-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain_operation.py
196 lines (162 loc) · 9.88 KB
/
main_operation.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
# coding: utf-8
from mt_enc_dec import MTEnDecoder
from semantic_logic_model import MTSemanticLogicED
import tensorflow as tf
import corpora
from global_variable import GlobalVariable, ValidOrTestParm, config
from word2vec import WordToVector
import numpy as np
import time
def init_variables(language, isTraining=False):
dataset = corpora.DataSet()
if isTraining:
dataset.loadDateSet(language, isTraining=isTraining, corpus_dir=None)
if type(GlobalVariable.wordToVector) == type(None):
GlobalVariable.wordToVector = WordToVector(GlobalVariable.bin_file_path)
if type(GlobalVariable.ph_embedding) == type(None):
GlobalVariable.ph_embedding = GlobalVariable.wordToVector.word2vector(GlobalVariable.placeholder)
GlobalVariable.corpus_sets = dataset.train_set
GlobalVariable.corpus_sets_num = len(GlobalVariable.corpus_sets)
GlobalVariable.shuffle_index = np.arange(GlobalVariable.corpus_sets_num)
np.random.shuffle(GlobalVariable.shuffle_index)
ValidOrTestParm.corpus_sets = dataset.valid_set
ValidOrTestParm.corpus_sets = dataset.valid_set
ValidOrTestParm.corpus_sets_num = len(ValidOrTestParm.corpus_sets)
else:
dataset.loadDateSet(language, corpus_dir=None)
ValidOrTestParm.corpus_sets = dataset.valid_set
ValidOrTestParm.corpus_sets_num = len(ValidOrTestParm.corpus_sets)
def run_epoch(session, model, eval_op=None, verbose=False):
"""Runs the model on the given data."""
start_time = time.time()
costs = 0.0
iters = 0
state = session.run(model.initial_state)
fetches = {
"cost": model.cost,
"final_state": model.final_state,
}
if eval_op is not None:
fetches["eval_op"] = eval_op
for step in range(model.input.epoch_size):
feed_dict = {}
for i, (c, h) in enumerate(model.initial_state):
feed_dict[c] = state[i].c
feed_dict[h] = state[i].h
vals = session.run(fetches, feed_dict)
cost = vals["cost"]
state = vals["final_state"]
costs += cost
iters += model.input.num_steps
if verbose and step % (model.input.epoch_size // 10) == 10:
print("%.3f perplexity: %.3f speed: %.0f wps" %
(step * 1.0 / model.input.epoch_size, np.exp(costs / iters),
iters * model.input.batch_size / (time.time() - start_time)))
return np.exp(costs / iters)
def main(mTSemLogModel, isTraining=True):
x = tf.placeholder(tf.float32, [None, mTSemLogModel.n_steps, mTSemLogModel.n_inputs])
# obtain the semantic representation code
# _, self.semantic_rpt = self.encode(x, self.weights_ed, self.biases_ed)
# shape of semantic_rpt:(batch_size, n_hidden_units)
_, mTSemLogModel.semantic_rpt = mTSemLogModel.encode(x, mTSemLogModel.key_map['AE_encoder_scope'], mTSemLogModel.key_map['AE_encoder'], isTraining=isTraining)
# obtain the logic representation code
# shape of logic_rpt:(batch_size, n_hidden_units)
_, mTSemLogModel.logic_rpt = mTSemLogModel.encode(x, mTSemLogModel.key_map['DL_encoder_scope'], mTSemLogModel.key_map['DL_encoder'], isTraining=isTraining)
# concat semantic representation and logic representation as the input dialog system decoder' input tenor
mTSemLogModel.semantic_logic_rpt = tf.concat([mTSemLogModel.semantic_rpt, mTSemLogModel.logic_rpt], 1)
# ##########################################################################
# This code executes single language auto-encode task.
# ##########################################################################
pred = mTSemLogModel.decode(mTSemLogModel.semantic_rpt, mTSemLogModel.key_map['AE_decoder_scope'], mTSemLogModel.key_map['AE_decoder'], isTraining=isTraining)
# raw_input represent x after reshaping, shape: (batch_size, self.n_inputs * self.n_steps)
raw_input = tf.reshape(x, (-1, mTSemLogModel.n_inputs * mTSemLogModel.n_steps))
# shaoe of pred after reshaping: (batch_size, self.n_inputs * self.n_steps)
pred = tf.reshape(pred, (-1, mTSemLogModel.n_inputs * mTSemLogModel.n_steps))
# loss函数，优化方法
cost = tf.reduce_mean(tf.pow(raw_input - pred, 2))
train_op_optimizer = tf.train.RMSPropOptimizer(mTSemLogModel.lr).minimize(cost)
# tf.summary.scalar("cost", cost)
# ##########################################################################
# This code executes single language dialog-system task.
# ##########################################################################
# placeholder for response sequences of shape: (batch_size, n_steps_response, embedding_len)
y = tf.placeholder(tf.float32, [None, mTSemLogModel.n_steps_dl_out, mTSemLogModel.n_inputs])
# pred_response_seq = self.decode(self.semantic_rpt, self.key_map['DL_decoder_scope'], self.key_map['DL_decoder'])
# concat_times=2 ==> semantic_logic_rpt
pred_response_seq = mTSemLogModel.decode(mTSemLogModel.semantic_logic_rpt, mTSemLogModel.key_map['DL_decoder_scope'], mTSemLogModel.key_map['DL_decoder'], concat_times=2,isTraining=isTraining)
# raw_response_seq represent y after reshaping, stand for next sequence from corpus.
raw_response_seq = tf.reshape(y, (-1, mTSemLogModel.n_inputs * mTSemLogModel.n_steps_dl_out))
pred_response_seq = tf.reshape(pred_response_seq, (-1, mTSemLogModel.n_inputs * mTSemLogModel.n_steps_dl_out))
# loss函数，优化方法
cost_DL = tf.reduce_mean(tf.pow(raw_response_seq - pred_response_seq, 2), name='cost_dialog')
train_op_optimizer_DL = tf.train.RMSPropOptimizer(mTSemLogModel.lr).minimize(cost_DL)
# tf.summary.scalar("cost_dialog", cost_DL)
saver = tf.train.Saver()
# os.environ["CUDA_VISIBLE_DEVICES"] = '0' # 指定第一块GPU可用
GPU_config = tf.ConfigProto()
GPU_config.gpu_options.per_process_gpu_memory_fraction = 0.99 # 程序最多只能占用指定gpu99%的显存
GPU_config.gpu_options.allow_growth = True # 程序按需申请内存
with tf.Session(config=GPU_config) as sess:
init = tf.global_variables_initializer()
sess.run(init)
if isTraining:
step = 0
save_count = 1
while step * mTSemLogModel.batch_size < mTSemLogModel.training_iters:
try:
batch_xs, batch_ys = corpora.corpus_next_batch(mTSemLogModel.batch_size)
# Training order: firstly, use common corpus to train AE, then use dialog corpus to train AE and DL
# And number of single AE training iteration-config['AE_completed'] is the separation of these two steps
if step * mTSemLogModel.batch_size <= config['AE_completed']:
sess.run([train_op_optimizer, cost], feed_dict={x: batch_xs})
if step % 200 == 0:
batch_val_xs, _ = corpora.get_next_batch(mTSemLogModel.batch_size, isValiding=True)
_, Y1_loss = sess.run([train_op_optimizer, cost], feed_dict={x: batch_val_xs})
print("single AE training -> single language auto-encoder loss: ", "{:.9f}".format(Y1_loss))
step += 1
continue
# And this is the second step, use dialog corpus to train AE and DL
# AE training
sess.run([train_op_optimizer, cost], feed_dict={ x: batch_xs })
if step % 200 == 0:
batch_val_xs, _ = corpora.get_next_batch(mTSemLogModel.batch_size, isValiding=True)
_, Y1_loss = sess.run([train_op_optimizer, cost], feed_dict={x: batch_val_xs})
print("single language auto-encoder loss: ", "{:.9f}".format(Y1_loss))
# DL training
sess.run([train_op_optimizer_DL, cost_DL], feed_dict={x: batch_xs, y: batch_ys })
if step % 200 == 0:
batch_val_xs, batch_val_ys = corpora.get_next_batch(mTSemLogModel.batch_size, isValiding=True)
_, Y2_loss = sess.run([train_op_optimizer_DL, cost_DL], feed_dict={x: batch_val_xs, y: batch_val_ys })
print("single language(English) dialog system loss: ", "{:.9f}".format(Y2_loss))
step += 1
if (step * mTSemLogModel.batch_size + 1) > (config['checkout_iters'] * save_count):
saver.save(sess, GlobalVariable.MTSemanticLogicED_save_path, global_step=step * mTSemLogModel.batch_size + 1)
save_count += 1
except Exception as e:
print('Exceptions occurs when feed data in batch_size, details like this:', "Exception: {0}".format(e))
# save_path = saver.save(sess, GlobalVariable.model_save_path)
saver.save(sess, GlobalVariable.model_save_path, global_step=55556666)
print("Optimization Finished!")
else:
# Restore model weights from previously saved model
print("Model restored from file: %s" % GlobalVariable.model_save_path)
saver.restore(sess, GlobalVariable.model_save_path)
loss_count = 0
total_loss = 0.0
while ValidOrTestParm.epochs_completed<2:
batch_test_xs, batch_test_ys = corpora.get_next_batch(mTSemLogModel.batch_size, isValiding=False)
if type(batch_test_xs) != type(None):
loss_count += 1
_, DL_loss = sess.run([train_op_optimizer_DL, cost_DL], feed_dict={x: batch_test_xs, y: batch_test_ys})
total_loss += DL_loss
if loss_count > 0:
print('average loss is:', total_loss*1.0/loss_count)
if __name__ == '__main__':
# train
init_variables('English', isTraining=True)
mTSemanticLogicED = MTSemanticLogicED()
main(mTSemanticLogicED, isTraining=True)
# test
# init_variables('English', isTraining=False)
# mTSemanticLogicED = MTSemanticLogicED()
# main(mTSemanticLogicED, isTraining=False)