-
Notifications
You must be signed in to change notification settings - Fork 2
/
cwgan.py
231 lines (210 loc) · 9.92 KB
/
cwgan.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
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
# encoding: utf-8
import tensorflow as tf
import numpy as np
import os
import shutil
import logging
import dataset
from utils import Translator, seq2seq_onehot2label
'''
model size:
embedding matrix: num_symbols * embedding_size
generator:
output_project = embedding_size * num_symbols
cell: ??? relate to LSTM struct
discriminator:
seq2state:
cell: ??? may be should use the cell of generator
state2logit:
state_size * h1_size + h1_size + h1_size * h2_size + h2_size +h2_size * 1 + 1
'''
logging.basicConfig(level=logging.DEBUG,
format='%(asctime)s %(filename)s[line:%(lineno)d] %(levelname)s %(message)s',
datefmt='%a, %d %b %Y %H:%M:%S',
filename='./log_file/emb_wgan.log',
filemode='w')
output_path = "./ckpt"
res_path = "./res"
'''
network config
'''
batch_size = 12
embedding_size = 96
num_layers = 2
num_symbols = 20000
state_size = 256
buckets = [(10,10),(20,20),(40,40)]
to_restore = False
max_len = buckets[-1][1]
learning_rate_dis = 5e-3#0.01
learning_rate_gen = 5e-4#0.01
CLIP_RANGE =[-0.1,0.1]
CRITIC = 10
gen_critic = 1
max_epoch = 100000
keep_prob = tf.constant(1.0,tf.float32, name="keep_prob")
encoder_inputs = []
decoder_inputs = []
target_weights = []
def generator(encoder_inputs,decoder_inputs,target_weights,bucket_id,seq_len):
def seq2seq_f(encoder,decoder):
cell = tf.contrib.rnn.BasicLSTMCell(embedding_size)
if num_layers > 1:
cell = tf.contrib.rnn.MultiRNNCell([cell] * num_layers)
w = tf.get_variable("proj_w", [embedding_size, num_symbols])
b = tf.get_variable("proj_b", [num_symbols])
output_projection = (w, b)
outputs, state = tf.contrib.legacy_seq2seq.embedding_attention_seq2seq(encoder,
decoder,cell,num_symbols,num_symbols,embedding_size,output_projection=output_projection,
feed_previous = True)
trans_output = []
for output in outputs:
trans_output.append(tf.matmul(output,w) + b)
return trans_output, state
targets = decoder_inputs
outputs, losses = tf.contrib.legacy_seq2seq.model_with_buckets(
encoder_inputs, decoder_inputs, targets,
target_weights, buckets, seq2seq_f,
softmax_loss_function=None,
per_example_loss=False, name='model_with_buckets')
patch = tf.convert_to_tensor([[0.0]*num_symbols] * batch_size)
def f0():
for _ in range(0,max_len-buckets[0][1]):
outputs[0].append(patch)
return tf.convert_to_tensor(outputs[0],dtype = tf.float32)
def f1():
for _ in range(0,max_len-buckets[1][1]):
outputs[1].append(patch)
return tf.convert_to_tensor(outputs[1],dtype = tf.float32)
def f2():
for _ in range(0,max_len-buckets[2][1]):
outputs[2].append(patch)
return tf.convert_to_tensor(outputs[2],dtype = tf.float32)
r = tf.case({tf.equal(bucket_id, 0): f0,
tf.equal(bucket_id, 1): f1},
default=f2, exclusive=True)
return tf.nn.softmax(tf.reshape(r,[max_len,batch_size,num_symbols]))
def discriminator(seq_raw, keep_prob,seq_len,reuse=False):
with tf.variable_scope('discriminator') as scope:
if reuse:
scope.reuse_variables()
cell = tf.contrib.rnn.BasicLSTMCell(state_size)
emb_matrix = tf.get_variable(name='emb_matrix',
shape=[num_symbols, embedding_size],
dtype=tf.float32,
initializer=tf.truncated_normal_initializer(stddev=0.1))
emb_ans = tf.reduce_mean(tf.multiply(
tf.reshape(seq_raw,[max_len, batch_size, num_symbols, 1]), emb_matrix), axis=2)
print(emb_ans)
if num_layers > 1:
cell = tf.contrib.rnn.MultiRNNCell([cell] * num_layers)
_,state = tf.nn.dynamic_rnn(cell, emb_ans,sequence_length=seq_len, initial_state=None, dtype=tf.float32,
time_major=True)
tmp_state = tf.convert_to_tensor(state[-1]) # 2*batch_size*emb_size
h_state = tf.slice(tmp_state, [1, 0, 0], [1, batch_size, state_size])
state = tf.reshape(h_state,[batch_size,-1])
h1_size = 32
w1 = tf.get_variable("w1", [state_size, h1_size], initializer=tf.truncated_normal_initializer(stddev=0.1))
b1 = tf.get_variable("b1", h1_size, initializer=tf.constant_initializer(0.0))
h1 = tf.nn.dropout(tf.nn.relu(tf.matmul(state, w1) + b1), keep_prob)
w3 = tf.get_variable("w3", [h1_size, 1], initializer=tf.truncated_normal_initializer())
b3 = tf.get_variable("b3", [1], initializer=tf.constant_initializer(0.0))
h3 = tf.matmul(h1, w3) + b3
return h3
def train():
for l in xrange(buckets[-1][0]):
encoder_inputs.append(tf.placeholder(tf.int32, shape=[batch_size],
name="encoder{0}".format(l)))
for l in xrange(buckets[-1][1]):
decoder_inputs.append(tf.placeholder(tf.int32, shape=[batch_size],
name="decoder{0}".format(l)))
target_weights.append(tf.placeholder(tf.float32, shape=[batch_size],
name="weight{0}".format(l)))
global_step = tf.Variable(0, name="global_step", trainable=False)
true_ans = tf.placeholder(tf.int32, [max_len ,batch_size], name = "true_ans")
seq_len = tf.placeholder(tf.int32, name="seq_len")
bucket_id = tf.placeholder(tf.int32, name="bucket_id")
#[seq_len * batch_size]
with tf.variable_scope('generator'):
fake_ans = generator(encoder_inputs,decoder_inputs,target_weights,
bucket_id,seq_len)
# 创建判别模型
true_ans_one_hot = tf.one_hot(true_ans,num_symbols,on_value=1.0,off_value=0.0,axis=-1)
y_data = discriminator(true_ans_one_hot,keep_prob=keep_prob,seq_len=seq_len)
y_generated = discriminator(fake_ans,keep_prob=keep_prob,seq_len=seq_len, reuse=True)
# 损失函数的设置
d_loss_real = tf.reduce_mean(y_data)
d_loss_fake = tf.reduce_mean(y_generated)
d_loss = d_loss_fake - d_loss_real
g_loss = tf.reduce_mean(-y_generated)
optimizer_dis = tf.train.RMSPropOptimizer(learning_rate_dis,name='RMSProp_dis')
optimizer_gen = tf.train.RMSPropOptimizer(learning_rate_gen, name='RMSProp_gen')
d_params = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,scope = "discriminator")
g_params = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,scope = "generator")
d_trainer = optimizer_dis.minimize(d_loss, var_list=d_params)
g_trainer = optimizer_gen.minimize(g_loss, var_list=g_params)
#clip discrim weights
d_clip = [tf.assign(v,tf.clip_by_value(v, CLIP_RANGE[0], CLIP_RANGE[1])) for v in d_params]
init = tf.global_variables_initializer()
# Create a saver.
saver = tf.train.Saver(var_list = None,max_to_keep = 5)
sess = tf.Session()
sess.run(init)
sess.run(d_clip)
#load previous variables
if to_restore == True:
print("reloading variables...")
logging.debug("reloading variables...")
ckpt = tf.train.get_checkpoint_state(output_path)
saver.restore(sess, ckpt.model_checkpoint_path)
if os.path.exists(output_path) == False:
os.mkdir(output_path)
get_data = dataset.DataProvider(pkl_path='./bdwm_data_token.pkl',
buckets_size=buckets,batch_size=batch_size)
translator = Translator('./dict.txt')
print("save ckpt")
saver.save(sess, os.path.join(output_path, 'refine_model.ckpt'), global_step=global_step)
for i in range(sess.run(global_step), max_epoch):
data_iterator = get_data.get_batch()
if i < 25 or i % 500 == 0:
citers = 100
else:
citers = CRITIC
for j in np.arange(citers):
print("epoch:%s, dis iter:%s" % (i, j))
logging.debug("epoch:%s, dis iter:%s" % (i, j))
try:
feed_dict, BUCKET_ID = data_iterator.next()
except StopIteration:
get_data = dataset.DataProvider(pkl_path='./bdwm_data_token.pkl',
buckets_size=buckets, batch_size=batch_size)
data_iterator = get_data.get_batch()
feed_dict, BUCKET_ID = data_iterator.next()
_,dis_loss,fake_value,true_value = sess.run([d_trainer,d_loss,d_loss_fake,d_loss_real],feed_dict=feed_dict)
sess.run(d_clip)
print("d_loss:{}".format(dis_loss))
print("fake:{} true:{}".format(fake_value,true_value))
logging.debug("d_loss:{}".format(dis_loss))
logging.debug("fake:{} true:{}".format(fake_value,true_value))
for j in np.arange(gen_critic):
print("epoch:%s, gen iter:%s" % (i, j))
logging.debug("epoch:%s, gen iter:%s" % (i, j))
try:
feed_dict, BUCKET_ID = data_iterator.next()
except StopIteration:
get_data = dataset.DataProvider(pkl_path='./bdwm_data_token.pkl',
buckets_size=buckets, batch_size=batch_size)
data_iterator = get_data.get_batch()
feed_dict, BUCKET_ID = data_iterator.next()
g_loss_val, _, d_loss_val = sess.run([g_loss, g_trainer, d_loss], feed_dict=feed_dict)
logging.debug("g_loss:{} d_loss:{}".format(g_loss_val, d_loss_val))
print("g_loss:{} d_loss:{}".format(g_loss_val, d_loss_val))
#get gen val for the true bucket
gen_val = sess.run([fake_ans], feed_dict=feed_dict)
print(gen_val)
#translator.translate_and_print(seq2seq_onehot2label(gen_val),logger = logging)
print("save ckpt")
logging.debug("save ckpt")
saver.save(sess,os.path.join(output_path,'model.ckpt'),global_step=global_step)
if __name__ == '__main__':
train()