-
Notifications
You must be signed in to change notification settings - Fork 200
/
Copy pathSVM5.m
145 lines (128 loc) · 6.18 KB
/
SVM5.m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
% This runs time tests on the SVM solver with the full rank and low rank
% representations. It also compares the cost of the sparse SVM evaluations
% to ignorantly performing a full summation, which would be the associated
% cost of using an RBF network. There is also an option to study the cost
% of different parameterizations for solving the quadratic program.
% To allow for the low-rank expansion parameter to be set
global GAUSSQR_PARAMETERS
% Choose a range of parameters to test over, or fixed parameters if testing
% over something else
epvec = logspace(-4,2,41);
bcvec = logspace(-4,4,40);
ep = .01;
bc = 1;
% Choose whether or not to plot the output, which you may not want to do if
% running the file remotely or within a batch script
% Set to 'off' to prevent plotting, 'on' to plot
% You can see a non-visible figure with handle h using figure(h)
% If this is 'off', instead of displaying the plot it automatically saves
% the figure using gqr_savefig
plots_on = 'on';
% Set the low rank parameter if desired
low_rank = 0;
GAUSSQR_PARAMETERS.DEFAULT_REGRESSION_FUNC = .1;
% Choose which test you want to run
% 1 - Fixed parameterization, increasing size, low-rank vs. standard
% 2 - Cost of solving optimization problem with various ep and bc values
test_opt = 2;
switch test_opt
case 1
% Create random training and test data
train_N_vec = round(logspace(1,3.5,30));
lowvec = zeros(length(train_N_vec),1);
lowvecS = zeros(length(train_N_vec),1);
fullvec = zeros(length(train_N_vec),1);
fullvecS = zeros(length(train_N_vec),1);
k = 1;
h_waitbar = waitbar(0,'Initializing','Visible',plots_on);
fprintf('LOW\t prep setup solve clean \tFULL\t prep setup solve clean\n')
for train_N=train_N_vec
[train_data,train_class] = SVM_setup(1,train_N,10);
SVMlow = gqr_fitsvm(train_data,train_class,ep,bc,1);
lowvec(k) = SVMlow.solve_time;
lowvecS(k) = SVMlow.prep_time + SVMlow.setup_time + SVMlow.postsolve_time;
SVMfull = gqr_fitsvm(train_data,train_class,ep,bc,0);
fullvec(k) = SVMfull.solve_time;
fullvecS(k) = SVMfull.prep_time + SVMfull.setup_time + SVMfull.postsolve_time;
progress = floor(100*k/length(train_N_vec))/100;
waitbar(progress,h_waitbar,sprintf('num points=%d, low rank time=%5.2f, full rank time=%5.2f',train_N,lowvec(k),fullvec(k)))
if strcmp(plots_on,'off')
fprintf('%5.2f%% complete, num points=%d, low rank time=%5.2f, full rank time=%5.2f\n',progress*100,train_N,lowvec(k),fullvec(k))
else
fprintf('\t%5.2g %5.2g %5.2g %5.2g \t\t %5.2g %5.2g %5.2g %5.2g\n',...
SVMlow.prep_time,SVMlow.setup_time,SVMlow.solve_time,SVMlow.postsolve_time,...
SVMfull.prep_time,SVMfull.setup_time,SVMfull.solve_time,SVMfull.postsolve_time)
end
k = k + 1;
end
waitbar(1,h_waitbar,'Plotting')
h = figure('Visible',plots_on);
loglog(train_N_vec,lowvec,'linewidth',2)
hold on
loglog(train_N_vec,fullvec,'--','linewidth',2)
xlabel('number of training points')
ylabel('training time')
legend('Low rank','Full rank','location','southeast')
hold off
if strcmp(plots_on,'off')
gqr_savefig(h,sprintf('SVMtimetests%d',test_opt));
plot_command = 'loglog(train_N_vec,[lowvec,fullvec])';
save(sprintf('SVMtimetests%d',test_opt),'train_N_vec','lowvec','fullvec','plot_command');
end
close(h_waitbar)
case 2
% Create random training and test data
[train_data,train_class] = SVM_setup(1,400,10);
T = zeros(length(epvec),length(bcvec));
k_ep = 1;
h_waitbar = waitbar(0,'Initializing','Visible',plots_on);
fprintf(' ep bc prep setup solve clean\n')
for ep=epvec
k_bc = 1;
for bc=bcvec
SVM = gqr_fitsvm(train_data,train_class,ep,bc,low_rank);
T(k_ep,k_bc) = SVM.solve_time;
progress = floor(100*((k_ep-1)*length(bcvec)+k_bc)/(length(epvec)*length(bcvec)))/100;
waitbar(progress,h_waitbar,sprintf('compute time=%5.2f, \\epsilon=%5.2f C=%5.2f',T(k_ep,k_bc),ep,bc))
if strcmp(plots_on,'off')
fprintf('%5.2f%% complete, ep=%5.2f C=%5.2f, time=%5.2f\n',progress*100,ep,bc,T(k_ep,k_bc))
else
fprintf('%5.2g %5.2g %5.2g %5.2g %5.2g %5.2g\n',...
ep,bc,SVM.prep_time,SVM.setup_time,SVM.solve_time,SVM.postsolve_time)
end
k_bc = k_bc + 1;
end
k_ep = k_ep + 1;
end
waitbar(1,h_waitbar,'Plotting')
% Average values, to try to avoid fluctuations
T_avg = 1/9*(T(1:end-2,1:end-2) + T(2:end-1,1:end-2) + T(3:end,1:end-2)+...
T(1:end-2,2:end-1) + T(2:end-1,2:end-1) + T(3:end,2:end-1)+...
T(1:end-2,3:end) + T(2:end-1,3:end) + T(3:end,3:end));
[E,B] = meshgrid(epvec,bcvec);
E_avg = E(2:end-1,2:end-1);
B_avg = B(2:end-1,2:end-1);
h = figure('Visible',plots_on);
h_ev = surf(E_avg,B_avg,T_avg');
set(h_ev,'edgecolor','none')
set(gca,'xscale','log')
set(gca,'yscale','log')
set(gca,'xtick',[1e-2,1e1,1e4])
set(gca,'ytick',[1e-4,1e0,1e4])
xlabel('\epsilon')
ylabel('C')
zlabel('SVM training time')
shading interp
grid off
view([.5,-1,.8])
%colormap gray
colorbar
if strcmp(plots_on,'off')
gqr_savefig(h,sprintf('SVMtimetests%d',test_opt));
plot_command = 'surf(E_avg,B_avg,T_avg'')';
save(sprintf('SVMtimetests%d',test_opt),'E_avg','B_avg','T_avg','plot_command');
end
close(h_waitbar)
otherwise
error('Unacceptable test case test_opt=%d',test_opt)
end