-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
executable file
·169 lines (126 loc) · 6.01 KB
/
main.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
from src.config import get_params
from src.utils import init_experiment
from src.dataloader import get_dataloader, get_conll2003_dataloader, get_dataloader_for_bilstmtagger
from src.trainer import BaseTrainer
from src.model import BertTagger, BiLSTMTagger
from src.coach.dataloader import get_dataloader_for_coach
from src.coach.model import EntityPredictor
from src.coach.trainer import CoachTrainer
import torch
import numpy as np
from tqdm import tqdm
import random
def random_seed(seed):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
def train(params):
# initialize experiment
logger = init_experiment(params, logger_filename=params.logger_filename)
if params.bilstm:
# dataloader
# dataloader_train, dataloader_dev, dataloader_test, vocab = get_dataloader_for_bilstmtagger(params)
dataloader_train, dataloader_dev, dataloader_test = get_dataloader(params)
# bilstm-crf model
model = BiLSTMTagger(params)
model.cuda()
# trainer
trainer = BaseTrainer(params, model)
elif params.coach:
# dataloader
dataloader_train, dataloader_dev, dataloader_test, vocab = get_dataloader_for_coach(params)
# coach model
binary_tagger = BiLSTMTagger(params, vocab)
entity_predictor = EntityPredictor(params)
binary_tagger.cuda()
entity_predictor.cuda()
# trainer
trainer = CoachTrainer(params, binary_tagger, entity_predictor)
else:
# dataloader
dataloader_train, dataloader_dev, dataloader_test = get_dataloader(params)
# BERT-based NER Tagger
model = BertTagger(params)
model.cuda()
# trainer
trainer = BaseTrainer(params, model)
n_params = sum([p.nelement() for p in model.parameters()])
print('n_params', n_params)
logger.info("*** paramaters ***" + str(n_params))
if params.conll and not params.joint:
conll_trainloader, conll_devloader, conll_testloader = get_conll2003_dataloader(params.batch_size, params.tgt_dm)
trainer.train_conll(conll_trainloader, conll_devloader, conll_testloader, params.tgt_dm)
no_improvement_num = 0
best_f1 = 0
logger.info("Training on target domain ...")
dev_target = []
test_target = []
test_detail_results = []
dev_detail_results = []
for e in range(params.epoch):
logger.info("============== epoch %d ==============" % e)
pbar = tqdm(enumerate(dataloader_train), total=len(dataloader_train))
# pbar = enumerate(dataloader_train)
if params.bilstm:
loss_list = []
for i, (X, lengths, y) in pbar:
X, lengths = X.cuda(), lengths.cuda()
loss = trainer.train_step_for_bilstm(X, lengths, y)
loss_list.append(loss)
pbar.set_description("(Epoch {}) LOSS:{:.4f}".format(e, np.mean(loss_list)))
logger.info("Finish training epoch %d. loss: %.4f" % (e, np.mean(loss_list)))
elif params.coach:
loss_bin_list, loss_entity_list = [], []
for i, (X, lengths, y_bin, y_final) in pbar:
X, lengths = X.cuda(), lengths.cuda()
loss_bin, loss_entityname = trainer.train_step(X, lengths, y_bin, y_final)
loss_bin_list.append(loss_bin)
loss_entity_list.append(loss_entityname)
pbar.set_description("(Epoch {}) LOSS BIN:{:.4f}; LOSS ENTITY:{:.4f}".format(e, np.mean(loss_bin_list),
np.mean(loss_entity_list)))
logger.info("Finish training epoch %d. loss_bin: %.4f. loss_entity: %.4f" % (
e, np.mean(loss_bin_list), np.mean(loss_entity_list)))
else:
loss_list = []
for i, (X, y) in pbar:
X, y = X.cuda(), y.cuda()
loss = trainer.train_step(X, y)
loss_list.append(loss)
pbar.set_description("(Epoch {}) LOSS:{:.4f}".format(e, np.mean(loss_list)))
logger.info("Finish training epoch %d. loss: %.4f" % (e, np.mean(loss_list)))
logger.info("============== Evaluate epoch %d on Train Set ==============" % e)
f1_train, report_results = trainer.evaluate(dataloader_train, params.tgt_dm, use_bilstm=params.bilstm)
# logger.info("\n%s", report_results)
logger.info("Evaluate on Train Set. F1: %.4f." % f1_train)
logger.info("============== Evaluate epoch %d on Dev Set ==============" % e)
f1_dev, report_results = trainer.evaluate(dataloader_dev, params.tgt_dm, use_bilstm=params.bilstm)
# logger.info("\n%s", report_results)
logger.info("Evaluate on Dev Set. F1: %.4f." % f1_dev)
dev_target.append(f1_dev)
dev_detail_results.append(report_results)
logger.info("============== Evaluate epoch %d on Test Set ==============" % e)
f1_test, report_results = trainer.evaluate(dataloader_test, params.tgt_dm, use_bilstm=params.bilstm)
logger.info("\n%s", report_results)
logger.info("Evaluate on Test Set. F1: %.4f." % f1_test)
test_target.append(f1_test)
test_detail_results.append(report_results)
if f1_dev > best_f1:
logger.info("Found better model!!")
best_f1 = f1_dev
no_improvement_num = 0
# trainer.save_model()
else:
no_improvement_num += 1
logger.info("No better model found (%d/%d)" % (no_improvement_num, params.early_stop))
if no_improvement_num >= params.early_stop:
break
dev_max_index = dev_target.index(max(dev_target))
logger.info("Best model basd on dev data.")
logger.info("\n%s", test_detail_results[dev_max_index])
logger.info("Best model basd on dev data. F1: %.4f." % test_target[dev_max_index])
if __name__ == "__main__":
params = get_params()
random_seed(params.seed)
train(params)