-
Notifications
You must be signed in to change notification settings - Fork 13
/
model.py
executable file
·171 lines (149 loc) · 7.1 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
# coding: utf-8
from __future__ import print_function
import tensorflow as tf
import numpy as np
import time
import os
def pick_top_n(preds, vocab_size, top_n=5):
p = np.squeeze(preds)
# 将除了top_n个预测值的位置都置为0
p[np.argsort(p)[:-top_n]] = 0
# 归一化概率
p = p / np.sum(p)
# 随机选取一个字符
c = np.random.choice(vocab_size, 1, p=p)[0]
return c
class CharRNN:
def __init__(self, num_classes, num_seqs=64, num_steps=50,
lstm_size=128, num_layers=2, learning_rate=0.001,
grad_clip=5, sampling=False, train_keep_prob=0.5, use_embedding=False, embedding_size=128):
if sampling is True:
num_seqs, num_steps = 1, 1
else:
num_seqs, num_steps = num_seqs, num_steps
self.num_classes = num_classes
self.num_seqs = num_seqs
self.num_steps = num_steps
self.lstm_size = lstm_size
self.num_layers = num_layers
self.learning_rate = learning_rate
self.grad_clip = grad_clip
self.train_keep_prob = train_keep_prob
self.use_embedding = use_embedding
self.embedding_size = embedding_size
tf.compat.v1.reset_default_graph()
self.build_inputs()
self.build_lstm()
self.build_loss()
self.build_optimizer()
self.saver = tf.compat.v1.train.Saver()
def build_inputs(self):
with tf.name_scope('inputs'):
self.inputs = tf.compat.v1.placeholder(tf.int32, shape=(
self.num_seqs, self.num_steps), name='inputs')
self.targets = tf.compat.v1.placeholder(tf.int32, shape=(
self.num_seqs, self.num_steps), name='targets')
self.keep_prob = tf.compat.v1.placeholder(tf.float32, name='keep_prob')
# 对于中文，需要使用embedding层
# 英文字母没有必要用embedding层
if self.use_embedding is False:
self.lstm_inputs = tf.one_hot(self.inputs, self.num_classes)
else:
with tf.device("/cpu:0"):
embedding = tf.compat.v1.get_variable('embedding', [self.num_classes, self.embedding_size])
self.lstm_inputs = tf.nn.embedding_lookup(embedding, self.inputs)
def build_lstm(self):
# 创建单个cell并堆叠多层
def get_a_cell(lstm_size, keep_prob):
lstm = tf.compat.v1.nn.rnn_cell.BasicLSTMCell(lstm_size)
drop = tf.compat.v1.nn.rnn_cell.DropoutWrapper(lstm, output_keep_prob=keep_prob)
return drop
with tf.name_scope('lstm'):
cell = tf.compat.v1.nn.rnn_cell.MultiRNNCell(
[get_a_cell(self.lstm_size, self.keep_prob) for _ in range(self.num_layers)]
)
self.initial_state = cell.zero_state(self.num_seqs, tf.float32)
# 通过dynamic_rnn对cell展开时间维度
self.lstm_outputs, self.final_state = tf.compat.v1.nn.dynamic_rnn(cell, self.lstm_inputs, initial_state=self.initial_state)
# 通过lstm_outputs得到概率
seq_output = tf.concat(self.lstm_outputs, 1)
x = tf.reshape(seq_output, [-1, self.lstm_size])
with tf.compat.v1.variable_scope('softmax'):
softmax_w = tf.Variable(tf.compat.v1.truncated_normal([self.lstm_size, self.num_classes], stddev=0.1))
softmax_b = tf.Variable(tf.zeros(self.num_classes))
self.logits = tf.matmul(x, softmax_w) + softmax_b
self.proba_prediction = tf.nn.softmax(self.logits, name='predictions')
def build_loss(self):
with tf.name_scope('loss'):
y_one_hot = tf.one_hot(self.targets, self.num_classes)
y_reshaped = tf.reshape(y_one_hot, self.logits.get_shape())
loss = tf.nn.softmax_cross_entropy_with_logits(logits=self.logits, labels=y_reshaped)
self.loss = tf.reduce_mean(loss)
def build_optimizer(self):
# 使用clipping gradients
tvars = tf.compat.v1.trainable_variables()
grads, _ = tf.clip_by_global_norm(tf.gradients(self.loss, tvars), self.grad_clip)
train_op = tf.compat.v1.train.AdamOptimizer(self.learning_rate)
self.optimizer = train_op.apply_gradients(zip(grads, tvars))
def train(self, batch_generator, max_steps, save_path, save_every_n, log_every_n):
self.session = tf.compat.v1.Session()
with self.session as sess:
sess.run(tf.compat.v1.global_variables_initializer())
# Train network
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.targets: y,
self.keep_prob: self.train_keep_prob,
self.initial_state: new_state}
batch_loss, new_state, _ = sess.run([self.loss,
self.final_state,
self.optimizer],
feed_dict=feed)
end = time.time()
# control the print lines
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'), global_step=step)
if step >= max_steps:
break
self.saver.save(sess, os.path.join(save_path, 'model'), global_step=step)
def sample(self, n_samples, prime, vocab_size):
samples = [c for c in prime]
sess = self.session
new_state = sess.run(self.initial_state)
preds = np.ones((vocab_size, )) # for prime=[]
for c in prime:
x = np.zeros((1, 1))
# 输入单个字符
x[0, 0] = c
feed = {self.inputs: x,
self.keep_prob: 1.,
self.initial_state: new_state}
preds, new_state = sess.run([self.proba_prediction, self.final_state],
feed_dict=feed)
c = pick_top_n(preds, vocab_size)
# 添加字符到samples中
samples.append(c)
# 不断生成字符，直到达到指定数目
for i in range(n_samples):
x = np.zeros((1, 1))
x[0, 0] = c
feed = {self.inputs: x,
self.keep_prob: 1.,
self.initial_state: new_state}
preds, new_state = sess.run([self.proba_prediction, self.final_state],
feed_dict=feed)
c = pick_top_n(preds, vocab_size)
samples.append(c)
return np.array(samples)
def load(self, checkpoint):
self.session = tf.compat.v1.Session()
self.saver.restore(self.session, checkpoint)
print('Restored from: {}'.format(checkpoint))