24GU4G#FNP%Jq(m|7yL^ZG
z#845`T{JP_mzRgUJi)*00sIfpG0;swv-puIoCQ`?RK#E)*fs-)cAyKOLspMy;H`oL
z9aIZViqVm#;FX>uTVO?_q`8cYOe=VlO*huti_4l{KOu>mNxELRmuDiE*JcoZds}g9
z<)}JLe|mIpaX1vC-TL##YJD4G)*#JxR!gomz=H|UlcbP){2t)B>*9UrLj#}5zRYMB
z8DEC(o(Yiy%)_FhxPWEa)QceUMAd7<_EUB+eJ~xCQh2{`@ZyB(KJgkgD>a%I{-^=~
z=fUH!khTS(OjVtI$jIgy6~V|oDwqrZlKDmUmn(9dg7{Cglgva#*js0-PCE|_#g%Fd
z{K^s+<~+7VWNt<DT<koqRKww-<85JwZc*D9O+O~Mq38$CLBDa>4niMyC2f)*C^Yct
zll5vQY_PVj*z?Sg%L?9_h;5fOC@m@i%P<%ND4g#PY%F*lOI_o&V2k3Cl1}tk5&-mq
zEueQJwfgJVuTj%OLqj&3eTCxCRa(CliIKnfnR>yW-B5WCcK7Uneg>zi{{^-I@*^*|
z3vi_ZSbHmmkgb&~wY8^#KNS{t-n)2qVw(R)qmji>ROi7kaR5tV%5MED;oz;ZbuWJ-
zi*;WiX|lz8bTrXgaUV}#udV&1SbFlvGtYs~YuEqo99TJdVqm|a2Rq*o_5E%G!J!qh
z_Ux@br@ptILCo<5AAedLtKJjDLf1i;XW>sCaQ%p?=+#p+=NbG(9oM7fEJ9f6byD6t
z3NlI;%;>6{*@CnVPorbZ@RVwGmLh#kLP^Z97SSAPhJkG2Uhp9=kI}1sxr~Q0yl4*J
zd)qx0xN%QI-9n`dJpcoKQWspICv;v$)-@p@VGx(>Av6qx$#affwi8r~KBdi^a{rGl
z@Td-y#ZdqI5Q|W-nF9C!{DgPl6Fh+OYzLL@N>pqtMbE)cm4i66pJ6Jgue9v0oRw3E
zzy7+@yiq3H%eW3-TF36V{%2{zItuHq@A+ZEpiLN~N>kRnWUxc#A(%AMPq3*Z&kdG>
z`pU;oT&+zfa<y!6nX4lY*2R{8UJ{_?o)s4i{a!<|{55ilmifJWv9a;ldllb|$Tt_Q
zabz@FjXj6xbtGC^T=uB5|KDHJ4p<h17UKRdroICn%l3W$F|tZVWF<4oN&}@KJf!WB
zkUc6g%C78e*_F1e6w0PDGZQU}ln_ZoMM9DBKQ8rtfB*aQen0Q`ZFuhIzOU=N&ht2r
z<2b#)moNidbv9{OMw8|pzPzMF|6OJnH&y&7&k4=6l%*$@Zh!c+ED8Bp_T#ox4gWNi
zrBC+IBpreo6T&x?Sga#Q#KOg${k;918|W%JIxM96P`wtex=y6Q5*w27c(a2`=6CT=
zdW)tP#e7=-sI|A`u>Ro5@xW)|>M8f=?i_6+VdU>+rv8MtwDE#;ALy+BGfKiw<Hy@;
z+NW>OcE29uXQr1zE5T-ImFkr9BtIfZD@L?X+g%viV^E!)MGVN@q<W@4zs$cV09|-X
z@aKTHJSN;HQemm8ntOBtnn2+qolMTo!xyx_EQvn<Cx<rh@kD;DZji_=Q}c`AEEGj(
z7>S%8mAn?{?GQShL%W=alY#G(z9L#Xrria5cv0M+^v&auNhnCTZgIWd`mA{_AJ7F{
z?sovE0kVuh#!9I|yC@bLaPC?IFAtBXsA#hM-G%Rbm>r1-bzT}c9Dz4cQWJuYV@DC|
z)YKH3-*B<h0u}~ov^i+A;2_Q4PHJCUDIF!Xp;_*j&SlQ_Wiko7b(E0@oWRkeokr8S
z-(vUa9EB<z3?LzJ(P9u1PM(gUil4oF*-%JHJoUvbnnl<i4?eEXE9G&h;G*3Tzx<TQ
zMuFabicxcz@3k?o!#6l$g2oL#nTSSZs7LzneeUeNDlUF0oJFdT_4>=)!0OH=9K|9x
zzIBqw;T(00fvxy3i)Y0i9Mztg`bp8(K5ng>UKN?6K^_QI`7yQ;NR|yT5hXP3yKTn(
z+9KlI2c86gug!B+KU6`1>q(Ak#Eqs4;S&eA%px*sHyac&l(g9<w;f8pGh&IT+rewN
zR+?vH(%JNPOIRfb`!7o(!6er^l&|#Zde2st`?$tu^w~|&rR={`>pfBthW?|ZloUOM
zu=PP-z3%tz?+R$|;59jP?3ffO3Wys5z^RsznogTIGY@XQyBoN^ZDDa;GjFANL_VZr
z@ssTB44KqW%XS@d6Ny6w#_q)<k&azrH*z<zMNrS(tc*x?Shx{9PU+V}3^gEr5FBI>
zeR@#YD3sHKDPeI52_CO=<xwKyaxio}1pB@^;&4EmT|lQMXdZ>n`vAdl3Vv{pIlK^S
zmEY~YDs^c}<G);gn=i8)t3~RtbNruV|I4(uNe}3dhQq7N+J`wJ2NlM`m8gUKj+tFB
zxZT{FJ&0-cgH|l$1K`4Rk2{!|nVA$C-<4HRcnk&}v`eHNU<3Is_Tj<<(0zBB^<9GW
zW>Luk`pEh~%|+K!RKE1|@kcqwn&ZzDJIuW>Yf-b~S^6ov;oVqm-0oZ9V*;Y33$}uI
z1RmVk+eO`>mVkd|?&)tJjcV!Ya+0K{wN?LN*{OTdOR%h#uQiV%*aE?V!LnZmWdq*Q
zM{p^2>}LD5Rf1c-xkM<2Z-nY`a7~ubp5CDz`BLuqeOoMCc+w}PbQ{06s8|LgGAWBy
zmB>ze?9%d@)1EQ*<n&~NsPz|*8Oh_wrAjB}V9vzINXU7oN2{58zkQ3wiXiq9Po6#F
zp*3W|285HBHw4sMV)F>jLfq4S(x$%YHMHj0HKoNTD<&=tUZW=7+$-~_T+J9rJ&G;I
z>js8r53c${q4?5=)$UJO-WmSHld;rYQ{!^HbMY^YWH*(6IRcX03w9IIEV=!_w}RnG
zi7o@grlAB%0Z)#wWQna1iB@2rfeUlL#P<)1M(4?|hcns?%~vGsPA3+esJ)k5Xq3zl
z=9+zwv$CQmWilw~;1%pk<$Y=H>KP^*<(6-l-^$fRb`gv7(lC8%llkup7#CvVdw`uR
zB;_I^OE^AqP#!w>E+7-&m|G)9GAcj44#+BzV`qJI`ZHf!Jud~%t_zJFdF)(fu_Q{q
zI_%CaeqH)P6W2fz)`fT}Tm18nD3Ax=0I1zmwpm@A>w<+@rq<!(0!Z-cc;Nd56AXqL
zrr`M&fOXJr5lt7d_W`cNOhIQ0I`@hm&j;U$m;rU)39rsABjpT+hV9BE`m`!brsllB
zf{NdVqK>$A5Cbb<exC*g(!dfxrh~o>e(Q99f+0jvGt2!_7e&sOP$mV}wJ-fHJhNI~
zCLzv|+(qgZeE-wn3Ok}_Nkk;jG_-Vf!ay+u)iHU=pxWVv_-Nl^*kO~<Rodk<&(F*p
zrfj!h%6I<>;7OepgH7+xBwycoj{p{@<Rl~*z+lisWkvQ-*+``jF&({pnGZQu4_I-K
zsFsxXUtWVs-iM(g#gpgf&tj^0FS{Pup+^c4OaTOel8M6sY-Sgv7nZO{R588w+5ldf
zVscS`)*_ZP9NQ=)cA3p0JRqZ1=HzPXiMHdE07mX$?{V%<Koc(rvkp~g{c!Fhg}>39
zi;lD4T_dn9;(0_N1P{{VBZ@H;(&9s$T-BELk3-Havr|@%$kAy1e$?=$M#q$YYF0iK
zVSDM*J<Gy3&5Kb2Z0ff-r;G*n0MIz~*gAjo_?Cr7=?S1_6diObjmCwV;Rj*JL+XBI
z{ZTe>cH*yHI|mRo<yJ6UjxxTjka-)GR{!1wFfd@zj$**0E7o>LcvUQ)@XE$x6gD>e
zbu)Zo4;^KNd`HZUrh^NhI*FsSonnqoib%xZx(BqphkPRXJATt&qzSc$B=Vy~H~=N`
z<)n=u3hfg+5x+X?xwdEjF880|!<?s7hTT<(keRfE$D*|*(AW+63^xSSnL0X^)s}!6
zUBa3k7(IJvxs`IEo`o+{L~xJXCKK1m(ka4+Tz@3IZ7f1H4N))8=-DQ}<9snZP9aor
zt#FfDgG8rFE*y#{SHfJ{8CoqzxX{8ys3!B^qb`ALgS&v=7#~QazNl$f$2iHxJ~W}U
z`tI6*2R3{y3E$sBMuKO>=D$`&Y)borz|jPy(cx!Phm=Nt@?1Q*G3B1IL)%O|hqTRG
zUk2JCy`anUqf4&|O{hG}ZGEGDQqfpGDWh=cvjFpAxlJy|RhIJIu2%1TEB4~x-@Bqp
z#>CNCkr(vpRo?N}Qee>YLSjS#LybcC0!X^$U{gSha9br(_v^^Si?_-P+V18MEgL!#
z_HcCQ%NJJLXa<FQ7E+Rrl5f$ozMhwqQyb9}VA}pWHE*UfRZsG{&n_cBPb=2Y?W!;B
zrxLh@C!6-rEmnN9jh@4QzJ)#ZQJh5-m;aTM36A3Q;#>Qy@~;r6?Et6?1nnc&LOyX&
zPY;LUA_}}%>LG+!2PGpAfQSrDDkFu(BeEtp{mbfC<~uH=&`q)x5ALRZ3#2=C<cTbW
zwfXs{C8r`2@k@z%r*gOF8OaG-3oLAQtWxehPXh2pdPz}7+Lrj*T)e=jt#g~JUanc9
zRm>M>;vtBNB0L8Yvew<>id9Sa3c0zYY!S3omHyU|I6R&RVwb9cmBy7%rK+^GZDS&T
z)xzx(gGTCy5|giSj-!hxI@P_5WW|yzK6}}EVv~_uzR|2%+LDp}My-~in7v`EZg|cP
zYP+T4m;8~BFIj#KF7hA9A*Vfb!W6XKiLgj3aye_GnsW~YRGHu-pvt^sTgu{u<^-|s
z0CxKG>RQPUj>?N0mm}8|t&`V7I3B8!JBBu%knY`VJ_4+_=o&_<Iz{LCI5{oLauBTR
zZ&yhzHE14rRBzq()86m-;fqQ<$>k|53_&|<U#@v(pF}cM_cDM7G)V2vpXU~vVrOSZ
z4^N{2rL^NB=3wPy2yW3_MRuYwHsQ4m7@eHsOCjJ2p!={s7g4kfXr2MW%=&z-f`YaY
z|JqI5T!%G}9#*<}V<t84QF}C##LW`_TU9;`Oi{i<srO&K7TT-hby`8xuEZ3D8SV-t
z9=66N!<%Pk&kU~<wIeW7_SieJL>|b(SP6+#1Jc;wPcMOv{>ydavwH%XajG*CADQqd
zo}c;$kz*)3bp)^6=Ec&;0dOA}2%fH(q#6jXs4;h(et2b-G_ec^Inath6)j}^bry8A
z)@7}F-n@I9MS07)oL6FJen0H3So>jz^XcTc8JF3cRa<1=h$>Q_o%vRMQTOZbADe^2
zycRe9iuU7rKbOyGMYmmJtZe4x-n&VL-qTA%Q9n+dOG8q&k|C&}rhVPpECKR8aJA-#
zpOGZcTgNJiXHVh`dxFc^&b(5^ddmp$np!2+S;6;g`Lao~8G%(EXZ}|a(Y8FM6XR;}
zxFUZH=(Mi6#p+Tv+V#z7pas>j>@1|GI~I-o<i6N?_F!{c#lVPWO_J^hxnoJYNT|kT
zKmV|9HU9o5gG6oZ7Uk2K4Dt7_N|r9A-7S+mPQ7_=rQYJ^_n+vlCU5HqADpb8POq}g
z-?8dJn0CK-=~A+waltzmcGqS*`7W~{c83&%(zhxZ*3$ZAie5$?Pm8xOqE{1nx!ix>
zP*=p2smTF@#z%1nDa9QtE4x?9&s{6qxun3pQrn3m$>fR1I-h{HvBHA&y&3m|VlTTz
zK00Qk$9SxAO6BgSq?dQYE~E)=H~)^mB&KyJ%HlsRfaY}Z?>~l_2|8!2Df(=y_TNgZ
z@2EX)_}St~Lq((Sk7q^e9<*K$@3;i(WS=VC7A2DP;et;v>WV>CVxsZTb0^MCKsnU4
zk0eukv$BdtC*60tT9?Poln3s?|6g_~>AE4~$&)T0{P$08{_V9zBGUqa`Q4(1-hsZe
z)eL>c{Xd0Mg;Sh$Yjo3h@@+ZbzN_4o(^u>cg?(SC0T$7}EA@Uod0*_oCH0~YhBhh~
zB&mba@#NOWxsLH4lRMmx7;c?xJ}ox&CG3EhvXq9=++2`!v$TSq(Oh@e`{qZdR<FE%
zbG4-ZDeYC<%ysvhekGY02G;Tpq+D!TciWc!Yu_P5;X=0qOyynm+qL#A9LqbzdtX&A
ztg!p{GBa5rUj&$!6X%n-pBdSr<n{1-kqjDulAPT}L+@6oWL@7)OWqc1*<MC96fY=M
zuwSb`JlyLvK<_=(R;w;v;LGkMKWRQAP{9yd-BoRf(wbbQ`_Kab%LI)wjWHNApXn2$
zF;|msx(_sv9-ZFhe_=9J$*cP3HvP&W?Q+?stW8V7TPSmTBf7?2ymN&QK9O9h?{V&c
z4o*~7Y>l9}5migbk!A5n;`s(0!Ltg1a}?g=;Bt4DhyaZ&>LD9e!SUjSaM@MB|AES=
zdMc>;UPI0erHm7Dyrd<(a4~)i$V(ucL?Q}-P##`!E>jbPWR4_({yJAn_W3=$gYBz$
z(@NFoi(l;`UoO=|DnTb6DAG(~X^rKH$2pyC87E|x{jYpr><!4{y?3~fiSR!FBgEgh
z5e2WIJ}4?s@3jbzda#`hdPPGhzT12~O7U7r$|jnvM1oF>+G<|E{Z`G_ujR|yD%oFa
z`rDx--=A)isN*7n#VuDb4X34PTY;%@MdjqmOLB+B7m6rKk8}OzTf9e;wJp{nr7K>C
z!jn!tt!J2@tzuVLSBZn=Mj`4#!;?Hs<;QeZvII#5i0L-f4u&1k>A%Gj8C2Qqaap(3
zDB{=!I<^Q?d|mFI2V%W;zc+`7y;-}+HB;kn4;(+m-lly0Tedlvq*>!`-WEIXks|>O
z--+S&NK#cMnkJ^OHpmpxQCwE`>dJm)R1p`ap~=m(D3;eO=)V_09u|m8y&&I^ehym5
zFmw?^g?^Z(rO9RQafrnu7UY+VdXP_+UP=sUwASV;mJZxV%1T^aT!rrfM-1)->810W
zKbdc2spx$57<ps``&Mw?PlfK_98OV#H_7eR-jlyzc*5e$8m8yzeZcYy`sC<pVSOTc
zbd)mb>%S3|elU7}{rvStNAw9-&CG5M`6RSY>f$AXE<L}*gr85FvU&%GTgGO*gqviC
z<1ZNc4NPfoXp$M~8R+v_JQck<H>2{~$Rn0R85+VD<7!LZEM;~*km|4T%*l5-mxtrA
zNC7?8Q{tlH@0~MsnL;$A<UUyEvpii01f|PXa2mfO-ks>%b74m=Xocv|UO<)tR`;Vp
zFRFW3>u%fA4uQJHNnw#~R|h|Hd7C7GawvcvVs(9XKZ6YAZNZ;S+l}X5I~|=<DSQ{)
zac#)t#fSR?lG}g#bO{d-FOuUYxB366d~H<j9-TLn7#2A?caFiJDfevT^*{2yUEjwo
zetef!>vlCaIzpK%`t7$vy!!e}zq4))Kbv(e7TX+}miH%Lr5|9bUJ-9ZbzkxVx<)l1
z&b=U9ilV`V3afqVF|-ApUGfvv{%~jV_@B0-YZa6hn^N;Hq(^Cbhi_s3edl2gHXz!d
z-fr&^xZ#5v8-RcGGJ~n&bd)IJrxoVmj%GOPcMt!JGdnYOpE(#U2;D8Jy$uxu_d0Wq
zHYxCaj9}eWGqxv^P0?YD(`xE6+fPOBAr`Bt`l*uNedGEwd0Ptija_?wCPU?WNMg~J
z?rC=+wzbFl4s1BRT|D*f(MuxQW^PKJpJ&29<ux|1HkrG%fvea8!0Te-?E~&aZ6CUK
zwg(Xp$yMr$*K{=~hyixpIe#i~r}|B+u7^8bO***x>nD30aw+#JInA$8(Cz_sg+0);
zcj=lOe?L&@1Ln-Mtqy!Q*5A0?#&(iJhqa8)KK+WcTlf<%BGm)!nJBq2<scv|jQL1`
z?B)K8yVJ?GY$g3<<N$)q!A@TrL{9o1-vl?qTik0jx>83S-l-#u_5{Q^2JTJeGCLdY
z7@%>RYW?K&*B53%O^^40uCa9YPOvQom6ep0iNJzryTDhF`*&6sFY~z=jo%F})zYMA
zJ!ijmhpi)z8H6iO?BS}x6c!bBy@Eq<N%7nGpA3TR?cKRi;T>x5Jo^3nw<Sn+0KmvD
zq;08@W{C(t5ft=_t8%~C>YV4*=^dsEG#0oZZlCtt_U4-od~MkGsp_}8Chs_M7e_2p
z*#4@V8zgqt&qb=)82#P|g5scZG<+oH5miW+8*T;~|IQP8=1s_wOZU;b892XmiR4V0
z<Fn0D8r_)!R!puRHm9s;E@TcJk3Qa#e_19g{^~)?HXm_&ligR@cBL$H7Cp&e?Xyg)
zT`YAA03DHJa6pMcmxcyCc&NRDrzc)jlW7kd1r_$}Ur4<KoOkF@-LIQ|-KJ$IT6(U<
z^eoeBF3R#JtG?^%Y`Zu4&ShJTqLtnJa$#?=NBko>rVho9;@p~R3N`fxspK45io%{1
zE)v+jMDgx5U2MomgfwWvlPi&Q<IzPAuD7Npx~NbVf4#Nsp6Ut$j=xs|y^KUAVi6zC
zCBX^Zbji%nQ@Wcs%jsgsQphO83&+2L>=cDVS&QR#nI>PVkId1e3g38V9jGt94AJ!j
zoOcKt6EdR5paD_a7Y2r;dTPj&k`YE-CCA+8+&{ZfZaH6Er7qsPZ3ng0%T?xpmBSmA
zzePS(#md%}2)3g{)Xe|&m+W_vcTf<BlKh6kJC4G7bQnD}j_p}C6b;^#{o$>j^bFA}
z&XCw}b>AjK7apZj*RM^7*(+$$i055(+kSNh!R}NbIy(A)>!!M4vEvR1zp%@g4Gq`v
z`9Ns>15G?QC~N8X;Wvo#gV;<o5?4EP%X&Y3l9YS|FQv8Rqk(xoU?yi}$q<jU#s|Ff
zos-+y3irgH8D;0c7A-%0ty;pXugg%{WYr_l2mF)vTTY)YDgN!i#2#7bTDfg&x$)eu
z)~b|>dzX@uOqL-1{Qe;BgW9UcUQ?U35M`>m3j3#K&!^tI<li1(IW8c`8dAsS`Ks*R
z*wt3Ap`r<+=@+~0ZnSEQITW}3Zty<2Bb&?3JoAmR_4gVjXV)UW^SRk4>&iI{s%i34
zkM38te13U%eddzJlx6u6f5JY52AAB8=TYU_eKm2=p6V`4h;W2X2H8DvQXzrB*mvTL
z9lD@-g?&j!N%%>95x}4!1eP)vq?8sG76_=b<IBKM$DkbvNq=-M$N(iWUtk9I2I4X@
zY>1Z{W9^W8Ln{o|Dz-+9Ou#Ut^f*V8v575@s$Q6fyq#;&82ecenfzu`viIJHaWj#8
zY{@P>my@m^qC{UXHl3UD*`#_rxc}s|J#**QKnR<}CAlNwmbSXGg{%<cQh%^f_x|m!
z>#?Gn%9%~O>Tk5adXTNDf6D*#nN8&5dN+xed7cf6yh8V-<J%YEkLC8c&7!Lu*LxoE
zk42{%>*MF2INepHf?Q^`%MNTk3)2-!6)|kFu0Q9n((tdb!~J{ORAGHkT+W?4H`@Ma
z?vTzL*mYcjf-=Zf!=D&9@33P4p$STu?cj{DHTpvqy$<9asGT)+=1fjjgup091nXX7
zj8EP;7}=P_R?}x8^8mO%o@HJM_SKcR8EBYxq_wa#-e0w{`vlAI%BTpAnC0jFzq>1!
zYK5{rRr6lIc$;;pRZUP@@TYmNnXRcGOKoPx^~I)aV{U4_tfW1(?dZ}NJ8<+!>c>=x
zBfcwwKfTV*cgw2{&u?T|%E5E~(a;5do3~dQS(ad(nEZ@YS?6m_U!z&y`TJ-4FJ&uZ
zd+r+gy3XH4c2gX}pBEm}X7}};pJoDvZtH!0aX9E!Gpnw%j!x@$V&`L=JsH{n>Zp~9
z7B{WIMv*`H^%q?;Gx-HlsHGw)QpF2+Nfm~x#mPliReZBMx973>fUMixU<*f7o}HIK
z>Nu+g0v<UyRIlckm0^w1!v%r!yc*=`8T#eNaZV3vlTA-n-}~DJs$!cR*RMZ!U+l4M
z9H}CO!Tu$DUT9lSb=|QieGw}Yp}%Ls`;QAz>O#||OYc6~+|4S^IO%9#uS`Aiv~{=i
zlBEJ`OPRRL8i)f12xHTs4e<Mi_p7o!iH{KdxgY>jTX*gAA8%VOVq&TnLy>MUa3t7j
zHX)%9<tr$+y+cFiz~UDNbp^tu-=u#EjSvrURfbQ4mf-@LovK`e!UIi~gsVSnB4aD&
z;|)36<!|x6=cMgCy8gQ8g(hmd=ui=O(7Kp!MRXJRBAV!UQivHU@id1YKKrRtr-<VR
z5g&u4+w&NA;x2(+9$IKnE%ThuZL<e67cfk`$HKjnJN;<a;@PUyZmBzJ1cSMjARR^+
z2!T-nX(NTQLRgrz7KphW9w9S@h?+ov0QCW`+dPLaBh;Tr1%DcD9-5&`xCmHNBDWb;
zRMM-3{OQTiwIvg)n~$ZudmHY1X8v(i^12n=+^j(0A!82(24dlteey#kX-stc&-;Ss
zM_ijIJs&@cXGvjM6vsiMf`l7{l#oww5Oq21n!5Dg0fC;(RsfLaHKo<TblTOW_4TC*
z$Ax#2rORn|#-WB6g&Pj(!%!?cijHSwWDv_9@G^dYM3x3U500JXofq*11}EEETP;CX
zY&Z2&g0BI5Je@4gWeAf(n_R{rg^63v*Gj3cDXdK|HC^$<f}=1*QzP~MBql-;&p&iG
zm?)s^@)BD=R2W42kEho&;y+*Z_B@_3_0%BboP1~q;kR0z<_d3iWrRYQW$|Jno5k5D
zl_y5~v#<cjcNVnW!XX(prz|OlB;{;5yQ0Z6RAnAA5Q;(t92^__vcH{Z2f-jAFcyw*
zh?Xx_g)V&339#-+a~Mt-F!YUF1VW@JTuc-}oFVsAGH_GV&8@KKHuW3Bzk|Lt3n>JN
z^B{$56~6;yh1!O8=M}tu;Oh1ztG(ZKvyZPYX74Z-kxf>RgQ^k#{y3KF58Rr_0zj@M
zY9yl60_zFz^g+{Ish+|AQc`;lq(VAuSHsc6K}jGBpym!Ch$kkpXI@mB113moxC9#4
zjT<+fKJvpU1K+$KchqKCDa367D$AuT5q=FxocA0->RfPC0;&f%zS=sUa5gCc#CDW2
z3U`K&;3bJWECiI|%___%$A{0eUDZbifp!PCPqYpHOn?4@l*QM$B(z0kWf#Mphz0I2
zWGaF6P-`Yww2Y(gxo=_Tp!N#9&Rxtg)7ZIt_ntlE+?Rl0F}D&1{(5VOTp9#NQb`|x
zO;>*zAE|zc7>_i6U|iMn^{d>&Z6v^9wdxb!DbSgkiyaoEy~v)3(|(TXiG>riu+aX#
zd2Rz-Pn<y&g7&v2D~X*VmBuQS)-sf|2XFVjEzn=kOKWlXP$-bl!N;Vf{}-lbaS$S5
zLqlYiPHPTMnpK#`0*$^32w;S=SNHg_xTGW_JeV8MiV89`KeGru9tO(4sd?&HGLQ!|
zsz{U9%l-6-vW85&i^#@QgE0*LB9o^mL_AYeRP?{o&SY+ebiR<i*_}9{jWC23G=%n{
zXg5N`OguNi3{YF?{WX8f^!7Qm!p7hqv$gl8uRh4Cth}k#j#pm1%)VHVIOC&j25U*;
z+P<3S+|!XqV@<N8^~I<xX>YMfw1^wlVDLdv|CcKQPzNb+w8J5{7f0!1Jo!<`P#GwY
z(BwfN`lNs68@|Qx?=KefJEmVemsh=?nNR0YtdxzmQpVA!V-l0*rp(1PtU!X?a8R+q
z%9fIC;9YBjein>*Ix^xto*dD;xWZp(f6S!D*f))vp9xzwBN0Y<@X5{^6r?JjT_z)2
zSXfvK;Dn9>GnaaJ%+SqvN`LkTopQepFodCp;;GhL*W=!%+T3m0<0AG`UlzY=f7RG*
zm5h)=$}>F*T0p1gtq=<jzq`2Rpz*}d3l}Z`h4Fd*Yp$?v5R991w{J_Kf!1JO!cEzo
zz5*rb9-*1*gj_iL3ZB9kLI#N;%s6U%(I7udc3SFEBu86Aqi(jQ+AKx^8MapKct6@)
z@_nCsqU+*Z-d-qMzWLA{v9mls=rjS`Kz8NeU{&6@-51n6`{ItFcaYX%&;uOY*V@}Z
z45Y)FEzW_zG_k$G{DY9jBc(H?(7h5*?_T62oNr??e0%>z$>6b9jMloNo4%GjT>9>i
zG`1V(fEv|ZK6Pmt0z6o|o%>>cj*nB}^#!9YQ-$ZGsju9C5sg^W3nMdXYYDp##8y!?
zH4#XI9|v)44+m!UT~lX#OXeUw32#(tcZ`|h>eF(!xCIpSZ&u}hOa#c{!H+SCphs`O
zOFV!6Jh4I{&J+BKmK6L7jt|(>#6Zj13&J4rcQE&xI@$sgW*lD7`TGz?xFvjcHo@ov
za8)2NHY5CgXpo7IAjxIm6P9|HyLhorAUP9^KRNEFyDz-8CDkUfNhU>w-?#+Zd*k+)
zYJUEWOn=7aqVJuv8GXZ8^K#8z@YAmkV5%6jMqqK08H9%Ar(hOB?Al;C3@;9i{ri`G
zulQw7n)jsN=r#FuO@<O4eoOW{2+ZJ!K~^?4hio?<U>&jU#efUquVrO5vpFbyA<Ejv
zPn2==HDA5AT=;$Ky`8>kISE4OHhvoB2?WSSYtenyD;l3b%zhB-o*b%Grw}1NQ7{lO
z<>@c!m!X|Immq6MQgQ+=EF2C;Hjrn4aDO*6Xr4<zfE8sm;{AP>4XKk6|2}s><XXVh
zzMRUu)AD>{k?THk{4^m?L*H+I%BUM##n43I(Xifc{EUX)Yw5DX9S>@15|ITb50877
z4WUVH6pX)lGX_pQ;_KIQ!;(Y3;{kcb85KAlHU0eNJU$S;u#Jrb0F913UfQ4fZ7d79
zMMl?Cxh=sLR(aDhoZek{MjaO_Hf9G)5(aRyr41`^#!{L)JAI~<aMM@z-ZmH^s)8f(
zkYW*U8Wb|#&wmY|rlWw&Vqx+|-2TCTS)}K;_E$lh4i#`v@GGV4=%T+ZIJ)?z@8n@Z
zR{YV8O7I6x8x2TzF$RJX0QEBu=ucu|>QpmC(A4+qVHSlCBa<TtK_(AQQKVB%ra?fy
z6;31)Fy$)!Trf;c<hsoNHv{w>-g{j2#l@8mXNFO|+IjFg-~Y+^_S}>E<@0lYP@y*v
zn<RKSqI{@nYvaX|ynLC<?7pJ_$_k1lY$G7{DlfZ<XXkhhMs0+5OAMs&8xY_0^Nt*V
ze4jWyp|6`9O8tg6OaXO%UQd)neS2HM36BiB)uD{sMwTTrUMZ9GSOE`EZS=rZfy}-n
z9)Dzz0ov`DaZ(GYnAnbj-D18S)YlaV5L$3vTy&p+g{2;lXA({Xk<dYfPSO<EYyK_$
zRERSJ=s_!C28B(<+Wv((cwuOv$8adx;`n;RUKXAyqhn)%kjf(mg{EEXa^#=vSHlyC
zBscPAMdmJ?ROAo-Jpap&->K13u?O41&fcE;+K1+5Q8|lNe{YjihgT`(McjPr-9kRK
zziCY=zxDHT`|)9#xt`d@M_<ps^)a(ai*qdXkP2d*+8hGH1={IM3~a82$?$sqB(y~8
zMul#mRN@u4$ZETql1ME($5*V4DuVmA=62IFr%RmIkWPO0Hdw=HIrvN|<tdy~ZMN|q
z$&C2o8eW+Bfd#?xw0(F2!b9)}b?!W0D=SNk;^TZw()BVGiJREf4RGCi`<4Za42PA(
zC*i?SCGsxrf?9pq{|$f?Xiw{SM+QE9vP7;xuo+~R;NntSCGAyIXBP+yb%dEoJP(nR
z6IKMuPckqIOOg<b@npy>QgD_+++k2t2QVGAIZ|nT{T|4PxGq^7J{*DC#@KG|;DK$c
z<Q_q@Pb4wm#>Elf0;MkK@B4Hrv!6VdsY=<_kmuI7U2Li?!T8438Cg%qx3Y5c`i~Dg
znmjTA6q3IDPHbZRo1{hkhOIsA%LelM?<CfYbycJWA9<bq`b}%^Z2Kjc3;cYk=(KOu
zcB4`8Pk}FF6O`Cb5A~l+<{6)DZ&m7v^*3P(Np6!Unux}eW=iN>Y4v>_i(9NZn&o9N
zbAf+Et}q#@*oyf^X)kNqbGPRk-)gB3mOBnjtzu+;xVW)hDC(hpmKLALfpt^b5;9p1
zEv<5&wZmdQdZ7=5%LE9p95hKYw;Q%WrAtQ0Se`r?3v-@c95;l(CbSWW2Ocgb1(OQa
z!A}ad7c}<-hp~x?dN_R{3a*h<L*yotl@`ddVEI}<y(Dkg^`pgr1Z!xm7nriYe*UZl
zw+A#RDajA;?jsWx3!HYvUU819Jhe$mZL5LK-1W6`>JOgi?!C57TW79j;KR<?D<`gu
z_N*40T71)+gXYq;&ad=}G;8;jpQpyZ3Wa#zjwpFQSFBbU6rN!u_Rf8{bF0~C_vMqK
zn@pcsTA4nQ9({36X`y)cR$RsNzS4hrZ%Vhum~Waz-R%PwJB<Qj<7O1*H(VHK+j1mV
zSE5s5`Q-f5?!+*y!}pyiYPL+FW}I5f>F88Cb#-}8?Py%&vaeG{cUR$R*#{aI>N*S7
zD{cI+{&MGfrS>IzwM^Z1KHtH}rMqLf`Wvf&skxq`m8Y4H>h@S|_HU@Bb-r{>403xp
zXu9=RT3U2PuNvEVvEFKl-v@rL*rJiUSYJ&rX^OpmHqwGcZ$4>Ia!Dd*R2j1h{^P{n
zt&jc6HR3whi3DFeivKLse>Lyk9Uk07j{mmXr$)Nhn#xN9CuZhYNrk=xK9nIy7d}F{
zuL-nN!J=|WypDe4@rYnDl6c{wNuc@<gPNfi1Qm^)1Zd1MfYi1D)mu{8HaKD6XVoE@
zRD{h5g}CIxZ^vbR&bL~DQ;eGprWyZT`c5A|e#BhxaEOUb<ECMUvdnRnlz3Jd5{+vr
zCu5}cU0dDYH8B*@(_;x~2HTY=9sd7s6gsVVje<g`sRHKv-4vZbVFjO+MQ|VYxEvkL
z4xIOCGzfi*AI(k6VlwNZ-CB|>7RIIvX3}ZSw05hkS6v!v8VMFYuGZ#K4~HU$_N9ot
zytn3%$I#2m%d6Wjqje=<UInq6g25KQvLinw8SHwy_MF73#_(V=Y=?^`!K--cTF0N9
zX-^@!eb?J9$t=|@b|xmK(TRx=fS=7Y@)7n_xXoas(xJy9+kL_@-v-)rK}@*|ymV>#
z-hKNP>FDUIknuItRvq?<!L-RrwUH^|eb)#7`+2Z&#lacbybeEWa241`>VI8p16wg6
z0q!9@w>*!oRP&2MjY>h>gwdnlcR$jSL^QW@bR<f8JG;Dn8qv#=rvANv=x_+k(7wa_
zWQM|oUr1;%nL8KWdJAphc%5J<q}cS31c*SXz6?K}Fsq0^J%C3|98W3P*=8ra!MM8b
z^qxsYMTHE2pb#g8&(BM&xs~?+?^P)&5^*!^+eQalB*=^*=uN`VRr&tmQ7p+-V4qns
zg0PzJGlhZI+;q!O(!{H7R8kUK9*NRkOlKB;dA*>|k((fre?$V<QVH>%h>eJ-#;0%D
zv;XsdTmUk47X0vbSJl+gF!q`bU46!{ovawu<k-Dre3&^(Sk`@Et|*S_0GRp`k385u
zfG^VnWi8Q7|3I}x9Ii-hSakd~afKo(M>J8W$c)K|j+8XqAXh~ZKKUc^2#p5OER{M9
zX1ly&FBl(3t(l#_v7)M-?{^=EWuJD>%o+8xe;cQ^7`Z1tObkZQ0Mbs6ePV@8P5?1h
zA}n8E7i3sHBNG$(se`Z@tArweEPUZa7TAJmLYCTc-^m28;L+}e{a4p7vhbhvNVlQt
z=Yu9#i>L@6#o+7LO%h_Zfg>cf|3)|*DEs*bMB45{jMfLb%S)IR+lIltd+!y+l;sDx
z+RVxW^KRd|b<ywdEpoM`g8P^XRV}f(O0b!SYcl-47au%$aPQUiTAQIGVTIQ&XfYTO
z%MHkhPkcRBz^l*6sAQI~vlFizn2Z$19)f-p9b(Xy)`7A<5W_{Q$6ySM)g?%_t*tH6
z3xhAeF@$yo7yt_k3z_&nAi`=5;OIB}_+`6J!q?5&#U(R@O7##D3RX(b%NrklA)Fl3
zUGz^MBDfMebR1;(&MWxKa`N(4HZ~0T4`-}3zJLEt2|)6r%(6=tqZ9&+<g{TB1^=Jm
z7|qrU;_L~#2{ONzj8#tQ<E*&&&nO__y9z321c|Cu+72%!hO4^DFfzQV=(xt6g?5T~
z>O!)Js_USD%DQ#xAQ9+47E$wCLK#FmJK9#lD;}XQ8BY)Mi&#+~F;gfITb{kZSy4^^
zdU^SM{e-Rr!{(5b6ft7YhxN8tI3Z(y4{xxNnDXTOWxhm3I-^n^yAJ)gd;f0J*41!e
zg+U<)X5?vO+a>fNoojAyH+<)h|58*^(t>G~{RXM{bE&D~8~X>Ose#Co=qWGMv)aeU
z$CK~exwF0-LyE{e>+bIE!FPqyU|M6^mMtAOcbmK;r-91VC0CRdH@{|xih^%O&mb-3
z;J+0su@3e`MJGS3gd27LV<1L%3r^xBL0yOz#6goES^Hcf=tb6UVmr<|a^aBKZXPyh
zUfJ}@QqGXa3tA!k0muUfH-r)gO(dxXu^`ls9XnQH^S!f#fy~oE9~7e{?AyK&bG@`Q
zOqi$G*z~!+#}Xjq0)h&nj6ONBRWa3zjNOBQUsL-A?A!2qYa#+8RVeHaqA~C;qjRS-
zJW79md+-4s(a8nw9$JLcPYligt0ATHC~-msV%L_nVAut{<B+#Abv_MUQkaHRA#PxV
z6}?b!+sg)K7WPGzf0houW*&UcXH;%9NIYPxR&9S68~${ltwTj=YTYMSzbmM`TXhBa
z`7^qD6gO<p1T4rdoh2Yn`79rmGW+4!dCmzfQ1Oa|1hEkK1q6D5Mg&Ggu=@D;z`mRO
z0}A!!r-x$J$jAiZD06Yi@9@AFSu#eFD0JbFcC5th8`RAB)}U}4DBdcSI;~-7$VaSM
zQDSn!LNsETbPX(PAUgps9sfWRLficPjN8vMaFk*KnV<M7p*WUl*Y5;Jn(WG>O&9IY
z?4`aegvlKVf0{u~T;c)jHA5=Cs?Ei+qut)=A4Ik=6&+{z`d->m0l~8QMevH6=T8S(
zJ09ptX_}vaXC%o1vkr^rvxw{+(^%$frat7qynNX;db5hR&BV+XVUz0YtO(P=Fo+r7
z=m?{SB?oaUf}LEu=vKd(O*glm2}C0{e4mBA`;sL~@}R^f!zMBO;`Bh`speb@HA)r!
zYF2hkPL2YRmLq==my=t9E#BlroD$>XxxiJoaGJ<)QN_>04e38l!VkrhLNl3OJf4LN
z$`ci)b=5LidHft*JuB1qnQ7+bdVXxKHm9a7Zatgl<clZwMMYMxNQzg&TwPolCy4nl
z@K1t16X(#54#6GA-87(qCvx6RVzk=NIA(A&#KfhEsTcU2Gfb>WaoOC`!nH}P?G)_5
z8c^FNm6=&sh+z5d1it66j+fSTXo|jNe5k01a`I$V_sZrLI!5EU;DwLH2d^X6&>AUK
zh--8dR3QY+0hDHGZ!hDlni~4Nv5^&i@nlvA;%UGurfn30bZBU&F^Pl3h`xmW)bvaL
zxgbdJcSv0j!yNCyD^w4rT+w^%s;X6MA6=aD*AC4){O{w!hvG2|7-)#=Q-6d0Zrgcf
zpWn_`u3xvow=&AJq$A@Su%WGi|FrR5&Zy3`tE~9CVjQ_RQO&?;To|vo4xnQ%yf-0!
zD!1P&PCVgWyx4!V6E5GF@Ubwxl9<(jxDHVi3VT{tjqG30fjsH`_OEWS;=2G*I&@+N
zEb+A-+rcjd-=GKQ)o=9ih~E;-%B%W5Jctw%r=<VnpkR8d-idNp>?1Gm&2)oMO+Zyu
zm7F;HZ)_sdq0G1G6n};tNf=f9%$HxBsiC3RC`c-dFU{F<RK9o9&-PusI5O_i)V)x0
zlylKE+p`YIQmW7?YyqI${~PRoi|HFAtA*1H86<-=axuZ}jGlc+k)F0V`s>RQ4vw9>
zc15_}T%qQ-0sz8VkNpx75<~<JV?a$9LBOqsY)u1`W2t<b@g9+$V&t%dJ5Q?WS(x#P
zi|a+^e%M0V4aR>kEl_;jy8Qp{Mg;@bd?wlqe}g*j{Abb$VZ{>f<RS9)JxPV|o*-^V
zh-_p48Qh=`*;uEBYGR#{*8#(z-kzQScv<kla34qkR-`(xpFN4~I7Gm6t%4mGnT0Vp
zG-TlBYJU%pJ_IctR~xnv`jV28aO6~2xC>}3ulPz@V(miapIHx!3kYmsT}7UllCE|B
z!%kl&x+*M5EbUqd!be{pQ0qkjtO0BMJco7#co*F6Oz?+YZ~yM9bB`&Oq#b}h4fOAF
za&q>c>A2b?j{f}QKhT3(@W-ddWu_oekiog`9v%eO?1bMb4pjZ`7M?yw-MA@$dsxs<
z0|M`QhL&Z3Pd%tiCo7IwN}QYlSn;jaGO4AJ3y&607wEgGv9WPTA0g5am;^xI#*^Kq
z@0t&;=4tY!igUq6wMp2p#po)M1xgj}HcLX428!$mhrhpnf(z^mO)+H*%w{yDv{9X;
z;OHo@u(T9~5k|s?M|N}lpFhJRj)52kr={6UY-6z>hx=O3&`{QTh2@b}$Bu<!Z=`{>
zT`Urn5%YI~`oh*5+KA)@<Gb~rmOxT73KQ*G1aO=}*_Yh7mE3XKDypdDUUG8+D<5<{
zl8oK$_n{tz=!dlPU>EHb^nc&eeiXt58eGib+l*Djg>yOxUr%*8Esgvt@)4^AdX~%C
z*x1<S_zXI90Bh?b`Jg_4(wW12oz6J@|B`hsT967n5r|PA)&rDH%Q0tC!Cf*{8`HG}
z;T{?W&K9CE=7ol_jU3P!mm!oI?0@ubC;>NbuJru+_AFLd8oX@P=4UoMd;WZC`oOq_
zBj+0iJija3eO1RNCKhETp{~NxRfi-wZW_OVVu|;jAEeAU8j~>(OlNj{+`SwwH-T;b
z)8lK1e;np$>da1yOUW!%nJ*10l%C_h<<~jO*jehz*DH3JF-s(NOS=TuWWyr$A+_+B
zhOX<hO|yT_oE=i5e){%j)0y#miW4K-g8eQ`2DaYLdfIe4^ytsZkeHnNq7xf7yc+Gl
zGorp{=apXxMNun$KkH^Q6VAkM6u2rn{ZJjWUwm->OQC6Hu0fFWft#YT8-zsmny=RJ
zJ$!b_;RCPNof@*U8X6er-DRxI<ekfP%kRX^C2It(@=tyrsA+w<cE`|bIZ{l{Dq~s(
ziLdZ_fS1zcW|wfm1WX6$t>sXo0s~QmxjsL9%7r5Ylabr=!qJG}?yByC0T};Kh3!ks
zMNlx3qz1vM4|~15AdUskcJcJ|^y$866_hLTI<v^E%YQr;!bCjoTXV?$f545b2F5vY
z(<&!At3Z%QFK6{~D)XCF#naAmO8k{o>NSyN`TWcDt&-y#itay9|De8YQRbtQ!(2v2
z*HdE43PmUUyIlD%DK8)V_C_;q`jO-IV1sP=-j5w-SEkig2JhSAFfQ05sWZDxZ6@@X
zR86_IZrhi^A<ID{l_JHH3OcicWvrw8N_5Zng`BbQ=0C%JGWyinkj!||iCitoo@lNd
z#glLJ<EGo^BYOV~$*ztY5Eqjb+`O+ZZ<8}usizB%)J}CeY_%5nu4*49B`H$)z_yog
zGStur6fKU(@cTJA844;4>SMuGHa0fIQ&XcMnTfP;v6ptja0gq0xXY7r3kPK|*4QNl
zvvW`b5L6cHvb5PHg0X?gYFlzz8Uv|khgx^VO_!FIVqD2i%(Y?6%uP!p6XR2Sy{&R$
z*4+JoVoZGqVZsK;wOv<fvE_sZW=3Y3CGiUiBzt%ZjsN2Oto_9Hz)Z-S$(l9}j*@WY
z$@_tA99#V8oK#s1=g0Hi?A}-J&zqKxy{WZK!s?o^zK5REKKGF>US+3{sd^##Z}QIW
z0u)&jtM2y7_BVCOev@_%wlX@tPMK1j-m1yP=~qQ3if>)>vB(xbOxw;~_+3J516yWE
zgzgTy4Gcm)_Z4-?ELU9mVRjA<X&^IjMo=jf_&%c0?kqpE49!hvXXgN_I88x)ck5f1
z6x)T+nRQ`|t<Ytwy=>05>1798mn8?#!%%Vc-XQRAa1G_4ZvTjaxrr$ml?>2OxVy3^
zyQh~UpE$}8pOH^y-7fq|NwZA4>Gv_Aw|~!%2&@TYk-yGTSU``4%-x6n^V><0iko5;
z4#}Vy!`ABd{zMt~C0Xjo+RmZ}g}D;M#te37gTupp;C~pEdCGVD{c)mjI%fK$mWJ+$
z!m+NT#~d4)S#h!^jpx>#JME_Y;;936w;f{6diFU>O;SFsT7PBGM14=t(;{vz+Ux}c
z9^wOzvn&v`5!r)@Wb7<{`U1d0%LtV_T)+rPJ|zIa6UHXUJ}|5B;@$D(P$+Ko|BghO
zL~(DJMO(MZn$7H(WyJKFlWSX2e$L_J7e=f7j0Ep<sX`y2WQiyWj!gi1`7mivaKGU-
zjKUC%k2>2VUn!kge7B<GwU0Km?!*aGi)<s$^P4e}C|2RHApr!(E7uHs9(iL$<G6sw
zMcO4&wJ$_#CB2k83apqLmZ5>q&|2kcw|mb2IL*fF;KPoJrNaAs=`N7W2L5A#u(GuQ
zYg=PyW;=$S;oJl9MbgMOL_<u+$x8|8_9hsW7~C@wbHAv17ySkNJe=`?tuLQH&o2Dw
z03HlHRim)ChS(P{G$d{>&|-c#IVib)?EJds6Q(8o)aye<oQeWmTnJT50OkfJCJLj?
zt$&{g>M%vFqoX4a|2cevW98O>n3$!{o;@>vXo1PopmHFzFV57T?wkd?0)QY31O!;F
zYNsASZ-o3?6W6zUM{(;bxBD1<$;a(Mnb-XKwPxh<=;3Y8keQ$wsK<I-lsS(sr2g*}
z*)Lnqyf}8`2q_qqm6e-YGJSS=dJOO;cbgtMb%K)c=%)V8=&xTLQHAg$sDy@U*G}<o
zMPR9~uSZ`oNOU4HG7_3AIh}7!q-H^ekMMFT<UKFjB6NdD9s<wDqD!y2+O}DzFf34@
zxDra{26?PEoRBcNbk^~3D3(?@3(2-xa?G|MuBT%8C@Lye!4oRsJ#d34{1(aDaPJ`r
z3A|C!(LbPJyja$B3$Fl=i_Aa;C}e46MTd{Wxsir^_w?`46xR>%WdfH29SEkCD_1J6
zUmu7j6$+?P6uwzUqGDsi5ZwqMlcLFtonr%=(9;kLm(0S3sCRfc63A*55)Syp6T@<h
zI|7<^$5Cb_)XgZ>@ef=DoBGhoi^b`*Cf4a)!74vg^sY*H1z09$F+aWr5B>J-+k>r{
zgpp|ejU!Mnoo}1%Qz9h9R0F#=_XJcuzq0uH`U1ZA2>M?Ci}dg1Sxu($X)ih{H|M>u
zap^Mk&#kTeO^?xZgDjc)Y!((DWZJ<c(gqG>WoO3%A_i(^F<2oqsr@P4jYoxBBaz?7
ztkS3Cg>nz}G8uSB(gc~@d5o`HXvBJtY&&vf^n(DElKYR!ruuO@?O=H0!$tWzfqu$q
z&8E5O&GF&kf)yUaZwNAg(w%HQlv(Fs{qpfgUjv0q`@)tJMSh2-TOCYgi6k2VfwY>C
z69P;q9`iwl4cGN5`~}j$+vwE611;%8*M2rUxAd7j3-|ykTqawO3>Wi*D?NE-1neXB
zjbvsD-Y6yD)2Ab__7lU{4iGp=YEN<~;LKuJk1+nH=z7(kPfL>kTHf>R8#{smQs9rk
zQnX+-kJ^j@&xu&WIdv3OgQ_IT9noBV#*IQaGkK2h){}X#B*DPUb-%eOTVjNdH@g$6
zWC{gPR6Pzv2Ef0>8I4p?MaP#RiM9XsGFmXS>1JWoyxLlnud#Si)pd1@xZiqWRDl9j
z0uUfFRZ5m}ZfoM*`4_tHI#0I@!+xCg-%MjNodAs*JlvoQ7Y-)aqrZvfg8*ub-<V*9
zR`>xBe3{X1z>w6SvuD+_XQyc2;3Fn5Ha3=L)3B5{l#|LCtAR93F-vC&WIDvh3cgt4
z?)v}LQONW=qZ_OqCawJO;d7796~Q8IRYdx<Ux1qlx`gHU(DU0h*Q<LW>FeK)XuQMA
zOBHjH^uF&F0VaWGZ}qLCn2aL&`t|F@_t+#qfI=dh>ZVZ2|9meai<xv^p~S{Emigs`
z|3=Ztns2lOKatGwfG(pN;**QfQ4o^i5VC=e9i!|7DLpVE)xa0{0ZL&n9vYBB36u{o
zo(+S4Odl$yLvMa#eqh491q7b>C0N9#bdFtt!4l~};(p*fkZzjeOvY)T|3Hj>z^nph
z&P4pH@e-Y34A{RumKZByk-{c$%5Zu^%H#Bcfv@<33lri1T9APW;}d&E+P0z#hjrbP
zc&fXaERZulKf4YbvNfk0dVa`bAXGJ18f9)h&ZbbT4YD@|J4-fDCI&=wOY4pR>?brW
zRJHUJpj3p>u|!Dup4XgZ#tZwF)J2}(J6Z8SiCi9N6v90f9*ejZ2Gir`YBWe@Z6JC>
za<LA5?%NQ)zl;u>vNF5e@|dLr8A3`=rfYoq^a(?x2TOi6Z-K?WDhiVCntsXdtmVY=
z6o+l6UN{9+LL@v9RzKp4tS25ClngnuO(>kh>*wcxwBcDmiR9wa1P8gv>SxcEkdXl>
zwk}#6H4H~WM4@1lk+=#*POgj^6Qv5gQe7h>7Sw(~#MU?fTNXm{*CZImAws%LXlIsq
zSHboPKK_O!3Q8vuw%NuCZ(+i02h6rQ>XsJ;PNeaD#|R`7Su5_mD5r0|es|~vyNCIK
zn>)J%Da7gwlsVeF6>=K@GXqtv-0hJ$H#Rz2h13HR>p9_9hZSOj)dKu@mgC<^tD~5?
zIpX1EPhsnLeqZD0fHTWXQs5rnLUunrWqR_PN$(_L6`WtIQHrD3KI;9BIUnn%|1n|*
z3vAL{ZBU_7C>Yz7AcW@A>5`E$jJ5>I((i|JS_31A(U&6eR#;fLza5Srl#AU-ieP-A
zNkPh#v7xlA-yj)KC{Qc8ln7!j8{En?@oFinN2Uz!%YA%w*x1+r&%}IatL;cv8E7Eg
z6=hAzKqp`Z2rdWR`m6Uk?>BuuB;fHmtJE#6p#1_L?SVE>Eo>Tt-Wi5X%tDbU^@)29
zh<yQQ;1NGIyryKly|<OV-X-3B+)AFN)<xHNF-a33I@T9@T*L<v-LoIRf7sxpPzNsR
zKQ|>!6v*!S3yEy?)j^Zg>mjn^8`&*e>T<>ykh;79?S`IGR9r0EX>MfnyQc4Hj_!FK
z$^OT(s!94X*`q7w%W|pWYfkkaq|wmbIChI1`~w7oPLEArf7D|;Y%#A~xsu_y_9;y1
z7VRi*Vcn6wOLJTt-~Q?6o<{?9k0R==!>Qs4Onygy4xRZ!m*sY-tK^4_Tb}A?k*dva
z`3c8>*iB+R60661n8)np<wa^^B*1KlDa7>pH^_KsYDl=7idW~W8%1N%N4`s+cK^OR
zxuzw7=+k_JQT<~QjInaba5un1D+z5BA;|pALiZrryzFy~#!k^xK`HJC%)|n6kd<1{
z{~M16g<^I=1?ZK4hpLeQ$!wj&;v+7Ke*~Y}9Y4;B^BC&^)9wdL{@|Iz&-RkZxjY%;
z2Pcr|U#;QmYPk26VE`bYS^tKlf+Xg6Rm{bf@qKas_-iJL^Uu%w5{>~X*m2~DB!bw@
z_W`*}hh9WCnIFhX3q`}G&U)S4%+CYp51<z$^C02kMNt`<EV3b3;rf9ygftt$kk;A!
z{nMvloDtSjWBo=r8_=^OqvRjf*%cqg<T?1@3K9Ai$WCiJ0h4^F$ABf$dB!i^xkQFS
zErc3k)I}Ukfu-ib!j3$DB=`@#e<@(;!>#ddxdzn?X(h*nn?2QOJnEGgq!eUv`&n0)
zFolpJG0!u(^FA6Z=s5t41I>rN?9_<Dg9?C<WG)SIFHF`oun}0r&5A7<33_NBz)YF<
z*yY)sFF`TFk=uX}lvQBy?&kry<;eRijvj8sfD{UaWK=lYUEjkj1GBvtDa2=)(t{qn
zY&*}(#>Q&2PR%E+^dxS*aGNY{HxJU0tRa>_IHq0SW4+rVMYh3AHKl6@yGm&9hId^>
zen(mTXk&h<5{R{UAfY(6Gs{;JZxcW*Yk!Gmc9Y~1No|u7^HaP5aSTSruuLr=cZ@na
z^?!}B)z}xHTS+LsWNH-}p*?VXKpu`!ACv%Bc}EIK6o^F_sVxBRYG@I%l6aO!L8_14
z@j({hpm6)cXqAz6D%r231&r`<7A;8_wMu{pi&>6nR*>n;09Z37&xBQ!G4gg!=NMq;
zpo_@Q&(E?fc{#d2%y6$HQ+@tcgB)l^5}=U$Krj%<7>OCYAp(9lV^}+xa!phUV6-~_
z`ff%h<`FEd<py5KPJjWun7fOX0}DnF5JP`rxqaitz496tWB{b+ibofW-b$MP{2Xx!
z1rka!JEGA+GXcpqamEh}3`AQk4D4cSh~oIS`)opXe2REbV>mdeZGbvNlkNbr0Ai#8
z<^>)4Z8Jcm$sY#h5(f$ec+YIeg_Ne<l0$bU4tzyo`GX*VNRbF+iH4;!Opr-G1iha_
zUi*9n{+=cj%O`PS#PgfzpkaUtQ*=gj)D7=4QOHODtn@32R`i5CjiI_^3?_L~QBhH+
zmAn~#f$V#zG&U$mQ(jUmOXRKSJp=|u;tmS3B-}wcC;;i}2#*u5O@u>ayME5EuG{;B
zkQ^R-@{n}EzP^1ZxN~ym)|<+Z^#TR!H`s3iuAb$8)AVsgk&tt-w{U33b2M71#d|?c
zBz>B@I_8I5)~S8UW`|9N1Y!YU@Rk33v4ui`_zP|2MGI4QKq#?-7p>9XsZ`7q6ej3b
z%%r+n>;MC$x8g<9<EHIf*8CayIXU}lTN|-I!ZdASz=pvF`Y|Sl4<|Upz{S?#iv!Mf
zQOYi=x;{Y+NKhq0D>Ps=WD>~iQH&RFnuS^f?4t;r+}P1k!dsR<e*Bn>NFuSFL<gX2
zhQ|c(?y%?*N^s7H65I%tGF;8i>dTOMc|ay|%Rffy2qWx6A%ajHgmAiMAOod7vhpBy
zA?1sIR)uCgll2&op%i~+(w!P+)$J%hgJHdSr$)@-Kv|6;q2$$-g#r^N$vGMKiJWke
zZYj?XV8^6|YKl^4Ok^u~Ce=(<O0UtTB&p5C%3X=D#YO!AjJwHKrnK%Vcp-;=m#-5*
z4PpOH_Ch1No4U2&b2LTzxbi-@u*W}M;jSnd1DNWCql5tiTOT6<bOtjAkD%Z8WfPuW
zA7~h4Mkq2zl%2HRLqVAERAM+nc)Q4(b^XZ5Q^Z2HU$Y<5dCtD-U)%=@59SfVFbj1?
zHT=8M1qRT5z)K-x2vJ8OHZTCtKlH)_+mTph!=LG$Wy|Z=#3(o4s+I{PMe=4IKW0)@
z?az%uXT6#P3j8sYIjI1}n|uYQ3ySn-WxQXUtG6b%f(^P8o_iQM2xm}>USH<6M8iW@
z?*>e}X(qNWqmW>M?ok*J<f9V$$h}4%3!NkcQgT89lejp5v8<*cu@T`J`PM;s%V)|k
znw6@*jC3kdsV+wB%Fv)?wmZv~xW8L%`?THCsnxuPplFAz^Z5}dynDBU--9H_TYgp6
z6>MVTJ=X*U1P)y*MlH`9<8b^q8ANH{hRJYQ?rP#5`ja<!BdCA~J>;eI(!Y1=4ouWF
zH*<!DhUV(mc6V340%9w;p>4Qsfs7lT)rh?muEWMm8P6LJ$aaxFyU?W~C)q#xl)gmj
z>k}V8_bF0#{JIomY1?_Yc`NP@cX$eUP%qX!Djq?!1wz1i7<7)HEhIx(o@d5ZbWV>X
z1nrmaYAL-MRQ31sG8CqIf;a$s29d~!MpIB`=9+ui+0b1r1}mtxs)~YD#*aahpz(#o
z>}l;;U@GC`+V&eB6!CS^7T<X(MESYQqXSD8&RZ_7ww3<@McP*c-cpz}?krGaQ%lK1
zz{t0HoN3LN$O9Au*{qUGIrBey8UE@X5E^RGpOEd!g^3qAeo*V0EzS%VB(u1P@gbgg
z!uK*H)MShnX!gZ%QA3efFl6cySShw{aXOUp1$udSxcT9;TlqS!d4Z;rs$JLJviw_V
z#|}BJ&!>_`%%iI(ex4%a6&+vtm0M2-pizkdB3Z9)|Leihphw*7_PJG0BQ;bWM<i0f
zu-mt#0Dnx@_+s#41OXC3$H4sEob+HSbzJ;kAEiY$;$>7#La^@miGJ2dDNl_Cu~gNp
zj0_g=nTXy1exM>awIw!h7JwD-lg37Mq?*g2RYX84e!<7f73OK?1m3^>rZe@aC%1t+
zmLgi5c@c>d877QY>*@wTF+^~IG#<6TX_Y&bJqe=@f$jD5_NJ@=#;}5%4+wiTFJk=H
zN{VMaeftj$xqK<^42OTcNEkiak<WL?lsI0vR|=B;I!=sgG_bMhLhxC}MRs&R9&r-y
z*t2J`px}`IUQn{ud4@8(PZT>SPW_yn>`O9Ty3Z142$(jjRSf<QYot#Xx8DlJ54>f~
zPsl+rg9No6S{ugQpiB&5*R6c=^5srIPk{7R8_fb^A%vO0H=uN(e}eWY?3n>7%iGPO
zZ>K)=^vHU`0@f?r-qOk?@!gMJ;YO*r&aCvV%;wfs&OATCqM~BPyGhq>gucxPqknT?
z=U)9A&Jzb;(UO1gXsyy}=KPsmkt2Jt(WNi<{plNVayq_O?@_o6U-Y?@D~=2=TEDqQ
zp<8o2NE&oH-X<=;Z%b1}wsU9r(Sp1z_B}iIIQ{pBlmL2RlZVM^$Jb)P(u6%g4Irxz
zR=QlOZ1H*+t%F<(ea+w)jn+Z1w!~5P5ajXA^3QzRu)VkVoipM<U26$`YxC=6pKmIF
zzEcbJqNLF*tb2(y7|Kmxt_Z)Va`8;S(;{^MVHDi+^YpwgT0DGVyF_VCUdeVpHr4RJ
z@(zn*tNUGKUg@nLRg}tq<G`?^DNX;@onzwP{Q`GPB=sEkTE3FH*sbvAr=17lrp5)V
z(}Jh8V!of4wClZf)?`<&v;7zIDLToFnNFV-@4e|eN^52$HmRuyHT?P(aOu9{i0r$P
z^)nBLSGX64lzbFdu?S7_`?E~dxTf``>}os3k-M=4-}^q+zWQt#wtQR5yL%6Ssbwrh
zcNIMb^j=iKN?ej#F)2vq3<KuVI64D(Qrp=%&((*(Eig0zsIn9~OLU@ZQ2CL;U=LmT
zh*25^XB_iI!_#xXVe<0xX*X@a_5e|=3a2F<1!8+LM4}lIR^!ukRF9|0>Omy;LvDJp
z>`qo$*%q9gVaPPET)j%Sx<aG9vvY;8symsw06aAmBOZDX>WOtKdJGsSK&iqbMtVm=
z4bF{(0n9sE1B{8#NPyoV>s>f40g4#hjYS6pX%ipPGcw5MZ<8q7xVQ#vKqLjs>O)9H
ztUv*<`%qaSI)OefRBB6$oZcJ6O{2^I7DFLWLdom+ZrHE^MI>3#S2vtkRlxSD=ag%F
zGcAM5>)Je{^vKBL*>RRDLT_uzBeI+02Nx}gzHZslE>e3tJ3(|}XW6ql>AT7`O$Jwf
z6+h_wnrAb+v8Z?h@2$!+mlwUAt`D<wIxeN-J1q8Eu<p-LpY6Jbqoe#htIt+MY$&}l
z!3%VZtL>nIf&whN$N;($re5qlC<WoA{loQ7Sa9$TP+ffgT&nzqO_z+YMNvHVj34=d
z4XQA^*SFb;T?QDUIKgFOIiM{eL%uPO4ga%2i}IoGA7#>UM-F$*zl!P+1!d-bOije&
z#1P)U3OMvVcu>hhoj`ibl%AosyMz#lyo`>5(F3Rq&|pwRE&yXMPUu^KMv-g@%rP9Q
zGKMD!rV8vCZD?kK6B1qnSd%Rvx&d!~Gh`zzZvPdeOi}o<FDbrmSh<YGZkP3%K5_m}
z?@WK4E)P#XE66^A1s~5_oBH!GE!y4OeQV2#XMb1;1$JnilM^9b^Z*yfUCmz>RYTg3
zniYpca(RCkc(afVE1n+hVY@8pgUW*w^E4@Fq*b8@!G-^iCDXqhF8lRlqzGa<8Hw`H
zZD1)*y-<WYpvq~XBJ#AZgKv12u(HxXYV}-#&I2}1|6}}JItmN{bT`L_ge*djGRtx7
z`OR%Y1eXK&+4Rl!$dO<`devZjl9ZaTL@DWpvw%FmEpQAn=<Gg}Z!sP`tZr~IOqMl;
z96n5Y#V=VVcNw{_{}!F|nSo^JRCY4t6Ts8O=V9OophvU+%*&>NA7r%j+O-GXHJ=th
z69SpijWKV$YdVJAA6{~E2bUbBkB`H7!G^{s=F*5X<>Ry<&=ZjJ$3O;?byrg;0L25+
z(^o-mVo}<SLKsKc#SfR2cs9?22`#da#IOhGNNOAyPHDbqC${Df-S;~?S@7ru6&bQ`
z4;~o;QUTJLDEuOE;v^W#VEeKVhy-?hz}PZZaImw7=t!c0H?XH&5QsSn@TO>wk#1oM
zW+lzP8d#O}k_=eEl%Tr0IwY(Vr0!%AI>-z$NK&zS7{J+i2rMAPki(#Ue&vV1$;b;e
z*f0I%Vec5UA4=Q#6q=;)2HVkl0ed7#ZbR#8M^;3}lYnK7B}ztsB2>@=D<Kop?;J=X
zU?KKH>8X*W7>IE>C;jf-yX$YKr2LTi5iwNZzjlKc3KxKJR%AH4vf~xkt%#GDqmdvK
zJimomOPs6D+K;-AD#=(m&COMD>*1YKF0x-%pdT9&qjBrV(blEP`?l0HWo&cbIDzbr
z%1-$(08mSrFzycA+*%DVh;Y$J#Y#~Gp9@D$4+0Yc>tk8@wY#ajZXO;XXV22XLy%}=
z(P~V=%CbX?hgev{kAjTE29O3(6}PGa7p=k#RVX;e7K8jUzh(8P)U|8I8V_x~ke-&g
zfAQdMb{VyRj7ua>HLCkSC!T6}`Pzl>QRYV}7cv)rHQBV7Ohpz{M8lbk(7LkalmetQ
z7Y*02L;Q|r4pDIdYDqU-3&aeaDsW6nm93X4k>7&|UL9`X;IJwJ)$PAebS3aG)dud!
z#4gC_<n)d2=3P)bDFJZ100>{A?WSS?Pyxpd+)+bBC-$JvsjDlh<CHqR8VLxxZy}}%
z4gZg^_m0c4{r|_Wq#>d-5G6^IGFloM(iWv9Nn2XlC@Cd{1}P~T8nlx%X`&L*PNgM@
zC>0G_zvq$r^Zoqs`~LC09{1z%e&6nNUDtV@$8o$~&-Frq|GiaICJ$xL@AbNFTL)rx
za&~c)+MfAf3|qv|W3&Qo+jh+`zpLR?&P$^2eDVnEhbM_P2TKm~rO2>Z#J8F`IIN+7
z%p*fFoJU#(pE`7|2c#e>8;++Ekv_mZ(I;_#{r1fqs;AvzVp*X<ic4&WfrZ!z`0^^<
zC3NQ4N<(p~XX2h&0GlA47mTn`gCv8XK1!GwKMn^`O7kx=G?6GiK~f>J>@q+KGi=r3
zo63OmHeEK_S%WEZE7L`9zb1_(q$X{Wz<w){@JbME3W@YxyY)Z_qMi@;TcwgZ&}-G6
zZ$%VJVNKjNcJ-Co*bVf`nq9YUpRYPAAapA^N{0@&l!+V^;wy-%3}DoWoc&ah5(3!B
z^#|Gl*m^47_VsNA=LU!O6?_6npn)OJe)Nuu3caSBWOq?4^hD=-<Fq-M>&iTf7Yv%9
zr^GhU`1H&%ASH(TdqEzcSAiv3(|YZlhB)8%!Y~&;pK3or`f#G70#ONKO?XK*UXt$5
zaO#0;DZky)P}Hs>ri4>BOO>Y@-~=GSA2>;bYoQ{<b{zyA5;<4z=CDG?n}r4f1-8&2
z+BKr^#e6;l7*|TlC3$HXM25f2Y96xt5%P@w{#7-(9oH=+@~^4w`gTFR{!3!PEjce^
z#(_R$$E6t%Q@VY7md>UG>_5;YC+`(m##58_0^tY)5=iD*;Aq;tTcB)*@}apEm5{@c
z$?S$&N0!JeYvfLvVI<nJS);MJd3}nfPaA$IL+)&(8t?hPPd~Ns5>^N}*xqvJ9SP`z
zsV|QU3TOzXgFM=_>R+fS%avStl1lmEA0P-37EXx^CdHt1EYTLDqw=%Ash<z^Ba&aL
zTTo$s`La>kT>x7snH@!`#4XI6OyN%YJ5<$Dp_3$VwdC@y-zkT26zHYDofo?w+<Ogs
z4K^8N4-*yBz15cM64O$c+Eqq75mc1qm5tW6{tGvAn+Ogl>>3k4BU;mXLN2Z;d0srp
znXY&o@(rBc`hOL0Y7hlBxXN8e_v%Ano{y^Q6K1l{0^Xvc0L-lc^KEWs<_8K3_0{V8
zH-k+BfffJojo#qc^H#B6OnKx&Y=VE5zq^UAJH!6K48sx@u<KD5xh5l&j5ZY>Jd*j~
z`6I#QUb9KF^r8$^zTp<D+mD%7p+tlNF=KB|Zf-ahd{s-!F+aWcwZ0P_H^C&3(b43{
zbVBP!rV*lj+W`#zN|=grF87i6P>Vx*+{GEqoXmvXx9}8v;5g@>YJP$>s&2w?-X&5%
z$kZ4Q!x7;l?zT-VWm!Ra1U10x_w_{~`}OR3>K|?jvN8a|u|%UI)D6guKvc!go~^$5
z<JD~LB8T(Th(4K~vO&b8656-`)SDJieGp(ex{y0#7<Z%pj67$KXRce!nhZm)r(4!Q
z#*s3-V<3S)n(Xx+k$Op(*6c&>2zh1DX5VAu-m`}*KRRFi!%@KWmPBK@cCA6DjfDj%
zUrDf%pH<Xg#e8Nb<bnM_KcT6|QO!lh^Fhu>5SRza%sS%CH(oY&?1j8jCrVBHX;X_G
zJ*D5Neh1#26iWYIY33}~{G7fVSJMiq4SF6iUII(=t%wNSK@Z^9ICZA%WKO~6L<H=3
zilZDcYe*R_9uQ~)$Ege$Ioh&-MB~)u!8yE^vRgu;hR9W+xTKs1P<x>bcXU;?ULVKk
zy>sYJ=h=L>Ony>Azje#cb#+4eN3pnnpCcoBw&HO0l`FB1>%%~2W8>AAwt@E@f;WuQ
z?sJ>LcjbpFX&n0Dib<CxPb>fe@)4y6^h%JPm_eM0!knZo=zDD$A;1tuctfPW4mb4q
z<R2oEhl3t@?+VuUpk#R$FSi=hh+)pL+hSK>765JmQL-7NBJc?v>*NA=hZe{GuoB2<
z<<t4)YOk-U7Ve}z^6OZmQlG|GHDzAE(=?l;IOl}5rn1S!jb{DlRQ7r^+|mQX#(L?i
z*ZGh9GRr#tcggy2^}Qv>^_S+p+1lEvWfb+^<_``J&u!j==y{?XLYYlskcR?~EZw<%
zI{?@l4*BPkE)=*4h(U>LR{(6BJv|Ao2o#7MBv=jGkR<^^Uf*k{sOvvfu|Tdavd4C#
z*KP!w9WC7l$JlO*m-Hy94XjT<^fnHn0sr+=)ZPwS33~fFvLSB)M3+9qiwm`Mgy&A>
zD=#SuO{}dX3-nH_xKB*8n&-`N3)cr^39Mr{w2F{5%xyRSbHjt(N4Kn#>;}{k)Dr#K
z@;`jopzv(#S@MjfCgwV756QIe-M(GH&^T$nTq5?0s3y8E64wb_=fR^#hRL<iGp@tN
z5d=Fk;hSY-p1F1-@c~n}zX=Wb-a>VU4gg}ogZn}JUxi+8+s>U?^@e{^LCXU<CsKfV
z2?!og=Olss&b|LkV*ECRhew?3H8Z-oZt5y_OtH3`^Y^Q@s#(9;>mvQd5bZ)zU^&}g
zj6GL=+qVh(VXPbQM26^I{J-e`6lGvLSn%aS_ce)T+&<;CWXH4u&6%w}RL5J0S)7bB
za~`PXkQ>8Zh(&r89ymN#nax{PQqT<|AVz(3OB5i}s^u*#9={f!egFlFU4k%Xmr>KL
z&3e551J2F+P%HLg$HyIsgCC%gIXLaQ>*{=GdASuu5$_);>*)TW!tL=m8n%<gJTJ9D
zkws4Pcm5_CMhP<i$n^DpI{}W$2gyw0(jdMFqN<?a2d`pAN|yS@t*0boui%e?ZYW}j
z<Xu|uGAVIQtqEJbW{tAC`YIxUVhrt?n}IsO4f6?3rG;R2*+!7E=!q0lhp{M#U5$`G
znGT!9Y?**KqdG9!b-{l1<|_p|@gQ34@%^DZrLOhYk>VF+E@OMfcJ+M>1Wg;k7hz(5
zw(ig57O@WdYm0@+54N?lN@I3}P8p_;`eX0nL3{F5Ksi(!pTr>7v_Y6R-E{p8it3Hh
z7GBJGdxg>FTO%VQA(%Zf>Md<BRSdF};LB+1)5AXKKfw}#rpw4$7Gi09WYOoomY^xI
z{=?BPL^%RLN2-BQ)E%2_;ScFNJYJXAA42E1Ya#n)d&3qKOL}fx2SSx?E@eOwmv;4|
z=CjMP7bIf88z)vU-U~n_<%Vq$PPf!)x*M2f0uchBh6-a%KzzjVh~--F?ihsGb#-bO
zRA5+o7BU_J2o_B5$T&hoZa^ZTGEVc+g4YUS6#be^=_Bx5na&P)#bms5OuCwtBfhE!
z*xK+g>D<iWb8Wk1LWa*bHxmpZ%F!^#qzXH0)G2KA`hT7&dKf*iBC?#sy^G@m2XQ2k
zncza*<QxPAsW)3H4p|5pZh>v1Extg4bU?pXTsuCq4GZ~!=RUEHd?x9&I-%!pPDn0l
zNQG%m)asZd_0sKE)~j$A-|F}7p-*F<=Nhpcx-ir+>rk_Da)tsOe*uttdU_fHwG8LZ
z3TQ*2>U+`My}8e4h7(X3tc95-cYa_`A*~r?;l$oZXoL?RHWF49&iP4MI2Zs^%%bu~
zK@ke*l~4lz$>?rHMWyO`;{hd76KoQVd6uf^B3`t$6?`lU0z5<ug&9~^<DuDAEPM(=
zQAsEOaK_*l)(_RAJ4IR!^l)<-@)}7(0IrPkENEgS92tn*E;cp^6+dDYbUTY2#>fmp
zXk~$D2%O6m_4x9-Vt4`<xCz{tk&#iNex?D@kl|1CM>1rJR6{vvU<QdA{0Wa6J^~OY
z3I)X}o#SY;emoiW0(L!J&v>9sPB~N`nI)L8n>`tI@7~-;OFYnwZf<TU`}Du{UL4sH
z9Us3o@ps0Hr2=T6(A8;>(F6LHhQm++tjD(Up#Hp;)*2wKgvZ?S3}-9albf*_e>JCY
z{?`Htv-e1E9Wn5|&BEN~fA{wZieG!pNg8b>J-4iouI+oL7~)?yGv9J&2PbeP43NIo
z&h^Oz<+LO>G76<Ms7utIL~TI)hc3hKIEi2Z)Dw#8q3P(?R>csK6JKsouh(i!hP3T^
zDu~$u{h)zBPZC`jb}nT|`(Xr!A9M=>Yz1?*{(05z%3jzDr|$%k6%?WG1IqUT$R2SK
zv+!#R*o)wd0A6|!cp7nQqOd3F6tft>OlGQ)^bUZe7kaPf7~gzth2Kb_u&}TYh6wCF
znQn=!E!Z$<`S~Al%&W8LJD^J>lewWc3<iLsq^e2>**!{zg!*#p_Y=fB4GA3SQt>cS
zrO6-hYyN{Bn}!6SAm!<W9~Ct)=};VI7HHlv4n!60A0<)4ZLaUvk3Pgjz{>~Gs8?Vc
zfEtp806|wW>^g*yh5Q_*cGK<hmPxTklPFHI5?27RmoK=BiHY;Jc`cGXs;B3q^*j(r
zs0B3xYe{#6QXdERD>4QYU1ot@x|<0ecl1$04%{NprQ#OTQFfjzWsMd&shc_W#u($8
zap0q>yXrcK11SJtD9&$UBEqwiQGIm;*N<@K!OBS-4aZxSd%-Z2CTb0buUs`^$&T$p
z%br$a8!X;$A-Clp4ta0)S&P+f)SDkzl8ReLj!XfB-pCxk|E`n^52X@I8hdLMrX~r=
z!l+IQRfwF;;#Ti-TpRRDmZ;|4-Z|{t_k(`TJ-;WxerGHUM4~j4Vor!ypRta4cs%B~
z#)gi$XE7y-h36ZbVp8td*u<Q;{^qb-$444AWkvSR81urutu-H+^p5sFzSx@^ojdu~
z?aQ3`XQ}h$zhvKkZECk9PD99;fP}qq-8VsiAE2=oFug#wyatU`b~pymAfc<S^yhmr
zSC_P7c+^PY&E@l3hX@s+;8R4YiiYyE<5e=w9i8p@(Si;#xdoM7QvIoXOy(gRJpyIT
zunx#Q1(JJm=+`wn4QML~dqaLN<Xfa~bN*#aG@lqWdE(mz#QdR_dh+gbN%VJKWET=B
z4^!kU<k%4ubUOY5EqS=`ktr0k1Ub_MK9WEHf?iuy%P;*of~VX9J*WPG=(GcAqhGiw
z?q3_Jpnk}8{`mg=&#>Vs7<;k6UINmFgMLz|<5~9wq)@?AE9E`ClPG@2#+-b1B8x~z
zN0Q%G@6ne9qvz-M{Pyw*oLR#2Gk8mtOA#U*@&~^xEG;)T^zApXh(tE;<W%R)8yW}$
zSK4NO_q!U4epJ5WzNgd0C5pc<w>C6xvOm3vjwi!s$5oz<T5L<q2bEjD8u`uNU-8*m
z^Ve#H66=ZBC*YrwusODWF^hq!U_DT{sQB&lD`4d^8v8&ZKropa!Euo?IR=>%n7{r4
zAyyzDhq&M{oXKiXq%a*~MqfcqA;ANbanJ$y32qUQcU`{)_Q6ea-rL&}n?u{|^=l*H
zVBC??Vg|Ky`xkEK{0j;x=Q{pJ3vj^cV{vzU2k<%9H}gGjJU-#%HNRtltrN1Ptd6bD
zXFDWxENu!Ys<~}WZcG%M!=S%(2BLK@%sBF67S$STgG9CfAWbF8zIXIF7z(8BZj+EG
zG(M&2$h4A*+*N$QQv-Q`Q1!aD5b|d8VK=7!HQ?O4CtVy?9Xaf_v4IV}o?VF2$E`05
z{59)#mih-67Jh2_azqpGXNLM|f$Ddke?DMh`wAh8#-+-@V2(YQX?%`Wz3A!fR~Cw>
zCIW}{-HQd=iQ=Kp^UPt$#Y~=gW>-(t)z!IbV)%&_iRDJOoBjL!QwPDPCr&HrN!aH?
z-UBx=AGpJn{a@Y+P!M8qGB6LjIJQdCt_zhhtLgPc?P@dZ6#qIn=+^zq>7?gl=??YI
zODDiuvgH`!R3utgGU*qKEktEd?qs)eThwtQ`4gO;8vA}kEuAPV&D#I<`-x`_RdYk%
zzFi|~O~euD7I0Xcllh5Im`Y-BfF^ks{!V}a(eCTybyz1bVJZm`%>V|<N2L^*NJJO$
z4z6u!EB||;jfuI9#Z`x@B<&`8(>Ppb&Wi~zzqkoIi_33{cDk?RHf*3<%abJ3yy-^X
zz6-k;e(!gF>Jsou*X1Y5^cv;J*SnW$Egu&|KU9>=$Uj#SV}Dp5szC~JlqT5(oI2c}
zc?edaaFS_fes%AL*=wtzvnd&RZ1QW4N#wV@ex3f&Em}_1`u#3?rIIg)ti2QUJww#&
z&Xv;sZ2eu{BW@J;J>|*wb;TDLuP4!mssy0vm-&+Wqe~)23t7<T%J{C;E7N#$cYGD{
z2w}+S6S#Wg6K%DuJB35r?eSY|EWv;Szq?k6yTqf(#8hB#3TE?zTbC-qhfE+b@Bum$
zLKs7c8-Q|zD(zVm2Y@ajd)v5%7C2E(b1fpq-e{cWkZ@lnpKnFQ=>~WnIKo}lzy3L=
zu^Ir_47Ia66)h%b>N-xpP#n6Rb2%U}qPurNna6G~BxmGcg~((RFw>~87&-64f(J@^
z`bb2?(jy-0Dl<Q&iHdbbn4OEZuu$XuAVlaVPMm=6^?7sc8z=UwSMT&~J+NcPtJ2dt
zG1$1MDG*3eD7%b5&X-IB&$LCbWJw<KGl}&d_DO`uA%BcQ0YoVvsf{o~oFn>wje|#Y
zJ7<0ama&KnHGKG^H2wOB2|u#&D;vK5G(bQ;;S@mCM`~CX<T}#P(S1dtKR&|80;1v<
z^4-<?`{I=Ei<GdV>8X9AwbOl`l(c(a61nM%94Gt21g@WxJq-j7@&j=gao>oK{+yB}
zxeeQiW%yk01v*5syaa=k9S#gfcB=Il+5Cleo-?@zf{-xjsFTf*I?_Ify%lEqCnBeM
zKRj0aK-AvJ@Z}aQ=xIAX^O_PuV9k!>#Jni`=kslS<IN=!R#sMtsKX=LP%KCT3*kGy
z03|urEF+`?5v}}5rNhIlE94*5&s={ncW(Q>4J@jQ)C{R&`M2Hn=2}h)2P7U<`&}Lz
z6VCO5T`B$Mm)j9Bl9SB8{<^<DkRB6v;^^SdOfSn-o_xOHvL_|B=!tygcnO>?stDqt
zH5+$Rq~pfI2B9mhr`@K14}n*xc$Q~T!Ki^=v%MDMdyhd8Ykbe?9M5$g&=z6XNBf~y
zfA!dW=~-uH%QqmIvWSzFo&I1s^zKG@xN5=ypDf9(s~H#`n^mp&(A7oMrxZU}>4?Rb
z=-!dLcibtV5RnYTfP+$QrE~tkDhh-h{+$kmz4IR-NdxdE^jT!C3$n9Q-&`Ll7i6Qb
zr}&|yr9#lklS}VGms&v1u74>W4}x{rmi_){oscr~$_v4^)+EJ^1heWbbxqIA&`}^&
zOg0H{Zip3U+j}F4zQDL)NIktbGczW@w{I>#!)Za`Z8jn|;C#}v9FB5vb*%;%O!@eH
z34Qho@DsKLU3QZD?XSaqPc*e?!8uTpT{X@oY`l@V`j?t3jMcQ25Rie2EVwk8o%8DD
z%YGEaIgfuUU|hj#{yhME84^r0gyDdC3EDDiNiEGD((6jJ`{*)+)fZs3u4*e-Jlq6!
zGEGdIBP5)YCOh{L?^GYhp4FCp_ZZh@i~Y3U`n-#7^}}^n6N+x|B8kv6IGn$HWr9uf
z^P*SJ50t6&8ES?~^YY~$iWgbU5w)+BcCCNt$DEq@1n6TVd(Tr#$E^j5gNzS3B1tum
zQPE+2&}O$oLyfTqyJ{oR?1N3fW^xGO4C`4~CZL{t0shX$t?f=(JamO9M6RNfAv5Pm
z2}8zoqu?Vs9>mCo8KEU^n2Eg{0vTm!stCXj$OeR~SAZvcpK-@cCbLm5PIsOL2QEiu
z34-G%B3&qTFyi1Kq!JW`Yy*;I3H*#2Pz^z|u$)QWGSqqKmIVzf02C76^Ta^VRvsh?
zuU)(L{Fj*1cK!N0GGd8yh6uxvgMeYO1h4`FKmtY1J>Tz}iKoT4<Jq~hL|aI*^HAkM
z4N2M<tvKm6>4(xC)2LK;!(mQnSu%!%JoIEE=DOy~Prriv8x&~FM9c|amFdNaL&e2v
zzs@oy)f+9^t^GSI7GixU6*3ztz=i3Sl;ZuC<3U?5RU*BgFjY|Yn~tazT*uR*kgiAK
zDsX&8w4Fuh2r6;j?5*p08XWFC`3rC~GNl}__xbb1R*V)XvO0ep65ObjIx9DHTI`(D
zt4(^!dcvO>HpLc=AN+Nv`QrG&2S=;r1y+_*H9yZ4WOIGWFs<}P;moC3!@42*)xQtj
z3>RZGE-^7k{+VewDavgay{e%3n~mGis<OwyX~da{#%tgje|#lUQd7bXhUyVx4MGq}
zW&z<7D3xH8A>3P@Hwu_M`QvQ&uFLF5KfuBqz#&caf^aFUqcqwT5O)f6_9g@Oap5jv
z%0l~bb)?V|oD0(Z9@jyV{z>!zIJwAHO8ID_odr=Xi8M%Tnm;T02EO<KRDd6f&TchD
zkP#?_0b)u)<M$ZT!dx{|VQ}~qxWTj%r1@@?dLTF>7?m=P`E&OJl&nFe6%HtnIBkv!
zMV%+>&*`HS+Op)2GY*|ghMtTD!_riKWE}%SKtFcXqv8c<Ig;)OM{iMB=DUI(;0;7P
z-JGIsRDN+kG1owsfMJ4(_@r#E;kYp5%)$e!tn(7h-v{m#uN{Kjw46K_^PCy(1uzC?
zXFI0m-`TvKsNEL%cFe44UM*5(Mevi~l6?8h7y`p^yyZ_NMdu*X*Jcj(8=H)S6g~)z
zIQi1@*hxVQ3sv<sJUd1s0<Cq7;Exe!xf3lrU6h9PNp0``%F9DFudSZQyc2+%?-dXg
zqDaI3qnO!da94cx91)&iHmeK7FBmBTuc~3!ERJm=WI<lm4H(Nv-jtZw)8=(!(3=im
zXRAd?eiJF7Xfva`PfQq`@|j{I>o+d0VRXORdJtN(&?!R`PyRND)Z@|g1x7oEYTjXS
zVnw9=zL4P-N0Gc2ijVOMNHgh6SJKk<dZv$zm~=y*Q;Uq7>)3&+G0sMM^K5%A4TTH7
z5wd@vYYBSv=yV_-->p_@<-Zo&POsIKdj*8JsgKvX@1vMZo%o2z_G__W7f(uDI`U^z
zn1s{$I?*p<{7gD=46dP((kqzT+@MXZ!PY(g(e1r@;ojBaj7b6rxqPY;ubJFL(%JO6
zPPI9zuT)H%WRL@tcg{h3F%IrIpgoylchU;s_Idy+Yi2>VtwftrSyS_{-Xqh?F;h=S
zS;b&s*d*Oxk*D<bSNc=WKm1gbpiA#1nu@QYorgMZX1=+iVYB*d5W8;ewhT#S#j8op
z7e9Y{GLXJX!R_>!JNKc0Px58~6#&=+R}LQOwMf%Qcfa~7O(!|gz)fN=X`c_<veYlc
zrC%Zz>93;d`z24Wqb<I2kZ<y~oU(PR5dYIA&sSB#CtRdPC5yZ5&2v9LO1?G{+vMQ)
z&S;Hzt~^i1LlVA+m4gD4EFQG1hTn>T6cb7tuU2J(-mOe*O*?2ecATP+sCnh(nc$HG
zh3NL;JHE_k{%m+GKV<06=tY}!)aC33o(xWve5z-ni+)NzBBiu))S>mqN7R&C{V)lH
z3n@0lV@6^kAYo#6z3{4=3Qx%Tl<lRG0OUy)1U%HOI>E79bXZBKh{$(A=71^+h63YB
z(A2Gf+O~Ot`&!z7Ile7@W$NZfW>4;7GHsBt{-3K|%7=Rb4)ome8RK|zJ@?bwDQerH
zCPlpYVgDwFmt!}+-#Q*idm@ly#maO@?x>SmzL7vkd_^;0AmkqGfrEi0bEJkTw(_p>
z_XCHa!I}J8$et-)+CwAaFW>EJj7mu={v<1dvg4`}JyT<K!N_Ja7IAT*5=QCSY2{C^
zJfvbKWMst`o%06y3$NMTe(mrwUQ?bY|0{#T=jpE;Qo&oTwk6piis3Ti^MMhIdax4a
zjS30!*msa~9Kt(=!RCbRRCh!klF8;Qb`WvN_~sw;u*Yqalr&7j>;M^^u6JKqBp6KB
zUe;CJmeC<Awn*=3zgmY$$0(NHV@s0CsEwO5b<T!|l`2j_DA9|b3NZKFCfgUJm6$}D
zhpS9pY4QDQs*OpS7IsRKn1<91Id2Qp*nq&uE(ZU}YJUDpllRif^YqhF!xUP%mCqmW
z=k8FG5DK=th40t<Mt)oJ_+C9<aXL;1la#PS%8acz5z(&7b3^5d1xBg|;8&;2O>p<k
zU_w$<;<0l827MF7Qkki!`T|O^v2dgSyKg}Hgg-)>s(X96U?1^svVEEV{=f|~`2|OG
zlr#stYCgBGxq^DI>kN%pV<M_`S;OWAa<&J|28BE@3bPXL7K)0PLX2<(!qBKqefY4t
zw)T+-h{^3cZoh`6hzw*`hhv>8MvD*EnJ4M0KV!A=WArZAzz?z>5{S|us125eHJ&uQ
zGF~?HZeBMhPa_jZy~Xh(4m@SXkHBwkMV}i^t)fv>0A4%-C!)9~P7zhih6J?Cr|yHP
zcu^WV{`VcmTaO(v2LuUQGA)G+R6ubTDaA&BFmYbvu^XDo;#OdlbXEW9o4u~$Q<2k3
zcIKlSF<!V!b3)6o+qqvM-IBR!7mkMki+AcPpA`gVmPNNBeBkLD$#<2uB8P6;r;s~r
z_ZA((V*qscsB~z;mZR)eu3VW27;ycg?HTbX8~Q<O)&g2I-@Q?YGv8~<4uF1#%U(Q|
zJUkJIFDpP$05w`SWaJ9z39&_=*_QE0#A@q{ws|Y5P6_NgUk`j{ePro7!WZnnCdboQ
zs?&A9a?%K9Bh+<Rzj$3Z=2Tyc+7`MkGJ6^x%Mh8cGM7QxuA;N+FxHcVaKWq<rn`c`
zwP<2(7s}w^MQ?`T5CJ57@vrF_gwVV|4~wKRR&wy5vpPRkGE6ibU}lW759{3~1p-7#
zJuV?AvM>+e`#)PKv-;skX$4Dh3>bi6_{n}Xy(4VV?ES`zhl2fsDbdd9gG}PKBA)T|
z?$`V}Evm~SPnqrAoEoOwYK9S*^SAyA7oM7B-$WDrJ<AEa3ZkO?Aq0ht@uSNRET-$w
zM}Wzqf!NL?a{M9k1<1dFqac}Z;H_QIxDdZTY}|1A9F6-jfwU_W595HGJI<U;nw#=n
zn(+m|cL?#Qfct=@gjA!yF+&M)G_#F3)bLw&Xlsz2SlJ^GPlK+W-pV*_!+}{GzVChA
zncJve@8>tWOw+q|m)Z8B+n`kfI#jx8P|5>ulZr1_9g$LI^nYbh1&mGYwcKjOn~EaG
z+Amyv{#+JH@A<;;?^9mjPSe)1=sO|ojzmKPWxPQ=a;UXXJXZl))P1xcFu{R8pIHc!
zqWyh+PdMCb|D~oABQe*7&&-rcsO*~%eM+7+5;n63<Eeb<{mAkOPC;VYE@>3UmoQA(
zwE|rlIo8sXnZ?hbs(HQNcLl|`xV>lZ8RgIz6TQTMv!e7&E(gBT;|W*iyq0E~Z?sd&
z*?P6eaQtQbXEz1k;;yz)*N)!aUhvX2Xo<`v?g?`urlpSz7lC)ZWb_x&s(pTa8)MhA
zveSlO4kRT!RvL+?0J>f;=Oo7A0z4k~Y$)fweq3Ig%fzQc=)n72v>+KLBz^L&{G4>p
z&JQ0H9vrP~82z%opCgonBuF{;`+2;D8~Zx29_fTAG6O<cx!FA6@g3Q}Uy~_*fa&#T
zUoB_^r+A3VuadT878BY~oM{hjPtwDeuijB-%gL;uEUNXaD9cm+q^vvzFs_Om)8A?Y
z>sIcUrumK0X@&3HFzqFTOO68xGcR#O3pf0MMDhieBjFy$XJH}=xOtNa1a%TXi~m9r
zf?>s3L1z2{s9lSe0^-CMUZ!ZPAg4j+h770k@Bbi30dqcxf`n2D%Uu()@i0M$$j-p%
zBPz2BiX*c;qx!fLad}1Md}yQK*LXa2_qCNj+_b3FLx|u76VZVSBMe*#GB%M@A{FlF
z%T6;RyyYJuUX4Bw4+7?06ct=PsCOIB5U?B)=Jh=fD!Z&x6}?`J=wjI@{TR%puCl!4
zWO%!#Q;Qj3w?BlI5ofV-lv@3PwG?_(Y*MRq23baWxjGm`wq*)1`r6EI7=4~_Q+hpg
zh3dYn#{;9Y8x+rZpShDzWVDm$)X*{np`x{fv2KWK39Tu3{WE9`lbe6xnyv{{x{M(9
zDqK)P5Tbh}ts}97pg$xAepByJal$yEE1-6E%N>I=npgm_G-o1DK57|2djOd(yxbJb
z^I(Dl4AOgYzaS}Bgu9*WNjO$N0ZX6#njB6^G`2izIM;sM7YII>)>k+q#^Ed^F$SQ<
zs{jELVKhlGrTD>Mn2c*6AR<x&@FWzMi;#w>k^@b1gH)1B9(4qKbt2;uK=iKyf>aHv
zmP|PPgl)+m1DtWlxIp?2Spu%A(n#9!1FEMAh9-&cm6b%5)gDE@x~7l&wXi>ghH=w>
z?Vsa)XT_zEa+)7|Vw;Ti*xEXA(*UiT4~{9wi`_oY_Y8yuIo83g4M0y%1nE<G#Nk68
z4p`f;JKSa%2KC5^fh#LNPu^QfA;Sy$A!$Q_TxGRak?-~Zi9ClCkx)KFL6e%AYE<mJ
z4UcvZ%3A}OT<u3MK>dv)b%jW{36BfBj95t!u5$6uQ1r(At|=b#5E>J1m!vV^Gyw=`
zR#0Q0itrcWeFfG)O~DZE)yVWDxAMY;3&e#8=O^iQ&|V^A0AfoZqq2b1Nku*VT~SXT
zVBEo6(=x=ZF;IxmjU+S>z?t~<Avpi?E>74*<KF|SN%)Ld));{ogNgynEPvNW?%hnz
z5shgLx~IP`tzd}J9T8BT(sr^rpyasqRWj8=qt>Xf#}-N@&Oj<+H!I?r1JtveYKLJm
zT}O-cK><NAPiF5xrvd#o@vf^RH<d2MbJNexR4g%~AO;%neDUvhJbfhfKQn3q0rUJg
zB&qRDp-U!sB1G)B#ed;?{vRgfNVdTs@c64$W{*X3jj+dIG9R&}o$wXCp5)Pe`H8g(
zp!*w-AMdtzd#r^kz_ewH$!M&yzBL?(XkQ`!GY(()m9V4?!9k8`*>#9z2tv7@Z&@^!
zR9b_EOI*E>!8?6BC|(w8VK={^t|KEG4Y1RGMLaAS>qSmQ3V}$1lY$`}oGTnYS?pX@
z8lHeWZ(=A$hq<D-MD$b)b~D)YEWptaYXc+(M|0G5*Z}tWb92M?OZwUJ#lQ9>`vZwH
zn>KAC(E+F%w$Ur4p4v*eFYjA!Y?+w}?&NWi=+duOAPMLxSVJnH7_cScsp-~uKQooR
zmC@4fKH_2b*fu6I)FSO0R&Jp(uH~aD&E>ziskCd}$PG0U<H)mZ;(^WB5@xgWBk7MC
z(JyM$Zk6+5B>`d<7Ccwia$W~5H2}{Mn@vsB#PoE|@1;yTpqC_Za}n9iSzy<ZOa8pz
zfsV1USFPKji7)aOto9zqa~mCc_*K-Z$oi#*ly|RRQ#rdmG`9wMR$#wgKM)L!u5QHC
z+`pVG`hKn}SjKvES~lCSGtxAs-A@fu(tG5(lG9}8l&|`C?4;Yk%VFmT(tZ?xMqfnE
zI0n<1*J8U7Hpr^n_g>ag?h-H!220UX<spP$M`Sqh#*yJk#9`iOD<&S%1O63jnpinf
z!xPYFQ!s$UI6F_g0gj0zflpXTY+U>9HGmIv2<_lq;UqF817$dH4sOVbfR03bMlb20
zyX=;4+rCUA=wu5?uv!>SN^fjP{mMKq8@V^C^;Yo6OTkXY`_V(&tG|XQ<Oj66jM;s9
zdF}i_;0BUjflf3TxMC6KiFUJJGc!e8GAMfkzeIOB)Bqd9$hvVTh5E4t6Dwiwg<;N%
zkiCdUoM1iV4&^@D9HLP@;l!SdO$EnUVw3GSC=h*nhr{`Kd5U^^tb_$Y+ZpZcBag}k
zpmlZCgGJwjqc6nFDk+4d-(~z%0j*18_6bQe;QAXoaKAN9zc2`Fvb|!OtBM){2~$KK
zpYrhGCwN`S-6mckOq85k(5=tAd-y{^;s#cRGaTV~gmD+_3XCKEF7L!MByrJhy>`ma
zz$U%xOfz)mkFsJmtyb0(mvEJ-Q{GRCxBJtZxhB8w@Ze1{K4*OA#;+$A>Deo5ns;CN
z%!Lh#Ol3m_*5fkGO(CJmgmJ*dm>9{-t#*(gKpCt~m$e_|hs_}$=N5wTkw=;gPw}!i
zAD|?<EQ`iOMdCXRwMzJZL23(X+n?gRe1PRoay?tqk1=oAf_>bxzi(+#5^hek5O%2|
zO@xf@IpKyQ2pMRase>fb5ed(u!wCTMG4aRrn<WldctN&cb9m<Mbf7nko3gHG$}c-q
z%=bN?!*k~&n(LQMs%><PgHAMP`F6)L7T@$>rF9crVZviZ2Y_rOLLm5jUG9d$`*XB{
zC8zsSSNT=AHHh}9>YcjJ?B=3<hpuXctgU-MyY8QOH{JDz@~U-xp+O^)z7R^Xn=Jgn
zrNur-{s(Z>kSI8EXhdlQLd*057jWX!D_~#1jSs;a40{7ojDH?au*V-`m#9R5dk8l2
zgQ=wpS)ZPGS<f-{YJ-d-LjCAdmp0vC-=zd+0YxFR%>sHVJfY)qp=DFwMc%l6w>Rv}
ztDi(mM50WH90UE=I?DLOL>1^vGJeG5`T2>kQ~XV3v0zw{UNk19!n9J%zWqmhyf3!<
z;Ly<Q-2uD7gb?K~#Qw<i3B<XEi34Ntc6utTE1yg4Ub)t8sE4k@e~-(vZPC998J0Ut
zt$FoADi|7G0s~D1h@XB6CnCzfjxPT6pAQkSi8_L$*zrzbya-AEs($g}VDn!<G!d<$
z*hJwQ<t?z+&=@<u@olPe!7u;Y+4s<N#`1>NZ(3*j?2}oS#1`%DwUXYuencI16O=od
z$JH%4=5N1mgZfA`>jv{$(eM2~JkLsA{%NR_$JLi3Pg!^F%;NkF53j7==bOjgg%wQB
zduE#?hPem5a!~C3*)i^B_s3^Q!##EClg|{2jwr4Kj9P(TL2qCdPv0IS`F#HgbJaxW
z=>&7!84~x7!Jc+;n=pQj=etyy>8XJd=!AsjOk;?E%py<kj04;>q{H#|UqO@_n1xv=
zUt!?k09Y2X<E)AeE2(VqOFnLl(J~k?ib4Nr32EjKVX@Z0LsA8?W1?>Q^L4)y|3?b|
z3jl>eKxy9Y&0!oB+YlY=#aLpp7XKhh3+E@1EDUjlV*&s~TlP=?;cCh=VSx0>EMZg_
zDz2_ic&Dbk<lw7VqgvZla?4W^wGn#8n^wzz%qh*=8NVB}C1P!^sBCY@k_@OiI&|;+
zl)mnCgHMFN$@U%EO07?6N>%!es}7|M@tgaO<euL-9`o!SgI&m)RNDIszgDzzcwZa&
z?lSgy&xg#^3#Xc*Ib6QZT1%xl;z6+6{b~Fguk8=J@=I@Wb}=s+x{ddb1gOc}F)$Aq
z$dxPQSdC;Nbf{{a$y-8Ijt$W&d(^b^>aa3%{H=&-C1a^j0gt0B>A1jd$AP03)1!V$
zsif*w^_EZf-UEK>`r_Yo3}V<~^n=`XaVpn$o2;x!#9!jjCb0E4zG)A00C1#OMie=Z
z`U3VAGY~P@vQl}cia-+)_Uy@#$g^k~AfStA%~`eo)uv*d2&z;qnp3rAqzt>T?ytM*
zbcwqaTk3$;7pg1bH||L5#<fVL@0l)(ZhW4a%R9ABUW_A||8US*KM-W1vW0AJOMB>d
zQK!blRP37AdElZOum43mPwN#dJ3s8GxO(vh)#zKNuaeG8yB~il9-Rvp|2m^zDo>PT
zzyyf)oTNN~GC~1Hnhrf7{gPCpYj(QmiqJ%Qcer2$5tlm2xrXRWS`D+`@X}2$)}aL?
z6%py?Hf`oM>|CdOiO*lnPN-y8l8?6nU)_VqP7v;CHFezm%*k{~p=0CXnW<q1r&<BJ
z6D1M48<O-3q6AOoWt@v}L}$GTUElT;_$_f>`*SCv6grQ>c@J1EQtv|WfLU^&)m1^)
zk>FP>3liw3TMM`uIGFLvXG6-Zwo$Ra)k<7wDbX8tWE_zdZ)oCoVAXRybLTE^#x9Mz
ze(Sv3(zI=sTQ81NnVuh@`O#gp4)nX9akdI9Rv^6AN=lC1)+$iVY)eA;Adz?g_^X$j
z#!j~#f-a11B-%)X<?KNjPT&IK&eGG1R8C)x5J;0C;B%;f&KWKTHoFgk;>)GZL1o_;
zV4}tK@z=*^W@=$lN(oastcHb)K4^#m@*cv#lgG*&O)r_r01Heo2}Z%WU4v%pKg+sZ
z3JDDpH-1;x&?s!hEB$xsjx;t%@x90W*RmU>pyPBmL^K=%Tl_03mK<xzC>NDvW(b3*
zDOZAmUi9>oY0R-Nmp3{uUXt?H&Y#=)A^xB(P1AyEx_+Ot-kt3Lc9WhxDJlv=--PIL
z&+a`~yGY}i^7_7qC_=R!8owuV&j1_ruTRGR@Rk6(>$khKo@?^PWo~*P(xusNEVhNc
zcBN1`KNcX6quvg1>;x4e<KeW>YT+1duO%r>*ryI6`~dwS#Sfz%QIHT}t1gMrN$Muv
z_%mk&aX{}##~j!PE#G)C*+%?dxOIeHx_Z!d!*hwPDFOC{>{U+d)Bw2X3gVsXpHNZJ
zOG<CA`{6uu8$Q;{q?JZfPSV`}W4EMaWVQ<mzAg<RBEts;nN$?Q0KK)@$K~~Po1i5Q
zah)BlcH-#bFvf4Y`T{=N*rUU(YJ^l!R*ulEcWCT=Br|QgxT-dH^lDJo29Pw+-w5&R
z?|gZyA~qm9MBBEC+;wg|<+y+^3F<~}sV|bb!1*J<+MhrEJW22*y^qI5tV%*bZx<5M
zsBXnu=xx$r-A2Cx(#;*|3HR>3iSiBRScUQ-O(zf2)t?u$bJ@ug{8a1nDK}-t%SD!}
zSC&3iaebHYA^ORx9^*BGQ<_sXY<cuSOLl9oZi;Jc{viOAImzx=y54!fO(Y9>1qB)q
zv{i=RK0uZB1TZss-CC)fjwH{vl#psxpwx$Sr(J^mNw=-X==k*H1586c5*^6$%^M5a
z%H!}(+Lhd`UxUqnYo}e-gct#xOEB6Z4bH&jri6F-ITKRvtfit{7GX9S^GmyY>js1D
z=Z@7YR^ERj9TMXIMpu;wL1g<%2NaVxE4|eQFThiJ7SJRK8N!Z?W&zhjS$RETmJ8Bg
zVkHqxYHGoRZ@IGm;I^DE{EB|SMBDM?C3*K-3}15&51ZXLSRr-6M;`nFe!V|ERj-YA
zI<&l4E*nvPC_pBTFTW$b!h6OE@R4pr7v6!9218bU5L*Jwd;-@o<ZcH62*D}h=?62r
zO;q%~bmr@}HvLael|wagu^(5i*jmf}SbtnaMCTw|`%Vt*T!HqR)03B@P@CjnjV0AA
zW_$8ZIXXB1!6#2JhGjffG&+djmWgkV63FO1bRQHy0AQcM^Vgu9a-N@c!T48}513T3
zI1%n^h9Ut?4jJ;{Yql+eY8iXvr%)6{ZmoZ`Eolx<N_FfZP0oPWGzUdG)f?+wb6y_7
zH*$Xu^o;B#5GAOOY+z@{gh6#(v}I`z`s2bmCJ{zOrqz-OOx*NDB7l&=D^TRaTET(M
zpLh`O-@gymAt;E;iypGU8q^UfVcM+%DH}|buRUP0OJ#{o2iir9)JLx#^<S#rH~ws1
zTI(@UT%{R6FGC!;;1;6ER)v>(Fk>}GGnh28eUhpG`c7<dOe`$a6ilfjxu=MwB+WHA
z_(TZRA(w!}I9h7I&S0YCNMd9scb#~1X^1FRsQ$?f8skCp(r;JvTj?m#mPa$`41NY~
z1B|b^bdquT{lpI=E{_(%SJ|ccUw8w8a4S#PEO{)?`Na~fmTs@DRgDON1l><lYW}{G
zR9`^at-Ud!#in-|M1v^l$w#+G3Pze$$g<c_(eFOPWbS&41Z!ps!(sc%;<8Het^fII
ze*5U{FGD#4aV@{cnj}6|zWQw&{*=%q)znx?n0E)B%^PkGuM^S$+AY2}#((YH=a}8A
z*~KK&nay4Aklz`Hcs)^R3<mLl?;EcO`kyIeD=On-u^w+;$&HogBUq7idfZvSN+{h@
zXZy_^Umu^k#A#Et!uK}EExrx(CdQG$9G8j9Nk(Qa%o1|X%P1J<p($;*Br}Qigvx7|
zPXKY6e+_O~M_2y#y$r{klX<CwT09d2`fls0CaG|rXsqHj>OcKNl@(@=)?gB(j?of^
zQ8kM`!Xa=MIRz<_Mfs0rZCwdze6_6@$nL`<0y9<s9E>2rhF+IiUcQ3EyBDXqIf(&7
z?ia?8+5fnZ{^gqgnstAss_(e<^}CLpIlq5p<$FQp5u?`1z`Jq%W=E@AEZ&g)QwR#=
zc3N{NRH1PZ3ONqg4eJKUp>#wPi)YjWPHXZ4QjZhO=uhbuLVp;S76ewfrM0?@#Npu^
zynW<A75XE%^i~kr2A;ASY{(1<k-f%Z&;twHWAI0S*;7?EU%{3<i=i{HwU0Mzine|D
zr8|mqyb0t)P;=p6W7Mc~eoM7>`Yma8ac5I@83G@*6@%h@fTRN=hbHqv&R@7-4$)Sl
z?cz*%35l{ss1)=B(Yyb-HPsJAN>iAd$~4~XzkEMpjTL*U24|`%1V2?sMJDs39J&jM
zERMvv?S`HMjX9auc&_L6!#p6>>*Bpe1Oe<s%N#&?L=5)?4iMO+Z+O$5_vu>2-FK%x
zB>WZL$$kT11<t=z6mtM5SV(RFj1{2>WQI=c;D`~qgUWhG$f3p<0ZKG=Ge6b8Jth4V
znFmfdAY=}ZkwAHF!`cM6M2A?7vfwIY5zsqmaOP_}BC`ceyfUs^3DChsDWI8D2===c
z(Nt#pvxRe7Lv>b~<{(q?PG0>6#(-6a`0`u#;8R6nrNEo*3}hS4#wQ}l*?ji7&cMIy
zbsjsf<!VO_J+h8n^NUyQ)bBjGbZgTkNy^;(T;Gxkpb+G+GL}65bm|vstEhN;@FLKF
zkU1%}wZl0DAWNKw>e;{>LW0z5I5yIrBMfUi6v7I47!kMa-7z8lC$@1k{d3`0w!JHo
zxNl1{h%-s9;rsEnMR6Owv$MV?*~Z@enVltiY*Hi>fjd}Vl+6dToj&;T8}F=aYUw=M
zEtrzNuF~O4&=K-1rkNAD1V>8B{XvQ!cZa~W+pVIwJb;%_f+(goRMndxOP%!4u;nn6
znIdq`(^CcgI;oGougTuD&xVW-{!r>6Op=bk?_+T(rR#kHH~sS7+45BsIC7LAw<ie2
zj~^yE?ZC(ah+6BPxYR7l-tY^K76S5NwY=e6LW4=5XGp);08$bs2`~=~=8_$!tW>Nd
zRQ1yQu>4&)ZwWHl7$q2HeO3S6LPnpFiv8(RF^ZqSb&@qgvSV-&jB53##2>`ge&N>F
z&uI#IDa*NrJuyZ}c_5SrG6iOJ@@es=s(2@&NWuY%^hA{rqnTG$h5f1O-rHDGk8gQx
zK?P<QGFaYpa#|_7%M+ZRmcNLXW&f*hM}vj%;AnC_GTjg63;Z$0oOcL?zQ^r@QH_kC
zhrkX>NE&$CiL;Bu=o3*0=&cnL=z{SN$w9<1KuS-rBqTqXSnJPYY6sY!P!Jb9$YjUg
zA<~qr0{lua1+>Ij0Dh<~<K(6?)3c8A3vs~Bp#KGl1Rn`O^>ITifP4eh=Y?;RTv&<%
z3e`2V2TOc@z&x-A5F^$sHw=E{6vZKaVD860Qjg>QT!wbAz?6k2LM6T_B(7wKKmo-#
zbqyz{-tK!PXLm4*J0`1yD&Y$f5(lKQI_GeCoig*Y;n2$gaiP3Z;tPu;uvgt~8EEPe
zB{gA{dVTqR@5d!>8$TzBy;;?!ig78R56Nik%!}{9B|_H}N|-NLM4obO#mDX{bi5I`
zdV>WxrJ?DaNTz!Q*mB!-9x@&inkY+b7e)nkYy@h?7DECCG39uO;|h=#C45t86@1<o
zWNHE?!X`h9+_cB1-*aJ}9u-=a^3hq;^XBjF@5jMi3nNq38z)9i=~l1pyCfvoF$eol
z$*-~L*rPq5+L>saa<wEE%JzSqGITroIX()R#3adL3nK1E$apIB7uzK0H~ZlUB-bxV
z`OQPrW3uCw)Ddl>f`X*`N}0lxD<}%NxJ1q@;ODU3)pRTN`g5ZdW%|q^=n?5P1Zn|!
zEc}mz67u2VL^njTP=~@Y+w92bC}NO7=$C$ux+JD5yY_{H2&1Bqe9g)RU>neFkqA)K
zoq-0KG?a^=f(h4%yrr=cPscfw?O04q8#Wl4ABopVRx!sV`B3iT9k~20N1t6BT%!%f
zFPx6?a%i>bh}5~rs9OQI_MBB<N<Cu`HKIA8QSt7aNp+Fu_wP&-vulD2W~aRWJR2-K
zqWYpqKVWc4E<oq!CC22cl@x)dO^u#)_ab~2q<zj{IMcFu8lBQZ%y^`&MA`vv5*qfX
zluM8|6ZynSI?en@V(Y{`HFWCBEoEh9z6|U8b^!skh`*%0B+*-)p)aBBJ|2B6OSH>N
zK8*go>VuntsbRXg*x{8}51AA@KhPN5w|6he>%{Fe@JfE}aHJeCVzdkw3(K*w8SvJZ
zk@o_5&eR7Y=ZB--67b=9WIU4ePU2em7iiod*OB80q=cG+o0(H>`M<^s%t*!@;;eA}
zA@K)Ymf(0NYSdZpd>IHAKGC>8v7+_XSY2Dsrj`w5t1A@Fs*hBRe^%ByG0XOR;d{=X
z#dDT(1uy;%Su%a@%WAhXDWdT2^4qs_<B}SE@BB}P@0J5`k1Tuq(`|FI%l@8g{t~ll
zYRu=q35vn7$mrWliP55kGPby+B(Gd~>ZFMYZzgD_&r_+fW#{=F)*4u0Gy}jAQWj%r
zK{#14I7{3T2TZ4jMQ*>Y=Pb$-tf;-A#g{oQw-+e9l(aNSJ|m${g!iF577_KzEjVkv
zH-@m&;+urKqaSC~yY`HEo3b9F_dog|{diQz6}xc4XLhsPd!Mb~Rkh>(d+F<UoE%fa
zJWluvR95Lw^KVd9AGwjVhC`KKO8S|G?kT?O&Mr@a?q&WR3;20!Go?tE#e|K=s>mca
zNuat@kx`OA$T-uNKS^NxC*5*AiHdT|Pi-b<aUN>gHGbe107m|lZvX!gCu3)(pJZq6
zBK%MQ-MUZ2Y>0kw9Pk!|yjeANGE!0sxgCb*!|((d$efLN5>pTDd-nKUr~%+8M3OCi
zY6RoR-x`AlHITMZC}@(E;zdq;QsynN(KsH6!tTFU59tqewy#G$hvCq$n5}>p3!blx
z-F(nPK*XISkc31uGGDzM3)R&U=LO&+I+#pqxx9bGKY2d;=>~;J-pJ^Z_Rau`T}mO?
zIf@yiH$!2$$MYXn8_@;b+ZQ-iXmtjp2six-v>R6;;&KhoMhXRrE)?BAITHd>1_wW^
zr+PQA`?%JJgeR-`++O|k{qL&r@O&arN9c8x<SNj>d#vi1$d)KSKTeGgu?LFN?U){{
z8n&eh%jpU>Swi|@ebj=tBk4l~X!?&^hleTz?2tJUy=I;p8o&_8gj_EiY$TKwX+uIY
zvkUE9kfwmSy#ZnNIv73_u#aOPq*;&Z_)mBWu92x`L}HBdir7nGoqJdfLw^^^QX)yh
zjo5sMZjqcls1~7XB_8SG;$m+vxilvzD-YQ;v3(xdJ8}B2sqz^aSg!1j-+CRgKA*<v
zx{Rp$e<L>$Z17^!73EfLJcs^08#=Z>n~iPN87YII_!Bt6mpCd(bwzxpq>`cAdXcC{
z%pqAA|NVO<1$)OQaQn;C>H%t#lNhFTLTHi_6h>sDgd&H&vij+FLf!X%8d&Gkp8|u%
zF!x*kk>~M-iVmml!d>{}R21-nwRo?a_wR?`Ol7cyrFf!xwP-)0c1S;rayz>Fz)x@^
z<GH>It8qUGrcI<~BwAr%Zl)HgZV#Gl$G&4@Nxdo9R3%EgVkMyFD%<iIIl=TrVTqW$
zeTG3VO*mcG+~4z0vV^Kw|30aTs{?;n3N-dXK5CzyEJlwj+OC3_cMdFmB7GMX&1IMy
z4!nAmNDu`C1S$#4-exBpS0ZUyvu0Va;Z-D3d)x-XC`?;#c62i+@I-2F>b@B`0>TYX
z`LKbh2?SC9)OeTQy*Zj*&|u|+f+6O4CMM_ZVYdIPklBasC)p^6?S`+Z#>&&2^FCE~
zD6rKf?GEd`w;yh09=z4LVnYdiXh?!L&PJRNHAiwgURdolhxrGbh@fD)yXIydyYr;n
z^7`pzF2wL#(oDE0DEeSxs|&{#Ot*!lu+QVZO0ptA?h$UChIWH=z2K0LtY#x<WU##%
zoXO&*r)OX&eWb-JmrcF&gCjqp!e%Xd%r56UH@^O`r4Fpy^lbV3Zk>&3urb=n4jqV6
z#9!Kg2hHnjSae>HcCF2O85p7xVP>Lv$Fz;du3kgS=0{!2gYn!iW*byD3;Y-HlC(^l
zCka>}-y0tvOb|Mne^}tTuJ0cJA3Oh<v<AK-zERq8PtNC)eiW+9Im`frcz9NQd&`b;
z{z0SFWfqc4@b5Jf0!`Wz#6&?DMoaQHz<^oFEh0L78wS)?TPZWXd39gj{PHFURUaRN
zeiXXNNOVY@Zx5ns(N%pgcho}S;OW9j9X^tk_3!0QqNHh{%6&oFI9N>FN4wa7S(C``
zZ5}(5XlpI<xcAR}@fz7>l37^0g;~dX89~e$txgYFAiuPKXY{}x{uAyVPXezuDE_zm
zBzf1!(c4#ns3mnJ30<pf@Q4A56$}fu#{G~q^QL<mi3a<r{VD4TDnM+dQLXkk8~)p0
zQ_2Sj7(~7K9U7e1#%|rdpKx~halkJwJ@x#B?rLZm!-3Wr`(nD>8QX&A%<6h4LFBrx
z+Ts^q_5L7})UhkUV<-0&tW)aQdWDAl$B_j#+GFX<T_x#rf`fUnMHHApfJbIepu~RA
z@tug~A&%7GypdKFaf1Q<vCj6!u3zRU8y<2OeU=V0>A7s%7Y~p@MnL?8bB~#S_;vZ5
zWlRERLYr8xOF9j@Il@UpQbU`ndkq>7(o!&sUr6G9aj?AX@zlJANvR8iX)I;AKe!>J
zV^_$%=Sa|Im3`8-6yME@bz~K{-h`}zi_8nYbg7(o3VY~{ezV0aG#c*w&%1SRU|>&5
zm;!##chGFmU%>J77VSAkx|YJKlcUs|OX6DdzMfrUv)Xf^ZOhXQG{wTjJO}U@bPgu#
zrY9O?4t)D2sO|%v42s+3Qe2!B3gE}cDO*Ww-&%YuYiOzQ&u=feo$RYI14zI_7R!(?
zH_)DS@9Nyg`;FE`iPO2kNORze6}8N!e~HG7p_raQCL%*;si`Xu1FKM!#{WW-4tV^D
zXaIUN7<2ItJcKJ1DHM_<_UVf9RA0R2W_TKt3RG1;2(=P8R3$Ch_iXz1#$9XJS9=bB
zyXh$QLYuWQ>VF?CW9S<h0Z#_R^^Bh0jxQ6;bJml&NBhBn4TisNbF)|lPmb_<Y=P99
zx(2ZJag}`z0F=9s3WEzrUV8R;<}LA^%UYR(M+!*COF8w3Nx)hHz$9Y2%$`!b@@Wcv
zcL`_Mxc!@khW^-q5G7Mrv7lbvzx}U8Zeh|ncFeAMB4E|}JZ2i%FIKCQUseB~LP3YK
zpK~Kmh9uUuLb^H~|K@YWT^BUaoX6gb{Fp!c=l6!<*o2)Pk$i7dSU>!i3xyB8{rT#H
z%2O(xlSzs>x7PD-OM3Y5<x%>85>`6hB8AgC?~vhS|6V%Tq-7rols(D*<a#srhNy8j
zhA8}J@jX>ej!jyltIN_?;xDOjKGM1U`jC}ialPDsfAd!4@-F(IH6<uHbn|@N!R6xl
zpz`4mZu;P9pY<9s4AA|3Qfl9mzFk}c=9lgA=V{X=B&1A@gqE-DL4WQ_i##>tZIjnx
zb@*8pow9zXpP^WSFRUa8lpZS}6UZ_8{Mo=vWP-~NV;*3x(2etWWU0Kc*ppmY%rdYp
zP{Agz^8Ho$tLvG?caR(c($r(~8N5u#r2y21jD!c6P=zuBT2R<GT`#otZPl{b_xH#W
z$I|=WxLVbdJ#$-q-)rsde66x#l?;^<{i<yWv^P2Zlh`IQ{QN)Zdpmj;T(>{8D<?5K
zD#I~*SAi~#z5Z`qVVOE@qbN@J3dEUstxlc(voP?!Rg7u{-~QvK?`s28x&FM1`8h1N
zE#K;Z+q;RqPlK{+2X)SzNk=3>Xe*|JQ!df>#!xk;e8`gBNwxdYF*Q-^4K^dQCDdJB
zlgC0wUGg0a?tL?A=q+lf^2iv!ujite9^uFps61&SM0v`GO12TYumZGoXhYP&qqG37
z@_-*_Xm~g?Eln*dD#{DKC~@!UAJTEZN`~W)!GUU(A9Eq6H_tK%EPH4x#!pZ#3G`mW
zhnbCEY)bZcI4i7QDy(OGJWU~ispVFzczA8LQiqrRdF?u$XT5i)PM68e*tFzPD6~NO
zh%l35pN&@8{hyGz=0^WA$w7IuF?$_;>pYVf{k_XzkIMdzWL^&r-c45Zk=*B#*U^F7
zr$E|Xerc||Pe@kwquK(HT7Z>D%*}1?!~k7xg;*mCQIy+>tRXTX;cSWo<!--YzWw<Q
z^~qmr9d63$XfkO24zE~DP^#`~oh$jCRQy<BtH}A2j!GN@7aDkWgs(nU8^4&b^PK$~
z`FwY?jy(-?Ekc2TE5$A@7_(C9UjF?bEx?_~k*GjTRt_s0T1y@EP46@D(J7QlDUK@*
zSGq^<bdNqPr{W5K)Yd`g*Zy$Z4A*b9$znOzXAQFDUY~Du58mo-8GkQqaO1OMQQDL0
z?1rVcO(iXClWI#9<t4)#!Uym^<n^$L+kqAr+yfM+@PPxpY72lM?9eTPirJ^T1mQ+I
z^*Y%^a24C9mZpm;z9D2n4>(I`t1mjq-;D>?3G(sPs4YN0oL@EFmgXE~Z~a%vW9>El
zKF@>Jc}5&5Ybjzos1z4tU59x+Mmt}tEm+@>{Qj`p_n~$V$MA)QL6_G<f{lzt7iCy3
zEkqRk?(az-=`1^$+`4qc_vBOknIje2dpNr{r&pN0wwumevHYQHnz&b*y}jl}r*CpA
zH-ASb0;duKpBOnzKDm3GZ}akxPzv&6*HRLv>ZP8kMzFHnv6WgyF}eF@v~@%0cYVF8
zUYCDAWxHt0v|{%stqDUR0eRt5qP`yV68RT}D6MVfE%!WT-)g7EnRE$w&xosXg{!9K
z1Wca4|1kOAU+w2v{;SU*{Hp$2Z<n`8o0_tLeZwst`JW1LyI!|f-FJLAD5m}@M~v-R
zvtIxFTX)uff6-*6H`5AR_n~*XJ~0U*>LXlt%%m78Cii>Z7=~#Ze#^Jq`nB>~w%yj&
zLj#LC4=u$v9eO7gRkBX~$?Ny*r={_wqF|;eA(TXB+uPF#2<Q^OG=JW5ii(nbujP!*
z6#Fp`s#bZ9D-@2CUrH8Sy-lt&j69@xZ*Fm+Bx>_e%LqzX_^nOoOwZ<zw!8TL+hk$q
z;o7}9btQ+b#{wd<>;}b+J?-gQZc!+QBU`?;0;MT{Uz=0*{82<dp*eko$38bHiD!SD
zwjAiNSUGRcxOeB;^bKE2)}3J2q^6S(Z}!)SJK$psn!pA(%s3Tmt;9s6$FtiJ7ti(!
zuA8m9slnMl7ia|dWSekZig|syPd;b~<c~7+RC!imni;8W7#n7#m**B#|B>$Z&o5@G
z%O9u}k~ohyyNc3)`;7EeROFP72ziWsJdSrlw#>1kM{i<;$>$;cnIgfR2H)zvSw4^2
z+r05#XLB~XqN3tf7ZQeqrKF4?^%oKp+@mWG2NZBa3Ej;K3rULfiA_`VKOc>muX<9~
zsOm%2zqd0>f^ke%QMhFT+4@G<IdX$5+dEl>J;(IYy%TtE_&-;9?KEm7y`GY|Y7X{M
zF^Gud7Jj#oN2C#UjXRouj~@1c>?-0_>u(>*)b+s0&5aNhXZrXe;MND;)T_;u!#CP2
zhF-1ovq=v<TO5(3DqQm7u0i*f`8y~2#OZvpE_$(2oXxqygF2S*_$aU_1w5t-<&rDy
znn7>dRIxZEdhiVX+~?Y=BS*u>FFmx}$ks2ew#}4f$a2&|n&ICgC&COc>IudYSf4)q
z9pb0uXIPA0)IRMS{lA|mUL!`oxeh$*5Qc|<n>Umx2mNyLT#twJrTO9*Grw(I;kE87
z|NStwD2^+P1o>`-TJPLwXTB;#p^Z4*PPA=N*Vb-F_8J=-o3+%A#^NWd!_RaMRD84Q
z9k{WM*3TxbI7vaIsKq^mR_BZK&ge%U+}Hg3rL8O+SEhczXxnsfom|Q~`T3&|C>SF@
z1d(oO5SD62AgZkm6+DptjoJ)iImxGI(h}#hG=y(xsZ#lQ&3GtYHP~|D#3`#>NzT;X
zsM6`ox`IQ0qTUQm6l;O%CNBlK@x7dhsWD<_A`ldw2u-qjME59YIJt+2A4$W>gegQ<
zSyqJR7W15X#1Q_*E-&5Qb}Lyqy?-vyy&iHI`PwPU%GP>V$FVdfOixtfB%R)mE)n&~
zq8)V>*aH4tAiuu~0gvNpXlaFZ?>-7^dLtUh4B!qKXv(Ew_kgF!{)amKQ`%bB8TANh
znwg03=!5|=%F!jI1)jE|9=0wwyMRKU5+%dwQvGt%#fO#)+}M1CJRNWI(sW&|>u7Ir
z=K1fcZNQyK%*xWjor=cbxCd}3>vQ@-E&m%*O?#xG)3#8Bz8Q7V7pJ$ZW?_6p<#&~3
z^B#rHz!>Rc^QncWvs=fVI%2}yJ2g%Q>HW~a8+@RB)4^^p#f&$Iy^XjCPWekFzz6r~
z`hub1kn?(q0fk98-rsa|bf~Sx5ADTAsO;z<OQ{{>BU5l$p?Ud+SY2}ayrw_5%-EhR
ze=DEidit8b&x<w*?k@K&6x-eGHx6OLJ9FW0cfE-0!;pOM*de;@^(I03Ke~j}86NNd
zeEPfNzmJgtBNJF+{R~8BPUpg&Z`Y9%haJ#i=LxKN=tpu04IwC~zBVY?PLmrqI^b~?
zEWj$hh%6xqa@eeP@bQhr+UGRknYNbZXiVGtZtNECt1DB_>eMK<i}Rz39F68#h4J%x
z3BG08%v79hHIi@Lvox)4Tj}^Y-kf)EdH4cXByrVuXhZ7j>)mli+ST%F-Q^B?VJ;B$
z+4Gsl=Q?I6?F9GjI|dnf3+{I-CQ0t!vKDa+csMq`p`$1q!w0(ze&i6)_%^JP!>Uvv
z6X#{^23a0fPr8k)uV2GT$v!xI)5v;*)yTbw+@VRS!?h(=<%L|6PcYkVy~FZ~j+M%6
z%K+gEQE?jRo3Wm#hzN__ZAB$sGh^9MVSR_L8qZh@I`YW)_#E6rVEuaVy-LR7uC|c<
zmt1^m=`elY_qNA(8IT9OiTB*b?dmuU82L+NaiUTvM-4HGSQjLl2c!>w{{9^){hnj5
zoBs<Qlt=KQ*g+Z9RpRytkqbF+vXRnjci6cfRAN%G0TkoSXYZA%xHS%WetTJR@alT9
z{^U)ZoN^=_J{%pKJ(oX`Q~uYEhEB3!-Mv3^bN0wcEnQliYlC;gxE5|JF;^i9<*hmG
zHrz|<H=;6`+5GbQfthdJBkk&YYOVGCo@X;J?_!Cogarj*75n}a(+qIRczkVC1ofI*
zDl08*Ecwwn0NFKb9%gOMESO%mXZXvP5pceS08l<VOGVzii9n)WX&TB?6WKTTgkp)m
z&mQQA|1sry!6btBS$NTaa-Yz-t%eWW;&x8z2G8(z?xcFS9Y2_OEj=qs&)7H?-zIT#
z()IXG_8mKR{D0M*iC4~h_xG=a%p9}GaAe95DMN+^q@rjZG?+DzP)S6`OcEKQ43*NP
z5_L+03`Zd}7j+URNhL%XBK=;Uob#+_t!MrIfTwk@``pK^>-rA+vxoP7@Ar;mI7KCM
z+JV<E_R)KPSW#ii*lB)iOGT?aqy4fooP&@D4@G>tpg|&u*s9z~s&JNXRzc`zQkFgs
z&xh$|TV6KWmlZAN`*z^k(YasDJ#(URx0(AbyjeW=*9-B<s?6i1CeTJ8lV0RU|NUe@
zKootV!TKe%-rfWQ^*TVS;_Vx;n4>t(HjcVqVeVWT$_}&Dt3zwfNF;@WNe*huW8;b=
zuKqaJyuf#9#4yDRQVIIcrwwzp>vJhBf2zWk{$}DboF<<-IeCWG=c^k7_v|cg-@(0U
zdVnYo^p^1HbDQ6c)G?%+c#3O4pz>~{>;->+L`VIYqzn}@dEhqiNeVH^Q^$^}@98DU
z4apteercS7bwAbPTbJr@4ZG4mux5s2`H`ZPCX3xeQucl}Xr7+xGIreoh4Lzgk_T3M
zt|na5-a2<v%!K~FKdyKZ+iscn=K_?i(HGx5V+j3Z!<8%l+4iO0@j;49O^Db)LpD!g
zVL+Cux3HZg$VQ~Ne>)|?TSCatG4y73m7$Hzh4^Ntq`>C=b*bg=RDIRXZq?>$N;|qN
zy&Ct-LP?U_*WDt$?|`!*XRgkdn|)nxS=r9uklic$z3Mm3p-1>b<B%;MKmHjhKi$f$
zVqm{FR+rWu``P-*QsdC(oPM@DoaYrOPB3-aS^4uvhHb0i1v5-$@!82McGq;$(}X5P
zN6(;c4Hjd`u3l`SaHYSEFhWt62vJBt*=e8cW3F9O@{v_}XxD-c*}xdJ^-m{XY66^g
zNMPrR+j;-q5B|+<pB{^s1U-x2{w>P3uv2=&Ba^5J%f@;qZu;WjQ&VQ)d($Jl_-Dnm
zi2+|CSKGyzuUl|qzn+JV&yG*gC;exsG(WM({QPiYT!cvNEbsNU#=7mtD=L;iFs|Bs
zLz&dFU{gRC@&)z;0dd|@<5xIvH<*^b>CDRBlH4D;JYV3H2#}WfP=mB;x8mR^`i!is
z&K+-t$dnI#s_)>z)K+u=65(E~TBO6mEl2nhna{P^YL>3bNNaffx<%JS0yuY?nrpYU
zHQgl<dYd#3#g{JV_bp&b<COHKka&@;Ony|gSzJpqu(~m;Bk_e0r9pb{yCkfI4>ebT
zI*pZ)!fT1xg|wTlj*bDmXqbC?3)@Rxq~QT0rdK>3`c}5TV&JJ3%7@6#2{M64`E<#i
z9z$nGK(nr)#o`xA!Z}_=<<#pJYe+B)ko8e!tgilf`Z8R+5$+6PMKH0aws_e_6Rt@w
ztGy)1RGe(@F6+)-``CjsdZ%}{a14N-if=&4P}`K&tXZSu2P0mEHvFpS1&PWmx|3xp
z5^+l5uxXg`)O1-Wefg4<#M!3p^K`e~D3x(NRaR*lwB#UXM*O^J&3^pY#MJaO0F^ST
z(A@`4R-ZO_d~4x!9tt(vz0Az0(9rqhyDFK<R+Dg#&+6TzJ4;EvP8Tj<f!7>!GneyY
zeX~~mn<;+1Kr>PN3X7yaHZ)9;3paru8`t!eBb>;EBk5-goZqE;_YJRBOC&qIJC5-3
zpv+smu<d$+M1($hZ{=}+v!<jUkX?rKI1p_7^(Hg)l>GMCL}`j^G~M_r=NW~J!W865
zBr)B(YqtGz?pyc2^K17FwYak1bV`FZ&!|hcZpuc(z79<)i($Ck4h1FgkZ6H9sP=wR
z$;QAUq)$$sK7E<##u^$MSkW<p(-^FiAvW3WHZQ*82&-IF#mx4|rtIu&69a=4q0_vY
z?_KZS_MLV+C-mRfIA~&Y(%GusU8ipSFnVslUZ#m@Luq14r{&RRE*QT4ADZs2w3smk
z6BfGUf`Yn$&U&&r;;@fTReXm5s}n6p>F~m|YW+{2zP5W{uI7jJ(l2B-ZU6h4<(12;
z#Fr&|(PXlRHOTDpUi_AW+$Z)ckC&GZJ$|C4SN{%QhA0FsssFx1CE<j}AJZFFhy$4b
zJz~x3)!KfP=qgwqNsO%i{uCYzYnLx6;}YswfPu7Wmrk7~%XQT|=j-dcep=v8(w@PQ
z7ybRi(P_-la*&R&P4nBZnw>`hPU%g>Yd<d>u$0hqscPZR%S$t~M-ll`AD6hui_2^k
z4T<Ks)@DUsnYHuZ@Q0pMM5#{|yhsO}{YUrNADfzrX#WsGU)9%dP?epu2e(h5xb=DV
z^=sFV;+uIdD=T~Lb(NJ$67@w|R4hWz$THEvByp9Q&9*S7v<9toR-*@?(Yebdc)<)4
zcdLC>)vrD$S8Y(aeno&ZUR}S#bBV<T&VeK-i<}zZ%(BYvYcXuN2=h8iSw*qF#v;a)
zg&o=~r8$<7arSJ2-4JF9%{)9(#JPmIZjVsSBjxszAIm9kZqa2#btVj1Uo*KnRJm@?
zSHFAze)&jIgWoEx<ZIRY;88*_03W$orsyyz$;qhC8HCZHT@elFU0{&giO~*G0yCc*
zW1VU$zE4wgH-;YX`9>mnJd8>!AB`u5JRZ4p3%Y&MxUWL$a3fQ1lkz+N=(W`$OGV#0
z5iuq>Jt}e?Zmog+w3M|?K9z2%=U*=$?}uJ4TrLI{7PBx^b@Qu(b6bCG@nj^ewlpys
zGSHI?5fZOOnNwxFvNuoI8aMl6#HzpxADdQ=>%%PsMW(zCa8zbDD6Ky|W~MJ!N+R}f
z0)<UO%L!0gbC_FLyil`GE>g9w@1I+<wT`D72y<&9Z6>leuk1V5WxN)>{+#SOVe3Wt
zaif3y2(**t!zaII?6!#MFN8$IF{7}tgea?$S?*e5EspE4UbnUi>2TpdnPKz`_41P7
zcJ8tbjk&j)sOZ>yCd*7vbFLehR5mfXsPz5=o?2&ofdAXN-K$7E#@`>a`1l`h<2v(w
z%4=AL^YFog<0L-tB@&rUx6CSsgNI9;2Sk;K%tk%DIj5v$I5}CR)dypf|LH=7c;@JV
z(GOvqrDY5gEGiS$t#G;IP7hbcD{odM6JJOs&Wi0N2u$51J5W$iILUb-hW+I-A;`vk
zS$&r;zJ}qwPP0**8YhK;3p3t^EL(?Y6ULFETKZq|20c911D;W&U<Pqs9;}QXJ9DN5
zbRk3N9*Y;wn`&ZV5y^XwOh_<h#<WY%o@Tnbhe?s_W=Kn-h@29vEFS_PPIUG4XRDLS
z`pL>hF}qrj=jD+%DdKu*2iMbCl~d$8a22;Yu-4h0I!ISrJC~+6=CrgcD@Pz^NB*i^
z4<9+AEZ6z*xum3|oj*TagwSCv-mG{d%;-7bK?Q6=nCaWU|60}2l3SBS`H1@@X-uyf
zsJ>^_HUSv~*?($kG9gom&G_k2U?8!&<-4<S<g@kD3Z~@lh5%J_czkRwLle-lMo{w9
zKx7n-ovpi(t=}%TPLcafvV7pIqG?J<fKKZqpj?S^KwWIAKC|ftIy#4H#;QlV0SCiV
zyiuHO*CQm@JjS%CreL8+mSYfzT1Na4Z?nc+;x*c9&|aUnI|6SD1C-?Ua=Q6#t!HE2
z!dg~&$KH4?>oZKv24bcP3CnbSGzU8NZZ~lmATDgZON5R4l?@S0$m4L8<hsV7vUif9
zFPG=p5D%hYU3QR>nZ8m%;x$eL*|XX8evco9%Mz}=UcX|Vz8G9{q>cyd{3h5I#j2B7
z3d6Zv|Hw|=+tk$bGGFOHb8)lk`D@SC%OWTLc!07Fc3_$29#4@7B~hFbl8C^g9VB+@
z;yZM9KAUbC*wC}7(-Omq&-4Cu2@9D=zCwnoI3p7iQc}8hytmZE!0f7gLQMkGxKUxA
zm6equU!b!st1nX+o0eupol^o~E&*Uign@rEVqaf>{{b6EN~<rrJ93tnDp{ggzu3*E
zW-X%F<N-Q&zP$MCO4W9`QH-fNh~{%BNwXbWdunro_q;T|wGZH;mc69L8XA9y+K#fm
z7LoI|(w^?711v4_sVQ=OBujlsq>{(_XGQj@?x9VlBk%?B6_LXIB3i8uQ(`HBJZ2RW
zk8{|ssJvMzsH!IK(g6l`m=;uQdRJcNGUSDyAR+>e6fupPKnGaZ6g4b+DJRQymn_W@
zUlZ~7Lq(IC*rmz4PHO&ezbkCPPJd_X_lmj;+uz-IaRx{8{6~k{JZw>Kp=4(-JlUu*
zV<4AWKo4p&Z*7T}$66IRiF$#*(Kf5z*JavXvmHZWJoo}6<f2Z`v5H^0kCDD;F+&F7
zFo0RsMZ??qqseOC+FX*st57X4@ao9;?(}cpz7@@Q*WjLj)Z~W0lj9vS^d#z^Zg<m^
z3@D6TH|tdLr`1W{B8;3Bsjm64+KteiXH$_`O6Ad6E?5-tWNCiV_E02+YRxj5Pl<dG
zhJxO?83r_4_xF+3TefCRm;hU_vZ#;{&3}_bAkM`LrDX+0`sCTOk?8G0`O!ag=#VIf
z#QEom$vhg~-dV%!M?{7^09M+1)44|C8kEOyK0bc_Y)bYzN(aX+a1`FLjmlaC4|&|V
zn5y2r_5rg=xv;J};q965G&V30zO(kptoNZa51K|p|3QsDKop!0AL`cseZy$i*So_S
zlQLTL7(t{rF%1vkER|MP(Tz&jfs@#!bLT0Bt)$e=_Y`-oxU!SfSqNNn`UXe0-XG~3
zL9o)Qt*)x7VcfzLQ0I8_jSDAaH=Wv3wLrY4wfEOvr_-QCxQ$ai?dKOFruV?3jFvAC
zGL5wwr<3{fwRAuVHe`<kr-&HxU{LnLs*7U+U#|AD0Pw&Yh~7mL(k`MjMG5K#$Lq5A
zCDJKb6^E0Q$6lX;G@f>YOQZ+>n0eZR*&f`jT~QMP?vB=u>G6$qB{}9TG0Vbk|5h^Y
z{+L2-Pw{Bg9sylO3bWaTO{#zdmv7FV*j%mF8cDv=^!Aw%po1Zl6e{86stURC=Vxxw
zg89sY$wS6RU=iN;R;u>T4HYXHORTo|Fhd+I$}Vy)-Nmo&zLw;k6qW4H4+qo+#<Uh(
z{CTputyP!1xSZ*R0Ln`&Xp{~~W)bhIF|gu?miS6ubrYodLHeDQtx@AXS>#~C0*tDv
zd&}%^?fcr??7$vmSu*EsStzHIa_ak>=^|M=*rrI_l_YmJzz_Jr&C+=G&AvIi?m`|A
zLm_OYwezI>aB)PeHHntda!GCe?^v+0diF;kmO8~<-@M^82$<!}nX#$mPOAMS3A)AJ
zJ%xiH;+qf#$pbTGv(nSj$jTOM=@oMUe@B#WdN@Ar4aBp_ax!*#lwJM$_6=n>;Tt9f
z@|eCCIWe0C&LVpi(_?C<w$HN*o;}o;u~Ehu%{ZT<?YTwr%~P237#(WUGw7pjoY60L
zv#5yR#R(J^zEC6Dkq$_lv-|owIgv8*SZF&29PKQLY4>kcVma3PUI+<~BRhv5w&iI@
zjt`;lmU+#`O9%Is3)7{MW1{27-74ds`3712jtM)a1?oZl6p@-KZo<E^aM6r5Cck*T
zPvX_Q05pUjul{c{e%56qF!R*(CT|`XT}kq$O5%puD&V!sS9NvUEmY3whd-otF*lE3
zsX*1MS7V7nY~d<R-=2i{ERL&xOKzayyu1gBtDA17J!-mLwRupqPxcom3ywSpzfdw|
zkObWa4<0DTOG%4b9lLknaHzU}2;yJuR1Hs5yQNk&xVwCOr8BY#0o_*(?neqcF3rI8
zitEhX-SRa{&6Lbq-kv2=mhmeW#8e`{I;qa*S@;DHT}XB=G5&Olw2&9`MNpt6+iC6v
z|E!4>);SN?UJW;*@QR%=Nx-$pP`TdgXJ&`pU89RRQYyxVKP4w|NS^?7!#cC$cV8|A
z=#ZUhQD9_jEFWGF3BmrEwkJ*rcQZ0ks!-AUfSa1QF3H4VGP7`)=v|L;QUlNgTQgnP
zAy9+um^yNR)ur9;NJ)3JD*ooU>bgb3>Z5&^+qHXn=B%NHMlj9p$L(*e89#C2se&3~
zuAyeEq;@%XC(J=!90lP~y}eI7UOV;Nqa(;%VH^To%dBMqy%jd7_SI_1oF?PMIEn0l
z|19w~;Wht-3MwG1QVb(1iA-OJEV+Hd{WYOSQ|P#-S8^|I2FDYLF{(`Z^y$;oW1VhQ
zH9M_+373y7FN}m;JWc3_`4qXY-ebNaB+5%B#6DkqZsU#qK~p|_7eJh{DE=Nkew_Y-
zEhhhTd<>WMX<R2Cn3~oC`)<Nxzk&Hg4e<33lm`TyGY6u!C=m)Ns5n}7u!21x?1}b#
z!EaJQ79ko>7YmU%lA6n_g8K0f4+PQ{XNjlnd{*R+Lx=9Xy^H8bv?(NFH&Ph}9PP%!
zl(*|AnqO5HBJ=wCNBO!}n{O=6YPWRBTenmJeTwX^h`3$^F3x3-Q&P%zQI^=9HeZ==
z7U?I$1q=42rF~VLSNYEu`U(=e=b}(-dABvgh*Sj7EEpl@`!y~K+#G|z$4PQw{$C*g
zihMy}`=XJ9n>ur}N1+Z7-@h54cC%aOJlbyKM~&K-vF?xlfbkJL_FB>qvi{YH#pk3k
z&k>X(wzl&I7qwHLdMi3~?eFw%P#(V$>7!-F$>RtWOc9U3>N5WP)X9@}7z)*ZRbsXX
z(V`H7PMU|nl6bB54Ei@0#qA7Zad(7m#?JkQ(yk{c8Y{S)U&$h-CClj5_ipWC_Le(=
zuy7(9(H;O@p^Gu^<x7KCjeGy&0=)D&*G&>I_w>LGqfx!0PFbhg(J~l=#6&WGHPHO+
zk?BbJ^kkN!vdk!Zg#SFyK56{NTQQ2%MI?&oKJz*fTxtxZ?f+QRXULE%Jk7BqMreow
zOG&e2OD`|jt0L5cW7i~>*c8wX6x*af-|gV_enrO>Q}bf2X9@Q+ZCbHQ-vlM1*Oe;q
zqLhuOK$=OA3n?-_9UU}zxfd?$Pp9%IaufkRL(5Txb31Z$he;KTJ~&#h?4@3(f(TLf
zK}z$fs!Fe~^F{Fe6xbJSlZP*zBX+g^oS@MysHn2YXlfptgFLZNVU5km%g`i{`S)Gy
z=X@;yNu;o;mb*>|K*tb~rlKhn*f*?VJLbp~7SX8yA~LO81B709nScuF5Hi&Lg_J1q
zp{L}L;$I>c&8~L~JSro}Y?yLrZ3zA4hmRg9%Sp$uxhJcTO!dE)FJ3e>;u2;ho6smQ
zE7M#~lg)#Aq^h#=5=gDKAG5@q;8FC26qJ@$@fSrTLXKfV;N#0gX5h(3E)yh=57Lrm
z;y;#_M=hnY!uavC-V+*f1E>l|ckDJWh#qLJOSc3~v`$O9Pyc(lswkwqvf0`-dqa7O
zNk&WkMp2cB2OJz6OmK>@)M{L^tN8?uZIlju$7A$1`DDFJt?_2Y#-WrLi^&M+c%CkI
zDgiHx-nm5Tg&}2KQ+?T&G=Jne1UZMOH9@UQkjabVv1gIZN)Yq50y(7m+mx#dj3`bU
zI#TJlOMWJa6ijp;OKf*pi)JSXqA>>sw4@D0S(j5xConqm3n|dXw1)lQMC&H2SVkW{
z+-1YFYYQlo#ks#?8ZB;_fxv<x?Q?^~Y>vp4E!pkGN{$GioN)r+oj}_Vlw9x+_0B5o
zTM}orAI|w=cBJId?o&n<3b6XT_#>XkFWBv?6bZ8n^p-lpp}#yQ_kKl1Msd8EwX4Z0
zE$OJ4fQH4f*5h?R@`^<4{q46EGx~X*ep1<hu#3j#)bfW<o)ilj1p74vu7!1S!Y_|(
zr=EVPtndi$cOl<DMedTtrW;o1zxcmPuI!Nx;Kf0IbgMm&D7W@}+{{jAijy|l3Q&jI
zRscC*xR#wsE)agC6(DM8(X)dYok#^#U!H0v{0UrGphdlf9lSi|?ddg<RR<^TzRV>P
z8h7BX2HQY@tlBE+mNyI4@~P5dc~WPON*69qv=rwHt4&zR=NPbsN5C6Qfe1{1q%(MC
zr=WW;4Z2UX*|8&SF5^O7oiy7z@i7_{=<KhHg!%;Bb(k<$L#JPKyLjQwHUdu9(<|b(
zY5RcUgcm%!A1Mj?Iho6|p%)3<S_G~HY9M~>61k~FCb%$~2UVMEh($EQhDWzZ?DB~U
zZ{4=X_@~qOQ&3d&C+ahAcfFgG@_6M=dH;AO9dafM=r*P;!wYKKq<X&ere~_?Y31iL
zG7aiRio9q`FZJs{;Fsw)MbHP&@*QRda3PTNbLS>7^GLJpMTlLmkbB-xr9by;yDH_Q
zWWvIMlUALO2UOCd?|=B{(P_BZodfx$$LD4mu|2gIZ*(#+Fq(A%c5H^%)z$UmBFU)1
z`Mo}cXdLXX2z*;=)^?{cxMmk5{i1oomoEjKy8qyT0YL~Ac~Q0#WwQZTBc}g9P`@yV
zvZRe>lth5Dro>>;pa{7~5*GFAdM%>?Dq0YuP{ASH=iv6kd(-HTpFY5=f3`!S!nA3o
z0DV+FNP-0ja4)N!S66X)?^E^m<|Lh+EG{r%Gys%H_?SO0LZxxvfdhd@_x2a1t$<IC
zmesof?^me*E}1-1Ny(6%y!G1@D>8Ov0Dc2Xl^*0RkCxqQ*Fb8c4C{H@wuU<4{e!Bu
zGrR{4XXfLJY8j;s<f!G_s?RsWKZrhx*J$}?Kt`_7k~#M(7AH@D79z4MbSAD0c1qR|
z@Z?A-^%Ewbt_T2XPx4>H9}ycx)du)uInJi#l8I1C`_*Ca(h~bS$yy3$>|_s5E#R~%
zjU53AVFR@>Nze0{a8v{nn}b$b$NZ2RoL`YAbHbw`E$QR~{@V9ycdNsU^ZFcSN3uE5
zHGpa!%CyFVm(C(00wr?1?H2r(``!`J#{<oO`s^V9yT%)CJF_XS4Fx+lbBhcAd)(y7
zxz2{lGL*l?bh}^&4AL>gdPTzJ%4P^>C5c@ric6O`qc#{@DLhBvG#RaNeBd`qM2O5I
zgD!#Qn5ODmSe&(*WEPV>L;6-vjsNO|cX)EJchMz}uFrV~O@~Q&P4+5W;^K`f^>ki*
zx0>WH#LL#KU8^g|GAbxg;@L~YciRvXR48w_`o+W(Gwei6L0QBU%dXm5Um6BPS@UO^
z)2f-y(3BSbVyamtgK_H^BIM^_)xS@lkvgi|KFW93v^w^9tr2dUEkEH%MmYenr&)BD
zsIh=SGy%df%o>6`fLoN?4I+>@L%)#K`K-@SI@wo!(O=Os3caay1V4qH2_D|(-|ZyP
z5PY%}hCy8)tCVOv$R!CXpZcdQXixeKoTACH0zvXJH7(wjb%DW!{G{v)QUj*)fc4Hy
z)&SUm51{Uy&$x$dmMDuzk96`8<QFN0lH*|Y^k9vH9)hk<+}w47<WczPf=#l2QQb9x
zPx11NRXQ!ZTQ9<FZW<zVW8sH!X2iR~i$zb$&dG_JvXe2PYvvOYpG=xx_sT`}6c7wS
zy*T+u&#ZRpmCq2ZO?*%KFcBM^XJ30Ut1{Bn!bIZ3w$ECzEe)awe$zo!SA_xXLaN*d
zIwA{5(Bi&)nqE{U+EPJH&E@ZRn$>@&`vLf-PhT;*Ve4@orGQ*v+7+p82_Cm4HEjE9
z^EHX)M+(2VTPaMOXhIx5-rW5*aSN*!tz}>>SFAP~eE0GyDfMX4v!WxKmf1kQGoJib
zt)*tkV6~H>R8<_GV*4x9LEE<7;4#xFkLK;4pS{hHRJn+Dsz}b@hE8){DXAtDHHkS@
zD7Hsx2CAU`YjttQ$3x6Vl?|`n7}fUTBsVMT>c}S|WWUSm_!T)65k%GxFY382QTt2J
zM*oZcXFI-JV%iswK1;iPKa2Gu0T%_(cv_w^>tT(HtbPYcjDK8Q+`E*zNkIC>E!WHH
zxlf91&oIK|;0L0!xBQlu_tEV%;+kj&KP)R-MqSc{J;R_TMK`vrBQdpysUtgOW&7>V
z+I8IYe%_@Aj~>y#v4x>{=KOh4P+q#YyvcpG;JML8?K^1DHj8R!qCUkwp|YyJ-n#xP
zc?rte!EEYoKKv_6?q6Vd(OI#Bd!T(UU5DSYL&%E-LrHXN<VjVu>EW9aguIhICAn6>
zT<{K)GdGc!rhj}mzGLr}GJKUs-W1jm6jK%s9?{(^``-9Dh29vfja#ZU2-_)nvM$gU
z2T@DZ*|v(lm+02eWdwiCBd?8${oWEUAa$^87F)&$4jb><^4e;hqNiBSL#L#==^5Yj
ze7YH3UVOaV$sZFMV7p^PYI$t+h3)!&g&Dhc(GV`*`MrKX`a~G6u#!yd>>`j|ip3j_
zBeBGv-3%>jiMDOg<@|~4hIrP*te1B4=Ypy5F3FkzCeAn|zwC%_FKRi^)>u`aGsCs^
zq}ZMzH*@oR?gw$zJaVFF`dajtDBFij?PQ=x`DJWjG0r#3$n^?%Hu55cX#X^U2aZ)%
zHqK~md7tR{-L0iDT&)Og2+Iv`nKNw4+4a+C>))FWh=LGmQ~HNHT>~KWiM(LN-Tfvp
zo9f+C9aqBwiL-1~MX;8*5+U+QFj@m-98(#v5$zlA=cf@+(Pjku8sJczXgWas_CVgj
zrNnw`(Kk0m>l9<)#u8{+;*dj%1e&lYB_Im<mp2shU};5+03{+pD;nd7)&4cA<xyL@
z@^ysx8o@^43Yb1zZ}5E_K+20dM0O34QVC`?rSK>nweQbzsHkHpofJ^!C{_N3CXw#0
z=7NijtlLViH3u_vEzp}@6B&L%5GiV{ro(I!?SczCKI({@;)q|QKe;`xj#lv-3O6H^
zJ`q@x#|7!+1FZ{c0F_RpzF@5<aJ4|rx_<Q0qmV4D5tK{D^c?5F=JpE}{Tl-US0lXn
zg5e^2l_;l_Ii>nM+s*%s=@lo*6A6UqhMReAB+H}+trPX<qX>C2fE$?;(nW!UGsa;$
ze!c34ZgB`ljQT7(YUWrSKrzbACFgTQ{mERmiKp7;SNEgTk^xN&UDGH$YoA>Y2UptZ
z?k>Js;vkfV0hq6gHW3dnK|ig_{5`TWH-~aoqvQrds*;Zk4Gk@AfP!W*p=i4QmoHyR
z043O_6w2Y6w%%L^2X-pd(*JiFg{#izYyo?^Usje{`<*Tupob1y1xz7oMfr56IcINQ
zvu1Q#!z#!P`KePs0sU(kK`UaE)x_9XaDC|pb~-(9!5+HcF`FJaC6%b6Ruf)a$@wvj
zHUU&O$+}tQ_~E<5EVoEJ@Tj8oHo6t>CYqZimiPC*xA}Qx+l>|abJ*<33bB)mp}Tf{
z{rr7ATS6s4u9q|pj;<ttsvN~9JMi`qe?392O!hERhYWao;wD-ur8MAsK0$uLOedeS
z5yc#)gTEYN;xZS9i7)WlkM|MDiXG|d<{MV9kn_nPcwJX5)oyj=FBE%^(#l)?7A^}N
ztZ29Oh@&7iB)VhrHD#V%M%klELH1-)d3us{6bpi6tAE4>pSf_sCcGN}g&!5~Fq2uk
z3=EEQD7SV_YI|s2nTd!Fh`NrBEWckM&OBQBd2EuFsCR!feEx~e1b6hsEozbS7Svu5
z{wUTwel%o_(h)5e%5z~gD0;NKU8iP5Ji4&$^&ru0d7AwmXt;2s@9YrVm;r5Tz*`@u
zl$y(*id*x1cZP~G5n!JHk+}BJnP5cp(;M>_DNLH=lv^fJRmrWMQ<y@91?~9Z#qGA1
zpa9YI26oph>>6EUSs&e4|J<XZ^`~0a&+6yf@dM|f*fl<ynqyr5rP_H$0t=eyVj25v
zXP))si4&LTywvVOcVDokg*U-%7cp9BRi8#bCZ+z*vPR!sKjsN;u2V+K=dRN_P3-S2
zROM>m3kj1--0^KQ9@yHo$Rq`qOt$#dp*o|m%yV2;n;S{)%c7VRl$!Mf{H8v2Ib_kH
z1tW4(Jd+gf0JFzIjE<)20asp1?bWY$@6FHa1j!rnNNrhC99XNO`A9W2TSi(<5YBT@
z9-T7=kQE*l?5rV!Z%DU26tPRf<h$jtv@885rcaqK_1yLQ;JBOJm$~S7*td_i{`$fc
zz#dVI!=?n!e;Sd<<>hXs=F>Y$gr=LPs5*(&3r`kZX7ZX?_tkGR7Fjq>JYI>UdknoU
z$|8|R05TJPm$ap|{yrT?wEYG3KCC{BWzxoxo{r!U;aTCSi;Pupge18~82M=Xn|8ZJ
zZK~o}96{aU2T&f4d5vRMhb$LYwbishN&-hj#5Vw3=oR9(yhc~zYq0}UZrn9XZmS(S
zg~~Dp(Sl|!8YYY<DXM;^I2HasN;YS}X3|l;^QtDI=~sKQl5J19;ystxub;;nxqzLY
z0xJea_e@y*cZS#L106>kVX1(yV3lnjSN~V}DzWZ0k`Jcu{m!tB8>Z3cGr|;goCzZv
zo97F%Xv;5C`}SZ-BiX05%iDkR&>A$!T_46nzn^H6maM<3E)enqAU}?L^vI?^P9OO`
zAXn+X$1#}%E26lP0<M-)Vr00+A~EHxcX6&YMFt7+u>?>HYs?jt-~3YHC->P*2Q5Sa
z_KH@dOi&F|4{c9%xjIEZGBTIh>R4oNuN^d(R8<R6p+9j6t09N!F!Ei3BwYy91+7SW
z1x;imDNm;EkrK4UyjRKPPNLuClXAH!?8<T5vZpj91mCOsZ*BQN?j7naLXZxzEec>z
zA*oD{98zGD#2P~ji)W*(b1ECg!kxy@TxD$mkYKE4*SOfYtS8|)KNZ4B9XJv_$_S&Q
z-e3VDKlAYLNI&<<&?y_RzBZw8e1Jd2rx;ZMZSXO@PS2E2cPI>Sz1`B)?xtO@e|GO|
z@U|U0dh}(3VRmzd`ig~MTGj>4hQ;?=bK>a}i>{`?@0T{n>r5|qIU!JMJZHxkaZV(5
z1M%^QrR)=Ij~J^PhWO#bhYsZtA7V&7&K^C`{{`7X3K^bYaU>_t0OhU$)Qy?;niua6
z>IRQ4`slXVLm!`?Wro)i90ozBCV0us-H}|ZN?S+p9-w;_K_Wtn_uSSScNhnDpg}3p
zS}2vGlG{+KOc8ACvyIJjbl?OHhM2A(FF#*IGg=v9g3ac04jKr{lFHoz36_Njyb*2|
zf!_$YKRQ!vyB<DV9WSa9+lpKa*`itcLf3IgYXn=FW~d;&(Wkckyg|L-BMvJP*h>&t
zShCo6;6V6iS~YNaiZaEN2#8YTrWTUV2jrae1OB%>)9+_Xs$lGKF3eJGD_jcn9#T}P
zzNb;QCaKKvytCNLf=9f8Jvra)PK-Sz&oMA3h2%dJWBxy`CqHeeqnJ3!<K!P!l_@Cr
z90yE|P5byTY859kvwDrHBp_DQw4}d2?=o6{T3*ck5W|<$g{!qh;>2VVp0MUsB#ic`
z&WWaUUQlA&QD}-L6GbKDG~@oGbU=NK92}zPbg!5($kI<>7EoN95wX8QY<(kgCpw*r
z*pOxg>Bq0Jeu5~bw#~i}RVVT(@uZ8$vdb|vK{7y1D>(ICO0<@c{xPJvV~9YD0Heb~
zLq*lrOl*jOP>$W0_{<?1<5^EpJUhmvo03Hq&U&AAiIYV;F`-uaDMS-|Wg*yb5TebX
zFSbN-f!eEVdA(Rr{%I!o|M-L)hhigS76%XpL7)U_Z=x@40YM{3(4`nr(7Q0(f{uDJ
zjXO_qttKU|w0)4QFQkUzlN$XtW|MbsAqv+79~f%5eEEw}>lPM#`m4{9H(0jW`t?_A
zJey32jdqZFn@L<F=(^G2n@5o4sjS+`QSsS=6CkH~ShacF|C*Pt)@l7|ol<pr8PviU
ze;W^x5?Eec{s$1HG<maN(idf>l)sZz$+%dBCY%`wk_t-KqI)W;oO>$`NO%;#Rd0&_
zUIJg2&RHB!tOkMilSR$%Z|I_8A4_s7j*yQ3Pjfs&#4zPfn{3eC@&Ytz;+D=F%?}m8
z)ytl{>^`WPS#8?KmeMF1OUCKT+iXO+Cnz3q(HIy+5Q3p)nr^?|H=-wpi$&)?P%uUQ
zK?;Lw@PrJs$z;z5wf{Xkzb`v}cFRZq3}a|Z_#|r)dHj@#CE_I_(-Sexjf_ck`{ye^
z`CD&}&U$xDUO<#+-I9fQhJt2BKzuf-WT8tUfdg9P<1<s{aU6>f|BMiYlz-9iw`5#!
z{9=S&kQ|jjJ{Ybb_458j3D{hvxGRzh-_)@eb&D~67}4;{6Z&V5syK~F3YG`F6#sLB
z^>ToH8bwLSQGoa5XksK56zDEqT)1TC$36nZY0kF&X;{5QYvIB=++PuRbtEK~S~S|D
zKc_a!(FmbDRlxh?Cw+Mg0M5EVUGR{!&_QnqWB%{}z|}}J%vbGi2#d&WsXM*QDNn%L
zqK82(87Z3t>jwAu=5nBupG%EzKaPF#?ePYM4$wa=Q(Pi1cz(8^>@G?EVOYHTCvB3b
zxg)-LiZn<|YroZ`k_9;ka9M}946F^P!~JKRjITB-N&KW}S|==p!N*m;bKIU;w7u@N
zi{az|=*J1(UXKtBB->1;1;*5Ve;=?LMrC#S6QoWWTT4^F1dWYu?fK%pbzk%LBj&4D
z_UK|a+`m`;!!w<Y)U_0p9}V8^->WtE(c;M`;{sGowJP^1FU&MQbtrAG`!K0ZniC8=
z8!Vo8FQ~6p`xOK7TUu_vMf=+9Y15wy?@PK>&OVUyZf4H41NW*woy)Ozbi7MSu=myQ
z+zThQb#?O(<~(>H)UWow3M$1jv%PVg!*@!!6&pADvfgP}mwlRF4gLGgW-Hv-xznbZ
zDo>|;mXbt7Mb*B#GVIa4dkeH}lvPw-k|x<XId#mv*}J#hl6*wDF3}Nksb{FE5A2!w
zR#tz{JHH6=suKegk6~WHF)`gqG-SsY<5NF~)el63C2S(-I-tt~4aqUxFtN{tojeat
zED2p={YELHF!~r{1s{@G<0(y)MuEle!>j|Jyzfu`S~Avt6pKqXY=|+?)x2n#OD9jP
zPg%cVLmn=)eTNP&36;q%HGaA8(0O6ev)~rVWf4+=CjMX|?JnDI*qDVxYMj=&NtNvy
zqmRcqUeOKPy?FFpIy=fX($<dEw#m*?M23D%fVsQD)$@Aq8Z%~$(B=`$ZcC+KOQsHX
zTNG*R@ODAi+7&DQ0N<Ni^_#)g-h&2>Q~lh~@JD?76d0^SzYc8rZem^hVJF(jA(hzg
z3np<aSIx?#<16Rm;}a1Rvo9=c)M}-(9}?-iN^+A2%|@l%<t=mnG+#UE$an($^<=+<
z7jiQbCQNwo{=F{<iKJKg)eFatN!TDj!$w+CTUXZrti1}v?dtXG+atEz*p=9KPAeTl
z6ny{qL$(4Tu-!>X)1eL^&hhQ(_u@aJmZn>NS#hG;ch96x<A)EIaB}HK{ALqKw_I|I
zw=6aL@AqAvD28P18KtEonsMER&Ng=`IeRlPh!%=(M8w!fY%9W*&SN(P)XD?w9KQ49
zNe@D>5CH}r+D8npgoPTS+C=~ADJyS<iWA+`K0;9W=GT9^ioHCv(C>WQW}Owb#Ak=+
zN-36K9W-sJF`CuNZ=zMBxn;}sTr&%cL84cms7UK&J$mdWCMM2Xwrn4w=JZQ%#bZ8u
zA{HQWM+`*-RBSdd=XX5J3m>~!8H_xxonjSJnw!h4@d;<S6LRP8xg9aH9fNoJt=!l%
zgVw+IQ9V7q1zB2JTD}RT(~!J_R@|va4<W_~b!7S4wU_fYy1UN+zLr6p9x_YcSzF)D
zcXS+&8*j+1zT-GevA?UyRhjf*VTR-6dv})aeJI!M%RB1`V+Di2KRr9AQYMDvU4A|<
zp|bkOpVhn0%ykW}`vLjq7U14-G}?Hc(5pts%eN<TNM>Xl>~^TLW_~T`(Vfxp4$Z~8
zKQ}ckV9%382o&<NSTwVr^;i(BmOD5c{&S{`7gbbyo%tHZVyi8c##cs3mM&d7Jf)zX
ztA;_|>*qje3F@h0eGeT4#99v8ROW|k-JBbLJ}=DeC%{_X)2E%dJqAMzY~|jclOCRy
z31;_Wa@+cy-#t2lnY^I*h=`7s7IJBBH}191yDmhqB*@H-vASx;Lpk^6DXX{)Bl|Lb
ztg&oapxuJEq@2#|K70A}_HC)m%uKdDYU}81+dSrjgfk%2uE?VILVgUMLpD0{PvDp{
zN^82n+B-y85xY|Oh7~JTNV@HswA!7GXc8}CK-Vo=dSQkyc|cgg9uAx&2pbTKEw_K3
zu)o`&-~Sql9^w0&bSD%%dPv~&b&w>drgm+Sx<xTCe-ryTu;Y@{oY*&0GZ^;u0>7I#
z$$#GH6Tz1+cce1ub2I1Lg4Hu7OwpB%ijK}JEZhsG>|NRv(;+5yU(98jX1RX-Mx~5=
z3!Ww&WYM$Y!2=z0huHo@hxRZsGBWVo@SjP~y0obN2T-~D_U&CCXe5FZuc9L&n-lr-
z2G8@&>>j^Nc%GlO{2hk_10JBu;=`)M<y3?`=y@d5LQa-lk?8Li=iVKeuY;VvL};6a
z)GrPQ2>3vm@4b>IDFnml6FSPPGIi*Mx9KG8c_1Vb`jQEw7ho$xtL^P*g6=hp+>3xr
zgthhatCuJ(wB>t;vQ`Ese9+f-Ky1yxW~=)js7^Z{n!kJ0qSJOK-?DFbu-M3inFi~%
z+{6=}8Ed+%o!<GsQ{5Xvif<Ma6fC$H%teUhCTQJ=Uk++fRgl0VKm57@--~UH^_UxJ
z+-1v_PcomwyL!I%KXs~~p#9>dSE4>FR=2$g7+*c$_)F&wBkY&%n!yfubvvVy-A?P*
z?<EuzXSrVK5c9m#GCj`$0nK|B#gNfRf*M|`+T{+<>%Ai=-8d5(*dcs(aBOT30$TW{
z8pDwpyiG5srpBUp>zXLHoz~WJP8BzOpn=(Wrm<Yt!O>Br+5f?6-^LO8ZQ?kM%5)c2
zqK<^c>_e3f)`0~1EOy1;DJ_*{v-{*JQ#5>srr2Ii4sX}Pbli%S@-?boJ3DW3-DfpP
zCimyABC+fUuNxAhdhmh_2smEs4z5tF$%5?@bmPV#s_Wb2H_HDM<mZc3ZuA^S&YU@j
zENwo{6q@4lHETxc>R&<YPDoKf{}c<Y(8zp&qkW1rDOb(D$7f^QY%W~5aP{idTbL$G
z+-IkuSR&R`L5C`KXCdE_fm7O%r%hH;>K$b~r)#z5{s9P!L(p8w{s?hhWKp@W4#ciy
zwEOh+^|kf%`u($N)v9<#NQHJVX?drvUH4c`8t_>ut%CAmPlvAkm($3Ih>KfDYW*91
zza6avb$2@3Qy8NnNe~|epFmSrw+qO5Ps!xz(-(>zNbkzK;=fs_rO9^ji4(gkon7ow
zZu0x_;cOabAX5X4Hjk#-eU!!K%a+ys@OU#TLu3uqaKoohA27_b@wnE^)lU?Ezq@?*
zG3AK&o72>$G)+znnpK=-I_5i5qy||jpBzxh!TIo>>>M2O?X8*%@0X1EbG&uY6JRKr
zpc@O-*LkWaD+h&z?b~{(O~EQ}%ScwoQaX%9ql~7>RoL=E>Nhbl(az3J-N>j<X=$n0
zBcIoy_{q}%oe#_EcO8=*t#)wM^k_EC`clWd;0TtaEbN<C5OVX+kPW+(o1WW!8CKC6
zH~;tBB-3KI`jFY~y@Yf+j!Mf%b{{G2imH>A^XIpx`M95d6HDhMr5o2idH7IMORFO%
zcJEX0XEap@!au!BFx=;c8Rl7)tovb|-lMjfhYD6T>ndx|uUMz{=eC`kobEq<yny~p
z@lAUC$1Yy<C4Cvu@Nq`Yz6%?v{PwU3MeI^wuacCV;*!D0={D7s+`V&0^I-Q*ozy)%
zRK?m&u__k!+-<;M*2JN?JJDkoG~ktVneIP%qQ&-5iX8>rUQ}slO`Au|UF+m@lyM#*
zj~LS6i5ZA}p$1k~N0?f<`{+?O*iZ`z#$tLXU?JR=8M}VYMg)1pDcuH&PPenecbxpD
zf?_R7o-ctsW&2f1c**mQN7(!&JFHRLa@n#x_*IY2&l<p{P%_WR*t2H`5tZjx%}KOK
zbDtvAjiLVvg#o*qN00sj+KgxKqrta4bVXQ7_z#*qf58iOCLW`<C=C|&mF~*?obux0
zj-+<1NePOI(yiW<_~)ORqfgxa`|lysH$E?nIL?GXPqfVT@9&tt@{iUf=2R<{K<TxO
zDw5JZp`bIM=@KNpgwYD1pisJ`cz{IgRAf)-8M_?Kf0L7DV}(;s4s{Kwz6~5KRPP7{
z#2HuF?eCw@F*Y^5^VeVP*)U^0cdMRWcVf-UH*Y%4nw1)^x|M%vARgGil`F~P(V&IC
z8vqMGHGA@tW1Qx2!$|RYU`jIJz;|x$DMbzp?N=#M7+Ia_qq4HHOU=w=SlWk+_o?$H
zZ(_Z{z!}cV!#iIK2#_r-F8=3~nayw3C55k9bbB=^X~3xIO<~zTQnpn07&2rCdWpgD
zOI9vCXFGpMPow(K+fSbkL{E|4*?`h+23?zZ(Iw`mZ*KdvmQU)^S?j-;H|$ho<T!%x
zZYUK#?O&3BBz_KiKK$>pvhVRezcjNBHfDLBZVZXPc1@F|mi`Ykf;>AgnBuQbSs;Yz
zAj@#NS_`X>?%6Ys|C{}pRET2n0ys*oa<RDmF~_yp`PjqOwnG2J9`)~<NSIelI<ZHx
z+}^(P^yx`5H^b3wd{I?ZR8@Ne*1>IfUD;GbH}4f*l9!Pxe_aBZd{bTVL|l^Fw;gRa
zx#)<WDfEsRp3Mmnk9#Esp}lt%DNH1zf+He2iPfDp9yOy7a<7{+$<yIgU7zm}!u~jD
z)Iwc7;p%jw=GP=@=N1=_KCGyyc)(a;UwC-+s=?B*9L0~p*22=#!^Op>k1nbWD=I1y
ztI)sn9_6tZOBY*w`S6kCfq)fL8d<bHcX2O2Q;bdceX%JLMu+sROC`H^F5RA0`la?M
zUEdjOMxm{Nj_-^a11UIBM~&+hU6<S#c|2~(6^}GIv|O!<HhigQ=}me<h6nIqU`r6c
zMZ>lMUS8+;i4!N1|7*aE(%<^&*LAY~dJpaK6?Dp%MjD@)y-kVvf$nqULk5UFeyHhn
zKBe0Mbs;ui9@;L`d(Nt4nGyCyoGFI7&YV5_lFu#rTGgWLs4>287qR~e2X*z@wOVCU
z|4lVUBW)`@nte5w*;0@p&YOkjl_bQiw?Y&gO7#>b$Hu4*{dItT4;f&M1}Q#~l@ohj
zyjVW3=K<}d=H>@E-;V-Ry0SrV|Ink)se<A`E8N}PUl1}x#t?j9{(QN~8rnTwUfhxr
z%TgiY$T1)3^ya0@EUHJO+irnlA3J{hkJ?%Xp=b{patov0_~zw*&wzjCZ8_XovzBD7
zFIR?Ieb~b1S5GIDvA&p*%VE_1zdd>KgsNDJwV-eb_TVu1%eOF9IA`<uLkh8r*{K-G
z|M_|I>{j~`u2ErO9ca4jK7U@$alYCJ=0aLOfR;&17VGH+8!I5L%m(v4_-L`oNaFnI
zwVrmEA+d4r=H#|q<=1~To?Es1*qcE@*p+riCGL;da3DN<3_Ur1f7P4d|NCCL?(h7s
zU;6cML!9=ijBHzl^xws${5HDN|F3^=yj8mM$9A88FCG~xkQto?`txHom+$@`;qmTU

diff --git a/.github/img/network.pptx b/.github/img/network.pptx
index e78394680c38c1bcf4aeb040d09d0db63a844c34..be6de9e31cdd0c670bba05b2be9ae8df3eb51573 100644
GIT binary patch
delta 17045
zcmZ8|1y~$Qv-aWyg1fuByGw9)f(Q5Dx>)c8StP-Ny9Act?(XjH?#@qg{`=kU-05d_
z_NjWiYNofQy1KS{W&~<>1PVh%9ts*001JQz003kFa8XBIM<oU=ByiB7gB4ZkHt_)^
zt`>`JENLRLEjvb&L?(}9jT)-X&2Eqz^Q8p@jViUS8Qw=7e_Zb|LTlD*?M7XB)?>`5
z%}&bIDkm756<J3Jy_IvTVU0e@PfbCc5E7u!9<TKHR63n%^NYAR52<gA&+kJKQWz@7
zAp;V+_QOEe(E8Z>LZHPu4*{S|y95cJeqXd9QeIG|@fzUAjXvU&w&p5*k!K>>^vKBC
zkA+H11G^77JZ<%ZMRw;B3pT_cj<pO1<wvI*f6}-dQ=%@BF#RO1z}AAHaa+oJ_q5TF
zO9GAl0q@cuCrC!A>u8<^7E26KY>_s{h1|xf$vqFL>7Nlao`J$BWNPcQ(h;q8*ij-9
zA{HM<qIBTQV3b-|4}Wzh!TmM^F?pSKapus4nvn!O$iky5&eGaP?Zr-c610EsR3?K)
z+_Z;AR%YnS2&)yd0drA2VxFDkLw+;7j=xq~z8%W9*|HG&jubhM_~RD%BXXW7jLje^
zcE318VV3HXO@dJDetw9;Fgp8kQf!eH6U9`K(&vN%Bg3Y3g1Mh2{`E>zSYM@$30I$l
z^gl@+21kW%!4iulfKu~c1&)mXpg0yxFg!FT6k^chLrx$H0I;PB0g6G!0<tlC+S?r&
z$j1=$qP4*F0ddcaDZ_5uuN%!PqP_5?w{&8+NCVb8!?#Y@$t}azk0jW9owC&(V4=HU
z(c9plAU?B0lK9HJ#8NzmT?zJ4dDOF5=bfp-MM#_|h!h_@HZ+tp?mVL?NlTm-8Y#ef
z)hq^jSoa=UE;sS?-FOK61PaTA?|XvShrRb%M;{u~3!Y_ngBx3)x*s<MKqjLu%9bF3
zeTUsl#$FG1ck1Qf7uxiu)R@N~kBf$e(GwGXOgA3hQr<^$L`WXEc{Y{_$W=;@zRu~p
zm(9&$wItiiovI=Xsp63@tV{-jG;qP^pY#^)`18I-y>wxzi!_hr0vpRMm%};@ulevq
z=pUXB=mj2r;ffHhvK)xK-t76WBY7(s60yjCP-09^zd<>+zn(oRY?<S`ynDG~Z+5BD
zjxJW7<vIOSyBYFY{QxX@C<Cf_FYol8>{l!YD<>^23WX;xHFK01Tb558uoIEIZ;)Nr
z6Fy$tXR&N7%s2482VSoV&RgrdJE|!ydg{D)7cioHhxjQCKE95D+=^5N6;U3&M{&p%
zoxQxJGy~SHK>y0N9>@tMQQ39n#&?bKjXzC1b-z?Cc`8EUa|^U$Dzw?eiMGiTw|*D0
zKm!3UTgu6H?#En+AVG|KSA`W}dwB&u=~h}?`~it$9>?5w2sB*rV?VE2DgsaV=PXkx
za}87o7ferSm;EAd_s!1<fJTEOhr~ed^dpUk++;NocY;m))kAs`eX*jTbn<w;#W@o5
z*v(5a#S5_zc|c2a;$COcqb7#usK45`S)KK}Ud!I`_{OexWR5W)FA}_f-!5alm3{wN
zz9^)JK2Jko9N5FHjA*kh;DMKCz7T%UQ>~UC6ZX6yH`v>G)2Vy0lKDLW<zf@RDuyWi
z_v|&%GzH3FLFdbd=0>AmG7n;Hm{}86<l;&Zl{4sr=uonfeADa6iO-CJTCXb8Q(E^X
zY<pZp_*Wgidqktq&S(L4pCGk8B{l4P$wjR;f@eY=(LmRvWWdulHwASHD~bWVsBPKU
z+YBx3m}Sf1!2R^~+|n<aK)l&kzUJIcKb|Ak!R$><g^cCFpbgom(kmsV;f`f)g_8Bc
zpK;10?n@lo%@2>Dw5Odwf{d5CMYfURrj<-uWxmR$$B|#Yju#{T_lC<y4d-qV&!rwB
zOcQY6pv52y7+@_H6O(6wgP-PFX|j3kf$WS>UsOKJu>Duor2Q0WQAJc;Q_^qY;gCDu
z+bWYQ#F?YH<4j&OW~KaneHz?Ros(cDIH!JV%$l2Z7A|n9mub#ebzc`3-Jl}@7z^7D
zak@(rE=~q_M-3$!Ee8xet`Vrqwi4zQMNi1cJ&5Z&#%<YIk!<Ya77uQ^jarPx4)jW3
zXmPZxvaT0Owog~>V)fn!tUJt%4;jq_!Cy;BdOoE{)|mtoawy0lek@IKsL*KN^fu*F
zm%P5BBtiD>H`W;yr|Ev#bkgBmRzR0q)IIC4PlT!F=xb!jcm@TM^8vjKcDFzF+m`8-
z!KhC|@{C|>$uf`zxKPb_sM-}Lia3t#b<zDchQTW73Fgrn1pX9RCH$>JRdTkZnWK&Y
zWwR*`_FR6t<=$w*wP?|un;!~Gzm5z|{}PU2tsq*OTFX!M&ci%}G>c%pXv?9clPffR
zahelR+R19R!bs$Z5ecXnYJ}?!%TS~~@X(c>$AOoZrwm!dgcjM;6$u}X8C|JzE-C#Y
z-JSl!RDu3b!{ck+rT0~M5M(z@<0s}O3!(TKy^*}i)Cmd*4V98B^YUKr>u~k%b|;JS
z$(+TN*2=CtZ;Hd&OJNM|i)?Ar$(2x>TK5kqDYV@7mnk*B8|y%I6mr1%1q=%-xH(S3
zN}khb@(ujd(7Kf-sN!&2JNymr%Wu3n7nRm^s4anETj#hq<OuEGif|69VcmtVgLKEI
z5SQ+ZtJNqYJ{}8WK(P5^q83}BTAk$L=_@pkl}~AXNakxUn?mfUQ++21@@|)7o-?n<
z0R;5lNDA&4wk;uk_kjSquo$0war?Y5N&_yFoc@^yR%~4M*Q#p3=<FQt^{UXk<s7ts
zcDY{2ez4^F`Q)xQ5wB6FI9+mZlC7db3!nS75^rcfFeE@+EJ-p|EOkCs?2{RLKPg$J
ziW92X`I=J|?2qUbxe$aa-*AMf$ezMam|$M)w6iwIKvBTEXqI<C?Z6Nn>-GSX=we-&
zZ3VWqlU1hf!s@lQtZ%ApnG?p*2xc0%IUTrIJ%u&A(o`c}c6Rl>7ByDi!o}yYvq)pP
zB?WQI#BD^g-c$}}4%_vSR+*2}SNm<rV<IfPN2bw7Ez|Q%W-2`xv@>s$U`QZ~v(U5?
zyX%4$rV?~2gDeM9sCZ`EcxTI5F_&t&sqABCYfk9QWE7ez;D?UWM{Ai>Bjjk}mMv^p
zC3X>grA;`-LV&cE=n5XE&xOTAK!;p-zwkbzOnmW!_}eYWXM5K8qUx66PK!^7=uDRa
zlc>@bap6&X<*VJW2xd{yCsp3z1)#*6MZakrz%Tl3Z3v9Ldtjq)PyEqS2(9vkjXqk{
z<ex`g*ud3vxxyz}ut!HhfU&o5>He2a>5{RV@f%e^T4GETWX2Mrc2ic8cGKaz-`;49
zSixxQYm{1Os3MqXZ1kVHQb?=Q`rA|@&`K?SXwv4?Ro)~i=~^omaaDaY2FezVwARSc
z4-{HK>U6N$umya9V0>?E0yu<yyWXz7P+^X`siS6kvi`i=X<EkT!~EN>%9xHCt)XD@
z#mK_Q5n?6c-kRe;mT3Jiam2DX2{f3#tUrxsM+M55H^BZI{(r&PlX?Dl4AQ?Hfq!}$
zt$)nt5)a(MRTsy6Bfqs3U*=Q(_JnPx{!q>>l`R!al+U!E_h1O==^?ijjH(pz{@g`(
zFkvKNLz)B<4DtwC^QTcEHCi>=NH{u7I#<3gHca2+)!1@XpdtQH)%L#9VM05VL&GMZ
zLd#_RX%(mm1=yv)%;fZp56Du;Qs;j(Gw-UDP`W(V*|1$5Z_S$`detzG%wD_>gSXW!
zN@F7dnxb~(oRv!9lrx+Pl7!Gp58o|qxy1>j^X#!B4FyKARr2Wy6tkJ~k>K5Q^UK`Z
zdL{M0TOnfDn`AR7RDW@SOA}aIg4jhqWcAp~_=M8?)rE*aD|y!f(Sl#!^E{@)Q1EfY
z=X{pL|78~C<Nyo@BB3P)x+kq@F$zrLiEi#Phuf+F^DUpPnE8G$O*wGl?nb4bXEP;A
zm-Ixq`xoYx;rQu|&GFTd++4<K8yvZZ2A4VqPg2ZIU`rh{{oc1s^k|EoLTYUpVC4!P
zB9FHKG`^Qrw9R(?JSI5hqJ7NKS&FSsL3gXGsxY|)_bgNZwow}wuJ+R9C|FN@I5{X6
z(q27k?=h*nYQK5Wd+_mjcJKm)St?I)?auFa=7=TX3?k4fTX6B*@-#bo8{o}hx@&_T
zXla2rp~2^;`7_nf5g*x+A+y3&n&>dQ@n~C~ZtkN*o^{Hyheh|DTe8DBRkN)zRW=?!
zqEm;QhDs%(mf?Nn<dB6tA4b)^E>F^J4&v75Nd+yJbZ9!~6jkBSc<;6uY%cO(5$NWP
zHxibEs$}3E;mp5V{@}=LvO3V`IMX^*4>kd&T{m6q&&`|YzE3L<I->Yxgfe%)-{zoT
zaveZ4AafQT8UoWZEDPBlP9i7WAql2s*VDg#1>C?m8S+1nXP}MNhFdz%SQ>^}Ru-3o
zQORbJkwR?Y%x^#J^4r?Wkdi{8^~%{i_Y{<e+Kd!6I6+$&C;q;HMdl#KM)P4at=I(W
zMQOT~+2<}La3hBo+n`X8MPyNv4;6lYKn>x>F4@o<g-Q)3h^8Yah<iTh%c%Zk7_>h>
za86;VFXq1yk!Zk_G!PmQ4c}EsHnPFRCRQp-P8MGF&J4rAqpxghO_S~G0_n)tChhl)
z;)p7wHm(R{bOO6{h*@=<$32MVd_ReS$PIwBFOyH^PD2+6*ov3=J$~HGTSPefp0%KD
z-PTR4ZPuYMjz)sVusnurruurI7jDr@y;GhLA0d+TJqhe}%~xvIsnAG^tQ_tHP7>D7
zcTI=5JG3OkS_2iEL~iMB^)Bs&RNY3XKlIRJ_>X;=Vi*{gGQdvYVn?Q_KxA>C5vqfo
zI63{sK*usLMQ&GXK(grYB;Q*Iu~v=EiYZNB;@%69Csem35BtU0dt-TY*(&k4BIt`e
z|K(uI^KBIHrPt^Dc~<20?r57w#LxZIWpw`4Mo+*j6Vd{5qxd4i@nd(AzMStt8aOs|
zqTmU3#@~zpC)@jSOUgv#!vQ`}=g?S|;cGqp%-AYJF!`||np53CDnHRA&WWC@j}3h0
zpv%m*t~g!+=X##*Eh&q;z;cyU^F&@M?Zmo2x5s1&>4T>85KCvJ-Hm$bni{+YA{n*}
z=9kJ616ETVObjBpj|?a3?T$pfhH=T7*j!qQ)cEqYCd1Hh0}EVe2wGLZeWmYszsjbL
z^L1`52ez)#G2!i?Z&D#{?AAXW1Pvq_P(|SFCUtkIl|{UVMkV2;h-SB!z}HRrUb_(#
z1c4Aa%d0g-Yh_5p(b||EI8_Z%^%<jFS9l@OLo8WmlgTqL*v?m{!RP~`u(er*WA%p7
z^@o_p<_wMkBbK6vP?fk{VCIA!_+vYmg<FG@RXD<TBqA;dBr2y=YL5R2lUlr)s1V*f
zy>?+}ru0%{ol%Vu4(8Z&eMWJHcb;V;Znp<qW#rj<+a;$kZ_$4A<41+f7Jm>gG!d(~
zn_a}ZI2uDVH7C1C=~5Tn`|?KFORS+M(+{Jc?Fo2LzH>>>>~icx+Y*;<%Q+T}3n<=P
zm?^h1=~svlV?8fu>$_ht!<U+7r5?eNqLY)CMEs<cMeqLtPZSj@n32&WVA)yg{DhnU
zt0K3sQ#w7d>G5ZUW?$U7vFZBMugb)s{e$A$uNUi&5V`f&S4qXEC-tuoe^<DIaIVnc
z5M*r>@GrzENF*kzbysH?0DueQEtLUc<HiJwDv3h_cOs#+DK&L<X}^7;{S@~_`)Oj=
zXN&hUkfhSarBe9W>+@;tEQ@>cb>wJ~VNniU`ECQix(|lCgP&D%58|jk+_YF+Uwy=|
z0&2%j{Fl+!r>buQO5j%)a=vdCBAMuYPx?0dC2A>xjU0`yrn(pleACcVoP}Ovkold=
zk)7^bKsPKH)*{l*8$x-HW)vr0#SZzLD-~VN4}I!2>;!s}S=5>Q;}FiQu-L2UUDK|q
z?3OM>ZVu*}=wJ_ys1pssqv;copG^zTx241Kn9D|Azp<#=YJMn^&9rCb(Pnc2^iBnC
zqak_Xcg*F7no}f~l&hSJO;`5h6qWZmN}I>^3T&Ll6jyf;JiZ%L{D4muSsos<kr7_T
zuzRqu7S%bB`_#e09P_1V?vny>jTL=eZfbcHv*7KC$Ls>|Zyo_@zv0tW5;OpS`Ibib
zJFFm4B}A~htTq%d^EdC@Mwh#-5}|XPN!;C`!^z5PAJm1Tu_m*z(7kib?PVi)-`sFp
zf1qSvCY!>VeNS`|B9#76G&vJWp?k(2t%hh(3wzX-WEbm~5%Mx2oLA=&67|4q9JK3s
zef8tv(V3!`C;#JmKx*>}->OLCP{?QbA5M&)QECLZe`#O<A--hT5aDyHC@!;e<s8e@
zfIX@ER(qRGe3`JF(29ma_-zRA`JiNA`r^$qK^n-hWrg`-H*GatQ5ss3StOKQjRgJ1
zjs-y93-^cJY!D7Z3Q4jwi&jYy$9p6(utZh8OiP-%yn5hCYz5^9vQ^pq@o~le{3-Kn
z-$ee2dy^?q^L-HM=tS)EQp7QSQyf9u9BPP)ss`mOq_*pDb6%g_*f5!S0&v~JmLe2K
zH+s#asyv$9;Kzuf)zR?k(ZD}>8YGvr)_|rrvsK6b@A37P@Ph_6=VnsD*vNDAIETk$
z>$=j>LRDI?ng>7ORK?Tj{`BWrxesgw@7_l$(-|>+HBQ~#3b2BO#cpSQwc;O(*!097
z4>xby+_n9g$Ymlz0!x)+J~$8TUtIkHrs;z=dnj370iAc^fWk(5S5N3SEj>_(3B6n5
z)Cb<p9($GV&SwG_Cs6~0cL*A658H0;19rvtKp1P1iIO;>`*GExz0L{sg|=4TkO(JP
z1IIgYT_a8)Iw|rG&p^Y6?fa@)=Orl?gv+m=gPNCKE`yU7fke0KR8cntMg%OnwmQ4K
zHGy{^B3{%-z^xsIt6%V0qkQk2TWNWTPmmi-mt)m84`$Jx2J%+1LGqaGR@FkTE>J(3
zWR1l^4a8w`b=ORix_env_qw^7KRzbThj>F@b|n;TcdVd^-+Ky@cPlUN4W9}AcnKtS
ztUjjOH~SSr|F7`{p=`o{(?3>10l#ygr^IFQb5MOtoimO@N(y4H^X`DUVEZ)zrStPy
zaQ@-?dZow2@b~2a!%RfQVVd^xO~9$Mq2gdjfJ$LQxn*6S>qy2d8H2uYD=`dJ)Ii4k
z?GlAj#}N&sn24vg;GN&ap84}$_f`Rs^@Xns(C3sD<0_3k(03yhe+uAE2z;;6CDMgG
z!L?=vk0fU|%6&G^(f)H^TYA4zr)hWc$+>7q3|rG4V~&3s!tEVIQmBxqkfMiO#=Ia7
zac&iT1e|J5nLy(E8=K9FQA7JDznT6Lx(6J8l8XEpRm(+csc-FIzAa^ijTHW0z5|=(
zCgOyPzX*@s9U#bjQ|tXgA}wwZxB6{>JSG(=mbnnfgN*LK9VFBHL}nJt+{R>}ZVSUc
zXzJqmBf$u!d!w#Bilyh=315{g)<Ls?zT@`A@#j78uP;XGKxNd>garVciT`Vq!J{Vl
zFu-U$m|<F#bUd_Pi3A<V;Y0s}D9bKchuF!sfg2OQmqX*nTaOCvUa#qb-(wb%JcQYh
zJM1`nA$x6{?+NOVYZRaa_-{)tcYh}q^iI#8QJlglcN|$#w$Hs1p0ACzZ}N2ZvcR<-
z(FWt(-VzhtU3TXuygRC+?~MLAxoNBueg@QEWQR9K=Hcpf3+YT3!O)#+nmHtn>i)1o
zpEfdJ_epJ<Js?k=NF@x_*k%r9YT9*iV`PWlX?39LS0|zx@iL9`uiDg#R?o)!Q?JzI
zB@GhT65oC;*>rO}%=zfXdfL`-Gbfg_LRdj03{(A~>Pm(}c~U+(qSLhK21-N32};3}
zK9&L%ruM@}GM0$DR`^oY{Bibn7=s)u7<T!MjWhf2{+l#Wei1k#c)guT1`(nSxcvC!
zGlqWj8LudREco40f8=kDHw&%;i&^190qdY-Nn}#^b<ldnF?AC5smb?Vd^Vne#$POd
z5x-V@o*y=Xo-dp5YZJ6bqk>eT$TRl?VLuY^ADQU4y}RahH62AMLs6ONNvLk~7m^Aw
zo#0q}G=u_^#6n*l?H=8eM0N5EcFsnTpyYO!SQd!<OtXYsQKNp3qi2VM!7D}x0i?00
zU8u}Na@C)GUkGbK=*cq*{m9+K!^GSz-Ws8Xy_b>qz<g#u2&)B8cBVi9sr6|gaFkvj
z)5bEdv>%L_%-CRCWbAnR`m}ZYV#H(6BsmGxx$%xL{E&Xym_*E5nqtce%j#;s)}8du
zbTWYTjEIGmT*;uUJ!-15zeH7ATO@|Hs6@fn+tvnd<^A^2t)V)SqPqR$0kp%up44F6
z{6<*kLYi;j!Pg&cQT9D&@Ld|iaSn>w)PlD!tvw5?|4eF0;&ZKYPyj$O!hf6VtX(`L
zkc4hWhV3e@&_Y|CIJ+x*;D)XH_H~ddo%d_$xC*L$+tsSc@0wpN{`FojgJHG3$zEI{
zs)lY9)yI8o0Wqe~nGh1F&}gj&C@Q$#q1*d2$CvzO)lfH$F!|AAf%EOR6s(UT@c=mV
zk&NdSuj}3UlF8vH4D-vK{reb477i0XV6o89hcWEq!3Zd#5Nb|<csY^+&wGr?a6-(i
zMa!U9z><E#tf|S0Gy~#r+jtyG>|tG71Sxc&eRmUV4FUPe{6^ngO;t@FhuZm`upFEZ
z@aZJ#wimeMei`3+2F12VWhvwafBy~7yD61`6@vA`0G09^68d5gtE3^SUS($a2eviM
zgSvBzDXwcX@l^igSMAW1;My*j(P1v8!o!!F#?!rqa4k{}tnPI)vDWtC6}63SmQBx)
zi7EKp#xfjz$&*f$?UDaKw;hhVO(VQ{1znu~4D%K$=*kHVG9?ZK=WGEX33hO%hk$ck
z6O{s!IfH(FW(E$sv!*L?y*JkFg9v3504(&eqq=5F_vJji8~dKhFq*AN^U-oegIpOW
ze&qRMIEEgb)?BfZPk3?4QOb>t1C|)t16O%N|JQRf;i{gLYz)QCZVCFVkYyu#DQ!{>
z8BHlXd9|<KOU<|W(wRanHkiQU9{Y(iV_yZDLt4IgF#H6X7&2lHW{n<;(au76`QKv0
z>>5qZ_pICkhB9!!SBVHS=LDbJ=8(#rz@$B74RY@ILzXY2!|*U-doISbr+3o_nEF>Y
zYn8Xo`8~-M(sk>tmM4s^ylY%NO;}W(bJYwiE>h1h0pTg%c#hB`S$<e;q62!<`Fj4=
zZ1JZ!awrADp9M|ekVA=Mq$ae)BfExEW@5;Fnr0<q$J)}-5mtVHU2%H<Bcr~!l<u>l
zfdK`4@ncbxIV9dsUbi@D%#URGStodR1B#xwh`(j+pfdA6Lt`tN5&*cy(R8fQ7XgJu
z86oz7sV*9mH(I~!9IU9~G36heKY9&h8i@D^Snl6jdYvv2?{q)0*Kbj;=vZo>#G=J)
zM6b<5oto1NJh63Xq9u7{JK+OoXG*K~kST>neioDZf~wz%E-gB{-(h7O5@2$GTm5c^
zYIr<rJao5}x~f)pOXHy4B>4;Xr^QGP#&{Jq*=KkEPl+0wx>|G((P|Wu4mE8&Dq=_D
zKp`rQ`@$Eu2Irj|>J`iyPM}7FjrR3}aC)mBf~#nm4X_~^*M=YXm~au0J^k|O%Nb|B
zi)Y#q7cy5N+qBS^PN)SPfsokG53a~4@K&3#q%EOo9qSN@Wkl?blQ`SFbPj~3iD^}|
zuj<WOy#j5!+cz`0Iys2+Gx^s0H5TLh7@SB@i{jX<u82?=wsl_@_%*PeLHce)So%De
z7~dly0>8x~xek)GO~K30^+1EZy3qjFqDkdB%zDOA`It;k^Hb}O${rGMwm3g^`+V}d
zcC9`?3`<XxZ7Ie{s9<Mz{ayA0aM%D-&E3&PiU29MXgydQxQLo@7?GDPUE4?B;XmyG
z?ij@%$#6=z#~gVHQy}#r(Vqd52>ECW`*1us%i}g1E-o+Cu6N-iU>w&J7v=-ApoU#P
zt?kJ`G)T;w@YPVvOFHPw2mqC~vOJneh-9NMYSIfp?1=a}>~6;B3V>;whCwYX0%Luu
zB-zAahe02;q%a7B4r|xf$e?DU`$O<8gX5@v(_P;E(fm}DKG&gwcX0srG^DyTa%L13
zL};+)S|NgXQW0K~DaTA6JHPZ0{{kP;;lx(p&pg>S5wt-y#il?4wRh9z;=u5!>#{yB
z<RmtWI3WUWBrwbJRZjpoYD8+@<h(qAts&?e8RPX1=<Z46AHVEyx$*_-&0%oi{>Ndw
zEjD;PMxcSL*qGIsqHZQbyu=%PJUFhBGdRky0<Engb7@;8?qc@~qGLTZ!vV!S3S(pU
z?V%1SzWszzDPwJOf<7(o7uBRk9zs0B&D9%6A67<$zkql9p#84stRGL?Cu=(kDqvfW
z=z^2=!y46*Xqq$A+Vb-j=ZpcS`cTklM65mUGk*Y+WEB($?Q|_iBN(lW4!`tHY*5(m
z?&l6kQ)@Co>AZz!_Nb5wNCE-;_>PtJReVUi@l}jx_eQ)i4$KHQ4wtJeEA}mXw&(E@
z*Q&L}$!UqDV8&qbRA;WR+;=jaZ&O1qzLm{))J4<$uSD$)(D6B?o2lSR=7>`-YxKv9
zLmqAo`Hk!401uh(D#2tqc7C7irg=?gi|hD&-%mI$ZFw6_M3MCDn+%posx+_?rG_t+
zj6M(kvrtCp;#eQ}^%g*=qku&DC_s_In7~8sh97fcvuX!D)b0=EHkS6&4Q`x`^Y-wJ
zWl%l&Oj&+s!ANv<-;=}@;is7HIlYDkR<k%7JpB;oWNYrET*ilfxa_>9a!fQo22<`l
zz~iVUF`rGO^|{a@ft3=<S<Yf+e2vdOiUn0;bm>N7-;vSZJw|~!ZZGc1PZ(i3OMxGF
zZ77&_f?@oFV|@)#Bb}hA#MqLzY4c7S@RT6MQ#}p#*4Eg5r*{Q0MyA9<s}As7o-6KH
zE<0e7^C0@XlasZF&fz7UdNrjOh?>#!$FSv-hT&2lZrMY}+u_t}VEbg><}%H@R3L=j
zDkFiSy@{L*`|g<5=qTp(E}eknwF{W2PqS%sTR&+jk{}RPAV4!cMsCyzU#Yk7in&{I
zk5JAixL0A!e83v@mCDe{kh8PcVxfY*BX9rEp6fe_t8<531ILnaU;!lZyt=WUm1y0s
zllL^u>pd4-W}Qf<DdXae0ZtC8*qX;n(@tC5X%=R4p_<<qv2xvt!g63<B2IyWH3#>o
zch#_&RuaZOW6&+7xEZC7-%72N=j~)uQP_$lA4(@XsI9f3m9kMSrDmvHOp-WrtJ_rV
zsphN(mwY$&im4x4;NPoi1(=J&>9p8G8aN9{czjgUEG-yDR{6PD6)A27=|Jw7NPCgT
zxWrh;nIp$$V!J*fJ!MnHzG?*oP@qMqa+Le)YEmyQ8H>WY&)`)_wPc%5;|=w~=u&m!
zYJN=<qUT}NGzBM1%HFtXtYg6)u{v6;U4{wbbbTta$-xxVE6bya+RU-<IhtXc?d#E7
zjTs1@UD}VK*&cVn_HTYS)N8|MvXL}~e(u1Wz`@L%j(&1{?tg!Q<<S5X*w_+4h+6jy
z-Akd~?otmpmVj4~stI)XUTn4yPfKueWnwaM`7x9trRrd*Yl%RCdNq_Z>8c0|)3^Uf
zA*LW=Bsnf69ZzJHA!aY?xcS0~Q$SsKQD~LW@!2n{-bcsbnQQ;vnA}8u>#h-Jd;G-<
zIZ<FBi)pZ&Y7qZSGNv5JEJ30Zc`%l;D^Q$;8|k3@vlDr{{4Plvy=ytC?V;+9R<>0B
zbcS9}KFe}M&CZx1>AeEHkTHJ?vYkXNm&7HfLC$!M5)5=4t(brHN~cx*2l{u8-OSpQ
zT@k9W8Osr$;^>V>oG%DzB3v*w3Zt;Ftu3kA+B@}$V{ZdBg@Uhvcgg1`j*iOGyIx;B
z613&j6&Jb40&0#RXA2NhvuyxI>;}qNw(=Qleuj<Evu*y9h^%6D<~M#5I|ARINo2=H
zm?zRjdf_StgS9t*-qxQF9#1wSdK7tUyWnyPsj5k?&80w7qBOKkNWr`AMX0Dz%El+|
zbBF;AiLc^qD`N|QaRY?&6Os>J)`uy`oXP&8I>j3h!|HCm`9FqT^1vEeU8-<-Q;jk4
z?JLnQ_glkztJX9X#a2K!z2;&)iFEtalp|cuRQaogpz@+g6bD6^mSWMpxzKG|;vDAW
zw*b;=cerp6+zY>}tMN=AX9{H~5^ik=%lS!Gdq`_7r96VN_Wcjd8oui2a~!djrHZrc
zC{@9<Vg^f@SqT~moulK?vkl<i8R1<`pxXKFo9|);iSY4Z<>GzKr8>;CrE>$J3ixSL
zm6b`?egl2w;{{qy@RGUdVMFN}sC=TM%Vt-Knhx}}qypqKPIEGDurfq_BCVkL?dx}b
zkZ7yfI<k@H)J&)vaZh{T&l@{STlUawNK{7UrNfe-srJjGGk%NAT(N2)O4c*xegvG6
zPg|E9W=&SR+5Z^L64uEn*Tjy$Eh|PqL)kAruu(BPQ4Bn3(mdTPBbOas3LmGD$Uel<
zXOe$!AZwj8C)LCrM`fbibM%#-r3zDWP?r(QWQ5-$6>gHGHg6t$yw4&cKlyx`+Znbx
zM0`ptER34(_FFtYWH$0VELX4G0d0M~=MFn=;=7NXP=E7{lr42iw~mH)L1=~h^e+*-
z__Sn3i3#8*QVz1&dvNb#{JX^Le2$b`X<MGrUtedQI7)q%p_|P>7n)q{&Gm$3`Xl{a
zXH5~FNj3{o+~oVBEhNw>Igb#K5G4?~yzsojRic?<le5un)1>}wA4=*-;v}tZRF^6X
zHh8W^Cc_il_;L|df7GtIKCu;PPRR}DG^1#uc#8pzp(ZDo0?0lG6dA+NN@wau7BH0O
z2wGvZo1Vy3MjAGra6_b%y8YIW_OLCY{bBrZRl1cTN?OBE%a+M80Zc7VC5xAc<bPb)
zqoJ`=FQV=_IToA8882R1lHwl^fn=NSIr@8AQIjjnOK-i$ZMp!XAbR$GZ&as0;Vt{Q
zMfn7*{>pe{$z_m#s7sEI1q{!G=V%cam~46y#Imy_TU9+tTH9Q#!nAOetp-6Z={gXO
zcki!IytLb$c>Fq>nI!<HCr_-`!hxG<1be8j1#wIpRs?0xZX2^g3Ni&BB=cq;7s0zM
zxt(Q5vW>an4G1?#N4jm}S??6WA=gL;0fE3W+hU~3;)9-UiUBcRDOyBFb`oD?wXC9M
zD2>hlc*D;syMT0y>x0C7PytxzW@CJE{-kG`0UV~^n>snXs?1sIZkgIT?DN8PP$=Q|
zFb$qxxp8h=r-E4$YayUTg&c%u%$)0kt+6*l6(!KmNAer|#Y}wm9l`+EuZ3>Wi3~j6
zf%j!t+z1i=x~88=P(=`|Q-*{!T9pMwXl2_b-UpXNfZN7~1eZG33+qnoiWA41`%dwa
zQbzj&Wj$ZEc!x_R!!(~8n3D0El6ONN<}YkYdoi`dFMQ67l7Uias|d25k7*=*B7vGB
zL2j>M`!mWFDBRyJBbn(cS{&)ikG+9yw+dK=s6X05RiHA=yi>{r#za@TFZpwGA#S2n
zWvs;cEnY1|gi*k+6z}f#Lj0c&U3a(bcDt`0?X4f4#@8n?!m9nfevEyD*%x)e#3D^s
zw3A3)rRCy*cg-~ssPm+rtC;_)>Q9ZmAZ2G>ZeiAK!h-=t0Pul_&<_)CYjFV9AurDR
zDhn+P958+b1xelhJTk7}ez{wl?)iz+>=lvy7<hPv{yT}kxANCciXD&keDiy+BSQ-~
z9k@Vbw9iD7?pdQnB^p=MZl#ycg(U+%?}`RAp6-j~5LbAI?#v^XUm3#%{0QoH+CQ5n
zc{%9(h*R!3*wGp5-Z6Ep%P!eKlH|#E!DzWUQ3jr~{PA4UAi`Ay&>rg(c#sJe0<^k@
z0J7s90I~AE2MusM{{Jftc!l+iiL`<M-eVnu0C#i7L4d}Yr``d4*{=-VYBHa|bLxT6
zz#*h}jxsXy-_iX04()a5`;8X|g-%)ry9!>ZwX7y}&L0jz1*^|bJ9T)>eUYJveYJg+
zYu%XE4GM%opE}(zbzOWKLJ2#Mbel_1a3t1(bk=~lK72=zC9$g5Ny+c<Bm*IRbcaac
z(Ke}jqZ`@`nhVcGw#d^@@P0l$Z*?n!fH)!<PuKOX7oD4vmM+(oHQA-rip6S+;#H=4
ze)DYM3<T8$ec|-j`R*tTu+83D*yC#}ye;A1w=U`l<zDi-n$G)uHLdyNcaUoeO;fAN
z)D7~5#4|r@JgUgpG?+;^oZgOc{=hjaCJmfEFWEY)?R|Y1Ufcv4xjkSfKU|gW0Hdrl
zOKSr?=h!kc#6<ZfII=w5i>Mn@<igDel5%eN4ml)8z{DY^Jr9h}jD28EOv`UFMd2)M
z-Jy|4NU919p@~Nsx`TB3=~NG@FLBATjXv=7Qm}@Bzaceegsa_eS(8neJ3qoQjot=_
z1WHdAQq?6XxJ~J0&!DTx%P@`G03)>It(XGW1R^^=k)*Y=Fj;?>Ze@URD-rw=a#yuY
zlxODdp+Ytun_vkH-|JGLV+(X6V(vdhmfX`TEpZN=X)v9vl*Hk3mHPd5;c=A$e9SK4
zc9kO3&z8q;fV`n<(38-9g)vwd!=gcvcfAfRzejA4#y}KXry6i+o0i=L3Z5OQK<9~9
z-`8(<5XC_|9|zVpUjB+4HW$Qh=RK=I7g((zhn5`D^?-(eL_faaJ;%cdj{~I-v8d;?
z9&3+JYgml`E;7F4xbvE}M$K4RCR&vTUD)NK?aEW!2B!h(F3Q<j@f(V$@+mt`$1cgB
zU%6j2zUd@NHFNMlR`7=aSw@!5qHK;3FFTuVIpWwx#n622u|=Z?<pdN+(E`-B^?vco
zRAy^PeT|+63@VZpj5L{TXTfP%8pdOfY5p>m=oeqpdxcO3$KF%SMs<cK(~+KiFf@d=
zOBr(SMV1MvTlDsMvZu3Wgk!8B(XJ$nml`C~ttO1$9u(s|fB)GAXw`9N3#hM=3pdgQ
z++m^ZvTBEyaCDi}mVW+OLDovS)nUzm%-35Vl*UO4=h-c7hjgDu`eBx=+$D{n@3%C)
zHm5;bUFg$}*obpEX#r9nmt?v8eNE_7eC!{<HSAWtp<e6%@Odm^OQ2l}-sDF7We>w{
z0Yk-Rw-=GVq-}nD!E|+J2DzzmBV6p$AqiF*Fh|?C4|iVdXC@9j#f>B~#vkr%M#Tts
zX^YstG78cbrvJJb4Kny$eEIcyK`iPfv#%}mm;Z4li~x6B&n&WDQ)>}W?3epk&@$kA
zBl5(cjb-uvA*qm&HFuQ!c~s;>)WU_&66^)g-u?yduUBDMLjW6dLqdX1w-7*4er@3F
z{45CY@RwE6cPK$L6V6m`Ie4%_p#v1qbC))bY4Zc-iSV9X^aC9!#=Vwh_$y!w{S|OM
zgYt}@a4aW++`OG*b7<p`2zJ~(x`YBw3N7slCqVqLE%r=R!B950v>=t;{o?zUjEsXW
z!<Ioq<*4QDFXUX-?`}vyS4!fhrAV=IaxfR{q>xsA$ar+oxFWfe6Yh7pVW|Y(lQ^`_
z?|7x2-xPUoMpmr<ypw<Fbor%Yc09X!Ia{&tX&7L=zfc?HhPxZbXs6zvp`+wr&d0{4
z2AsO>#`mw-Se7^zcw-z*Ho@_*rzP}?90(n<bQ+x{KW97||9S<)95%fI$_{sa6BE7y
zLR#@NMt=c&ukK#~`m>RMDd4Z^Cw>3h@FZ@A8`+^h5@L(zUjeEw)3#Q6WhdDWUuv@K
zk}h?W$f$YPrSHE|4#XFGt_Hy$FeIG40-mT|{K3^az~0N<XGbGu{?_%W7WrYlgB#wJ
zZO0+UEig3vnHVVaFd7m_sL_9oA<#gZU1%bOsRHpoCr}EDu&Ntck#h(Q)%^Pz0FxU!
zK|A#0lhQeDOH<IIQnk);Id#B0DiHAGV-}4KT0h>}jONeoQ;Zd_JkCW*G(|=c&I#Wh
zYI>Ot*S^xTL^DH&bSAxfe@n*m>YPU5@a(1!Qg>k+%ccwZfo3K+iGsg2hZ1%N?du<#
z52AsXf<Rr$TPL99TuhbH*zR!h&I0D{iSunqzpqrVW<i*n&|Xancw%I<z-6Ix+wGuR
zF3lmOVbN@rww$M(n6mj?_=mGcp%_uq%q<H98rvr_O&wGns3!x4stUt&Dc+(F7S5Ul
zw6Y!gnc0!_eF(JK0e?=B*RkiPa-S|k@~}=YNBJWx(5v9UY5C4fbjGzW#ny9aSvZ%B
zUQZmMA^~?7{<8!mJ3|i(Ref{3J@aVSf-CY^?x0h{b-!vdg;lRFONE4?KMUtqx{J>>
z9_zoHi-@fiS@d7R7N)ejc_x-tnie#FMpW+++q%$yWN!SdsdQY2d%9Z2^i<0e!wj<A
zu;<5P1~TQk)){zj6|<7TDc#$B^U-FxId}#9wA+6f3VV7sin%ATnNn9IgeR6pjK{QW
zWNRadsEB6Su5YmBED^!59+!2q7oh4moTV}|rw=ahZTTS5Vf|2iP;+BsV>`)-!|efD
zRJJZ$mV;WfH)yLF4>AfuF*{Ucd|~huX)D~O2P!9FsRKnWgnOz3WAS!1aX4Tu5k5r+
zw&5L#bZw^8DujdZG5Sf>4_8%d!Xo3&SF&u&glo%$YO*j@iKK6BxN0uoGF}1QJLN$7
z9Zmru-%hWuYG^O19<KoApI6hZ`(>^VUy{nGYs&=jf5zcWIXY#C3$L!We*SFxW9*0v
z3`E-D@&)EEzXArUftS-IKSN#ttHlo@!z3)<mLJj!f*G46T^E~EN=f}{4qpsVZeIb1
zzm2=iHsG44#}499uJ-w`3;5>J-rlpaRaonV{Yi(LL)=$@V#6yS+3FS0VvD}2JS)6f
z@(SqMd@=g;54A4i`tI5C6@YLDY}W1;d1e@U&LPw8oKW;jyr)?9N2!jrN}idmrQ;er
zz(_0hG6|4*VBXn-ALTmyy+!hl$x@hbDyzPZeTeJbIjfX3bWZ>(A0WWlumpO-!HPzW
zh!rXgf`%}${b9@fLHp3l=Ql3Ry#I&p9ez`K7M}a0!tBTsKFCPg%(oLV69rsQ>i`L~
zb3fd^9*Crz;9^2wRH~|)p}ttzCKJf2rLjFRdj&8)?Bve4rmk}*z5*~(!Ku?P1pXp!
zEYEB?dJ~Sp)~)A!jeN~vclOT&FW<+&LzYi3K+onhAJnA<*>!yeZjiQ%<xSyctU_$@
zafY})V`1*+B^Nat_tu3MtO@g`$Q9s5x83i0rWEeay|?>8%W%r#)^3YloL#?Kny1St
zLm3WD2gy5YlX`_e$}RgL@o3f>Xm>OyJj5?7>-t6M$1iT#uifuoqHjfSjv=fPL>B2o
z;Y}y53bkIS@7KxEnO+&7)Mvi}G=84MY^!NF1B;b!KU@?s?CXB$)%y~f$6O06lSM9J
zgkSi2f@b<9=;iqVTO7#oaU?ANAfI5!mILK#vGv}}@f9GX|MFv2xzM`S#nyYO@*tU&
z!G@<!QU{^pbTxnP`^%?Uwyb@g9)i)kwB?wq#g?dlx_qBydG`uv)w}|x2pbVJUJDq(
zT6@?3F%|VURK=TQ;Pd4xK;aO5Rbjg2udy8Xv*f-4lDMCqBK|O3+P#JDCms4`DsZ0p
z53?76z!m(?+6l)uaLjRP3rp?`frG!~D<Gfi{~K28x$-MOPQ<DY=Y$UU#b4$&CU{&C
zIk2!-@c8qlN-Fa}v2s7jCxG)+g?h>gT?g<!5>a|FHg4e8o|Knm``)Ku(J9j;k9w0r
z#mhcSwtI@hN==~Lb;`2`pw$^g>YfDe2g^-lM#UxFqB8Yq)yK4vNWAZK8pcZO{4W%O
zngiQ{E|fDbsMDuqxmFI0$L4ikrq)Vhi(G|6AYsMUzQK3<zP&M?gS}e(LeIyo;WR+>
z#l=B#3q&;@a&&gVpQ>^#?3^jl4@X{_&IsuVF~xfoiCoHfcRQ1EO-*54827Paj8Wg2
z;Fk(z-Uf*{yOH48sN4O4Yx>HNwEohxHePc`t8S@K$9c<p=HrKyiGgL_jf&HinOR$J
z5B`e=N9^2vzl_~xRU>BEjZ=d1_*vfIH=#jG{V&<G1T_x(Ti&x1-U=7X)>_)?=@W|Y
z9eQyG*@HYErq{suo>7m&VeTS~C866lNmF3q<|s(uvM|XVr47DPYLREg%E|$*S+kD4
zgI!;o@VIlUi`UT7u~<x;C)cFJS*uueLv3qyUE`xvZWO1qepnn-g<_u){BE8%R@1%<
z$@hTrcWSuCK?QJ;H%`2kO7VHBG?qgxheVzV0{M%MTFO$G+Fo<M9);Fwqh*Kc%y}O0
z_R_b&t*<!<BL2@j(!JN^W)kBTtSi8-g{wEec4gJNTLPS<-~Ww7ag4w6n`g0nW3_N4
zy>dts?sPS?7i-wcoAeL5=k6ab=<>(L7wYSjpC2l6p2jckc1x}Wlwqx0YEs>$D|}n`
zYb8RX9^WFQfnUd0x-Wms&vpg)W`DG;bw95LY6bB!VJ96%ZxYYCkH<FFUID;;oumic
zNB?q>YR9@}(2M@h6Pn-tLb|VjjusZ9-y+Yn&p%0x>hfN&?kQG;QAB4JU~>=^soL>F
zRt0mh`D7lB5k%e$@4_*7<`rN*#)a!UPu-KnGtA)Vk`Pl=q2Lpmq7HPl(u}r)?gCvv
za`cEq72s|1J-jpqBo?Y)8O5Fd0KEdNir!{54j#iw{F|R%>&M6F2Ho33>&g#BpUaEA
z9U0Pyou|G9oOu|*2@3hxbEWrW6>_TQNN<jaB{~Q3#*EDtw?n&+<)T5N!NSVYoYTbn
z9oVdBP|$dlfJa>?XMi#n`Z2*7#0|AcGxicHRy)#P!_tE3JE1KHugd6J@b+u5Cv*pk
zZD8qCdOC-r&2{W4gpKmNr7$b&^gL4qB{hzJPap5Qd=yM)cw}N(lo&wiK!Paynl<Hy
zLR*j=!zX;g7JcuZcz|7VzMwB<jV$SuoxQcSE)B;egt>3BKZ(*xsdmL(7Ez?OGNHKG
zC<4~#2tx$D+eQVO`y0Z682J#ui^EM2zyRx4v~&mnK;iA5eYCd?jZRLkEbi9k9x9(5
zom~Fxe2WGtIxe!l4?Tr?Ac0=i7Fa%TlTL$_s5C&Z6J1)-jHu+=L(GonzrW5gr?GD!
z_Qq&66wb}vJ~l>L6q7tQ=)_5;vRup6UT{MKS^@oA*58Ai<Zw?j*L_6&8yoI`)$gyB
zIxm0r$+Vd;;ja{69%PHo2V;5eL-9u1`&N_3;Wdg&ohx+AU}~G|Vrg0R?mpbbdAH76
zz4I`1V5;MckYpuC6TIPHQ)Tj%WRet(@1Ne4kv8_u-N*kX{rh=Fm@jL1Z=C^u=_~ns
z?zs;MiiqS#5OP^>q?o=RAQ9tjA#s5=C1i)*33YtipqJ3(No4ESmvyS?b?x>ZYZs~K
zBd1<_6y2=!@bY-X*n^R3N$BLX;?8M7D=CIrqXOCP5|OH$$=Y7;xDJ!lS15whj6U3c
z#sLMHlV^dG4GkO{61M^(0(}4HybU{L+3DdlP9CDk2Bz^<vlJ+rJJ};&0nS|vV?+tV
zSHa^Fp(e77wh=3%kEfudAD6rXvH~p6tusW&ZbTuuJcE>2z#Ja&SFw|rURm=8_>zBe
zz2M|gaVVG|%2WtbaKRKc1ke^k46_CCCCub>n0i_o3B*K|w`FWcYAqDg$kM-NVetdc
zY_llxi&|4SQz2z+uns*R{LNZcD{bRFO3Si6cLb97vEhx+so&-KPU?>jJSYvq`+*Z5
zBdav~wwWM15{Yq+tA5We>H65A*?&L#&H|HACTwLpSr3evNqjGghYU=~3>h;oF;h^f
zvOjju9CSq<gk)eUY8WyuT&S`8U5B_h;WrIqa&+z6V$~T4u^w29{V7PRSm)r!E7o-`
z!EM29>NDV)a3LY|UU3DcAeKeNhj?OF&1+%aQ~Vfnyo~g$e(*j1##CKNZW5D(&Cqs-
z%rcqVJ-&v4spPQ~Gd~a#UqB{1Jy(q8;WI80OQ&-1-qzyPG4In0Xo111dIZy!0o%3u
zat%#whs^$N3F^31{SklMa$=1|`51GI{#FomR#SncyIQN{PEZDcLxzC)GAjs{k7>tN
zbe~bIq-~?Kf}=cko+FRAPMSXZo<Q&=Tgs3Xsj|$xq!-h&LkCF5goJ(sw?Iqd&GW6r
zP+=7I)5W9HXE)j7(uB5dtf=oI(H^NmKy>x8{qeA~r#tM8FZ$shXf~ZZ6Enynq=V0T
zyO(R@pd(7clb;<vZw9Z}uOqQC9Ifh9i|mteiP+N!zTzUG^l#153JU4laY*RT1-4dI
zr8bd&o*a5Vd@A(uM@9?Ehu(3k5Z5be{s-S$EcwtUb1^A*DC!5DzA<%X!-M8by4oLs
z^H_#T6VhRHXHN8=v<2nI=obo}z8mecj>9_k)a!;9wEOcDrHivA_p*3hgFil{S(**r
zgUI->V8$YjTz$c!bCeK3-*{#DPFCEo)1cL?t(Cm}mo@R(zyT!$D4NLH*p|g?MYZ@k
z2M-t&Vj)ko8?q;@kyB&+S%+-+ddk*aKhI0*aVI);M~EJTGT88H$gRK)el#0y(PU5@
zFhub1yPQrj*84F8M*N;IIJ8mPS=b(FOXD=aAr`ec-6*Wp$wz4*Ta7Pr{JfS}A$gh&
z#SexWi%^c9)4BIfxX?n3?Uf;Xq4HTjxumN#shd<SZmG+l-}S}@L2olsGX;5+LQx$5
zgwd?}Ah)iC_Y^T>BEeE-)h0Nq@6{+C_42MoSfsJA`~XpNBoZnzCq$j*Bf`1Pu5phZ
z?XF*opl|6<?!)U{;O6DdvESt#{(2|_Lmk9Ub%f&&OcnEMS-#T*T@ORSVGYsGO~qNO
zV*U6!;TMW#6J@a$B=$5{_Mu%24$`f=Dm2>1pvivtZWGaOzT%NwWJGcgoJT7OiqD0s
z@3w1#U%&c^!26*q$cT0lQ>6=;F@=|W1HP2sAp!ueuTTILOnC@MCcyvMi4MA4(1Lgb
z$t@BS{?|rk0M?sp|KEcJ4s@|Z1F~Aghup*l$-$9>_SaBBdhn?4{}%&+3$lVI2Q99m
zf^^;?BmXbsO*<}#bcqcVg8bijZ30lmn@rUue8_ToPzxe9sCxsI@LvVy|1AR$#007|
zAqF{bU=jSQT>4)ic97~S=AYte5cvic;lIj$|Am9&{VQy89gFZ^wX6REK?;M;H#i9Y
zRd)F=0I$TG6pFw1{h#{5w-F)x7idWKud1?5EW&?P#Qh5-tn?T3^alD@J=ecLnrfhN
zCkl|t7AoPtYB>G{%Fy{MDP;?b@PA7Z008oTfz}NFf`Zqv3IFroAL{=FjhX!gajav5
zey-s|-dcf7)+paH?$jVwZfH=&8Up0A^<UB;6T)A6_=AG7{YzS1!+-1a9+l}|YTw!Y
zWu9>$zL^Q(|6!uq|79+-BK}7hj>BIP(#rp;By#*q>R7=Cb#wl=IZ-%)%2stCgPi~J
z!OQrB|H<!j{okJ?{4;1-odV>(f%^B!D1H9+VOZmYr1yRMssqXD_m^_|R!A4TO$o{E
z4=UQG2FY*ZLk<H$mfO^X|49SBr2^mT?*8j2%>=0y;(?%dP~U<ld<?aoe?lZzD|35u
zmcO5WOc=B{jQaLk{?9Q0V%ec2{J){q|8N0-&{5E@9STtL=zl-4+BnE@hw7iX95lDX
eNATCvzJ+GRkpF~c|J$Dt#JEcdySng)^8Wymlwx)O

delta 16876
zcmZ9z19)Uj(?2}1wXtoR8)LJvt&MGNl8H9laAMoG?QCpsw6X1f?tQ=Sx!?DhbDe9t
ztAACer%z3FcUAS`1jNDw1iGRO1SAFk8UPCb07wBn8w9OAHRv?pz!AG%7S!PLgeT<K
zDooa?#F_B!oUsxF<wJ<(5r7@xYAw=_rgI7~ait}Av=;H&p?9Gy&}E#u`6mQ>#6c%Y
z2<7XtehA#s&@mBVK`(^_s>^a?wr`@+u)`P(hxL<_+8rg($2r<7eux>f?(r)Qu&i~R
zGbF4t8dPS;X=cKlK!mPkfJKM4?UbT4%#51uw#Z5a%t2kO!^RF|8~Z0lnEtlOLCO;F
z-W~C*n_1l-g!}Z>@8CuY#ho0xV>7~?gMQWYc!dqWNu6<ujBzX4{YP!?1nFj3-G&?1
zg%iXVsFM#jF6wZJ{1;6JJ8E&Hlqv}4p&5>5YTnF>2vq0Kf!C;Hm{H&2AI8jT9%gTh
z2lu=`9UGQ+Phgu7CN4SUx!KfGo@XSRaSS+|h$#xeiR2#Q71*L-#^KqM<2XOi1{oT)
zM|N-7K`mcjuzj&@G&9Kp-&j}eo;L3==}@`-TvvkEqJ?A@`E5b;&#OV<H}xSS*(6Gn
zRyxvISGW{BN&%BrT2hs|AlW8LlTx)YSl7}Fg1y6$2ySnPXZBxJBVnz%)385x(kAO9
zB<lq|W@6U5$wQr-W^aH+uLvL|e2h9cNDhd)ODq@2x^}ub@H$g`4HN{g;{uxq6aZjK
z2@K?bj|tL)fd$%=521C!3<Gg43@JhyTc6fz8wLfC$d5g0+K|C<1b@ZF_cQQ{AxgEa
z;0W?8tJpzfg(6Y-p+JN0nm`IZ#JrG7JrmDdlSq7>(`h3uX$=*#DkN6B_0e0O*Smg$
z3CYX3FE@~d_H3N>ceflmHeYY&9=>zq%Mz3iJ@f#vjRU<7Stg(OGq7K!_XEFoy$n3>
zoQ_oQw-nWlxV6qMGml-~EiB}>Z4wuN+^-#loZnZ+Bb2mTs|GgItG;BlDNFS+N|vnP
z6$F;g3L5DCdGvO7pE{NDyf>Vi94(B=6_S-%yc&5k6i7+r>XfzI$rW-Om{VZpILtiO
zx}UDr@BzvXMo0K|wT@=`X#OOQB*wVy0zRLJY$18c>k~4|NXau~RNN7t+TJdnX!+E*
zZeMmBnzs5h>-;QLSmZu;DB1~n&wXlBdtDG>=&-vszJFLed!v>)J3SkcIon88ujbUS
zbWMr6hgr{eV9dHZeMkc7sm?NWzCGDzp42n82?1$~Yz1hb`7@Z~f}q&sA*I(a(Kt{k
zLVr-g?NFKV!QpBcs4E1noMF9cpAnCrz|7HRL_)8VW$m`@wXJ*4(E+W*<n>6j5lUAc
zpz^d!6t%%~?xX>tpXlXey|WNXdt+drJ8dMkR|H<mzt_y)xfOOxXyqeZwPv|z$+t7O
z9|6njH!a*pR3fj+CM`K{)#%H=uw)(5kcjmAhjs_`I}m@#-R6-^Mokx?EN2ZpRoX}!
z8S&{#Pak}|_GO)b;gt7l%=Jn{cIP9&SsbM6SzX}1;<2-^tyEjpY5TlfgU*FuSh8K-
zD9NIX*J9DRrc9Sp^OY7;rM_l-CqVmgLJufunZlLZ%2pA72r}wwoVgmQesxG}=cl3C
z>wED;TBeSgSUG-*I*OIri!gf6+vs4{q<AGU3|F+|KqD>c*R&2Z4fChSO+UIA8}-E^
zDgzc<S(Uw1<J_boBY7DK-egWfI%4_Rn1E||h^yLolL=dF2QCCXqvZ*p#ymJdXM+sj
zae?-muCglR7UZMa5qm~4_xh?D3Tx)${)cxC$_u@vT|bY7<e3j0CXx8pYY#jyLFV^$
z*&Zuo^y`%cTFtA$mYf8)>{o|*QQr&O4S>K@(58S(-z>2^-=J2)B0k0*D|O=i25#bc
z<mNK{W?7mZcPCR#)-6%K#FK-%gU(d>Z(w6?0hIlGMHPJdZS9Hr=aJTD&*__jmI|~%
zHoaNcVmLAx#C|k1rEUUrp1Uix;UCjfY`?OOT8FbXGb;(j(@p(3kX#w(U6+?Gj>_Bu
zN;5Pn)D|oTZc9t<2#^2_#ofo)1Lg5o8^;>V_Sx&`rd4fTP#|@u(Gx0?dq6CIw5{%z
z1#U8QL*uZm>-U?abL{%o1uL-8ADG$}9d}lb&)oR=SP+KIX|5dCNev{TK3HgiU8HDs
zOC&>xpc!!c{X%QV7V6u6r^3*qLuo;Ro4B6p%pdaM9n@RZ7Og21@r`xAf2`6WSeVD$
zN9n(Twh(xMttHn^rMnI46$=3+R{r?r2xa-0sNF4}j@xLs=0~%*EnK$awW|Zd)pmO5
z%l5je^lhWG%j48M&T^+q!oV6FbF|v+-oE2I>LaonFrL`_>6`G9=$jDj596RBSs2&J
zhzn|C6OEaQyij%G-rU0Ln>e*fg(ByvKVIZU;g=BuCTy7pXY_nMimw%PG(VJ;v^)d}
zTge3FveuA+lR+6Xs0zf!b}2cU7vg|M`R4V=&f1AS=Vl8#_DrN8RB1)$HzYb+o)q4E
z0btJ2`D7l~=zrc;b!16T8M`@Xytij?EZG=-x^15fiWDe|U8--&S>bQXgOS5XYxI<B
zt7~fp;UIklw$xGonnJF&<oRa5fh*C^$%W+D;R&OPwz$O6fBT2fQF70C)`8R;0%c@}
zl}Q{UfKY{b&JyXw`v|How*<TSvd5y}%cqtVA1d%~FEk{o(<D=C((H}NUO9Rt6;Vvi
zUOFY%L5_K!LaA;85_GGtotOat&}o>*uMWQ!Y?ga);2-_&HH72c`Yd@fov>0N8S^x^
zw%2kKSL$Gxyy$EFvNOxX)u;WV28r8AE5XOp`h3a~hqO|S`E`k$Y(>;B_{QwXvk-7#
zKLQxSxB|KANCSrlP2xmEg>!zB1Z?+gnIN&@b|k=J9z3FA7Gp*#jnJ4mGVpBqLxKAP
zKH(UA0#<^<)yxEd_2H=PRIpU9(7?%oN2LuF*8}DUi*Xb!StNh3(*0C2`U!8W>dqrc
zDeq}x(=udsHbsE;!$~9PipoDIg3*4>>;o6%Co1A@JZl4ujd^}T;xzP4^$RQw27#92
zoBv0rBnz=NbP&p(wzY9Drr=G#<ZEF{p$f7B{T1L)bj$pXVKFi@Q!_I^r6QE65em2d
z7>?&ilp|M>#S4z5{V8i?0#B}jlhWE3OB;r8L=)dej{!Lo913Hs`UQywg95T1w;oqO
z>PKVrKb=Svzg)W%F=_awiU9X!sK)-HmFQq+Od@c>IK7}|qbv4BR*cdFAowMUwbTqq
zn!s+IzfkCjyFG5K8bHq;sRWn(@mq27dT}rZi&-RHLyT3&S*=|x&f`23X)M^&=!t%+
zlX7M&#-L7_O_{N1e>}7?dRyc&+rz<7>l79%@DUcvda!Cf!mtGAXABs1O}O7fPH3nS
zDft3q&YmoE@YkP)oMYvpxC4<5O)kLEuhw5HVqGC+zPdhPswpc;N}zXt3Ra2$V?Z=C
z0w6$WYQP#AHN)I9=XkJ)LHt4P%4mUbqH&@LhvCKGbqOmwEkwddhW<eiP9uyg@o{y~
zg@Et%!ZW&D$D!>iii7=wbvjVvnFHV;8>Zz+4(Eh|?<#&=Q)R#QkU+p0aO{6c*6CyQ
zk~A>{#$W#fn%vW)POc@GSdIxM2qy^N7WsI_=Rh$DiR|=4=SEuc5VDY5j~KC7u_Wi>
zobDdp-2IalDR>EZiA$0nWIvdS6f2-ql<9&o!bybxp~82t{RjX|By{_I3INbz&;l`N
zZ}3Dst$N_SHV!@i=@<seKOJ^LZurW!7#)Enwke;<*cfAeeB#aczA2z1XgUaA0;>qs
zPbG?e9Q5UzT%UNqNOwiFXQLT;y^-pQS<b<rXQLmR9Ix#sTUJ});%heB1_(i_m0oe3
zWD5jzTccdYbpGFKVEf2C?wTQ#En@d9-#P5kSF2kEICWTNmwTs$I{CaFGu&?$iGAM|
zQO=I)LO||lL_n9s4Rr>cSzO_reWp-rWndxos|6Ep!`hr32hM&(#zih;f@IlXn452L
zeg(FV?$i=*Gx6PZtcKo+OK@PhW8kcLWdUROijLLnB;C0mZVA3`#{>&!;4iW`GXOt|
zw48OWOYNq>oU_I$dtW)04ms_;mXhr3Zpf=%QJt0Yv|zob7JJcF+SA!lrJly-Nzb4W
z&wbC`oAy&%+pC>tLWsG-9OwS>WnZ31BK8<O4T?D@_dR!qy_X*D5{4T#Ncj^D&@CkJ
z^1N`N9x_aq4H-NqRI!~7V*rk(^X+baI?K0QecCMhhQ~ljJe6je1D@v6hY1e7*S#<I
zYv{sH$<!1I{|Bqkaxaasa($7sinQ^OvsMSSHH*@d<**H2hsB3GxdBOa%bSUi<9vph
zo(*5zKU#kfX_!-=77*L#pXLm4faag8yjk7cG_fG|Em59fPtw42^q`j_Y0HVK#c<(>
zremtRn|Y;4a4yA0z9d^<Nl03ky%Bc{LQKd=8T7MnKVn@n$UprWyI@eCcf}&xUCHrx
zIL$tJJJdC=K!Sr5C)w7NuGoB-n^zp?RWpz3Ml7#g(C;yl>?iv+DS2Y53$(5&t!#*l
zPn#J8Q&KxMqM%@%6A!}0R3Q$S8vLfFd+cfG>=4K*VM1P86@4O&a7>w;SwSG=)`+<q
zlBu~o|E<I(c{o&<0Dqf7A053vmEu#tQYs7Av}wrNS7d$6*<z&ax#90U)%5iMs9eb+
z)VNxe%v#iNnOVDb#l-8wEnqf#w-(QM7R(SsmEkmpz114^j_AfU8R5B0t-!H+doDrn
zql1~d@q4z5L_H<<Ju?dVybyL}eGyDv56ydD_D17$P}GXzG=s}raNHjVN#v}1f<A$<
z5SWCkV-G1kbG<WN7AxsXbMeB&{=6uDgRvcOgr*{fwelk!1lR1qDIj(bep*aucyq18
zy_Cw~+Se2s*X4Cp?m7p0%)o`-7S8yrMhD~MJ@>g@A$gPzm=^qs+77j92WfNP<)(MR
zr+r|1-g(#C<4{sn*W+Q`bJq8XH;vg9TnJ1SbX1K51@9fY#x><_mHqS|9PxsC8r667
ztIeqMY1EVO&%DZu-9T*^8Ewm75;ehZcDNXp_W^!=5*u@u$(;_1xvMH-0KH5Q)8wa7
zo<cb0@sHJ4&Fo>K#Bl=Cx_B(|eQirxW#GZ_c5+sT@$yNLW={cXF_tJp3=O=oMw?-O
zH4POXa<lqcQNSi&R0kPz8G}`Msw_w1LT@S@a(sg|8AiFuEHHzG-B71wRj%f2Gk)Zd
zpB8lq_JkYaWVv@_4l04Z?dvD@3C4)90&R3$SQ0@8NgUIaV2)Z2R{H@cC<qLUD#!9-
zrKxs4GhcUph!RT(lRYZkI`4Y=^8lurL7j`QFv}0wE?rSr-ZO*xn-&9IN70!#-#?qn
zcj+tNV&;%3hyV+g%~}G?4Qz@n481=+rlXTXQTxZIYmRXY)5r#S@hhT@^4T|}RA}v&
z^h)PzBEzo?^zw6S?U+=3k;=rM9QECL-9V>jHBE!6GdJrjjXj3lMs?|iov)z`Mvy3H
zNu3%Fs&PUuVrf318n;e<T#m}zTj8=JVzK^NK5707_p_DeSVCu><T=OkSl0I^+hwoU
z%jSc{&Xrn_yNDqs!n%nNAQlx(OYTKnI_?XT_<<(J?&ZzyIxT)@<c@nbhoABcmw81Z
zQP;5`-kp^SAr^7bjzZ=<n{xv1HtccMAf|AtPPBc=(^=*-G3dfVs_wHDG#IGYaSjyY
z$Oh7|`&~z8%LoSA+=d5v`E-MryahnqmcK!Cw$z}ht^c0}4|?22trKzZ2LlQ0!q?e-
z2L}g<S<fMT5!PnfhXepvQ2_ulkb^QY2;Kw>lY<>kT4q6HQTd39)a|MAtF}(0T>^;H
zga=qK!&7;bKMYgbnmw$X9wHIHTL$qs5sreWRcaX{C8E81DXbrD{rI#!Pro?on;|LF
z<Hef$={k1t)sUo0N&D`52cK?4UA8{WD3AGhutWk>JZIUwzADsx@;Sq)hy-Z|!xyI(
zLW7RUhi^Qxy-SR}N?7~#j>-d;!dfr4+RDI%jHWA}_5?LucJ6OCBQXbFBkE(H#56vo
z4yqGX$eKtmX&B#0CSYs9_wI4TU6B|Q){O@qMz7ACEZ^NEWfdI0xzB9%*_%bXqYx9r
z&bI27Y$q>XU52Q&=x&q-h2Zb#B*cra86QiZH85+lPyZUqo#E&7ZYYC(9jPAm7-0ia
z$U(EAJ7oxcZ)eU_q_yGe^R#bvyGC4f80YV6U6h;R=b|-GLtjf9^Ivz8rwZPe>#zIR
z64xpvAnrbQqS%gKvtDQ+84e6~-&MnTha#dJbwZ$dc1xAz6RX!uo#Yi8D^S`WWKbHZ
z%?-pb1xYLf%w7)Vt2@$jM&Ai4>+J%cI-w9%*ug}Wpo2v*vJV6~Ff~*NbUdf10tVI+
zzF~H_t=oe0i5xkVz@9HC>$CJUcSib3VMP|VGc0Ov(3s05hSM3Tay>NU;i)L4?7zuw
zArXF4v80d?SIKdU1N)Lqy5VfgDW-;GlCIwDsCReR&y?#b_{6TxX6h*p!HWzOv?WJ)
zvXo0Y^umjf93dtv4#(XjG?iBO&V+58^*g6=N4U7O)q+=UrC4!9K4ajHI@oqOM|(Hq
zy64yQI6jVCo{%}bSPfj>KqoQ=d1WH7>e&<Jhd!?>=Bg+X`Q*+Vzn(5%W6x|Y1Vb{@
z=0=GMWHeJ%bD97kjEI+(=KO*7mWmCiN@{Xl9^$;PMX;t@a(@CRMd;Xm^9#?r22E4K
zxFSzks#oc!2zv6Z8Z#4p&N#Kx?O^-f?Ap$|{_-mBZ0+9oxZ1vN@L|a3iVQxdo!Sz^
z5a=m)+n(@t2B5OIZbF9tU^LXi0u!tawbUGDcfH;Qi^iF{O|qdB;X46N*+q3X)iTN-
z%))j;wUV~6hR$n$ag;+??v})Yq#7h*AvAz#%vtRxA({?*&29@72aB|f8}O~=E8R{b
zr=t+SZd1-sqK*uFImj=HkOeaqQQiuG9RI=V#30!tL9`BE`umQzD+VZ9<|cmphB8}`
zrB7@_0E1<qZ1{Onkzfqy=)aFgr-;Vo_X8EO(zaP~)+XQhe%qgLrj+z9LP>GD`6&A}
z4e3!}>MfcgCXf8-vT8OXsV)P1_PnX!@LDQFK_Pu+WzQ6x4}JkqkBMm*DQ80(VnCa-
z^NbhVf5n^$cW3Wr9G?*e)gIdM%O6Y0#*K423^_D{wVGoWY}*ppYFt+_?mVB@Q7#Xn
znb2~1AHlk_+KEQdw|()PPD`R`6>64eR<bUsvjqtc8LqaBGMO!zVzZA?<6UqD)fnR^
z1`5dQm27jD7-yVI>1nM+l+}Leuc|7o{~Z|^zgf5hH6@-u+vIOe#6WG9vLBpac`L2B
zL4eS@r27+D$|V6Q{b10=t~Y7p^@#Zm>K|Vq5_D2+dCi9fQ~mJ${f`C1Gr$DO;6>U|
zxlJ|1x7LJZbe64X=``38j@vCDfz9$6+N^w9fe~7Y35c+yoeL6Pk5d2*ojmcoNMcvn
zgYl1Gby+Z_@;Xnv0G%aU-GMPwqQoG!i8@}TlvoKcbES>2x_REq>8}7;B^t|Q*ItUr
zi}YrHnaj7Ao4Jl}jk0_QTav(yJn*{Mgu1QtW=Ts{dTG=8;MwZ*5rp&__x1FAQd7BF
zI@d1`$;<NSaHF+M1xD#%<K9l7N6#jFsim<z{cv_|d{Q1b+UqXeh=|$-^|gL&5r>gq
zs&&k*;@t39jH@`UY^Xo12oCFrB+|`uq{gJBhI`{(f9jv0yg%sS5WNE_em2Sncby>=
zP5R=T5@epGtBQ?%`(gOYTI*d!nE_iSzxE=#OFXE#kC7`;yu@=Fu`E@qtMp4!VO8V2
zgWD7(E%WjOY#?teTysy>r}I7tI&oKXj<d)GtSVgmipRyDi7Qa~&JkC|R{NE~5UUQU
zkz4rcytRrm3k5`B`~B2F5dh&h3M#B763Mbi_hl>;cyX#A2RL@=lZv3uv^jYIeXQVd
zeKqY9rvEGtnP-aop=^8+5-p>HtPT|CiW7HMb8ulFNsK?O80Hd|lRqI37K`d6H`3xP
z=e+OA3E;Ygv5Y34>NnMudRVRIaNe!*uu~xhb|coAjd%qWb3_9Wngwqbl6&VUNzJr<
z90}Ux_ihGK=deb3IbYv$e*TyXd5>D&@90sa-gwRP0>@M!p@s(~snD-u-iMbQnjkp6
z;uRi<=tS$o!<P$-H1;-E29Daxh5NL;z{TFfXzIMYeSsG*^ycV!r<o<j63Xp$-C}cH
zT|Ua*)t%SrF^yi};aMTzoQ9$BIIY%P>nFMal)N8mwdw(&ubyMk2?IBEW7lYFCPGyH
zTVn0eTK_u!PwLHJlFXYFOakALQ`xAZnu(;?)J#0VO?<){?A7miedC8E({=0D<68@x
z*r{zHc@<rQpNVBDtilL<@4va!12e^Lqnr+e50+7nx|4i>jeO!@)oSCWoI(zqDx7M4
zDFPvcgL;pQ5frkmjQ+Q#mr8M(N$VAA^^Kzw=@oq4O17{JLEeTf{$K>N4lXM9u43lX
zjq*P*W_bog+E>UuYNY7l><20p3+Ln<k~7xBTrz15Ct%^z(Xu(&C8HiS9b;V+ML!9)
z@4<NzYd5e1wWEQ!XK#0RCf5Sh2T|bG$EkZBSm49{L>J(V)s)s$uoz?4YK@KCX&{(J
z8ELr_?v5BmS#66=YWq%zmG@Z4*AuGsM`w=<FP96ehwDxj?-rj{_bs;hGpkQxiOlq+
zh`Lf5e%s_IOTj<={Jn8$OLJ)yGX&62&i6!^JFpOdJYl+jM0Jvu?+c6%pr=0HK^a+>
zqPQ5a>xL9?ZK_N-X{GcNrn8<4eue*D_c%#y&-I?+a#b3BXHYa>R?0cY8&x2-XmY4)
zbdPn#5a{E0j0eksNxH~tt8vTg`XiuHd5F(mQZH6$Fn@UGXkRfF|KiE0_ZxdG9Hv-<
zw!d{N#o&gMN#f(5YBo}Xpn-2)y%CIq7Y=2W0pxlpgeY4zO~AhEcMz)>CcN9Z3w{UG
zhxx(+DT<NRiSYP90vqy$>+?yzTT1dqDP568FhQA;xSh70#Ju-@%)`gC=Y#x07a?~7
zMwhgd<ppRq*ZBmf<K5k4OMAp}Fno>#3gmb-h#oINr**YvC;NTbCiOPb&&*-ZZ83L{
zDTcD|oNM3Nd>_7DT|7Rl>Y2Wrob?2XopU>V<=txiJb5dC4|MsC2R=3AE#+=YxFkq?
zp$qHd1fE_TchJl7IS%W9XMzpe6=}T2-x!aI#S15pY5KFFN52c9Z@uSdp3h@+z8bSX
z^PUZ&s*_fMz86l<RXc+1Cev@X);pBeNXG<2T&Tc02$*lgzcXzED(z8yJr8l(qUK#)
z6{;)!9^@7-CxizyHFmizw03J=c;Op@J=pjIh8JYlxeQ<96PH$;Mk`%wc-m!06#}=n
zzsy7^3?6QcR*EXNVr7KEZmxFqlmS8hPMGlb_SJt9KjsBYl)nRO!>$1VoU(KzvvE|k
z9*-p#WMldSTpwS5CiWD4L#h~ss-PB{xbyXO=juQ#_m=VJB$RgfSnLs6@OLYk5*~bV
z@he$sW$bM<ULD=`Kiz33yBo_%+aa;f?&y(y6pHg>>P&J2W)ODLq9~XP5jfN>^6x*J
zC#5Lb(V|C5F!j0(K_^-P$+D-6E5!PsOG|VL&FTq9+^{AgA77ife^w2EGm(9urEJB>
z%>g9<ty)BvxG}I=%z2JsqNPNS*s*Y$o4_CscXHa3)0=384f%Inm#c3Q*ZVUE16G`_
z!8xQBZeRGYM~dofEa|2QuTpktcEv_G_*!X~<6C9ptCXeqgnkOQGWdMj)xxXo;;+#Q
z<aQ$egdc!P7aqQ={X4v}r3oVisg*bX=N0psjcxx0Sp=GZ>_UQ4KK*P9Tx7>%hTh}f
z;2XUC2}FeBhsJWD{~g$NMs7&pV0eBnCR{2#!blCy52o6SaDyjr5P{VBd7ip$<wB8=
z!rANH)4*A+;UzC7gDc0ttu8hRQcJg_CQKpxXTQ(bO0sGe)}zY^Zr$Z_G<RjO1`A5d
zVRyhNiXpU97*+5Tpe&=0`d5&-yV<<~gS9AsQL0}mG$)F7x66`*<5DBAure1wg=w;s
zX?$3D4<=tE?O_>DL>KV~(Q{PaOe}(C4pNnxj|6$Jds`=<In_?Z%(gObqNDulw>u}Z
z>5T9n9V@T<AdR;7%m&Y}+~<nvOG>MXBN|S)dL*-bIy2<o$$ZrUxnK%nIIDBjHoH#q
zm~?{0d<z~N^O%q0#SB;cnlEok3F;okLPjlzU&zI$F)!8D&}&@xboKDP9W_(p4IT#e
zYGSa{evb5E$w4K#%(_nh*xz_P*M3I%XAuK8;4%HE{}HBst;=%qhXnpag4;rZOT%X$
zMDgrFvZlnf`Q9=_=#c`r-GoH2#beM1`dr<#Wo;60??7oydw{`V^J29plgTB8iN)A7
zN0C|EDJ;eqCx0$;zrFkhYl2Guf?~`ZVrz)WK012E#u`YYXvB<w2TA&EZ8uv<qqz*5
z>n%h#e^zb>)Roo#U<th19s+d<CdvDHct6=x3NNU~Cp1c{S?PRHmn~!Fhjug#yu-j%
z8t9!nGaAU&m>ea;Tf-m1Mg!>zrihJq#V2zMrd81~nlG>QV5!-tG`g0K)l0#~M_Ccm
zNz1w0d^vk1N;SI%*T<x&=JQRUc)NQr=<@vja$Y9Cxvfl`H4pSgs|dg4LPg<+#^8XZ
zk@BKaon)Ff(NKYNwCQu26Eyb;9`;-ayja)FD#(KJObp(hLZijJglb+=tjQq_w;ZUB
zTW7cCH7)BC7aA41VzYEd5h)q&rmt4d{<RXaa7Ofp?U!ltSoYqdZw$*+F<t}YVq8g-
zw#=Nis1;lZr8bap+6u0M64hHuDfuLj{h@#oY4*UqAloiM3WRFV*q-O|>kOHW0x$0B
z0tGDthmMk}y0VY*aorzQFT#~`VoF=%gfH<RH@LExEAe1ExDwV+#$`UZGSpbcMOsXw
zyHi}l`rxQ5TL!;%`@krya$|}UD6(krjg+Qc+!MK(?+};{lSvq!ts2TD)U4Y~8Anq2
z`NGYeNWG@z=MWhV%!c~6{b#ixzfj)eC=y}e+;t*q!5wf`r&`dXsnt!>D_v{qC_^h3
zrj0MGP!%6i?`&J!&UZ^?q9~4SZ_8cZm}`594bDFrV%!b=@!}7JUV=ftHwUvu(aLIz
z-Tz4bRx1Kt45Z-uDn<Gj`6l1{QnCHvrTNCZ2_EXGWf=lT0q#($i=Yles~%rk9LtJt
z*h0HIm*QS7aMOMCOJ)YV6<*zjMhjN=Sjf-{iw(Hz<%o011_k*xJQm?62U<Q`0t-WR
z*RhZ8Vl0T<^BcI^p-U92wcw?Q`1tbX4W-Hykq08l;Q*0H7S-V{=@9^3WZMLP;2m9q
zc}^15&aOFklY(@4YzIj9c{E}`(7S-DrO*n}E{xp+zJAvzXQ-BiFMHvI2sn!|n*zhU
z6PUa8U@gJJuC$vAF$SNU{sC<Gt<A|C|EGNr(J8_3?1pof8|`oNvLA4b<)GEOEhok~
zeQYyeV#2gmepFIK@N<Dbb>r9%rSVEnO+mtX<rJZrmN*Hgb#L#>>5zvb^V6&!Fi;KX
z3vXK0W$Lb-HTwJ((?l4!OQ;J~g!P6)R5K4NA`$o8x6H5$$lBHT0Ov&!FRfEqP9MVj
zefn`TM_OMcC7_ayZ#1;dTT{jOJ4g{96W+9YK=$Ed8Wa5V*V+mQHbMY#*guK0e_}H)
z9CpOVXR7T{R)p42@DLRYR`(J<uFIm<>FV5A<v}cW)At|XyL8QfC|f1E_>5d`nSzQ6
zueNxLh@CRA48d<=Ds!@!mu;f7EcTok(J!6-8bombgAxJLr9jkCPTZ0FQco_IX@6#a
zPG}a8sX`T_;H?sf*Bu26^bdbXS`uxA_D0MR2|Hbe7W_{2!2dc*^BF7R-aAc*DT#MD
zx6~F~r-(0`iL`sB8m74DFlt;SZNE2ASDo^bWVG3}D6aQo&v?6B7TGn)=gBGY-haL`
zKgV}uvzkB5>cJQC5-F{#0+XoVxyC6Q`p>@cvxpwph|*E>MW1Lf{ne_H)~@3<H>@iE
z#h^&<xi0Eqa+e9>3{-Q0wN-_%nU)PQGw7-;Ta$p><8pw9NS>Sya;D-)RpMde{ha-s
z`JbVM^U-|>P=^Kpoaz+BK0^VQeu7W`4EULX<|H6FA~Dx~R@CY7C4Co}7pcM~Yhk|&
zxe*Aww{j-X^wFlbm2~MNr88u5_LmPumhG$dKQ<iWsLpfl;{zJ%czWsPAPHRsTu@B;
zN`CtJdX9$jMj+tJ?)~C)F&4$<;RrmR_6Hi+d`;<&Dr5Wg!220F_y;VotBElX@W(;J
zSyxJP?U6n9$GCOJS&gRmZbR>3`!7i<HAV<czIiZLSg^uiJz+h$CPq8kN{IUo1_b_g
zvb|}aPYB=J9-ktb(%v(hJV)eiO{+(fY!c?Q5=COin=>MwB6y(h!K1CZXIV%a2W0Wv
z0$L&C^5!llMv7&&8%X2?wxeEL;O9TJcM@NvhQ3F3<C)?yqU%m^hO3j*Q7|EI1|M?!
z)-EM^T!I`0Y?c#X^CNQdNWY!VH*G)ukE(t}o?qbjP}Mrbe^eE;xdBy&@bv;5$O9og
z3ln#StUg?W1BGqc8V1|?p6UHgti6R>F(kyO|MGNZ@-9$4$OyG?ol7#83v1z(4YnAJ
z{Cx?_LG&|n#?rOeF~nRwD<xx%fdX1i^3Bv;=OW+~hN4H{`SbU;q3w(DtBc$1AqdOM
zqvgX%>irbIrg*r%A6Tmckn*Tf!0xc+1~&!^i1h_dh*E!Cn2mAv9zo=Dxd@q>Df`?+
zs8;imP6Y4OT7OJYSb^D4Kg<-j-*q!sKUTCDt*;mcC?n3azb>u53JH0(y)VjPGf&?b
zBzEZME-LLmR~En^)9>T`1Vhxr_oy)qSmGc-^d3U^8LFaTeL?1Gb7eI*FEmdsdqju6
zxskf!&n?FL+_R`EHmkZ?&*!hxsy^*x()h4d%|&gH?PW}L@hy#20Gw6LIbUZs0dBJ6
zy{_ht`>z>zw~v}#h5!JP;s2Te5BWH7AU<s{ud*zVVssfZVTh)sxYkGZ^7Y+GrbwPV
zf<sjtpXb$GyMnQz%D^u3{c7v<($sKV0Yh^g)|TH1@pm2$m_;`#q)5O0A_PWorjci>
z>vQ)6&S~hYoj)zPi=W%YE9^f`q;CJ9I`b}j{a4SsBbV!K@#S_;ueY7mF<DJumKFLn
zTD`=()3z{74jhtSj5Wk8QeQs@8H~iiUCM7D;2vwQK{L5&dQ6nv!DSa>mygMT3|7QY
zhqcQ8G&r=Wk%z6Zqnz>Tr>a0}^i1ytao6+q>m_m1diyTRiwL6Eh?WAOXs2;7k^Mfj
zEa0a%He{G_V1FYCVk~x$IzzFPHKIoXpItkytK4hyRN=%PN@C9aZ$`%+XJ<Ktqt>q1
zGyYblT+Crv3C9P`DSL;yrg)FbF&?5E4o>hxyW;_@*yjtWmTT<4<6Dw$&<g)y3$#C~
zAG#Z@j0KKo2`W-H1CyvjR-prffpy5vT~%}j38^vx*WFk$<T()ywT8ih8TbJ!!)&N7
z*_D>jN7TZ`c9jAV>KP6ho+@DidSCeNok`^_3N($%dQVA9t!k}T%39PQB@PywkEpl@
zqaTOyj^U?WGO$rf@F-8-tmwu46{kz-Mm6KC`b`+=Cl75{IneD>B0FWO>Vxgp4k>RK
z+jSm+h}Yk0G?9JmP()jhhp*e3Xn5XJC%Eo9VGQs3yW?Lk9dQX=x`?5aoS1*lq39K|
z-lIF5Cekz;)pW1=BAzj*gJE3foi;eV4xH_1h|19}Jh>Hk-^nk%`WCJE_7;}GtIB0u
zA*TKEs2y@vx8&h}`Xl!^E3?&CR&jNUZ{%(al<>oj??lB#O<l^rNq{9mmG-lUA1*Z#
zHK7}>srJsiL2BQ^Vf*Z?rlql)+A6+?s(P=2J%(rEy*($<TKF5D70Gq7Tf?rmPqSFz
z8$N(DfxOz{(<-1iEUNn}XbD<a?v}y^BAZGO?t@m65o8OPn^WlJ0$9a?sxg13-xvbj
zZ~4tC&I#gz#*!ozxW;U~jBV?@Yn&dHTUuF{2|m9$8#wPsJU&D_PNrZ&afY)YJBq5t
zBhb;MnjLHQ!lH^|OG1!CB7zL37Fc3kLWJ{>CV_G$_I1N0bIyHC2)4tGnC6-kIr&z%
ztB<j2%G0J=1qbY3BT-wH!8e7WYNUYo@xONSo)!-WQ=(HnRF-C5lqgH}=8vCU-tK6p
zz#2+l(QsMk{;BMNNtC^ggb&UpPi#5XF7nLfOsL*tAfs3qpQ4J69?tR#kYq=SAzgCr
zeh`aoPGdaNI%$Ta9d}U6R%j!xgyyYX+abL5?d;{kIa{;8*`{LExJJmM7aij)Vb%JH
zFd)j7?{Hb67t`&6p4AC4({RG!WC#5GNK7EU@#3+#MpD`sBY_w-sDUOt<ic$1g#?kY
z*JO8gttn&rWeT2R@_p*<Lkk)TQutlIOIj%u<z6Ci^yGoXw(F3|s0atPLH;?kur#(s
zjAhVvVjh)JiQ)L`WMw$dYMbNmLe^it_iFu7`SMK6CbIgxQv3FNu-taFoHiY;(|d5c
za`vZ_kh(KF-J(fNf)j`Y?lK$eFBStHDz<FI+ep5gy4D!u`<6Z|cJ3v_7l%cC!mm(d
zip5cF^2yLj6GFB8kV{9+#?Mx*PL|HrO=n}aRsit;Pd@KigW}|M)F7+F!1MdY_5ky^
z`(&a1TB(-#j4@PwF;+606F{$m!;nW4Dyo#kB&F4emdjX=E>E(rJTGImhDna-K`RyV
z1;2)2z8}AtG(nNf`X;2lwG*>tu!n7sq1Ifi%IK~jZ3D-_ZMubqpJ5|`O;TfuHK<_2
z8$IhC=AW$^{B12j8S=wYaN+#5#7mkJkU&EOcyR=HHmt$;9o9@|cM&NYpG%YK>*wdd
zap_(w?D%RVNL!m##A2^LNTdv6n}foUq+d0~a7aJL>!@w06%Yn}7oBI<WawvncW+Nr
zzI>08os!y&YkVgX@ZtA<vbq?{Wb<foZRVeMoIx0cv(aviM1m%1<P=7o1~dKzNCjG^
z^!NAoB)Z<E4+q>e=zB0Q3eLD^%8%KYD9^L`6{rwSL6dgRRYMf#T@t0qq|N$~Gme30
zy@y;M{2?<s=X!tLUjLI;hsDgRYT#P-VRuLP#@J|Tag&_Cr226n5+=))`WFVn)i9c^
zETNUlknoQ3b@JjL$30{wQPnUD0lCB)DD*dxJR*YNIC0oUH>3vDBgx$=!qfVONL~l!
zD(@*f9*ok78rQx;79h;lgbjB*ZFIf-4Q7HId`3S)mhg3kgLW_=K418{Rp2HG2!0Q>
zPSrOA3@%bIiqM!605GoW%eI36nlvxPG9F7|oCzLy|9qk)m3}GX?gw7M0iTXx-Xg!6
zl~`}3RGkzo7&llCr*RGXWEvU|jkKj-ulUid9A#3?LC!8~Er)*c+%jO>7Ty)1mxzND
zM?9TivqIulzqBiW1QkSv!~RJnz)w+N8J-5c6?Lx7e};Mc6kV(P8gb$alv=I2^UO4R
z@hR1B-Y$DEfBW<fs9rn4!g+SRf4M!rOJNtr^Kz@$UYF$EmBGijGEYQLOOcnARn5Bn
zd?}e#_RysGEI)>i@yu}!HvX%Kl6CSa+H34$@rCjoVBWdJAhh#zE_CemSRoZ6w8cDF
zRChvr`6>upxZU>SWZ(n>u{JxlDI>#Q<DCCtr=V*D&H)>=SFp;y7kB2^7VA*9FQA7?
zNB)ZKG?r$|FU{os9l#g9Cm{3=fCFZgpJEYTzC6AIbQZ^=svcaK<F3_PJEFK-bguCq
z8G674P))X4cK}&Km~1}c6KCNxr8r6w7=V9oDIi0Q2)Z0Qq&Tjb000WfpAvH`NeI3D
zD{~b|<7@>MzjatPxU|XUa}tg8N|2MJ2WNu$<@f&kOiMYg5l9solyQzM*0At6mJBv}
z0ZPrqv?3YxPJ@&&0b820$y}N$v}s?`hbGO&#NV1r5_<{^JlB-GSF3iu=U#cl11MP1
zpMh({%Xj1_pnBeG)5pdiywB-AM^^@DhVzzN>k3|sh2=MOb%$$r)sK2!HOfW!g_Vs<
z_oMWY7gAx&-U@}fdPk^I_M9DK-Qu!Zuxjd%Sb%o`_SciL#UdqnI;G`wHOtXY4v7ae
zx#6|o;WXL*+$ir;uP>F~`}N5}`eIh<DhnT>f$l{|4jUfE!V50Lsn#BA>w@{DblQ?o
z8l#BLLA<2?xR^WGaci5>TuJ8JwtO%a%BIbFUb|Jlc5QLaf4oF`KD5tp6pf57ZwB`a
z=6e=!bl5);TPqo0Y|1!$U0bUAQq5+2+Wrc2XuTa#qip<b+~wO~pxMz9oAtH!k|;Ay
z52!PG=#Y9kqPc_`SOiQX&z$d52{nSf{)#-L>Qu~as^%-S>(ikm$KgQYiNo|oDf46q
z;7ip>mHD=>>{*|tfM$6&qpKr(=)k!eS5r+FSmdBXeqeQ1+FFh!d03@w!NhR3-<Ub8
z)EaN!xYFitn+Mr7DstjuMcN;~he^}33`}^9^mLNUwcme}T5618gAh2%{bSf^BpVyU
z1MxZJA>CR1F!qzNs3!%yj}KmCy?}Uh)GdOZ<!y)MX#;F-=n;3DQbkjY!2KIDr}Lu*
zYtHiVAN*fdG0``xU5_sBfWnG*fSh&F97BU8XNTqQxtGw1yzfC*Xm@X~&mfTQ1Te&?
zBH5m`Z0D8h{7o<K9gta}`$GNpsxkd$@#yV#K6fgBYr0KdxPCHk1aaDAaCHFedI#v0
zLDLhXem<)VQ4IwaNe?&hPW11;+S)y^%P-2Oxi@JJJ=re4v`M{QYbHMVw%olMyipTh
zJyZWf>c6iVIu?4+7&z*reFp#!^+Qf^;X8@qtWT;&h>kCO@&m;?8#`L0D-9IW_f*CZ
z!KJ3%xrHS?tMsZ!9;B^#Fa8uDjrSxP6)S`V(6LZ)0Q`p`A;NDpsX)#md@+Z91i#^U
zb}La9$(v@nbjH?;dj!X>1esZ9(Ud&9el<v7XgCgIC`kbRxq<I9byU?h;6P8ib??rW
zkDN03f?OTT1$Dmu4V|@oY8S5EQ=ebJy)50i)CrI`)lAp6vzxP5SD*dRS;5EHa#(B2
zUD$HJQFEJfSHA<0y7GXtcv#<mz5}u*&#~*Gzu}ijd()em%~+i+d^_UVMW>)^Gml9e
zqoXBgBA81}sJa{aQQu-u25f4s)giQ#F^f#Ed3|zP1A9c5sazc?r2KJvR^dp&YM_Iq
zg+<uj$a}S|etclrTF`3Mi8Zr77q@UUcEY!oxJhnL+$AsVB<bs?zO<(IbqQypyz=q%
z9T1l#*Q+FhoqOOjj{TJX4sZd|#pOl3Dg#@E5~K}eM+N!j7&vKNfgM&~T(t}(=j<IK
zFc`4!&!=k9A5;IheCAeT*)Y~h1{SWr1IF-yk4Kv~6RkdCa&2ik3kA`tE&e)69ld1T
zYCzvtZt<aOs+LF;=}wjnVE><!51t#HzZ`=t9ty8dRe3MdR}br(B`Xi?GV>E3N}`?q
zfdPJqJqKnznSUsXS;b#PQMvS?eyM5N-&J@r*U3bfKA%W)lWg$lI$YGlT77=ISRm{K
z`P{wg0&UhmkQb}YSJbbp|DR611Hyq%O3#wQU5+v@^jA{t-bCRRE=ZER`_@T#LCbDx
z%D!KoNg$#NywVpg<*|P$YICgxMFG`9MCr!_QN$?lpz|>5I>wMh?K6dQqDP4J!pojC
z$`_BDp`Js%n1>&JjR5OaZ-FWDE1ujwal3z96o2)6%>#FgUA)BIuGFC2Nug@G2(I$6
zB^Prvw`S~TDuOckA(L}Mg35*h=^#nJ|7ty`&sQ^jes23B!jrhMp@q((9Sdwhjg8%X
z3Rk;~&7U)0G#)JrRXxe~(U9;iDfO5f;!0K9+Z80q`V|0YqT6F3xhE?1QhkfLw0ZPK
z-!xrWm3s*_nTFXOSNW}Ba8I-^ntY*lTVNqk{cs2q=zG`LrZd~Jm@w<nKrM4}|M@8V
z(MbW@W8EdOo#;a()*ppiEav9*I(G%~g-+0ndE)Wc+Sc^5I16Jpo<Xg6^r3|+sN~sx
zgXJ6knnzPzZsz*NqPqTu2yo!q8?gt46`p)riAV0n#u1KjyQZz3&2XGhy{`S*hBG(6
zA`tw-x+rAJH_+169%$+4eitE)W%08Kg$zbNEhZgp^6L$~=gc1=RxsU9`7Aw9@+g=m
zza1trD8J^)Pp0@R@LeRk%QbG4XNJ>MUROVyE6x;T=`0lLeLX{6i9836;7elozCPJS
zMjz`AhUES-YJUfWH5`1{yS9UA;H3J;A5?AD3QuZ<v%|{1NpGM_se|7*xfEryOC*Xu
zb?Qd=FGJ+TKMe1H#Qis~rUqw+tmj(;smHv%%SnwRw@k#TeOs=RNc~6PtX&8a_We8H
zIKgQ3pzgIx14z8G*lr;Gq6BmlI!v(hU>8OzT+ceC--=kx8R&U?2h7wk-mw0orxu%z
z->F}kgw_<SUcBnw0YAL;yR80EQ^EhGrv4T6NRn8oG69@$ot|G9oMLZRFnkk^Id?Z<
zD&EGjtymsw<swDxQ`I;2X^B<zlkb5@9EBTGWjIT4Vbr^j=0Abh^6Z1ox3BJC^om6X
zUF19L)Nj)&^KWK$9p5-l1+V*7S18&d#!`B-Zk6}+RBtvkcm#Gz2*G}zbs~Fs-K|*j
zM~W$06iG%7r3cdOKsplVRH+5Pp5}ZZgCs*lj8oWFpm?5}Z!eaVzt%9%JS+m1$L=-}
z!w7tKwP!3}3C^DW6^W0+3#}FmJDGH-t>Vo-YsZ*hm!vv}sGJiSl7=%{y&~z|_4Ov4
z*4Ux`lFK8sWYxNCb>>gAC!fZhwn1Y^JaG^ngbuz~@ez@d0yB}5&zdpr)hqoRqwP+8
zWB3d~=$3@U$Vl%`WGgS)nZ8-nBHu3s2evxs$^{+yX-#gxy5-gogu12icCb3bX;BEc
zYocZNR4@SGD=1?U7pO8~x5|RrfjumQ+D;u8IoJR-1{RRp0cinN*22E4`K=p?U?ukT
zme4jTayJIw{QFppu2B6#HH7%wMuN9nk~6ftSAm}16>z?4mC`&pNTt^UhEfRpDr9`T
z$(05fZ{tl;5`c`LbVXI5Xo5QmUEa-4{E-K+pNARNt7%Rb4&(-Jq)OX=B&!qU7n~TX
z`t`gnFY|D?w-F!%@dc9nQ`nFsO!`!>#*`=|MY8Xcb{txA=7`?NbaCnODmHcfA)B|@
zpdQRjh)LJw<qgetqJza}_4<_~^y1D~*tqEIOv5CcOpC7j2=9@)G3`Au18v5u$NS27
zJ#?22i~4|jMW9_I-tu&q;^U@LS={VA(pFZIMhsk7r3UHH21H4GZ+H9<(^9#FKVx2-
zK?M(ayjG?=9nJTb<nalWoC~GJm72!gd*@t1<yj{-7G*cv+st|)BRmB2Lx`mY)_rtC
zDBG@==?^bZ<?RHtQPMYN<t%za7kTxGrPom*#l`x0TTIkAeFhN6OQ1gNE3{{N6vXHS
z64&3LJyejE$VVbn(Hs>RP+?9OqciP+n5v`hG`IY#nauE~-kt=x#^0Vc22JQ~?Bl<h
z1E@LTP`3wQ!|=Y`8B*XNQ#oC9Gv}_q|F#};bN`V{QsjHyQDv~<ef?*;pu$%cI-4UY
z*K*!EN<#R7?RV~CJy-jD%MjLdaG2kKvheH*v@K)H{rPlM92`xb2oOCUwbfE$kRqb7
z;uZEucCnhF`qS5Wb7-;oI#V1cV)gv@Qk34M>!{gkqcJ~Nl4Kb=kup_Xg5-Bhr2MS=
zBJPe;$Xneb>Ar8#_dON-$PLt_K1^!a>p!UHn9%g(<C33$Ws};aJ%6>r#hWn`#Fjh?
zb=kF6Bn(4qC~7%?wE&)EhpI<Fub+&`yLo5&@{QhgMQ&skYyy>^q47>x2E4*4W8H`a
zL`Ny|*OOtg*FCKaNS|c!UVhVl6<+i}^=mHu=0#O_>PG)ZS_wyN<yH%kkz*K1s~oJL
z3{npw(ZDyv@y9QJEdoFLe6rN8o@0B>U8vcZ)s=c^Sx2g$YBRvm>~6>0NT)h?^X0|a
z=`FODmkwu7NBPrI)$T)~LRMgMRwLv)MWS}zI?Cb524nwm6YAvM<jhXQ6|yv!z*k`U
z#w(75Vsp61f_-!v&W%?nx}jY)&}ii?H)Q6eHsfa3*sWcR35YZs*muZ9U;mS^pIdz)
zLPNYVO)F<)4Wu{S^X$@`gcR~c%D#4e(*4<ZFs6&aV#K$%p9_obCO#s`i~MpIo}Gng
zYnkFyQdY!ICkOA6PGoUP5Oe=Uy<Z$9fOX%CbQ3yl&&}mHGIeJTf1r$b|7u@AU+0YY
zAJ-(*(oZKSuR5V63NT<$oQO;x3&zlSz-G?wM!}(rP0q2fDK!iPb$C;ZV3P$Xo}zMf
z-5(~Js2%nh)tA2M+)!unND<XZ`mO7JXOD8)kW|GUc@VzhDYPDZ*MFN2&1yv|36!%b
zk?AT+$uJN9WC4eTwA~jNY94P5VVuX#BJXz#52}qf_u+76$^+u)2tth5l#!^SZCH_~
z&^MZevk#sxV>sYI5;FAE1n~yT<YaM5)~ispE198CRY1ZG#RNd^F;FoEbZNmzt_UHN
z?!wUi&@sE9)(%1HsLlnu7O>wtg=c)Ce(-YfRnnWsE=|XT!H|M0l5fV?u9@ih*`PSF
z+9}{$p2c;1>ju1S_x!$Du1~(^U<$?Eoz8Vh2m&uH7o^?KU^|u^m|cVaBA>{4!%w3H
z;jGkYDZ7F>bCS0n)cGsMV_^JXk>F^z)W?epH<t(T%OZ8YCc2Zjbae0#x2eka+fC7y
z<KG6x2O%BOs55TDffFdkjG<*w_dtyoH~`@N9Ri>z1NI4=5%6E<$w4<O>R?YGnN=cC
z`6?bb6Ab7biwtzMN(U|r`%%OJF|Bbxe*F*uBE%*BpreA8*6_e{5kNO<RG@^9q8l=(
zW{sNQf1O?iV167||MxRa2>P)~4XWP71CL+?DQ!`JK3l<qR<;O020O^$u}ptSpdGxw
zh$Lo^(e_tR+7>c+2Fu?TMO%dbfb&`Zk~X*h2VBbbx7zcY>Muhz``_y17XC-BI39Qt
z2S~_)=|8+!Tz^{yb~(Taxc?Tdyx~FFyL8}WJRem-g8!3A{%`j`N;JG6M`a?A^=~YK
z|C6lwe?SU?e?j=avGD&*iTdZ*{{Q|6gB&-AL09XT1pj9D{5M8G`mdn85757vD*p{a
zlm81+{Q&(Z^@8)iK|Ly<ksWq|e<QH}4ba#9D@9}jli)wGM56x&B^dn$#e9JN4Uqje
zh|U7kv%yaAZ<Oc10U0)br7&$`68syg_-{~y!v}~Ql(2zH@UNHhzd>(qe<l6e#Qdu<
zB0hg#=EW{5=wcNfG`oifF69e4-y;Bt?oohu{Qz0)QGK}M@W3+*K~#HG1pjf!eK_T?
z|M$BAa2A6~_h>;&<ERM#(#Jt8e<(npKX~9_lc1_Ulm!1-;(P?N(Ei1Jl*VR24u8o1
hXKn>8|KY{|Yxq7^L#5#V4P^u|?o&W-to+0O{{RqJI#K`t