-
Notifications
You must be signed in to change notification settings - Fork 2
/
tan.py
69 lines (65 loc) · 3.74 KB
/
tan.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
# -*- coding: utf-8 -*-
# @Time : 2018/9/10 19:09
# @Author : Tianchiyue
# @File : tan.py
# @Software: PyCharm Community Edition
from sentiment_models.model import BaseModel
from sentiment_models.layers import ConnectAspectLayer, AtaeAttention, LocationAttentionLayer
from keras.layers import Bidirectional, LSTM, GRU, TimeDistributed, Dense, Flatten, Activation, \
Lambda, Input, Embedding, dot, concatenate,Dropout,add
import keras.backend as K
import tensorflow as tf
class TAN(BaseModel):
def build(self, embedding_matrix, aspect_embedding_matrix):
self.sentence_input = Input(shape=(self.config['max_length'],),
dtype='int32',
name='sentence_input')
self.aspect_input = Input((1,),
dtype='int32',
name='aspect_input')
word_embed = Embedding(embedding_matrix.shape[0],
embedding_matrix.shape[1],
trainable=self.config['embed_trainable'],
weights=[embedding_matrix],
mask_zero=True
)(self.sentence_input) # bsz, time_steps, emb_dims
aspect_emb = Embedding(aspect_embedding_matrix.shape[0],
aspect_embedding_matrix.shape[1],
trainable=self.config['aspect_emb_trainable'],
weights=[aspect_embedding_matrix])(self.aspect_input) # bsz, 1, emb_dims
target_augemented_embedding = ConnectAspectLayer()([word_embed, aspect_emb])
alpha = LocationAttentionLayer(name='tan')(target_augemented_embedding)
if self.config['bidirectional']:
if self.config['rnn'] == 'gru':
rnn_out = Bidirectional(GRU(self.config['rnn_output_size'],
return_sequences=True,
dropout=self.config['dropout_rate'],
recurrent_dropout=self.config['dropout_rate']))(word_embed)
else:
rnn_out = Bidirectional(LSTM(self.config['rnn_output_size'],
return_sequences=True,
dropout=self.config['dropout_rate'],
recurrent_dropout=self.config['dropout_rate']))(word_embed)
else:
if self.config['rnn'] == 'gru':
rnn_out = GRU(self.config['rnn_output_size'],
return_sequences=True,
dropout=self.config['dropout_rate'],
recurrent_dropout=self.config['dropout_rate'])(word_embed)
else:
rnn_out = LSTM(self.config['rnn_output_size'],
return_sequences=True,
dropout=self.config['dropout_rate'],
recurrent_dropout=self.config['dropout_rate'])(word_embed)
cr = dot([alpha, rnn_out], axes=1, name='attention_mul')
# attention_alpha = AtaeAttention(time_steps=self.config['max_length'])([rnn_out, aspect_emb])
# r = dot([attention_alpha, rnn_out], axes=[1,1], name='attention_mul')
# # r = Flatten()(r)
# r = Dropout(self.config['dropout_rate'])(r)
# r = Dense(self.config['rnn_output_size'],use_bias=False)(r)
# h = Lambda(lambda x: tf.slice(x, [0, self.config['max_length'] - 1, 0], [-1, 1, -1]))(rnn_out)
# h = Flatten()(h)
# h = Dense(self.config['rnn_output_size'], use_bias=False)(h)
# h_star = add([r,h])
# h_star = Activation('tanh')(h_star)
return cr