-
Notifications
You must be signed in to change notification settings - Fork 1
/
trainner.py
309 lines (239 loc) · 11.4 KB
/
trainner.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
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
# -*- coding: UTF-8 -*-
from __future__ import print_function
from __future__ import unicode_literals
import os
import sys
import time
import json,errno
import codecs,logging
import random, argparse
import numpy as np
from collections import Counter
import torch
import torch.nn as nn
import torch.optim as optim
from source.elmo.models.elmo import *
from source.elmo.models.layers import *
logging.basicConfig(level=logging.INFO, format='%(asctime)-15s %(levelname)s: %(message)s')
def model_config():
parser = argparse.ArgumentParser()
# Data
data_arg = parser.add_argument_group("Data")
data_arg.add_argument("--model_path", type=str, default="./output")
data_arg.add_argument("--config_file", type=str, default="config.json")
data_arg.add_argument("--train_file", type=str, default="./data/corpus.txt")
data_arg.add_argument("--embed_file", type=str, default="./data/cn.vector.txt")
# data_arg.add_argument('--valid_path', type=str, default=None)
# data_arg.add_argument('--test_path', type=str, default=None)
#training
train_arg = parser.add_argument_group("Training")
train_arg.add_argument("--seed", type=int, default=1024)
train_arg.add_argument("--gpu", type=int,default=-1)
train_arg.add_argument("--lr", type=float, default=0.001)
train_arg.add_argument("--lr_decay", type=float, default=0.9)
train_arg.add_argument("--batch_size", type=int, default=64)
train_arg.add_argument("--max_epoch", type=int, default=2)
train_arg.add_argument("--clip_grad", type=float, default=5)
train_arg.add_argument("--max_seq_len", type=int, default=30)
train_arg.add_argument('--max_char_token', type=int, default=10)
train_arg.add_argument('--char_dim', type=int, default=50)
train_arg.add_argument('--n_highway', type=int, default=2)
train_arg.add_argument('--encode_layers', type=int, default=2)
train_arg.add_argument("--activation", type=str, default="relu")
train_arg.add_argument('--encode_dim', type=int, default=1024)
train_arg.add_argument('--cell_clip', type=int, default=3)
train_arg.add_argument('--proj_clip', type=int, default=3)
train_arg.add_argument('--word_dim', type=int, default=300)
train_arg.add_argument('--projection_dim', type=int, default=150)
train_arg.add_argument('--dropout', type=int, default=0.1)
train_arg.add_argument("--min_count", type=int, default=1)
train_arg.add_argument("--eval_steps", type=int,default=1024)
train_arg.add_argument("--valid_size", type=int, default=64)
train_arg.add_argument("--max_vocab_size", type=int, default=150000)
train_arg.add_argument("--encode", type=str, default="elmo")
train_arg.add_argument("--save_nll_loss", default=False, action='store_true')
train_arg.add_argument("--optimizer", default='adam', choices=['sgd', 'adam', 'adagrad'])
train_arg.add_argument("--filters", type=list, default=[[1, 32], [2, 32], [3, 128], [4, 128]])
config = parser.parse_args()
return config
def get_truncated_vocab(dataset, min_count):
word_count = Counter()
for sentence in dataset:
word_count.update(sentence)
word_count = list(word_count.items())
word_count.sort(key=lambda x: x[1], reverse=True)
for i, (word, count) in enumerate(word_count):
if count < min_count:
break
# logging.info('Truncated word count: {0}.'.format(sum([count for word, count in word_count[i:]])))
logging.info('Original vocabulary size: {0}.'.format(len(word_count)))
return word_count[:i]
# def eval_model(model, valid):
# model.eval()
# if model.config['classifier']['name'].lower() == 'cnn_softmax' or \
# model.config['classifier']['name'].lower() == 'sampled_softmax':
# model.nll_loss.update_embedding_matrix()
# #print('emb mat size: ', model.nll_loss.embedding_matrix.size())
# total_loss, total_tag = 0.0, 0
# valid_w, valid_c, valid_lens, valid_masks = valid
# for w, c, lens, masks in zip(valid_w, valid_c, valid_lens, valid_masks):
# loss_forward, loss_backward = model.forward(w, c, masks)
# total_loss += loss_forward.data[0]
# n_tags = sum(lens)
# total_tag += n_tags
# model.train()
# return np.exp(total_loss / total_tag)
def train_epoch(epoch, config, model, configimizer,train, best_train):
model.train()
start_time = time.time()
cnt = 0
total_loss, total_tag = 0.0, 0
train_w, train_c, train_lens, train_masks = train
lst = list(range(len(train_w)))
random.shuffle(lst)
train_w = [train_w[l] for l in lst]
train_c = [train_c[l] for l in lst]
train_lens = [train_lens[l] for l in lst]
train_masks = [train_masks[l] for l in lst]
for w, c, lens, masks in zip(train_w, train_c, train_lens, train_masks):
cnt += 1
model.zero_grad()
loss_forward, loss_backward = model.forward(w, c, masks)
loss = (loss_forward + loss_backward) / 2.0
total_loss += loss_forward.item()
n_tags = sum(lens)
total_tag += n_tags
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), config.clip_grad)
configimizer.step()
if cnt * config.batch_size % 5000 == 0:
logging.info("Epoch={} step={} lr={:.6f} train_ppl={:.6f} time={:.2f}s".format(
epoch, cnt, configimizer.param_groups[0]['lr'],
np.exp(total_loss / total_tag), time.time() - start_time
))
start_time = time.time()
if cnt % config.eval_steps == 0 or cnt % len(train_w) == 0:
# if valid is None:
train_ppl = np.exp(total_loss / total_tag)
logging.info("Epoch={} step={} lr={:.6f} train_ppl={:.6f}".format(
epoch, cnt, configimizer.param_groups[0]['lr'], train_ppl))
if train_ppl < best_train:
best_train = train_ppl
# logging.info("New record achieved on training dataset!")
model.save_model(config.model_path, config.save_nll_loss)
return best_train
def train():
config = model_config()
with open(config.config_file, 'r') as fin:
elmo_config = json.load(fin)
# set seed.
torch.manual_seed(config.seed)
random.seed(config.seed)
if config.gpu >= 0:
torch.cuda.set_device(config.gpu)
if config.seed > 0:
torch.cuda.manual_seed(config.seed)
use_cuda = config.gpu >= 0 and torch.cuda.is_available()
train_data = read_data(config.train_file, config.max_char_token, config.max_seq_len)
logging.info('training instance: {}, training tokens: {}.'.format(len(train_data),
sum([len(s) - 1 for s in train_data])))
valid_data = None
test_data = None
word_lexicon = {}
# Maintain the vocabulary. vocabulary is used in either WordEmbeddingInput or softmax classification
vocab = get_truncated_vocab(train_data, config.min_count)
# Ensure index of '<oov>' is 0
for special_word in ['<oov>', '<bos>', '<eos>', '<pad>']:
if special_word not in word_lexicon:
word_lexicon[special_word] = len(word_lexicon)
for word, _ in vocab:
if word not in word_lexicon:
word_lexicon[word] = len(word_lexicon)
# Word Embedding
if config.word_dim > 0:
word_embedder = WordEmbed(config.embed_file, config.word_dim, word_lexicon, fix_emb=False)
logging.info('Word embedding size: {0}'.format(len(word_embedder.word2id)))
else:
word_embedder = None
logging.info('Vocabulary size: {0}'.format(len(word_lexicon)))
# Character Lexicon
if config.char_dim> 0:
char_lexicon = {}
for sentence in train_data:
for word in sentence:
for ch in word:
if ch not in char_lexicon:
char_lexicon[ch] = len(char_lexicon)
for special_char in ['<bos>', '<eos>', '<oov>', '<pad>', '<bow>', '<eow>']:
if special_char not in char_lexicon:
char_lexicon[special_char] = len(char_lexicon)
char_embedder = CharEmbed(config.char_dim, char_lexicon, fix_emb=False)
logging.info('Char embedding size: {0}'.format(len(char_embedder.word2id)))
else:
char_lexicon = None
char_embedder = None
train = gen_mini_batches(
train_data, config.batch_size, word_lexicon, char_lexicon,config.max_char_token)
if config.eval_steps is None:
config.eval_steps = len(train[0])
# logging.info('Evaluate every {0} batches.'.format(config.eval_steps))
valid = None
test = None
# if valid_data is not None:
# valid = create_batches(
# valid_data, config.batch_size, word_lexicon, char_lexicon, config, sort=False, shuffle=False, use_cuda=use_cuda)
# else:
# valid = None
# if test_data is not None:
# test = create_batches(
# test_data, config.batch_size, word_lexicon, char_lexicon, config, sort=False, shuffle=False, use_cuda=use_cuda)
# else:
# test = None
label_to_ix = word_lexicon
logging.info('vocab size: {0}'.format(len(label_to_ix)))
nclasses = len(label_to_ix)
# for s in train_w:
# for i in range(s.view(-1).size(0)):
# if s.view(-1)[i] >= nclasses:
# print(s.view(-1)[i])
model = Model(word_embedder=word_embedder, char_embedder=char_embedder, n_class=nclasses,
encode=config.encode, char_dim=config.char_dim,filters=config.filters,
n_highway=config.n_highway, encode_layers=config.encode_layers,
activation=config.activation, encode_dim=config.encode_dim,
cell_clip=config.cell_clip, proj_clip=config.proj_clip,
projection_dim=config.projection_dim, dropout=config.dropout, use_cuda=use_cuda)
logging.info(str(model))
if use_cuda:
model = model.cuda()
need_grad = lambda x: x.requires_grad
# optimizer = optim.SGD(filter(need_grad, model.parameters()), lr=config.lr)
if config.optimizer.lower() == 'adam':
optimizer = optim.Adam(filter(need_grad, model.parameters()), lr=config.lr)
elif config.optimizer.lower() == 'sgd':
optimizer = optim.SGD(filter(need_grad, model.parameters()), lr=config.lr)
elif config.optimizer.lower() == 'adagrad':
optimizer = optim.Adagrad(filter(need_grad, model.parameters()), lr=config.lr)
else:
raise ValueError('Unknown configimizer {}'.format(config.configimizer.lower()))
try:
os.makedirs(config.model_path)
except OSError as exception:
if exception.errno != errno.EEXIST:
raise
if elmo_config['token_embedder']['char_dim'] > 0:
with codecs.open(os.path.join(config.model_path, 'char.dic'), 'w', encoding='utf-8') as fpo:
for ch, i in char_embedder.word2id.items():
print('{0}\t{1}'.format(ch, i), file=fpo)
with codecs.open(os.path.join(config.model_path, 'word.dic'), 'w', encoding='utf-8') as fpo:
for w, i in word_lexicon.items():
print('{0}\t{1}'.format(w, i), file=fpo)
json.dump(vars(config), codecs.open(os.path.join(config.model_path, 'config.json'), 'w', encoding='utf-8'))
best_train = 1e+8
for epoch in range(config.max_epoch):
best_train = train_epoch(epoch, config, model, optimizer, train, best_train)
if config.lr_decay > 0:
optimizer.param_groups[0]['lr'] *= config.lr_decay
logging.info("best train ppl: {:.6f}.".format(best_train))
if __name__ == "__main__":
# if len(sys.argv) > 1 and sys.argv[1] == 'train':
train()