-
Notifications
You must be signed in to change notification settings - Fork 11
/
models.py
137 lines (111 loc) · 5.97 KB
/
models.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
from layers import *
import tensorflow as tf
import numpy as np
flags = tf.app.flags
FLAGS = flags.FLAGS
class Model(object):
def __init__(self, **kwargs):
allowed_kwargs = {'name', 'logging'}
for kwarg in kwargs.keys():
assert kwarg in allowed_kwargs, 'Invalid keyword argument: ' + kwarg
name = kwargs.get('name')
if not name:
name = self.__class__.__name__.lower()
self.name = name
logging = kwargs.get('logging', False)
self.logging = logging
self.vars = {}
self.placeholders = {}
self.layers = []
self.activations_state = []
self.activations_influence = []
self.n_nodes = None
self.inputs_state = None
self.inputs_influence = None
self.outputs = None
self.FLAGS = None
self.loss = 0
self.opt_op = None
def _build(self):
raise NotImplementedError
def build(self):
""" Wrapper for _build() """
with tf.variable_scope(self.name):
self._build()
# Build sequential layer model, make the output prediction
self.activations_state.append(self.inputs_state)
self.activations_influence.append(self.inputs_influence)
for layer in self.layers:
hidden_state,hidden_influence = layer(self.activations_state[-1],self.activations_influence[-1])
self.activations_state.append(hidden_state)
self.activations_influence.append(hidden_influence)
self.outputs = tf.nn.tanh(self.activations_state[-1])
self.outputs = tf.multiply(self.outputs, 1 - self.placeholders["Xs"]) + self.placeholders["Xs"]
# Store model variables for easy access
variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=self.name)
self.vars = {var.name: var for var in variables}
# Build metrics
self._loss()
var_list1 = [var for var in tf.trainable_variables() if not 'graph_' in var.name]
var_list2 = [var for var in tf.trainable_variables() if 'graph_' in var.name]
opt1 = tf.train.AdamOptimizer(learning_rate=self.FLAGS.learning_rate)
opt2 = tf.train.AdamOptimizer(learning_rate=self.FLAGS.graph_learning_rate)
grads = tf.gradients(self.loss, var_list1 + var_list2)
grads1 = [tf.clip_by_norm(grad, self.FLAGS.max_grad_norm) for grad in grads[:len(var_list1)]]
grads2 = [tf.clip_by_norm(grad, self.FLAGS.max_grad_norm) for grad in grads[len(var_list1):]]
train_op1 = opt1.apply_gradients(zip(grads1, var_list1))
train_op2 = opt2.apply_gradients(zip(grads2, var_list2))
self.opt_op = tf.group(train_op1, train_op2)
def _loss(self):
raise NotImplementedError
class CoupledGNN(Model):
def __init__(self, FLAGS,init_values,placeholders,node_input_features, n_nodes, **kwargs):
super(CoupledGNN, self).__init__(**kwargs)
self.FLAGS = FLAGS
self.n_nodes = n_nodes
self.placeholders = placeholders
self.n_layers = FLAGS.n_layers
self.influence_dim = len(node_input_features[0])
self.values = tf.convert_to_tensor(init_values, dtype=tf.float32)
self.indices = self.placeholders['support_indices']
#initialize self activation parameters
with tf.variable_scope(self.name):
self.initializer_layer = tf.random_uniform_initializer(minval=0.0, maxval=0.01, dtype=tf.float32)
self.self_activation = tf.get_variable(name='graph_self_activation',shape=[self.n_nodes, 1],initializer=self.initializer_layer)
#get input influence of each user
self.input_features = tf.convert_to_tensor(node_input_features, dtype=tf.float32)
self.inputs_influence = tf.contrib.layers.instance_norm(tf.transpose(tf.tile(
tf.reshape(self.input_features,[self.n_nodes, self.influence_dim,1]),
multiples=[1,1,self.FLAGS.batch_size]),perm=[2,0,1])
,data_format="NHWC")
#get input state of each user
inputs_state = self.placeholders['Xs']+ tf.tile(
tf.reshape(self.self_activation,[1,self.n_nodes,1]),
multiples=[self.FLAGS.batch_size,1,1])
self.inputs_state = tf.multiply(inputs_state, 1 - self.placeholders["Xs"]) + self.placeholders["Xs"]
self.build()
def _loss(self):
# regularization of l2
for var in tf.trainable_variables():
self.loss += self.FLAGS.reg_l2 * tf.nn.l2_loss(var)
#mean relative square loss
self.popularity_pre = tf.reduce_sum(self.outputs,axis=[1,2])
self.popularity_true = tf.reduce_sum(self.placeholders['y'],axis=[1])
self.error = tf.reduce_mean(tf.square((self.popularity_pre-self.popularity_true)/self.popularity_true))
#regularization of cross entropy
self.activation_pre = tf.minimum(tf.maximum(tf.reduce_sum(self.outputs,axis=2),1e-3),1-(1e-3))
self.cross_entropy = tf.reduce_mean(
-tf.multiply(self.placeholders['y'],tf.log(self.activation_pre))\
-tf.multiply(1-self.placeholders['y'],tf.log(1-self.activation_pre)))
#the total loss to be optimized
self.loss += self.error +self.FLAGS.reg_cross_entropy*self.cross_entropy
def _build(self):
for i in range(self.n_layers):
self.layers.append(GraphConvolution(influence_dim=self.influence_dim,
flags=self.FLAGS,
n_nodes = self.n_nodes,
placeholders=self.placeholders,
L_values = self.values,
L_indices = self.indices,
self_activation = self.self_activation,
dropout=True))