-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
212 lines (182 loc) · 12.1 KB
/
model.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
import tensorflow as tf
import numpy as np
class HHP:
def __init__(self, global_step, learning_rate, batch_size, neg_size, nbr_size, node_size, node_dim, num_node_types,
num_edge_types,
norm_rate):
# init parameters
self.global_step = global_step
self.node_size = node_size
self.node_dim = node_dim
self.num_node_types = num_node_types
self.num_edge_types = num_edge_types
self.batch_size = batch_size
self.neg_size = neg_size
self.norm_rate = norm_rate
self.learning_rate = learning_rate
self.nbr_size = nbr_size
self.l2_regular = tf.keras.regularizers.L2(self.norm_rate)
with tf.compat.v1.compat.v1.variable_scope('parameters', reuse = tf.compat.v1.AUTO_REUSE):
init_range_embed = np.sqrt(3.0 / (self.node_size + self.node_dim))
self.embedding = tf.compat.v1.get_variable('embedding_table',
initializer=tf.compat.v1.random_uniform([self.node_size, self.node_dim],
minval=-init_range_embed,
maxval=init_range_embed, dtype=tf.float32), trainable = True)
init_range_edge_type = np.sqrt(3.0 / (self.node_dim + self.node_dim))
self.edge_type_embed = tf.compat.v1.get_variable('edge_type_r',
initializer=tf.compat.v1.random_uniform(
[self.num_edge_types, self.node_dim],
minval=-init_range_edge_type,
maxval=init_range_edge_type,
dtype=tf.float32), trainable=True)
self.optimizer = tf.keras.optimizers.Adam(learning_rate=self.learning_rate)
# print("var",tf.compat.v1.trainable_variables())
# print("global",tf.compat.v1.global_variables())
# print("vars .",tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES)
def compute_loss(self,batch_data):
e_types = batch_data[0]
s_ids, s_types, s_negs, s_nbr_infos = batch_data[1]
t_ids, t_types, t_negs, t_nbr_infos = batch_data[2]
basic_info = self.construct_mu(s_ids, s_types, t_ids, t_types, e_types, t_negs, s_negs, neg_size=self.neg_size)
mu, neg_mus_st, neg_mus_ts, s_embed, t_embed, neg_embed_s_list, neg_embed_t_list = basic_info
pos_loss_st, neg_loss_st = self.construct_mutual_influence(s_embed, s_nbr_infos, t_embed, neg_embed_t_list)
pos_loss_ts, neg_loss_ts = self.construct_mutual_influence(t_embed, t_nbr_infos, s_embed, neg_embed_s_list)
lambda_st_pos = mu + pos_loss_st + pos_loss_ts
lambda_st_neg = neg_mus_st + neg_loss_st
lambda_ts_neg = neg_mus_ts + neg_loss_ts
loss = -tf.reduce_mean(tf.compat.v1.log(tf.sigmoid(lambda_st_pos) + 1e-6)) \
- tf.reduce_mean(tf.compat.v1.log(tf.sigmoid(-lambda_st_neg) + 1e-6)) \
- tf.reduce_mean(tf.compat.v1.log(tf.sigmoid(-lambda_ts_neg) + 1e-6)) \
+ self.norm_rate * tf.reduce_sum(tf.pow(self.edge_type_embed, 2))
# print("vars",tf.compat.v1.trainable_variables())
return loss
@tf.function
def train_step(self, optimizer, data):
with tf.GradientTape() as tape:
loss = self.compute_loss(data)
gradients = tape.gradient(loss, tf.compat.v1.trainable_variables())
optimizer.apply_gradients(zip(gradients, tf.compat.v1.trainable_variables()))
return loss
def train(self, batch_data, epochs):
for e in range(epochs):
if e%100 == 0: print("Steps",e)
print(self.optimizer,"opt")
loss = self.train_step(self.optimizer,batch_data)
print("Loss",loss)
def construct_node_latent_embed(self, node_ids, node_types, node_size, type_size):
# return tf.gather(self.embedding, node_ids)
with tf.compat.v1.variable_scope('node_type_embed', reuse=tf.compat.v1.AUTO_REUSE):
indices = tf.range(node_size) * type_size + node_types
embedding = tf.gather(self.embedding, node_ids-1)
new_matrix = tf.reshape(tf.compat.v1.unsorted_segment_sum(embedding, indices, node_size * type_size),
[node_size, type_size, self.node_dim])
embed_typed = tf.compat.v1.layers.dense(new_matrix, self.node_dim, activation=tf.nn.leaky_relu,
kernel_regularizer=self.l2_regular, reuse=tf.compat.v1.AUTO_REUSE,
name='node_type_embed')
node_final_embeds = tf.gather(tf.reshape(embed_typed, [node_size * type_size, self.node_dim]), indices)
return node_final_embeds
def construct_mu(self, s_ids, s_types, t_ids, t_types, e_types, t_neg_ids, s_neg_ids, neg_size):
with tf.compat.v1.variable_scope('mu_layer', reuse=tf.compat.v1.AUTO_REUSE):
s_embed = self.construct_node_latent_embed(s_ids, s_types, self.batch_size, self.num_node_types)
t_embed = self.construct_node_latent_embed(t_ids, t_types, self.batch_size, self.num_node_types)
e_embed = tf.gather(self.edge_type_embed, e_types)
mu = self.g_func(s_embed + e_embed, t_embed, 'l2')
neg_mus_st = [] # mu between s and neg-t
neg_embed_s_list = []
neg_embed_t_list = []
neg_mus_ts = []
for i in range(neg_size):
neg_t_embed = self.construct_node_latent_embed(t_neg_ids[:, i], t_types, self.batch_size,
self.num_node_types)
neg_embed_t_list.append(neg_t_embed)
neg_mu_t_i = tf.reshape(self.g_func(s_embed, neg_t_embed, 'l2'), [-1, 1])
neg_mus_st.append(neg_mu_t_i)
neg_s_embed = self.construct_node_latent_embed(s_neg_ids[:, i], s_types, self.batch_size,
self.num_node_types)
neg_embed_s_list.append(neg_s_embed)
neg_mu_s_i = tf.reshape(self.g_func(t_embed, neg_s_embed, 'l2'), [-1, 1])
neg_mus_ts.append(neg_mu_s_i)
return mu, tf.concat(neg_mus_st, axis=-1), tf.concat(neg_mus_ts, axis=-1), \
s_embed, t_embed, neg_embed_s_list, neg_embed_t_list
def construct_mutual_influence(self, node_embed, node_nbr_infos, target_embed, neg_embed):
# construct_mutual_influence(s_embed, s_nbr_infos, t_embed, neg_embed_t_list)
with tf.compat.v1.variable_scope('multual_influence', reuse=tf.compat.v1.AUTO_REUSE):
pos_info = []
neg_info = []
att_info = []
mask = []
for i in range(self.num_edge_types):
nbr_ids, nbr_masks, nbr_weights, nbr_flag = node_nbr_infos[i]
pos_g, neg_g, hete_att = tf.cond(
tf.reduce_all(nbr_flag > 0),
false_fn=lambda: [tf.zeros(shape=[self.batch_size, 1], dtype=tf.float32),
tf.zeros(shape=[self.batch_size, self.neg_size], dtype=tf.float32),
tf.zeros(shape=[self.batch_size, 1], dtype=tf.float32)],
true_fn=lambda: self.edge_type_distance(node_embed, nbr_ids, i, nbr_weights,
nbr_masks, target_embed, neg_embed)
)
pos_info.append(pos_g)
neg_info.append(neg_g)
att_info.append(hete_att)
mask.append(nbr_flag)
mask = tf.cast(tf.reshape(tf.concat(mask, axis=-1), [self.batch_size, self.num_edge_types]), tf.bool)
att_info = tf.concat(att_info, axis=-1)
padding = tf.fill([self.batch_size, self.num_edge_types], -2 ** 32 + 1.0)
padding2 = tf.fill([self.batch_size, self.num_edge_types], 0.0)
att_v1 = tf.nn.softmax(tf.where(mask, att_info, padding), axis=1)
norm_att = tf.nn.softmax(tf.where(mask, att_v1, padding2), axis=1)
pos_info = tf.concat(pos_info, axis=-1)
neg_info = tf.reshape(tf.concat(neg_info, axis=-1), [self.batch_size, self.num_edge_types, self.neg_size])
neg_info = tf.transpose(neg_info, [0, 2, 1])
pos_loss = tf.reduce_sum(tf.multiply(norm_att, pos_info), axis=-1)
neg_loss = tf.reduce_sum(tf.matmul(neg_info, tf.expand_dims(norm_att, axis=2)), axis=-1)
return pos_loss, neg_loss
def edge_type_distance(self, node_embed, ids, e_type, weight, mask, target_embed, neg_embed):
# edge_type_distance(node_embed, nbr_ids, i, nbr_weights, nbr_masks, target_embed, neg_embed)
with tf.compat.v1.variable_scope('multual_influence_{}'.format(e_type), reuse=tf.compat.v1.AUTO_REUSE):
nbr_embed = tf.compat.v1.layers.dense(tf.gather(self.embedding, ids), self.node_dim, reuse=tf.compat.v1.AUTO_REUSE,
kernel_regularizer=self.l2_regular, name="edge_type_{}".format(e_type))
# ^ nbr_embed shape [batch_size,max_nbr,node_dim]
edge_embed = tf.reshape(tf.gather(self.edge_type_embed, [e_type]), (1, 1, self.node_dim))
# dimension - 1x1xd
node_embed = tf.expand_dims(node_embed, axis=1)
nbr_distance = self.g_func(node_embed + edge_embed, nbr_embed, opt='l2')
paddings = tf.fill(tf.shape(nbr_distance), tf.constant(-2 ** 32 + 1, dtype=tf.float32))
paddings2 = tf.fill(tf.shape(nbr_distance), tf.constant(0, dtype=tf.float32))
nbr_distance2 = tf.where(tf.cast(mask, dtype=tf.bool), nbr_distance, paddings)
atts = tf.nn.softmax(nbr_distance2, axis=-1)
atts_2 = tf.where(tf.cast(mask, dtype=tf.bool), atts, paddings2)
weight = tf.cast(weight, dtype = tf.float32)
new_weight = tf.multiply(atts_2, weight)
mutual_subs = self.g_func(nbr_embed, tf.expand_dims(target_embed, axis=1), 'l2')
mutual_neg_subs = [self.g_func(nbr_embed, tf.expand_dims(neg_embed[i], axis=1), 'l2') for i in
range(self.neg_size)]
avg_embed = tf.reduce_sum(tf.matmul(tf.expand_dims(atts_2, 1), nbr_embed), axis=1)
avg_weight_1 = tf.reduce_sum(weight, axis=1)
nbr_numbers = tf.clip_by_value(tf.cast(tf.reduce_sum(mask, axis=1), tf.float32), 1.0, self.nbr_size)
ave_weight = tf.reshape(avg_weight_1 / nbr_numbers, [-1, 1])
avg_info = tf.multiply(ave_weight, avg_embed)
hete_att = tf.compat.v1.layers.dense(avg_info, 1, tf.nn.leaky_relu, kernel_regularizer=self.l2_regular,
reuse=tf.compat.v1.AUTO_REUSE, name='hete_att_{}'.format(e_type))
pos_mutual_influ = tf.reduce_sum(tf.multiply(new_weight, mutual_subs), axis=-1)
neg_mutual_influ = [
tf.reshape(tf.reduce_sum(tf.multiply(new_weight, mutual_neg_subs[i]), axis=-1), [self.batch_size, 1])
for i in
range(self.neg_size)]
return [tf.reshape(pos_mutual_influ, [self.batch_size, 1]), \
tf.concat(tf.reshape(neg_mutual_influ, [self.batch_size, self.neg_size]), axis=1), \
tf.reshape(hete_att, [self.batch_size, 1])]
def g_func(self, x, y, opt='l2'):
if opt == 'l2':
return -tf.reduce_sum((x - y) ** 2, axis=-1)
elif opt == 'l1':
return -tf.reduce_sum(tf.abs(x - y), axis=-1)
else:
return -tf.reduce_sum((x - y) ** 2, axis=-1)
def init_saver(self):
var_list = tf.trainable_variables()
g_list = tf.global_variables()
bn_moving_vars = [g for g in g_list if 'moving_mean' in g.name]
bn_moving_vars += [g for g in g_list if 'moving_variance' in g.name]
bn_moving_vars += [g for g in g_list if 'global_step' in g.name]
self.saver = tf.train.Saver(var_list=var_list + bn_moving_vars, max_to_keep=1)