-
Notifications
You must be signed in to change notification settings - Fork 38
/
contextnet.py
117 lines (96 loc) · 4.47 KB
/
contextnet.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
"""
论文:ContextNet: A Click-Through Rate Prediction Framework Using Contextual information to Refine Feature Embedding
地址:https://arxiv.org/pdf/2107.1202
"""
import tensorflow as tf
from typing import List, Union
from functools import partial
from ..utils.core import dnn_layer
class ContextNet:
def __init__(self,
num_block: int,
agg_dim: int,
ffn_type: str,
embedding_ln: bool = True,
l2_reg: float = 0.,
dropout: float = 0.
):
"""
:param num_block: ContextNet Block的层数
:param agg_dim: Contextual Embedding中的Aggregation模块的输出维度
:param ffn_type: ContextNet Block使用Point-wise FFN或者Single-layer FFN
:param l2_reg:
:param dropout:
"""
self.num_block = num_block
self.agg_dim = agg_dim
self.embedding_ln = embedding_ln
self.l2_reg = l2_reg
self.dnn_layer = partial(dnn_layer, use_bias=False, dropout=dropout, use_bn=False, l2_reg=l2_reg)
if ffn_type == 'PFFN':
self.ffn_func = self.point_wise
elif ffn_type == 'FFN':
self.ffn_func = self.single_layer
else:
raise TypeError('ffn_type only support: "PFFN" or "FFN"')
def __call__(self,
inputs: Union[List[tf.Tensor], tf.Tensor],
is_training: bool = True):
"""
:param inputs: [bs, num_feature, dim] or list of [bs, dim]
:param is_training:
:return:
"""
if isinstance(inputs, list):
inputs = tf.stack(inputs, axis=1)
assert len(inputs.shape) == 3
if self.embedding_ln:
inputs = tf.contrib.layers.layer_norm(inputs=inputs,
begin_norm_axis=-1,
begin_params_axis=-1)
# stack ContextNet block
output = inputs
for _ in range(self.num_block):
output = self.contextnet_block(output, is_training)
# flatten
output = tf.layers.flatten(output)
output = tf.layers.dense(output, 1, activation=tf.nn.sigmoid,
kernel_regularizer=tf.contrib.layers.l2_regularizer(self.l2_reg),
kernel_initializer=tf.glorot_normal_initializer())
return tf.reshape(output, [-1])
def get_contextual_embedding(self, embeddings, is_training):
shape = embeddings.shape.as_list()
num_feature = shape[1]
dim = shape[2]
# 所有特征共享Aggregation参数
agg = self.dnn_layer(embeddings, self.agg_dim, activation=tf.nn.relu, is_training=is_training)
# Project参数不共享
agg = tf.reshape(agg, [-1, num_feature * self.agg_dim])
project = self.dnn_layer(agg, num_feature * dim, is_training=is_training)
return tf.reshape(project, [-1, num_feature, dim])
def contextnet_block(self, embeddings, is_training):
contextual_embedding = self.get_contextual_embedding(embeddings, is_training)
merge = contextual_embedding * embeddings
output = self.ffn_func(merge, is_training)
return output
def point_wise(self, inputs, is_training):
dim = inputs.shape.as_list()[-1]
# 整个网络参数共享
with tf.variable_scope('point-wise', reuse=tf.AUTO_REUSE):
ffn1 = self.dnn_layer(inputs, dim, activation=tf.nn.relu, scope='ffn1', is_training=is_training)
ffn2 = self.dnn_layer(ffn1, dim, scope='ffn2', is_training=is_training)
output = tf.contrib.layers.layer_norm(inputs=ffn2,
begin_norm_axis=-1,
begin_params_axis=-1,
scope='point-wise-ln')
return output + inputs
def single_layer(self, inputs, is_training):
dim = inputs.shape.as_list()[-1]
# 整个网络参数共享
with tf.variable_scope('single-layer', reuse=tf.AUTO_REUSE):
ffn = self.dnn_layer(inputs, dim, scope='ffn', is_training=is_training)
output = tf.contrib.layers.layer_norm(inputs=ffn,
begin_norm_axis=-1,
begin_params_axis=-1,
scope='single-layer-ln')
return output