-
Notifications
You must be signed in to change notification settings - Fork 2
/
BCSH_batch.py
225 lines (193 loc) · 8.2 KB
/
BCSH_batch.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
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
# -*- coding: utf-8 -*-
"""
@author: yingwenjie
"""
import scipy
import scipy.io
import numpy as np
from utils import *
def BCSH_batch(tfidf, bits=2, rand_time=8, batch=100000, iters=100, lambd=0.0005, word=[], arg_dir="./data/tmp"):
[m, n] = tfidf.shape
print(m, n)
"""
外层循环哈希码，每次计算bits位的哈希码【每次计算的bits哈希码是独立的，且内层迭代完后不再更新】
内层循环为计算bits位哈希码进行的收敛次数
"""
summy = np.ones((1, n))
for i in range(rand_time):
print("===============" + str(i) + "================")
"""
shuffler
"""
col_rand_array = np.arange(tfidf.shape[0])
np.random.shuffle(col_rand_array)
weight = tfidf[col_rand_array[0:batch], :]
[m, n] = weight.shape
"""
dropout
1、概率dropout高频词
2、随机dropout高频词
"""
for k in range(len(word)):
print(str(k) + "\t" + str(summy[0,k]) + "\t" + str(word[k])) #打印高频词及其概率
num = np.random.randint(0, 2, (m, 1))
#weight = weight / summy #概率dropout
#weight = scipy.sparse.csr_matrix(weight)
#summy[summy >= 3] = 0
#summy[summy < 3] = 1
#dropout = scipy.sparse.csr_matrix(num) * scipy.sparse.csr_matrix(summy)
#print(weight)
#weight = weight.multiply(dropout)
#print(weight)
#print(weight.nonzero())
#weight = weight.tolil()
#for k in range(n):
# if summy[0,k] > 10:
# weight[:,k] *= 0.1
"""
随机初始化哈希码B
"""
B = np.random.randint(0, 2, (bits, m))
B[B == 0] = -1
for it in range(iters):
tempB1 = B.copy()
tempB0 = -B
tempB1[tempB1 < 0] = 0
tempB0[tempB0 < 0] = 0
PX1_B1 = scipy.sparse.csr_matrix(tempB1) * weight
PX1_B1 = PX1_B1.toarray().astype(float)
ALL1 = np.sum(PX1_B1, 1, keepdims=True)
for r in range(bits):
PX1_B1[r, :] = (PX1_B1[r, :]+1)/(ALL1[r, 0]+n)
PX1_B0 = scipy.sparse.csr_matrix(tempB0) * weight
PX1_B0 = PX1_B0.toarray().astype(float)
ALL0 = np.sum(PX1_B0, 1, keepdims=True)
for r in range(bits):
PX1_B0[r, :] = (PX1_B0[r, :]+1)/(ALL0[r, 0]+n)
"""
归一化高频词在0，1哈希码下的条件概率, 效果不行
"""
#PX1 = PX1_B1 + PX1_B0
#PX1_B1 = PX1_B1 / PX1
#PX1_B0 = PX1_B0 / PX1
#PB1 = np.sum(tempB1,1,keepdims=True)/m #计算哈希码为1的概率，查看位平衡效果
logPX1_B1 = np.log2(PX1_B1)
logPX1_B0 = np.log2(PX1_B0)
logPB1 = weight * scipy.sparse.csr_matrix(logPX1_B1.transpose())
logPB0 = weight * scipy.sparse.csr_matrix(logPX1_B0.transpose())
logPB1 = logPB1.toarray().transpose()
logPB0 = logPB0.toarray().transpose()
tmp = (logPB1 - logPB0) ### 规范化非常重要
tmp[tmp > 32] = 32
PXB1 = np.power(2, tmp)
PXB1 = PXB1 / (1 + PXB1)
Fx = PXB1 * 2 -1
#Y = Update(B,bits)
old_B = B.copy()
##位平衡约束
alpha1 = 0.0
bit_sum = np.sum(B, axis=1)
bit_sum = bit_sum.reshape(bits, 1)
m_one = np.full((1, m), 1.0/m)
bit_balance = np.dot(bit_sum, m_one)
Fx -= alpha1 * bit_balance
#Fx_ban = -bit_balance
##位独立约束
alpha2 = 0.0
if i != 0:
bit_un = np.dot(Bs, B.transpose())
bit_un = bit_un.reshape(i * bits, bits)
bit_un_res = np.dot(bit_un.transpose(), Bs)
bit_un_res = bit_un_res.reshape(bits, m)
Fx -= alpha2 * bit_un_res / m
#Fx_un = bit_un_res / (m)
#bit_un_sum = np.sum(bit_un, axis=0)
#bit_un_sum = bit_un_sum.reshape(bits,1)
#m_one = np.full((1,m),1.0/m)
#bit_un = np.dot(bit_un_sum,m_one)
#Fx -= alpha2 * bit_un
#if i == 0:
# Fx_un = np.zeros((bits, m))
#for ii in range(bits):
# for j in range(m):
#if((np.power((1 - Fx[ii,j]),2) + alpha1 * np.power((1 - Fx_ban[ii,j]),2) + alpha2 * np.power((1 - Fx_un[ii,j]), 2) ) <= (np.power((-1 - Fx[ii,j]),2) + alpha1 * np.power((-1 - Fx_ban[ii,j]),2)) + alpha2 * np.power((-1 - Fx_un[ii,j]), 2)):
# if(np.power((1 - Fx[ii,j]),2) + alpha2 * np.power((1 - Fx_un[ii,j]), 2)) <= (np.power((-1 - Fx[ii,j]),2) + alpha2 * np.power((-1 - Fx_un[ii,j]), 2)):
# B[ii,j]=1
# else:
# B[ii,j]=-1
B[Fx < 0] = -1
B[Fx > 0] = 1
updateB = sum(sum(B!=old_B))
print('update-----------------' + str(lambd))
print(updateB)
loss1 = np.trace(np.dot((B - Fx), (B - Fx).transpose()))
#loss2 = np.trace(np.dot((B - Y),(B - Y).transpose()))
#loss2 = np.trace(np.dot((B - Fx_ban),(B - Fx_ban).transpose()))
#loss3 = np.trace(np.dot((B - Fx_un),(B - Fx_un).transpose()))
loss2 = np.sum(B, axis=1)
loss3 = 0
if i !=0:
bit_un = np.dot(Bs, B.transpose())
loss3 = np.trace(np.dot(bit_un, bit_un.transpose()))
Loss = loss1 + alpha1 * loss2 + alpha2 * loss3
print('Loss=' + str(Loss) + "\t" + "loss1: " + str(loss1) + "\t" + "loss2: " + str(loss2) + "\t" + "loss3: " + str(loss3))
if(i==0):
logPX1_B1s = logPX1_B1
logPX1_B0s = logPX1_B0
Bs = B
else:
logPX1_B1s = np.r_[logPX1_B1s, logPX1_B1]
logPX1_B0s = np.r_[logPX1_B0s, logPX1_B0]
Bs = np.r_[Bs, B]
PX1_B0s = np.power(2, logPX1_B0s)
PX1_B1s = np.power(2, logPX1_B1s)
#summy = np.sum(PX1_B1s + PX1_B0s,0).reshape(1,n) * n / (i + 1)
loss2 = np.sum(Bs, axis=1)
loss3 = (np.dot(Bs, Bs.transpose()))
loss4 = sum(sum(abs(np.dot(Bs, Bs.transpose())))) - np.trace(np.dot(Bs, Bs.transpose()))
print("loss2" + "\t" + str(loss2))
print("loss3" + "\t" + str(loss3))
print("loss4" + "\t" + str(loss4))
oldi = 0
mini_batch = int(batch / 20)
for i in range(1,tfidf.shape[0] + 1):
if i % mini_batch == 0:
weight = tfidf[oldi:i, :]
logPB1 = weight * scipy.sparse.csr_matrix(logPX1_B1s.transpose())
logPB0 = weight * scipy.sparse.csr_matrix(logPX1_B0s.transpose())
logPB1 = logPB1.toarray().transpose()
logPB0 = logPB0.toarray().transpose()
tmp = (logPB1 - logPB0) ### 规范化很重要
tmp[tmp > 32] = 32
PXB1 = np.power(2, tmp)
PXB1 = PXB1 / (1 + PXB1)
Fx = PXB1 * 2 - 1
B = np.zeros(shape=Fx.shape)
B[Fx < 0] = -1
B[Fx > 0] = 1
oldi = i
if (i == mini_batch):
Bs = B
else:
Bs = np.c_[Bs,B]
i += 1
if i % mini_batch != 0:
weight = tfidf[oldi:i, :]
logPB1 = weight * scipy.sparse.csr_matrix(logPX1_B1s.transpose())
logPB0 = weight * scipy.sparse.csr_matrix(logPX1_B0s.transpose())
logPB1 = logPB1.toarray().transpose()
logPB0 = logPB0.toarray().transpose()
tmp = (logPB1 - logPB0) ### 规范化很重要
tmp[tmp > 32] = 32
PXB1 = np.power(2, tmp)
PXB1 = PXB1 / (1 + PXB1)
Fx = PXB1 * 2 - 1
B = np.zeros(shape=Fx.shape)
B[Fx < 0] = -1
B[Fx > 0] = 1
if (i < mini_batch):
Bs = B
else:
Bs = np.c_[Bs, B]
scipy.io.savemat(arg_dir + '/arg.mat',{'B': Bs,'logPX1_B1': logPX1_B1s,'logPX1_B0': logPX1_B0s})
return Bs