-
Notifications
You must be signed in to change notification settings - Fork 11
/
train.py
268 lines (212 loc) · 11.3 KB
/
train.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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import time
from six.moves import xrange
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "1"
import numpy as np
import tensorflow as tf
import utils
from model import Config, BiRNN
tf.flags.DEFINE_string("source_train_path", "",
"Path to the file containing the source sentences to "
"train the model.")
tf.flags.DEFINE_string("target_train_path", "",
"Path to the file containing the target sentences to "
"train the model.")
tf.flags.DEFINE_string("source_valid_path", "",
"Path to the file containing the source sentences to "
"evaluate the model.")
tf.flags.DEFINE_string("target_valid_path", "",
"Path to the file containing the target sentences to "
"evaluate the model.")
tf.flags.DEFINE_string("checkpoint_dir", "./tflogs",
"Directory to save checkpoints and summaries of the model.")
tf.flags.DEFINE_integer("source_vocab_size", 100000,
"Number of the most frequent words to keep in the source "
"vocabulary.")
tf.flags.DEFINE_integer("target_vocab_size", 100000,
"Number of the most frequent words to keep in target "
"vocabulary.")
tf.flags.DEFINE_float("learning_rate", 2e-4,
"Learning rate.")
tf.flags.DEFINE_float("max_gradient_norm", 5.0,
"Clip gradient to this norm.")
tf.flags.DEFINE_float("decision_threshold", 0.99,
"Decision threshold to predict a positive label.")
tf.flags.DEFINE_integer("embedding_size", 300,
"Size of each word embedding.")
tf.flags.DEFINE_integer("state_size", 300,
"Size of the recurrent state in the BiRNN encoder.")
tf.flags.DEFINE_integer("hidden_size", 128,
"Size of the hidden layer in the feed-forward neural "
"network.")
tf.flags.DEFINE_integer("num_layers", 1,
"Number of layers in the BiRNN encoder.")
tf.flags.DEFINE_string("source_embeddings_path", None,
"Pretrained embeddings to initialize the source embeddings "
"matrix.")
tf.flags.DEFINE_string("target_embeddings_path", None,
"Pretrained embeddings to initialize the target embeddings "
"matrix.")
tf.flags.DEFINE_boolean("fix_pretrained", False,
"If true fix pretrained embeddings.")
tf.flags.DEFINE_boolean("use_lstm", False,
"If true use LSTM cells. Otherwise use GRU cells.")
tf.flags.DEFINE_boolean("use_mean_pooling", False,
"If true use mean pooling for final sentence representation.")
tf.flags.DEFINE_boolean("use_max_pooling", False,
"If true use max pooling for final sentence representation.")
tf.flags.DEFINE_integer("batch_size", 128,
"Batch size to use during training.")
tf.flags.DEFINE_integer("num_epochs", 15,
"Number of epochs to train the model.")
tf.flags.DEFINE_integer("num_negative", 5,
"Number of negative examples to sample per pair of "
"parallel sentences in training dataset.")
tf.flags.DEFINE_float("keep_prob_input", 0.8,
"Keep probability for dropout applied at the embedding layer.")
tf.flags.DEFINE_float("keep_prob_output", 0.7,
"Keep probability for dropout applied at the prediction layer.")
tf.flags.DEFINE_integer("steps_per_checkpoint", 200,
"Number of steps to save a model checkpoint.")
FLAGS = tf.flags.FLAGS
def eval_epoch(sess, model, data_iterator, summary_writer):
"""Evaluate model for one epoch."""
sess.run(tf.local_variables_initializer())
num_iter = int(np.ceil(data_iterator.size / FLAGS.batch_size))
epoch_loss = 0
for step in xrange(num_iter):
source, target, label = data_iterator.next_batch(FLAGS.batch_size)
source_len = utils.sequence_length(source)
target_len = utils.sequence_length(target)
feed_dict = {model.x_source: source,
model.x_target: target,
model.labels: label,
model.source_seq_length: source_len,
model.target_seq_length: target_len,
model.decision_threshold: FLAGS.decision_threshold}
loss_value, epoch_accuracy,\
epoch_precision, epoch_recall = sess.run([model.mean_loss,
model.accuracy[1],
model.precision[1],
model.recall[1]],
feed_dict=feed_dict)
epoch_loss += loss_value
if step % FLAGS.steps_per_checkpoint == 0:
summary = sess.run(model.summaries, feed_dict=feed_dict)
summary_writer.add_summary(summary, global_step=data_iterator.global_step)
epoch_loss /= step
epoch_f1 = utils.f1_score(epoch_precision, epoch_recall)
print(" Testing: Loss = {:.6f}, Accuracy = {:.4f}, "
"Precision = {:.4f}, Recall = {:.4f}, F1 = {:.4f}"
.format(epoch_loss, epoch_accuracy,
epoch_precision, epoch_recall, epoch_f1))
def main(_):
assert FLAGS.source_train_path, ("--source_train_path is required.")
assert FLAGS.target_train_path, ("--target_train_path is required.")
# Create vocabularies.
source_vocab_path = os.path.join(os.path.dirname(FLAGS.source_train_path),
"vocabulary.source")
target_vocab_path = os.path.join(os.path.dirname(FLAGS.source_train_path),
"vocabulary.target")
utils.create_vocabulary(source_vocab_path, FLAGS.source_train_path, FLAGS.source_vocab_size)
utils.create_vocabulary(target_vocab_path, FLAGS.target_train_path, FLAGS.target_vocab_size)
# Read vocabularies.
source_vocab, rev_source_vocab = utils.initialize_vocabulary(source_vocab_path)
target_vocab, rev_target_vocab = utils.initialize_vocabulary(target_vocab_path)
# Read parallel sentences.
parallel_data = utils.read_data(FLAGS.source_train_path, FLAGS.target_train_path,
source_vocab, target_vocab)
# Read validation data set.
if FLAGS.source_valid_path and FLAGS.target_valid_path:
valid_data = utils.read_data(FLAGS.source_valid_path, FLAGS.target_valid_path,
source_vocab, target_vocab)
# Initialize BiRNN.
config = Config(len(source_vocab),
len(target_vocab),
FLAGS.embedding_size,
FLAGS.state_size,
FLAGS.hidden_size,
FLAGS.num_layers,
FLAGS.learning_rate,
FLAGS.max_gradient_norm,
FLAGS.use_lstm,
FLAGS.use_mean_pooling,
FLAGS.use_max_pooling,
FLAGS.source_embeddings_path,
FLAGS.target_embeddings_path,
FLAGS.fix_pretrained)
model = BiRNN(config)
# Build graph.
model.build_graph()
# Train model.
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
train_iterator = utils.TrainingIterator(parallel_data, FLAGS.num_negative)
train_summary_writer = tf.summary.FileWriter(os.path.join(FLAGS.checkpoint_dir, "train"), sess.graph)
if FLAGS.source_valid_path and FLAGS.target_valid_path:
valid_iterator = utils.EvalIterator(valid_data)
valid_summary_writer = tf.summary.FileWriter(os.path.join(FLAGS.checkpoint_dir, "valid"), sess.graph)
epoch_loss = 0
epoch_completed = 0
batch_completed = 0
num_iter = int(np.ceil(train_iterator.size / FLAGS.batch_size * FLAGS.num_epochs))
start_time = time.time()
print("Training model on {} sentence pairs per epoch:".
format(train_iterator.size, valid_iterator.size))
for step in xrange(num_iter):
source, target, label = train_iterator.next_batch(FLAGS.batch_size)
source_len = utils.sequence_length(source)
target_len = utils.sequence_length(target)
feed_dict = {model.x_source: source,
model.x_target: target,
model.labels: label,
model.source_seq_length: source_len,
model.target_seq_length: target_len,
model.input_dropout: FLAGS.keep_prob_input,
model.output_dropout: FLAGS.keep_prob_output,
model.decision_threshold: FLAGS.decision_threshold}
_, loss_value, epoch_accuracy,\
epoch_precision, epoch_recall = sess.run([model.train_op,
model.mean_loss,
model.accuracy[1],
model.precision[1],
model.recall[1]],
feed_dict=feed_dict)
epoch_loss += loss_value
batch_completed += 1
# Write the model's training summaries.
if step % FLAGS.steps_per_checkpoint == 0:
summary = sess.run(model.summaries, feed_dict=feed_dict)
train_summary_writer.add_summary(summary, global_step=step)
# End of current epoch.
if train_iterator.epoch_completed > epoch_completed:
epoch_time = time.time() - start_time
epoch_loss /= batch_completed
epoch_f1 = utils.f1_score(epoch_precision, epoch_recall)
epoch_completed += 1
print("Epoch {} in {:.0f} sec\n"
" Training: Loss = {:.6f}, Accuracy = {:.4f}, "
"Precision = {:.4f}, Recall = {:.4f}, F1 = {:.4f}"
.format(epoch_completed, epoch_time,
epoch_loss, epoch_accuracy,
epoch_precision, epoch_recall, epoch_f1))
# Save a model checkpoint.
checkpoint_path = os.path.join(FLAGS.checkpoint_dir, "model.ckpt")
model.saver.save(sess, checkpoint_path, global_step=step)
# Evaluate model on the validation set.
if FLAGS.source_valid_path and FLAGS.target_valid_path:
eval_epoch(sess, model, valid_iterator, valid_summary_writer)
# Initialize local variables for new epoch.
batch_completed = 0
epoch_loss = 0
sess.run(tf.local_variables_initializer())
start_time = time.time()
print("Training done with {} steps.".format(num_iter))
train_summary_writer.close()
valid_summary_writer.close()
if __name__ == "__main__":
tf.app.run()