-
Notifications
You must be signed in to change notification settings - Fork 2
/
pbg.py
173 lines (152 loc) · 6.43 KB
/
pbg.py
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
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Wed Jan 17 18:01:50 2018
@author: thiagodepaulo
"""
import numpy as np
import time
from sklearn.base import BaseEstimator, ClassifierMixin
from os.path import join
from util import RandMatrices
#from scipy.sparse import csr_matrix
class PBG(BaseEstimator, ClassifierMixin):
def __init__(self, n_components, alpha=0.05, beta=0.0001, local_max_itr=50,
global_max_itr=50, local_threshold = 1e-6, global_threshold = 1e-6,
max_time=18000, save_interval=-1, out_dir='.', out_A='A', out_B='B',
calc_q=False, debug=False, rand_init=False):
self.n_components = n_components
self.alpha = alpha
self.beta = beta
self.local_max_itr = local_max_itr
self.global_max_itr = global_max_itr
self.local_threshold = local_threshold
self.global_threshold = global_threshold
self.max_time = max_time # seconds
self.save_interval = save_interval
self.out_dir = out_dir
self.out_A = out_A
self.out_B = out_B
self.calc_q = calc_q
self.debug = debug
self.rand_init = rand_init
def normalizedbycolumn_map(self, B):
n = len(B.values()[0])
col_sum = np.zeros(n)
for key in B:
vet = B[key]
for i in range(n):
col_sum[i] += vet[i]
for key in B:
vet = B[key]
for i in range(n):
vet[i] /= col_sum[i]
vet[i] = self.beta + vet[i]
return B
def normalizebycolumn_plus_beta(self, B):
if isinstance(B, dict):
return self.normalizedbycolumn_map(B)
nrow, ncol = B.shape
for i in range(ncol):
B[:,i] /= B[:,i].sum()
return self.beta + B
def Q2(self, X, D, A, B, alpha):
CONST = 0.0000001
_sum = 0
for d_j in D:
for w_i, f_ji in zip(X.indices[X.indptr[d_j]:X.indptr[d_j+1]],
X.data[X.indptr[d_j]:X.indptr[d_j+1]]):
AB_ji = A[d_j]*B[w_i]
C_ji = (AB_ji / AB_ji.sum())
_sum += sum((f_ji * C_ji) * (np.log((AB_ji + CONST) / (C_ji + CONST))))
_sum -= sum((alpha - A[d_j]) * np.log(A[d_j] + CONST) - A[d_j]*(np.log(A[d_j] + CONST ) - 1))
return _sum
def Q(self, X, D, A, B):
_sum = 0
for d_j in D:
for w_i, f_ji in zip(X.indices[X.indptr[d_j]:X.indptr[d_j+1]],
X.data[X.indptr[d_j]:X.indptr[d_j+1]]):
sumAjBi = sum(A[d_j]*B[w_i])
_sum += f_ji * np.log( f_ji / (sumAjBi)) - f_ji + (sumAjBi)
return _sum
def global_propag(self, Xcrc, W, A, B):
for w_i in W:
nB_i = np.zeros(self.n_components)
for d_j, f_ji in zip(Xcrc.indices[Xcrc.indptr[w_i]:Xcrc.indptr[w_i+1]],
Xcrc.data[Xcrc.indptr[w_i]:Xcrc.indptr[w_i+1]]):
H = (A[d_j] * B[w_i] )
nB_i += f_ji * (H / H.sum())
B[w_i] = nB_i
# B = self.beta + self.normalizebycolumn(B)
return self.normalizebycolumn_plus_beta(B)
def local_propag(self, X, d_j, A_j, B):
nA_j = np.zeros(len(A_j))
for w_i, f_ji in zip(X.indices[X.indptr[d_j]:X.indptr[d_j+1]],
X.data[X.indptr[d_j]:X.indptr[d_j+1]]):
H = (A_j * B[w_i])
nA_j += f_ji * (H / H.sum())
nA_j += self.alpha
return nA_j
def suppress(self, A_j, y_j):
aux = A_j[y_j]
A_j.fill(0)
A_j[y_j] = aux
def bgp(self, X, Xcrc, W, D, A, B, labelled=None):
global_niter = 0
t0 = time.time()
while global_niter <= self.global_max_itr :
global_niter += 1
if time.time() - t0 > self.max_time:
break
if self.save_interval!= None and self.save_interval % global_niter == 0 :
self.save_matrices(A,B,global_niter)
for d_j in D:
local_niter = 0
if self.debug and global_niter % 10 == 0: print(d_j)
while local_niter <= self.local_max_itr:
local_niter += 1
oldA_j = np.array(A[d_j])
A[d_j] = self.local_propag(X, d_j, A[d_j], B)
mean_change = np.mean(abs(A[d_j] - oldA_j))
if mean_change <= self.local_threshold:
#if self.debug: print('convergiu itr %s' %local_niter)
break
if (labelled is not None) and (labelled[d_j] != -1):
self.suppress(A[d_j], labelled[d_j])
self.global_propag(Xcrc, W, A, B)
if self.calc_q:
q = self.Q2(X, D, A, B, self.alpha)
if self.debug: print('itr %s Q %s' % (global_niter, q))
#if abs(q - oldq) <= self.GLOBAL_CONV_THRESHOLD:
# print '\t\t **GLOBAL convergiu em %s iteracoes' %global_niter
# break
#oldq = q
def fit(self, X, y=None):
rand = RandMatrices()
#D -> set of documents indices, W-> set of word indices, K-> number of topics
D, W, K = range(X.shape[0]), range(X.shape[1]), self.n_components
A,B = rand.create_rand_matrices(D ,W ,K ) if self.rand_init else rand.create_label_init_matrices(X, D, W, K, y, self.beta,-1)
# convert matriz
Xcsc = X.tocsc()
self.bgp(X, Xcsc, W, D, A, B, labelled=y)
self.components_ = B.transpose()
if y is not None:
# label construction
#truct a categorical distribution for classification only
classes = np.unique(y)
classes = (classes[classes != -1])
self.classes_ = classes
# atribui índices de classes aos exemplos não rotulados
self.transduction_ = self.classes_[np.argmax(A, axis=1)].ravel()
# cria distribuição dos rótulos (normaliza a matriz A)
normalizer = np.atleast_2d(np.sum(A, axis=1)).T
self.label_distributions_ = A / normalizer
return self
def transform(self, X):
return None
def save_matrices(self, A, B, global_niter):
np.save(join(self.out_dir, self.out_A+'_'+str(global_niter)), A)
np.save(join(self.out_dir, self.out_B+'_'+str(global_niter)), B)
#normalizer = np.atleast_2d(np.sum(probabilities, axis=1)).T
# probabilities /= normalizer
#