-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
273 lines (255 loc) · 12.6 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
# -*- coding: utf-8 -*-
"""
@Author: winton
@File: training.py
@Time: 2021/4/9 9:22 AM
@Description:
"""
import argparse
import json
import logging
import os
import torch
from torch import nn, optim
from torch.nn.functional import cross_entropy
from torch.optim import lr_scheduler
from torch.utils.data import DataLoader
import numpy as np
from widget.data import DataSource, PAD_ID
from widget.model import Model
def cal_performance(pred, gold, padding_id, smoothing=False):
""" Apply label smoothing if needed """
loss = cal_loss(pred, gold, smoothing, padding_id)
pred = pred.max(1)[1]
gold = gold.contiguous().view(-1)
non_pad_mask = gold.ne(padding_id)
n_correct = pred.eq(gold)
n_correct = n_correct.masked_select(non_pad_mask).sum().item()
return loss, n_correct
def cal_loss(pred, gold, smoothing, padding_id):
""" Calculate cross entropy loss, apply label smoothing if needed. """
gold = gold.contiguous().view(-1)
if smoothing:
eps = 0.1
n_class = pred.size(1)
one_hot = torch.zeros_like(pred).scatter(1, gold.view(-1, 1), 1)
one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)
log_prb = torch.log_softmax(pred, dim=1)
non_pad_mask = gold.ne(padding_id)
loss = -(one_hot * log_prb).sum(dim=1)
loss = loss.masked_select(non_pad_mask).sum() # average later
else:
loss = cross_entropy(pred, gold, ignore_index=padding_id, reduction='sum')
return loss
def main(args):
torch.set_default_tensor_type(torch.FloatTensor)
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
# Load config
config = json.load(open(args.config_file_path, 'r'))
if config['training']['label_smoothing'] == 1:
label_smoothing = True
else:
label_smoothing = False
# Set logger (console and file)
logger_format = logging.Formatter('%(asctime)s - %(filename)s[line:%(lineno)d] - %(levelname)s: %(message)s')
logger = logging.getLogger('transformer')
sh = logging.StreamHandler()
sh.setFormatter(logger_format)
sh.setLevel(logging.DEBUG)
logger.addHandler(sh)
fh = logging.FileHandler(os.path.join(args.model_path, f'training_{args.task}.log'), 'a', encoding='utf-8')
fh.setLevel(logging.INFO)
fh.setFormatter(logger_format)
logger.addHandler(fh)
logger.setLevel(logging.INFO)
logger.info(json.dumps(config, indent=2))
# Set device and seed
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
torch.manual_seed(config['training']['seed']) # Seed for reproducing
# load data
train_dataset = DataSource(args.config_file_path, args.task, 'train',
args.version, args.context_size)
valid_dataset = DataSource(args.config_file_path, args.task, 'valid',
args.version, args.context_size)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=3)
valid_loader = DataLoader(valid_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=3)
knowledge_data = train_dataset.encode_knowledge_pair(config['data']['knowledge_path']).to(device)
vocab_size = len(train_dataset.vocab)
logger.info(f"Total batches={len(train_dataset) // args.batch_size}")
# Define widget
model = Model(task=args.task,
vocab_size=vocab_size,
max_text_len=config['data']['text_length'],
image_size=config['model']['image_size'],
embedding_size=config['model']['word_embedding_size'],
text_n_layers=config['model']['text_n_layers'],
text_n_head=config['model']['text_n_head'],
text_d_k=config['model']['text_d_k'],
text_d_v=config['model']['text_d_v'],
text_d_model=config['model']['text_d_model'],
text_d_inner=config['model']['text_d_inner'],
co_n_layers=config['model']['co_n_layers'],
co_n_head=config['model']['co_n_head'],
co_d_k=config['model']['co_d_k'],
co_d_v=config['model']['co_d_v'],
co_d_model=config['model']['co_d_model'],
co_d_inner=config['model']['co_d_inner'],
de_n_layers=config['model']['de_n_layers'],
de_n_head=config['model']['de_n_head'],
de_d_k=config['model']['de_d_k'],
de_d_v=config['model']['de_d_v'],
de_d_model=config['model']['de_d_model'],
de_d_inner=config['model']['de_d_inner'],
dropout_rate=config['model']['dropout_rate'],
padding_id=PAD_ID,
tgt_emb_prj_weight_sharing=True,
use_knowledge=config['model']['use_knowledge'],
knowledge_data=knowledge_data
)
model.to(device)
# model = nn.DataParallel(model)
# logger.info(model)
# Define optimizer
optimizer = optim.Adam(model.parameters(),
lr=config['training']['lr'],
weight_decay=config['training']['lr_decay'])
# Define learning rate scheduler
# scheduler = lr_scheduler.ReduceLROnPlateau(optimizer, mode='min',
# factor=0.5, patience=1,
# threshold=0.1, threshold_mode='rel',
# cooldown=0, min_lr=1e-8,
# verbose=True)
# optimizer = ScheduledOptim(
# optim.Adam(
# filter(lambda x: x.requires_grad, widget.parameters()),
# betas=(0.9, 0.98), eps=1e-09),
# config['widget']['text_d_model'], config['training']['warmup_steps'])
# Train
total_batch = 0
min_val_loss = None
bad_loss_cnt = 0
for epoch in range(config['training']['num_epochs']):
total_losses = [0, 0]
n_word_total_list = [0, 0]
n_word_correct_list = [0, 0]
# scheduler.step(sum(valid_losses))
for train_batch in train_loader:
total_batch += 1
if args.task == 'text':
text_input, text_pos, text_turn, text_speaker, \
image_input, image_pos, image_turn, image_speaker, \
query_input, query_pos = map(lambda x: x.to(device), train_batch)
gold = query_input[:, 1:]
optimizer.zero_grad()
dec_output_probs = model((text_input, text_pos, text_turn, text_speaker,
image_input, image_pos, image_turn, image_speaker),
(query_input[:, :-1], query_pos[:, :-1]))
backward_loss = []
for idx, dec_output_prob in enumerate(dec_output_probs):
loss, n_correct = cal_performance(dec_output_prob, gold, PAD_ID,
smoothing=label_smoothing)
backward_loss.append(loss)
batch_loss = loss.item()
total_losses[idx] += batch_loss
non_pad_mask = gold.ne(PAD_ID)
n_word = non_pad_mask.sum().item()
n_word_total_list[idx] += n_word
n_word_correct_list[idx] += n_correct
if total_batch % config['training']['log_batch'] == 0 or total_batch < config['training'][
'log_batch']:
logger.info(f'Epoch [{epoch + 1}], Batch [{total_batch}], '
f'Loss {idx + 1}: {batch_loss / n_word:.6}, '
f'Accuracy {idx + 1}: {100 * n_correct / n_word:.3f} %')
backward_loss = sum(backward_loss)
backward_loss.backward()
optimizer.step()
# optimizer.step_and_update_lr()
# Gradient clipping to avoid exploding gradients
nn.utils.clip_grad_norm_(model.parameters(), config['training']['max_gradient_norm'])
else:
pass
if args.task == 'text':
for idx, (total_loss, n_word_total, n_word_correct) in enumerate(
zip(total_losses, n_word_total_list, n_word_correct_list)):
train_loss = total_loss / n_word_total
train_accu = n_word_correct / n_word_total
logger.info(f'Epoch [{epoch + 1}], '
f'Train Loss {idx + 1}: {train_loss:.6}, '
f'Train Accuracy {idx + 1}: {100 * train_accu:.3f} %')
# Evaluate
valid_losses = [0, 0]
n_word_total_list = [0, 0]
n_word_correct_list = [0, 0]
if epoch % config['training']['evaluate_epoch'] == 0:
model.eval()
if args.task == 'text':
for valid_batch in valid_loader:
text_input, text_pos, text_turn, text_speaker, \
image_input, image_pos, image_turn, image_speaker, \
query_input, query_pos = map(lambda x: x.to(device), valid_batch)
gold = query_input[:, 1:]
dec_output_probs = model((text_input, text_pos, text_turn, text_speaker,
image_input, image_pos, image_turn, image_speaker),
(query_input[:, :-1], query_pos[:, :-1]))
for idx, dec_output_prob in enumerate(dec_output_probs):
loss_val, n_correct = cal_performance(dec_output_prob, gold, PAD_ID,
smoothing=label_smoothing)
valid_losses[idx] += loss_val.item()
non_pad_mask = gold.ne(PAD_ID)
n_word = non_pad_mask.sum().item()
n_word_total_list[idx] += n_word
n_word_correct_list[idx] += n_correct
for idx, (valid_loss, n_word_total, n_word_correct) in enumerate(
zip(valid_losses, n_word_total_list, n_word_correct_list)):
logger.info(f'Epoch [{epoch + 1}] '
f'Valid Loss {idx + 1}: {valid_loss / n_word_total:.6}, '
f'Valid Accuracy {idx + 1}: {100 * n_word_correct / n_word_total:.3f} %, '
f'Patience: {bad_loss_cnt}')
valid_loss = sum(valid_losses)
model.train()
# Save widget each epoch
save_dict = {
'task': args.task,
'epoch': epoch,
'iteration': total_batch,
'valid_loss': valid_loss,
'widget': model.state_dict(),
'optimizer': optimizer.state_dict()
}
torch.save(save_dict,
os.path.join(args.model_path, f'{args.task}_model_{epoch + 1}.pth'))
if min_val_loss is None or valid_loss < min_val_loss:
min_val_loss = valid_loss
bad_loss_cnt = 0
# Save the best widget
torch.save(save_dict,
os.path.join(args.model_path, f'best_{args.task}_model.pth'))
else:
bad_loss_cnt += 1
if bad_loss_cnt >= config['training']['patience']:
return 0
else:
pass
return 0
if __name__ == '__main__':
_parser = argparse.ArgumentParser()
# cuda device
_parser.add_argument('-g', '--gpu', default='0', help='choose which GPU to use')
# path
_parser.add_argument('--config_file_path', help='path to json config', required=True)
_parser.add_argument('--model_path', type=str, default='./models/', help='path for saving trained models')
# widget
_parser.add_argument('--task', type=str, default='text', help='task type(only support text now).')
_parser.add_argument('--version', type=int, choices=[1, 2], help='dataset version.', required=True)
_parser.add_argument('--context_size', type=int, help='context size.', required=True)
_parser.add_argument('--batch_size', type=int, help='batch size.', required=True)
_args = _parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = _args.gpu
exit(main(_args))