-
Notifications
You must be signed in to change notification settings - Fork 5
/
MultiInterestSenetTT.py
107 lines (80 loc) · 5.08 KB
/
MultiInterestSenetTT.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
#!/usr/bin/env python
# coding: utf-8
# author: stefan 2022-03-02
"""
多兴趣SENet双塔模型Multi-Interest-Senet-Two-Towers (MISTT)
基于SENet，把User侧和Item侧的Embedding，打成多兴趣的。就是说，比如在用户侧塔，可以配置不同的SENet模块及对应的DNN结构，来强化不同方面兴趣的
Embedding表达。Item侧也可以如此办理，或者Item侧如果信息比较单一，可以仍然只打出一个Item Embedding，只需要维度上能和User侧多兴趣Embedding对齐即可
"""
from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import Adam
from layers.tool_layers import *
from layers.model_layers import Tower, SENetLayer
# build model
def buildMISTT(temperature, city_dict, shangquan_dict, comm_dict, tower_num_layer, tower_num_layer_units):
# define input
...
user_feature_judge_1 = SENetLayer(last_shape=int(user_emb_feature_pooling.shape[-1]), reduction=16,
name='user_embedding_senet_1')(user_emb_feature_pooling)
user_feature_judge_2 = SENetLayer(last_shape=int(user_emb_feature_pooling.shape[-1]), reduction=16,
name='user_embedding_senet_2')(user_emb_feature_pooling)
user_embedding_senet_1 = []
user_embedding_senet_2 = []
for i in range(user_feature_judge_1.shape[1]): # 2个senet特征权重维度一样，取其中一个即可
x_1 = tf.slice(user_feature_judge_1, [0, i], [-1, 1]) # 取出senet1激活tensor中的第i个激活值
x_2 = tf.slice(user_feature_judge_2, [0, i], [-1, 1])
emb_with_jugde_1 = tf.multiply(user_emb_feature_list[i], x_1)
emb_with_jugde_2 = tf.multiply(user_emb_feature_list[i], x_2)
user_embedding_senet_1.append(emb_with_jugde_1)
user_embedding_senet_2.append(emb_with_jugde_2)
user_feature_1 = concatenate(user_embedding_senet_1, axis=1, name='user_feature_1')
user_feature_2 = concatenate(user_embedding_senet_2, axis=1, name='user_feature_2')
# 计算user tower的2个输出
user_tower_out_1 = Tower(layer_num=tower_num_layer, layer_units=tower_num_layer_units,
activation=tf.nn.leaky_relu, name='user_tower_1')(user_feature_1)
user_tower_out_2 = Tower(layer_num=tower_num_layer, layer_units=tower_num_layer_units,
activation=tf.nn.leaky_relu, name='user_tower_2')(user_feature_2)
user_tower_out = Add(name="user_tower")([user_tower_out_1, user_tower_out_2])
# item feature
...
item_embed_features_judge_1 = SENetLayer(last_shape=int(item_embed_features_pooling.shape[-1]), reduction=16,
name='item_embedding_senet_1')(item_embed_features_pooling)
item_embed_features_judge_2 = SENetLayer(last_shape=int(item_embed_features_pooling.shape[-1]), reduction=16,
name='item_embedding_senet_2')(item_embed_features_pooling)
item_embedding_senet_1 = []
item_embedding_senet_2 = []
for i in range(item_embed_features_judge_1.shape[1]):
x_1 = tf.slice(item_embed_features_judge_1, [0, i], [-1, 1]) # 取出senet激活tensor中的第i个激活值
x_2 = tf.slice(item_embed_features_judge_2, [0, i], [-1, 1])
emb_with_jugde_1 = tf.multiply(item_embedding_features_list[i], x_1)
emb_with_jugde_2 = tf.multiply(item_embedding_features_list[i], x_2)
item_embedding_senet_1.append(emb_with_jugde_1)
item_embedding_senet_2.append(emb_with_jugde_2)
item_feature_1 = concatenate(item_embedding_senet_1 + [item_prop_age, item_age_power2], axis=1,
name='item_feature_1')
item_feature_2 = concatenate(item_embedding_senet_2 + [item_prop_age, item_age_power2], axis=1,
name='item_feature_2')
# 计算2个item tower的输出
item_tower_out_1 = Tower(layer_num=tower_num_layer, layer_units=tower_num_layer_units,
activation=tf.nn.leaky_relu, name='item_tower_1')(item_feature_1)
item_tower_out_2 = Tower(layer_num=tower_num_layer, layer_units=tower_num_layer_units,
activation=tf.nn.leaky_relu, name='item_tower_2')(item_feature_2)
item_tower_out = Add(name="item_tower")([item_tower_out_1, item_tower_out_2])
# 计算内积，添加温度系数
inner_product = tf.reduce_sum(tf.multiply(user_tower_out, item_tower_out), axis=1, keepdims=True)
out_with_tempreture = tf.keras.activations.sigmoid(inner_product / temperature)
out = Reshape((1,))(out_with_tempreture)
user_input = [...]
item_input = [...]
user_output = user_tower_out
item_output = item_tower_out
user_model = Model(inputs=user_input, outputs=user_output)
item_model = Model(inputs=item_input, outputs=item_output)
all_model = Model(inputs=user_input + item_input, outputs=out) # 模型评估以out为评估对象
all_model.compile(
loss='binary_crossentropy',
optimizer=Adam(1e-3),
metrics=[tf.keras.metrics.AUC()]
)
all_model.summary()
return all_model, user_model, item_model