-
Notifications
You must be signed in to change notification settings - Fork 4
/
train.py
162 lines (136 loc) · 6.33 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
import argparse
import copy
import datetime
import logging
import numpy as np
import torch
import torchtext
from torch import nn, optim
from tensorboardX import SummaryWriter
from modules.classifier import Classifier
from utils.tools import load_config, load_data
from utils.tools import set_logger, set_seed
from utils.tools import save_model, print_params, print_args
from test import test
def train(args, data, cv=None):
model = Classifier(args, data.WORD.vocab.vectors).to(torch.device(args.device))
print_params(model)
parameters = filter(lambda p: p.requires_grad, model.parameters())
if args.optimizer.endswith('W'):
optimizer = getattr(optim, args.optimizer[:-1])(parameters, lr=args.learning_rate)
else:
optimizer = getattr(optim, args.optimizer)(parameters,
lr=args.learning_rate,
weight_decay=args.weight_decay)
criterion = nn.CrossEntropyLoss()
log_dir = f'runs/{args.dataset}/{args.model_time}'
if cv: log_dir += f'/{cv}'
writer = SummaryWriter(log_dir=log_dir)
loss, iterations = 0, 0
dev_loss, dev_acc = 0, 0
max_dev_acc, max_test_acc = 0, 0
iterator = data.train_iter
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer,
T_max=args.epoch * len(iterator),
eta_min=args.eta_min)
for e in range(1, args.epoch + 1):
train_acc, train_size = 0, 0
logging.info(f'Epoch: {e}')
for i, batch in enumerate(iterator):
iterations += 1
model.train()
pred = model(batch)
train_acc += (pred.max(dim=1)[1] == batch.label).sum().float()
train_size += len(pred)
optimizer.zero_grad()
batch_loss = criterion(pred, batch.label)
loss += batch_loss.item()
if args.optimizer.endswith('W'):
for group in optimizer.param_groups:
for param in group['params']:
param.data = param.data.add(-args.weight_decay * args.learning_rate, param.data)
batch_loss.backward()
nn.utils.clip_grad_norm_(parameters, max_norm=args.max_grad_norm)
optimizer.step()
scheduler.step(iterations - 1)
if iterations % args.print_freq == 0:
c = iterations // args.print_freq
writer.add_scalar('loss/train', loss, c)
if hasattr(data, 'dev_iter'):
dev_loss, dev_acc = test(model, data, mode='dev')
writer.add_scalar('loss/dev', dev_loss, c)
writer.add_scalar('acc/dev', dev_acc, c)
test_loss, test_acc = test(model, data, mode='test')
writer.add_scalar('loss/test', test_loss, c)
writer.add_scalar('acc/test', test_acc, c)
logging.info(f'L: {loss:.3f} / DL: {dev_loss:.3f} / TL: {test_loss:.3f} '
f'/ DAcc: {dev_acc:.3f} / TAcc: {test_acc:.3f}')
loss = 0
if hasattr(data, 'dev_iter'):
if dev_acc > max_dev_acc:
max_dev_acc = dev_acc
max_test_acc = test_acc
best_model = copy.deepcopy(model)
else:
if test_acc > max_test_acc:
max_test_acc = test_acc
best_model = copy.deepcopy(model)
train_acc = (train_acc / train_size).cpu().item()
logging.info(f'Train_acc: {train_acc:.3f}')
writer.add_scalar('acc/train', train_acc, e)
writer.close()
logging.info(f'max dev acc: {max_dev_acc:.3f} / max test acc: {max_test_acc:.3f}')
return best_model, max_dev_acc, max_test_acc
def cross_validation(args, data):
fields = [('text', data.WORD), ('label', data.LABEL), ('transitions', data.PARSE),
('word_tag', data.WORD_TAG), ('cons_tag', data.CONS_TAG)]
results = []
for cv in range(10):
logging.info(f'---CV {cv+1}---')
split_start = (len(data.examples) // 10) * cv
split_end = (len(data.examples) // 10) * (cv + 1)
data.train = torchtext.data.Dataset(
examples=data.examples[:split_start] + data.examples[split_end:],
fields=fields)
data.test = torchtext.data.Dataset(
examples=data.examples[split_start:split_end],
fields=fields)
data.train_iter, data.test_iter = torchtext.data.BucketIterator.splits(
(data.train, data.test),
batch_sizes=[args.batch_size] * 2,
device=args.device,
sort_key=lambda x: len(x.text))
best_model, _, max_test_acc = train(args, data, cv + 1)
save_model(args, best_model, None, max_test_acc, cv=cv + 1)
results.append(max_test_acc)
logging.info(f'Averaged test acc: {np.mean(results):.3f}')
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', default='SST2',
help='options: SST2, SST5, MR, SUBJ, TREC')
parser.add_argument('--optimizer', default='AdadeltaW',
help='options: Adadelta, AdadeltaW, Adam, AdamW')
parser.add_argument('--use-leafLSTM', default=1, type=int,
help='options: 0==FF, 1==LSTM, 2==bi-LSTM')
parser.add_argument('--gpu', default=0, type=int)
args = parser.parse_args()
setattr(args, 'device', f'cuda:{args.gpu}' if torch.cuda.is_available() and args.gpu >= 0 else 'cpu')
setattr(args, 'model_time', datetime.datetime.now().strftime('%Y%m%d-%H:%M:%S'))
set_logger(args)
logging.info(f'Start time: {args.model_time}')
logging.info('Loading saved config from yaml...')
args = load_config(args)
set_seed(args.random_seed)
logging.info(f'Dataset: {args.dataset}')
logging.info('Loading data...')
args, data = load_data(args)
logging.info('Training start!')
print_args(args)
if not hasattr(data, 'test_iter'): # cross validation (no test data)
cross_validation(args, data)
else:
best_model, max_dev_acc, max_test_acc = train(args, data)
save_model(args, best_model, max_dev_acc, max_test_acc)
logging.info('Training finished!')
if __name__ == '__main__':
main()