-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathgan.py
221 lines (186 loc) · 9.86 KB
/
gan.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
# -*- coding: utf-8 -*-
import tensorflow as tf
#tf.disable_v2_behavior()
import pickle
print("banben",tf.__version__)
class alignGAN():
#########################################################################################
# parameters init
#########################################################################################
def xavier_init(self, size):
in_dim = size[0]
xavier_stddev = 1. / tf.sqrt(in_dim / 2.)
aaa = tf.random_normal(shape=size, stddev=xavier_stddev, dtype=tf.float64)
return tf.to_float(aaa)
def init_paras_1hidden(self, sess):
h1_dim = 64
h2_dim = 1
input_dim = tf.to_int32(self.input_dim).eval(session=sess)
h1_dim = tf.to_int32(h1_dim).eval(session=sess)
h2_dim = tf.to_int32(h2_dim).eval(session=sess)
W1 = tf.Variable(self.xavier_init([input_dim, h1_dim]))
W2 = tf.Variable(self.xavier_init([h1_dim, h2_dim]))
b1 = tf.Variable(tf.zeros(shape=[h1_dim]))
b2 = tf.Variable(tf.zeros(shape=[h2_dim]))
return W1, W2, b1, b2
def __init__(self, args, sess):
self.input_dim = args.emb_dim
self.train_batch_size = args.train_batch_size
self.learning_rate = 1e-4
self.lambda_c = 0.2
if args.param_theta == 'None':
self.param_theta = None
else:
self.param_theta = pickle.load(open(args.param_theta))
if args.param_G == 'None':
self.param_G = None
else:
self.param_G = pickle.load(open(args.param_G))
self.build_model(sess)
def variable_summaries(self, var, name):
"""Attach a lot of summaries to a Tensor (for TensorBoard visualization)."""
with tf.name_scope(name + 'summaries'):
mean = tf.reduce_mean(var)
tf.summary.scalar(name + 'mean', mean)
with tf.name_scope(name + 'stddev'):
stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean)))
tf.summary.scalar(name + 'stddev', stddev)
tf.summary.scalar(name + 'max', tf.reduce_max(var))
tf.summary.scalar(name + 'min', tf.reduce_min(var))
tf.summary.histogram(name + 'histogram', var)
def writee(self,num_2_acc_top_10):
with open('num_2_acc_top_10','w') as f:
f.write(num_2_acc_top_10+'\n')
def build_model(self, sess):
#for train
self.left_embedding_D = tf.placeholder(tf.float32, shape=(self.train_batch_size,self.input_dim))
self.right_embedding_D = tf.placeholder(tf.float32, shape=(self.train_batch_size,self.input_dim))
self.left_embedding_A = tf.placeholder(tf.float32, shape=(None,self.input_dim))
self.right_embedding_A = tf.placeholder(tf.float32, shape=(None,self.input_dim))
#for test
self.node_left = tf.placeholder(tf.int32,name="node_left")
self.node_right = tf.placeholder(tf.int32, name="node_right")
self.batch_test_left = tf.placeholder(tf.float32,shape=(None,self.input_dim))
self.right_test_embedding = tf.placeholder(tf.float32,shape=(None,self.input_dim))
self.batch_test_right = tf.placeholder(tf.float32, shape=(None,self.input_dim))
self.left_test_embedding = tf.placeholder(tf.float32, shape=(None,self.input_dim))
self.neibor=tf.placeholder(dtype=tf.float32, shape=[None,128],name='neibor')
self.ips0 = tf.placeholder(dtype=tf.float32, shape=[None, 128], name='if_psu_1')#if pseudo 0?
self.ips1 = tf.placeholder(dtype=tf.float32, shape=[None, 128], name='if_psu_2')#if pseudo anchor1
self.ifps = tf.placeholder(dtype=tf.float32, shape=[None, 128], name='if_ps')
with tf.variable_scope('gan'):
if self.param_theta == None:
W1,W2,b1,b2 = self.init_paras_1hidden(sess)
self.theta = [W1,W2,b1,b2]
else:
self.theta = tf.Variable(self.param_theta,name='theta')
self.params_theta = self.theta #discriminator参数
with tf.variable_scope('sub'):
if self.param_G == None:
init_orthogonal = tf.orthogonal_initializer(gain=1.0, seed=None, dtype=tf.float32)
self.G = tf.get_variable('G', shape=[self.input_dim, self.input_dim], initializer=init_orthogonal)
else:
self.G = tf.Variable(self.param_G[0], name="G")
self.w0 = tf.get_variable('w_0', initializer=[[0.015] * 128] * 128)
self.w1 = tf.get_variable('w_1', initializer=[[0.015] * 128] * 128)
self.variable_summaries(self.G, 'G')
self.params_G = [self.G] #G矩阵参数
#########################################################################################
# draw graph for tensorflow
#########################################################################################
#PSML
self.new_add = tf.matmul(tf.multiply(self.neibor, self.ips0), self.w0) \
+ tf.matmul(tf.multiply(self.neibor, self.ips1), self.w1)
self.g_mapping = tf.transpose(tf.matmul(self.G, self.left_embedding_D, transpose_a=False,transpose_b=True))
self.g_mapping= self.g_mapping+tf.multiply(tf.nn.relu(self.new_add),self.ifps)
#print("shape",self.g_mapping) #60*100
self.d_fake = self.discriminator(self.g_mapping,self.theta)
self.d_real = self.discriminator(self.right_embedding_D,self.theta)
#实现式（2）
self.d_loss = tf.reduce_mean(self.d_real) - tf.reduce_mean(self.d_fake)
#实现式（3）
self.g_loss = -tf.reduce_mean(self.d_fake)
#实现式（4）
shape = tf.to_float(tf.shape(self.left_embedding_A))[0]
self.a_mapping = tf.transpose(tf.matmul(self.G, self.left_embedding_A, transpose_a=False, transpose_b=True))
Eu_distance_a = tf.sqrt(tf.reduce_sum(tf.square(self.a_mapping-self.right_embedding_A),1))
self.a_loss = (self.lambda_c/shape)*(tf.reduce_sum(Eu_distance_a))
tf.summary.scalar('d_loss', self.d_loss)
tf.summary.scalar('g_loss', self.g_loss)
self.d_optim = tf.train.RMSPropOptimizer(learning_rate=self.learning_rate)\
.minimize(-self.d_loss, var_list=self.params_theta)
self.g_optim = tf.train.RMSPropOptimizer(learning_rate=self.learning_rate) \
.minimize(self.g_loss+self.a_loss, var_list=self.params_G)
self.params_theta_clip = [p.assign(tf.clip_by_value(p, -0.01, 0.01)) for p in self.params_theta]
self.params_theta = self.params_theta_clip
# for test
self.l2r_acc,self.r2l_acc,self.node1_left,self.top30= self.get_acc()
self.merged_summary_op = tf.summary.merge_all()
self.test_summary_op = tf.summary.scalar('l2r_acc_top50', self.l2r_acc[4])
# with open('num_2_acc_top_10', 'w') as f:
# print type(num_2_acc_top_10)
# f.write(num_2_acc_top_10 + '\n')
# train_log.write(str(epoch_all) + "\t"
# + "train accuracy:\t" + buf + '\n')
def get_acc(self):
l2r_test_mapping = tf.transpose(tf.matmul(self.G, self.batch_test_left, transpose_a=False, transpose_b=True))
self.test_all_score_l2r = tf.matmul(l2r_test_mapping,self.right_test_embedding,transpose_a=False,transpose_b=True)
l2r_acc_intop_10 = tf.reduce_mean(tf.cast(tf.nn.in_top_k(self.test_all_score_l2r, self.node_right, 10), tf.float32))
num_l2r_acc_top_10=tf.nn.top_k(self.test_all_score_l2r,10)
l2r_acc_intop_20 = tf.reduce_mean(tf.cast(tf.nn.in_top_k(self.test_all_score_l2r, self.node_right, 20), tf.float32))
num_l2r_acc_top_20=tf.nn.top_k(self.test_all_score_l2r,20)
l2r_acc_intop_30 = tf.reduce_mean(tf.cast(tf.nn.in_top_k(self.test_all_score_l2r, self.node_right, 30), tf.float32))
num_l2r_acc_top_30=tf.nn.top_k(self.test_all_score_l2r,30)
l2r_acc_intop_40 = tf.reduce_mean(tf.cast(tf.nn.in_top_k(self.test_all_score_l2r, self.node_right, 40), tf.float32))
num_l2r_acc_top_40=tf.nn.top_k(self.test_all_score_l2r,40)
#print(123123123123123)
l2r_acc_intop_50 = tf.reduce_mean(tf.cast(tf.nn.in_top_k(self.test_all_score_l2r, self.node_right, 50), tf.float32))
num_l2r_acc_top_50=tf.nn.top_k(self.test_all_score_l2r,50)
# f.append(a)
# f.append(b)
# f.append(ccc)
# f.append(d)
# f.append(e)
#sess=tf.Session()
#tf.initialize_all_variables().run()
#sess.run(tf.initialize_all_variables())
#print("node:right :",sess.run(self.node_right))
#print("num_l2r_acc_top_10 :",sess.run(num_l2r_acc_top_10))
#print("num_l2r_acc_top_20: ",sess.run(num_l2r_acc_top_20))
#print("num_l2r_acc_top_30 :",sess.run(num_l2r_acc_top_30))
#print("num_l2r_acc_top_40 :",sess.run(num_l2r_acc_top_40))
#print("num_l2r_acc_top_50",sess.run(num_l2r_acc_top_50))
l2r_acc = [l2r_acc_intop_10, l2r_acc_intop_20, l2r_acc_intop_30,l2r_acc_intop_40, l2r_acc_intop_50]
r2l_test_mapping = tf.transpose(tf.matmul(self.G, self.left_test_embedding, transpose_a=False, transpose_b=True))
#print("this is ",r2l_test_mapping)
self.test_all_score_r2l = tf.matmul(self.batch_test_right,r2l_test_mapping, transpose_a=False,transpose_b=True)
#print("321321321",self.test_all_score_r2l)
r2l_acc_intop_10 = tf.reduce_mean(tf.cast(tf.nn.in_top_k(self.test_all_score_r2l, self.node_left, 10), tf.float32))
num_2_acc_top_10 = tf.nn.top_k(self.test_all_score_r2l, 10)
r2l_acc_intop_20 = tf.reduce_mean(tf.cast(tf.nn.in_top_k(self.test_all_score_r2l, self.node_left, 20), tf.float32))
num_2_acc_top_20 = tf.nn.top_k(self.test_all_score_r2l, 20)
r2l_acc_intop_30 = tf.reduce_mean(tf.cast(tf.nn.in_top_k(self.test_all_score_r2l, self.node_left, 30), tf.float32))
#top30=tf.nn.in_top_k(self.test_all_score_r2l, self.node_left, 30)
#r2l_acc_intop_30 = tf.reduce_mean(tf.cast(top30, tf.float32))
#num_2_acc_top_30 = tf.nn.top_k(self.test_all_score_r2l, 30)
r2l_acc_intop_40 = tf.reduce_mean(tf.cast(tf.nn.in_top_k(self.test_all_score_r2l, self.node_left, 40), tf.float32))
num_2_acc_top_40 = tf.nn.top_k(self.test_all_score_r2l, 40)
r2l_acc_intop_50 = tf.reduce_mean(tf.cast(tf.nn.in_top_k(self.test_all_score_r2l, self.node_left, 50), tf.float32))
num_2_acc_top_50 = tf.nn.top_k(self.test_all_score_r2l, 50)
r2l_acc = [r2l_acc_intop_10, r2l_acc_intop_20, r2l_acc_intop_30,r2l_acc_intop_40, r2l_acc_intop_50]
pp=tf.convert_to_tensor([1,2,3])
return l2r_acc,r2l_acc,self.node_left,pp
def discriminator(self, batch_embedding,theta):
W1 = theta[0]
W2 = theta[1]
b1 = theta[2]
b2 = theta[3]
D_h1 = tf.nn.relu(tf.matmul(batch_embedding, W1) + b1)
out = tf.matmul(D_h1, W2) + b2
return out
def save_model_theta(self, sess, filename):
param = sess.run(self.params_theta)
pickle.dump(param, open(filename, 'wb'))
def save_model_G(self, sess, filename):
param = sess.run(self.params_G)
pickle.dump(param, open(filename, 'wb'))