-
Notifications
You must be signed in to change notification settings - Fork 4
/
train.py
284 lines (224 loc) · 14.4 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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
#!/usr/bin/env python
# -*- coding: utf-8 -*-
import datetime
import io
import logging
import os
import sys
import time
import tensorflow as tf
from glob import glob
import numpy as np
from six.moves import cPickle
import dataset
from common import utils
from network import Network
from common import metrics
if __name__ == "__main__":
# Set random seed
np.random.seed(42)
# Parse arguments
import argparse
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument("dataset", default='', type=str,
help="Path to dataset configuration file storing files for train, dev and test.")
parser.add_argument("--exp_name", default='', type=str, help="Experiment name.")
parser.add_argument("--batch_size", default=100, type=int, help="Batch size.")
parser.add_argument("--embedding", default=200, type=int,
help="Embedding dimension. One hot is used if <1. It is highly recommended that embedding == rnn_cell_dim")
parser.add_argument("--input_char_vocab", default='', type=str,
help="Path to file storing input char vocabulary. If no provided, is automatically computed from data.")
parser.add_argument("--target_char_vocab", default='', type=str,
help="Path to file storing target char vocabulary. If no provided, is automatically computed from data.")
parser.add_argument("--num_top_chars", default=-1, type=int,
help="Take only num_top_chars most occuring characters. All other will be considered UNK")
parser.add_argument("--max_chars", default=200, type=int, help="Maximum number of characters in a sentence.")
parser.add_argument("--epochs", default=10, type=int, help="Number of epochs.")
parser.add_argument("--logdir", default="logs", type=str, help="Logdir name.")
parser.add_argument("--savedir", default="save", type=str, help="Savedir name.")
parser.add_argument("--train_perc", default=1.0, type=float, help="Percentage of total samples used for training.")
parser.add_argument("--validation_perc", default=0.0, type=float,
help="Percentage of total samples used for validation set.")
parser.add_argument("--test_perc", default=0.0, type=float, help="Percentage of total samples used for testing.")
parser.add_argument("--keep_prob", default=0.8, type=float, help="Dropout keep probability used for training.")
parser.add_argument("--rnn_cell", default="gru", type=str, help="RNN cell type.")
parser.add_argument("--rnn_cell_dim", default=200, type=int, help="RNN cell dimension.")
parser.add_argument("--num_layers", default=1, type=int, help="Number of layers.")
parser.add_argument("--learning_rate", default=1e-4, type=float, help="Learning rate.")
parser.add_argument("--threads", default=8, type=int, help="Maximum number of threads to use.")
parser.add_argument('--use_residual', action='store_true', default=False,
help="If set, residual connections will be used in the model.")
parser.add_argument("--save_every", default=2000, type=int, help="Interval for saving models.")
parser.add_argument("--log_every", default=1000, type=int, help="Interval for logging models (Tensorboard).")
parser.add_argument("--num_test", default=1000, type=int, help="Number of samples to test on.")
parser.add_argument("--restore", type=str,
help="Restore model from this checkpoint and continue training from it. Can be a shell-style wildcard expandable exactly to one directory.")
parser.add_argument("--num_sentences", default=-1, type=int,
help="Number of sentences to read from train file (-1 == read all sentences).")
args = parser.parse_args()
experiment_name = args.exp_name
experiment_name += '_layers{}_dim{}_embedding{}_lr{}'.format(args.num_layers, args.rnn_cell_dim,
args.embedding, args.learning_rate)
# create save directory for current experiment's data (if not exists)
save_data_dir = os.path.join(args.savedir, experiment_name)
if not os.path.exists(save_data_dir):
os.makedirs(save_data_dir)
# create subdir of save data directory to store trained models
timestamp = datetime.datetime.now().strftime("%Y-%m-%d_%H%M%S")
save_model_dir = os.path.join(save_data_dir, timestamp)
os.makedirs(save_model_dir)
# configure logger
logging.basicConfig(filename=os.path.join(save_model_dir, 'experiment_log.log'), level=logging.DEBUG,
format='%(asctime)s %(message)s')
logging.info('Experiment started at: {} and its name: {}'.format(timestamp, experiment_name))
logging.info('Experiment arguments: {}'.format(str(args)))
dataset_files = utils.parse_dataset_file(args.dataset)
print(dataset_files)
# load train input and train target sentences
print('Loading train data')
num_sentences_limit = args.num_sentences if args.num_sentences > 0 else float("inf")
input_sentences, target_sentences = [], []
with io.open(dataset_files['train_inputs'], 'r', encoding='utf8') as inputs_reader:
with io.open(dataset_files['train_targets'], 'r', encoding='utf8') as targets_reader:
num_lines_read = 0
for input_line, target_line in zip(inputs_reader, targets_reader):
input_sentences.append(input_line.strip())
target_sentences.append(target_line.strip())
num_lines_read += 1
if num_lines_read > num_sentences_limit:
break
input_char_vocab, target_char_vocab = None, None
if args.input_char_vocab != '':
input_char_vocab = utils.load_vocabulary(args.input_char_vocab)
if args.target_char_vocab != '':
target_char_vocab = utils.load_vocabulary(args.target_char_vocab)
if args.restore:
checkpoint_path = glob(args.restore) # expand possible wildcard
if len(checkpoint_path) == 0:
raise ValueError('Restore parameter provided ({}), but no such folder exists.'.format(args.restore))
elif len(checkpoint_path) > 1:
raise ValueError('Restore parameter provided ({}), but multiple such folders exist.'.format(args.restore))
checkpoint_path = checkpoint_path[0]
with open(os.path.join(checkpoint_path, 'vocab.pkl'), 'rb') as f:
input_char_vocab, target_char_vocab = cPickle.load(f)
dataset = dataset.ParalelSentencesDataset(args.batch_size, args.max_chars, input_sentences,
target_sentences, args.train_perc, args.validation_perc,
args.test_perc, input_char_vocab, target_char_vocab,
args.num_top_chars)
if 'dev_inputs' in dataset_files:
print('Loading validation data')
dev_input_sentences, dev_target_sentences = [], []
with io.open(dataset_files['dev_inputs'], 'r', encoding='utf8') as reader:
dev_input_sentences = reader.read().splitlines()
with io.open(dataset_files['dev_targets'], 'r', encoding='utf8') as reader:
dev_target_sentences = reader.read().splitlines()
dataset.add_validation_set(dev_input_sentences, dev_target_sentences)
if 'test_inputs' in dataset_files:
print('Loading test data')
test_input_sentences, test_target_sentences = [], []
with io.open(dataset_files['test_inputs'], 'r', encoding='utf8') as reader:
test_input_sentences = reader.read().splitlines()
with io.open(dataset_files['test_targets'], 'r', encoding='utf8') as reader:
test_target_sentences = reader.read().splitlines()
dataset.add_test_set(test_input_sentences, test_target_sentences)
print('Building dataset')
dataset.build()
input_char_vocab = dataset.input_char_vocabulary
target_char_vocab = dataset.target_char_vocabulary
# dump current configuration and used vocabulary to this model's folder
with open(os.path.join(save_model_dir, 'vocab.pkl'), 'wb') as f:
cPickle.dump((input_char_vocab, target_char_vocab), f)
with open(os.path.join(save_model_dir, 'config.pkl'), 'wb') as f:
cPickle.dump(args, f)
evaluation_metrics = {'char_accuracy': metrics.c2c_per_char_accuracy,
'word_accuracy': metrics.c2c_per_word_accuracy}
print('Creating network')
# Construct the network
network = Network(
input_alphabet_size=len(input_char_vocab.keys()),
target_alphabet_size=len(target_char_vocab.keys()),
cell_type=args.rnn_cell,
rnn_cell_dim=args.rnn_cell_dim,
num_layers=args.num_layers,
embedding_dim=args.embedding,
threads=args.threads,
learning_rate=args.learning_rate,
use_residual_connections=args.use_residual,
)
if args.restore:
logging.info('Restoring model from: {}'.format(checkpoint_path))
network.restore(checkpoint_path)
# split validation and testing data into "batches" (to fit into the memory)
evaluation_sets = {}
for evaluation_set_name, evaluation_set_fn in dataset.get_evaluation_sets():
evaluation_batch_size = dataset.batch_size * 3
evaluation_set_sentences, evaluation_set_sentence_lens, evaluation_set_target_sentences = evaluation_set_fn()
num_eval_samples = len(evaluation_set_sentences)
num_eval_bins = np.ceil(num_eval_samples / evaluation_batch_size)
# print('{} : {} : {}'.format(evaluation_set_name, num_eval_samples, num_eval_bins))
eval_set_input_sentences = np.array_split(evaluation_set_sentences, num_eval_bins)
# print('{} : {} : {}'.format(evaluation_set_name, len(eval_set_input_sentences),
# eval_set_input_sentences[0].shape))
eval_set_sentence_lens = np.array_split(evaluation_set_sentence_lens, num_eval_bins)
eval_set_target_sentences = np.array_split(evaluation_set_target_sentences, num_eval_bins)
evaluation_sets[evaluation_set_name] = [eval_set_input_sentences, eval_set_sentence_lens,
eval_set_target_sentences]
summary_writer = tf.summary.FileWriter("{}/{}-{}".format(args.logdir, timestamp, experiment_name), flush_secs=10)
# Train
print('Training')
sys.stdout.flush()
for epoch in range(args.epochs):
dataset.reset_batch_pointer()
for batch_ind in range(dataset.num_batches):
step_number = epoch * dataset.num_batches + batch_ind
start = time.time()
input_sentences, sentence_lens, target_sentences = dataset.next_batch()
network.train(input_sentences, sentence_lens, target_sentences, args.keep_prob)
end = time.time()
if step_number % args.log_every == 0:
eval_time_start = time.time()
string_summary = "{}/{}, epoch: {}, time/batch = {:.3f}".format(step_number,
args.epochs * dataset.num_batches,
epoch, end - start)
for eval_set_name in evaluation_sets.keys():
print('Evaluating {}'.format(eval_set_name))
string_summary += "\n {}".format(eval_set_name)
eval_set_input_sentences, eval_set_sentence_lens, eval_set_target_sentences = evaluation_sets[
eval_set_name]
eval_set_start_time = time.time()
predictions, lengths, targets = [], [], []
for eval_set_input_sentence, eval_set_sentence_len, eval_set_target_sentence in zip(
eval_set_input_sentences, eval_set_sentence_lens, eval_set_target_sentences):
eval_partial_set_result = network.session.run(network.predictions,
feed_dict={
network.input_sentences: eval_set_input_sentence,
network.sentence_lens: eval_set_sentence_len,
network.target_sentences: eval_set_target_sentence})
# eval_partial_set_result is a linearized list of dimension [ sum_i(eval_set_sentence_len[i])]
len_cumsum = np.cumsum(eval_set_sentence_len)[:-1]
eval_partial_set_result = np.array_split(eval_partial_set_result, len_cumsum)
predictions += list(eval_partial_set_result)
lengths += list(eval_set_sentence_len)
targets += list(eval_set_target_sentence)
eval_set_middle_time = time.time()
summaries = []
for metric_name, metric_fn in evaluation_metrics.items():
print(' --metric: {}'.format(metric_name))
metric_start_time = time.time()
result = metric_fn(predictions, lengths, targets, target_char_vocab)
metric_end_time = time.time()
string_summary += "\n {}:{:.6f}:{}".format(metric_name, result,
metric_end_time - metric_start_time)
summaries.append(
tf.Summary.Value(tag='{}_{}'.format(eval_set_name, metric_name), simple_value=result))
eval_set_end_time = time.time()
string_summary += "\n {}:{}:{}".format(eval_set_name, eval_set_end_time - eval_set_start_time,
eval_set_middle_time - eval_set_start_time)
summary_writer.add_summary(tf.Summary(value=summaries), global_step=step_number)
eval_time_end = time.time()
string_summary += "\n Evaluation took : {}".format(eval_time_end - eval_time_start)
logging.info(string_summary)
print(string_summary)
if step_number % args.save_every == 0:
checkpoint_path = os.path.join(save_model_dir, 'model.ckpt')
network.saver.save(network.session, checkpoint_path, global_step=step_number)