-
Notifications
You must be signed in to change notification settings - Fork 7
/
model.py
185 lines (159 loc) · 7.32 KB
/
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
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
#! /usr/vin/env python3
# coding=utf-8
# author: tudou
import tensorflow as tf
import numpy as np
import time
import os
def get_n_top(probs,vocab,n=5):
p=np.squeeze(probs)
#### 1. 只选取概率最高的n个词,所以需要将其他位置的概率置为0
#### 2. 并重新计算n个词的取值概率,即归一化
# 置0
p[np.argsort(p)[:-n]]=0
# 归一化
p=p/np.sum(p)
c=np.random.choice(vocab, 1 , p=p)
return c
class Char_RNN(object):
def __init__(self,num_classes,num_seqs=32,num_steps=50,lstm_size=128,num_layers=2,
use_embedding=False,embedding_size=64,
is_train=False,learning_rate=0.001,sampling=False):
if sampling is True:
self.num_seqs,self.num_steps=1,1
else:
self.num_seqs = num_seqs
self.num_steps = num_steps
self.num_classes = num_classes
self.lstm_size = lstm_size
self.num_layers = num_layers
self.use_embeding = use_embedding
self.embedding_size = embedding_size
self.is_train = is_train
self.learning_rate = learning_rate
if self.is_train is True:
self.keep_probs = 0.5
else:
self.keep_probs = 1.0
tf.reset_default_graph()
self.build_inputs()
self.build_lstm()
self.build_loss()
self.build_optimizer()
self.saver=tf.train.Saver()
def build_inputs(self):
with tf.variable_scope('inputs'):
#### 输入要转化为one_hot向量,要用int而不是float
self.inputs=tf.placeholder(tf.int32,shape=(self.num_seqs,self.num_steps))
self.labels=tf.placeholder(tf.int32,shape=(self.num_seqs,self.num_steps))
self.keep_prob=tf.placeholder(tf.float32,shape=None)
if self.use_embeding is True:
#### tf embedding层
### 1. 要在 TensorFlow 中创建 embeddings,我们首先将文本拆分成单词,然后为词汇表中的每个单词分配一个整数
### 2. 利用整数组成的向量训练embedding层,输出shape为vocabulary_size,embedding_size
### 3. 利用embedding_lookup(embedding层,整数向量)获得inputs的分布式表示
embedding=tf.get_variable('embeddings',shape=(self.num_classes,self.embedding_size))
self.lstm_inputs=tf.nn.embedding_lookup(embedding,self.inputs)
else:
self.lstm_inputs=tf.one_hot(self.inputs,self.num_classes)
def build_lstm(self):
def _get_cell():
cell=tf.nn.rnn_cell.BasicLSTMCell(self.lstm_size)
drop=tf.nn.rnn_cell.DropoutWrapper(cell,self.keep_prob)
return drop
with tf.variable_scope('lstm'):
multi_cell=tf.nn.rnn_cell.MultiRNNCell([_get_cell() for _ in range(self.num_layers)])
self.initial_state=multi_cell.zero_state(self.num_seqs,tf.float32)
outputs,self.final_state=tf.nn.dynamic_rnn(multi_cell,self.lstm_inputs,
initial_state=self.initial_state)
#### dynamic_rnn输出为(h1,h2,h3,...),所以将它们拼接成一个矩阵
self.outputs=tf.concat(outputs,1)
x=tf.reshape(self.outputs,[-1,self.lstm_size])
with tf.variable_scope("softmax"):
w=tf.get_variable('softmax_w',shape=[self.lstm_size,self.num_classes],dtype=tf.float32,
initializer=tf.truncated_normal_initializer)
b=tf.get_variable('softmax_b',self.num_classes,dtype=tf.float32)
tf.summary.histogram('softmax_w',w)
tf.summary.histogram('softmax_b',b)
self.logits=tf.matmul(x,w)+b
self.prob=tf.nn.softmax(self.logits)
tf.summary.histogram('prob',self.prob)
def build_loss(self):
with tf.variable_scope('loss'):
#### softmax标签要处理
one_hot_labels=tf.one_hot(self.labels,self.num_classes)
labels=tf.reshape(one_hot_labels,self.logits.get_shape())
loss=tf.nn.softmax_cross_entropy_with_logits(labels=labels,logits=self.logits)
self.loss=tf.reduce_mean(loss)
tf.summary.scalar('loss',self.loss)
def build_optimizer(self):
#### rnn容易导致梯度消失,所以需要进行梯度裁剪
optimizer=tf.train.AdamOptimizer(self.learning_rate)
gradient_vars=optimizer.compute_gradients(self.loss)
cropped_gradient=[(tf.clip_by_value(grad,-1,1),var) for grad,var in gradient_vars]
self.train_op=optimizer.apply_gradients(cropped_gradient)
def train(self,batch_generator,max_steps,save_path,log_every_n,save_every_n):
#### 标准流程
self.sess=tf.Session()
mergerd=tf.summary.merge_all()
init=(tf.global_variables_initializer(),tf.local_variables_initializer())
with self.sess as sess:
trian_writer = tf.summary.FileWriter(save_path + '/logdir', sess.graph)
sess.run(init)
step=0
new_state=sess.run(self.initial_state)
for x,y in batch_generator:
step+=1
start = time.time()
feed={
self.inputs:x,
self.labels:y,
self.keep_prob:self.keep_probs,
self.initial_state:new_state
}
####
batch_loss,new_state,_,summ=sess.run([self.loss,self.final_state,self.train_op,mergerd],
feed_dict=feed)
trian_writer.add_summary(summ,step)
end=time.time()
if step%log_every_n==0:
print('step: {}/{}... '.format(step, max_steps),
'loss: {:.4f}... '.format(batch_loss),
'{:.4f} sec/batch'.format((end - start)))
if step%save_every_n==0:
self.saver.save(sess,os.path.join(save_path,'model'),step)
if step>max_steps:
break
self.saver.save(sess,os.path.join(save_path,'model'),step)
def sample(self,n_samples,prime,vacab_size):
samples=[c for c in prime]
sess=self.sess
new_state=sess.run(self.initial_state)
#### 初始化概率
preds=np.ones((vacab_size,))
for c in prime:
#### 输入单个字符
x=np.zeros((1,1))
x[0,0]=c
feed={self.inputs:x,
self.keep_prob:self.keep_probs,
self.initial_state:new_state
}
preds,new_state=sess.run([self.prob,self.final_state],feed_dict=feed)
c=get_n_top(preds,vocab=vacab_size)
samples.append(c)
for i in range(n_samples):
x = np.zeros((1, 1))
x[0, 0] = c
feed = {self.inputs: x,
self.keep_prob: self.keep_probs,
self.initial_state: new_state
}
preds, new_state = sess.run([self.prob, self.final_state], feed_dict=feed)
c = get_n_top(preds, vocab=vacab_size)
samples.append(c)
return np.array(samples)
def load(self,checkpoint):
self.sess=tf.Session()
self.saver.restore(self.sess,checkpoint)
print("restore from {}".format(checkpoint))