-
Notifications
You must be signed in to change notification settings - Fork 4
/
rnn_model.py
159 lines (130 loc) · 5.08 KB
/
rnn_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
#!/bin/env python
#-*- encoding: utf-8 -*-
import os
import time
import numpy as np
import tensorflow as tf
from utils import model_helper
import config
class ModelParas(object):
embedding_size = config.embedding_size
sequence_length = None
cell_num_units = 512
num_layers = 1
batch_size = 64
dropout = 0.0
learning_rate = 0.01
decay = 0.99
lrshrink = 5
uniform_init_scale = 0.04
clip_gradient_norm = 5.0
epochs = 20
class Model(object):
def __init__(self, paras, sess, mode, emb_matrix):
self.paras = paras
self.sess = sess
self.mode = mode
# Model variable
with tf.device('/cpu:0'):
self.embeddings = tf.get_variable(
name = 'embeddings',
shape = emb_matrix.shape,
dtype = tf.float32,
initializer = tf.constant_initializer(emb_matrix))
self.global_step = tf.get_variable(
name = 'global_step',
dtype = tf.int32,
initializer = 1,
trainable = False)
self._build_graph()
def _create_placeholder(self):
self.lr = tf.placeholder(tf.float32, [], name = 'learning_rate')
self.sents = tf.placeholder(tf.int32, [None, None], name = 'sents')
with tf.device('/cpu:0'):
self.emb_sents = tf.nn.embedding_lookup(
self.embeddings, self.sents)
self.sent_lengths = tf.placeholder(tf.int32, [None], name = 'sent_lengths')
self.labels = tf.placeholder(tf.int32, [None], name = 'labels')
def _inference(self):
with tf.variable_scope('encoder') as varscope:
cells_fw = model_helper.create_rnn_cell(
'lstm',
self.paras.cell_num_units,
self.paras.num_layers,
self.paras.dropout,
self.mode)
cells_bw = model_helper.create_rnn_cell(
'lstm',
self.paras.cell_num_units,
self.paras.num_layers,
self.paras.dropout,
self.mode)
outputs, output_states = tf.nn.bidirectional_dynamic_rnn(
cells_fw,
cells_bw,
inputs = self.emb_sents,
sequence_length = self.sent_lengths,
dtype = tf.float32,
scope = varscope)
# states_fw: (batch_size, sent_len, cell_size)
states_fw, states_bw = outputs
concat_states = tf.concat([states_fw, states_bw], axis = 2)
# sent_states: (batch_size, 2 * cell_size)
self.sent_states = tf.reduce_max(concat_states, axis = 1)
with tf.variable_scope('classify_layer') as varscope:
hidden1 = tf.contrib.layers.fully_connected(
inputs = self.sent_states,
num_outputs = 512)
hidden2 = tf.contrib.layers.fully_connected(
inputs = hidden1,
num_outputs = 5)
self.predicts = tf.reduce_max(tf.contrib.layers.fully_connected(
inputs = hidden2,
activation_fn = None,
num_outputs = 1), axis = 1)
self.mse = tf.reduce_mean(tf.cast(
tf.squared_difference(
self.labels,
tf.cast(tf.round(self.predicts), tf.int32)),
tf.float32))
with tf.variable_scope('accuracy') as varscope:
correct_prediction = tf.equal(self.labels,
tf.cast(tf.round(self.predicts), tf.int32))
self.accuracy = tf.reduce_mean(tf.cast(
correct_prediction, tf.float32))
def _create_loss(self):
with tf.variable_scope('loss') as varscope:
self.loss = tf.reduce_mean(
tf.losses.mean_squared_error(
labels = tf.cast(self.labels, tf.float32),
predictions = self.predicts))
def _create_optimizer(self):
self.optimizer = tf.contrib.layers.optimize_loss(
loss = self.loss,
global_step = self.global_step,
learning_rate = self.lr,
optimizer = 'SGD',
clip_gradients = self.paras.clip_gradient_norm)
def _create_summary(self):
log_path = os.path.join(config.model_path, 'tensorboard')
self.train_writer = tf.summary.FileWriter(
os.path.join(log_path, 'train'), self.sess.graph)
self.test_writer = tf.summary.FileWriter(
os.path.join(log_path, 'test'), self.sess.graph)
with tf.name_scope('summaries') as scope:
tf.summary.scalar('loss', self.loss)
tf.summary.scalar('accuracy', self.accuracy)
def _build_graph(self):
self._create_placeholder()
self._inference()
self._create_loss()
self._create_optimizer()
self._create_summary()
print 'Build graph done'
def test():
sess = tf.Session()
paras = ModelParas()
emb_matrix = NlpUtil.build_emb_matrix()
Model(paras, sess, tf.contrib.learn.ModeKeys.TRAIN)
if __name__ == '__main__':
pass