-
Notifications
You must be signed in to change notification settings - Fork 30
/
bert_crf_model.py
executable file
·294 lines (248 loc) · 11.9 KB
/
bert_crf_model.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
import logging
import math
from copy import deepcopy
from typing import Any, List, Optional, Union, Callable
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import DataLoader, Dataset
from tqdm.auto import trange
from transformers import AdamW, get_linear_schedule_with_warmup, AutoTokenizer
from ..backbone import BertSeqTagger
from ..basemodel import BaseTorchSeqModel
from ..dataset.seqdataset import BaseSeqDataset
from ..utils import construct_collate_fn_trunc_pad
logger = logging.getLogger(__name__)
collate_fn = construct_collate_fn_trunc_pad('mask')
class BERTTorchSeqDataset(Dataset):
def __init__(self, dataset: BaseSeqDataset, tokenizer, max_seq_length, use_crf, n_data: Optional[int] = 0):
self.id2label = deepcopy(dataset.id2label)
self.label2id = deepcopy(dataset.label2id)
self.n_class = len(self.id2label)
if not use_crf:
self.dum_label = 'X'
self.label2id[self.dum_label] = len(self.id2label)
self.id2label.append(self.dum_label)
self.tokenizer = tokenizer
self.max_seq_length = max_seq_length # set to -1 when test
self.use_crf = use_crf
corpus = list(map(lambda x: x["text"], dataset.examples))
self.seq_len = list(map(len, corpus))
input_ids_tensor, input_mask_tensor, predict_mask_tensor = self.convert_corpus_to_tensor(corpus)
self.input_ids_tensor = input_ids_tensor
self.input_mask_tensor = input_mask_tensor
self.predict_mask_tensor = predict_mask_tensor
n_data_ = len(input_ids_tensor)
self.n_data_ = n_data_
if n_data > 0:
self.n_data = math.ceil(n_data / n_data_) * n_data_
else:
self.n_data = n_data_
def __len__(self):
return self.n_data
def convert_corpus_to_tensor(self, corpus):
input_ids_list = []
input_mask_list = []
predict_mask_list = []
max_seq_length = 0
for words in corpus:
predict_mask = []
input_mask = []
tokens = []
for i, w in enumerate(words):
sub_words = self.tokenizer.tokenize(w)
if not sub_words:
sub_words = [self.tokenizer.unk_token]
if self.use_crf:
''' if crf is used, then the padded token will be ignored '''
tokens.append(sub_words[0])
else:
tokens.extend(sub_words)
for j in range(len(sub_words)):
if j == 0:
input_mask.append(1)
predict_mask.append(1)
elif not self.use_crf: # These padding will hurt performance
''' '##xxx' -> 'X' (see bert paper, for non-crf model only) '''
input_mask.append(1)
predict_mask.append(0)
max_seq_length = max(max_seq_length, len(tokens))
input_ids_list.append(self.tokenizer.convert_tokens_to_ids(tokens))
input_mask_list.append(input_mask)
predict_mask_list.append(predict_mask)
max_seq_length = min(max_seq_length, self.max_seq_length)
n = len(input_ids_list)
for i in range(n):
ni = len(input_ids_list[i])
if ni > max_seq_length:
logger.info(f'Example is too long, length is {ni}, truncated to {max_seq_length}!')
input_ids_list[i] = input_ids_list[i][:max_seq_length]
input_mask_list[i] = input_mask_list[i][:max_seq_length]
predict_mask_list[i] = predict_mask_list[i][:max_seq_length]
else:
input_ids_list[i].extend([self.tokenizer.pad_token_id] * (max_seq_length - ni))
input_mask_list[i].extend([0] * (max_seq_length - ni))
predict_mask_list[i].extend([0] * (max_seq_length - ni))
input_ids_tensor = torch.LongTensor(input_ids_list)
input_mask_tensor = torch.LongTensor(input_mask_list)
predict_mask_tensor = torch.LongTensor(predict_mask_list)
return input_ids_tensor, input_mask_tensor, predict_mask_tensor
def prepare_labels(self, labels):
O_id = self.label2id['O']
if self.use_crf:
n, max_seq_len = self.predict_mask_tensor.shape
prepared_labels = np.ones((n, max_seq_len), dtype=int) * O_id
for i, labels_i in enumerate(labels):
ni = len(labels_i)
if ni > max_seq_len:
prepared_labels[i, :] = labels_i[:max_seq_len]
else:
prepared_labels[i, :ni] = labels_i
else:
prepared_labels = []
add_label_id = self.label2id[self.dum_label]
for labels_i, mask in zip(labels, self.predict_mask_tensor):
pre_labels = []
cnt = 0
n = len(labels_i)
for idx, flag in enumerate(mask):
if flag:
pre_labels.append(labels_i[cnt])
cnt += 1
else:
if n == cnt:
pre_labels.append(O_id)
else:
pre_labels.append(add_label_id)
prepared_labels.append(pre_labels)
return torch.LongTensor(prepared_labels)
def __getitem__(self, idx):
idx = idx % self.n_data_
d = {
'ids' : idx,
'input_ids' : self.input_ids_tensor[idx],
'attention_mask': self.input_mask_tensor[idx],
'mask' : self.predict_mask_tensor[idx],
}
return d
class BERTTaggerModel(BaseTorchSeqModel):
def __init__(self,
model_name: Optional[str] = 'bert-base-uncased',
lr: Optional[float] = 2e-5,
l2: Optional[float] = 1e-6,
max_tokens: Optional[int] = 512,
batch_size: Optional[int] = 32,
real_batch_size: Optional[int] = 32,
test_batch_size: Optional[int] = 128,
n_steps: Optional[int] = 10000,
use_crf: Optional[bool] = False,
fine_tune_layers: Optional[int] = -1,
lr_crf: Optional[float] = 5e-5,
l2_crf: Optional[float] = 1e-8,
):
super().__init__()
self.hyperparas = {
'fine_tune_layers': fine_tune_layers,
'model_name' : model_name,
'lr' : lr,
'l2' : l2,
'max_tokens' : max_tokens,
'batch_size' : batch_size,
'real_batch_size' : real_batch_size,
'test_batch_size' : test_batch_size,
'n_steps' : n_steps,
'use_crf' : use_crf,
'lr_crf' : lr_crf,
'l2_crf' : l2_crf,
}
self.model = None
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
def _init_valid_dataloader(self, dataset_valid: BaseSeqDataset) -> DataLoader:
torch_dataset = BERTTorchSeqDataset(dataset_valid, self.tokenizer, 512, self.hyperparas['use_crf'])
valid_dataloader = DataLoader(torch_dataset, batch_size=self.hyperparas['test_batch_size'], shuffle=False,
collate_fn=collate_fn)
return valid_dataloader
def fit(self,
dataset_train: BaseSeqDataset,
y_train: Optional[List[List]] = None,
dataset_valid: Optional[BaseSeqDataset] = None,
y_valid: Optional[List[List]] = None,
evaluation_step: Optional[int] = 50,
metric: Optional[Union[str, Callable]] = 'f1_seq',
strict: Optional[bool] = True,
direction: Optional[str] = 'auto',
patience: Optional[int] = 20,
tolerance: Optional[float] = -1.0,
device: Optional[torch.device] = None,
verbose: Optional[bool] = True,
**kwargs: Any):
if not verbose:
logger.setLevel(logging.ERROR)
self._update_hyperparas(**kwargs)
hyperparas = self.hyperparas
if hyperparas['real_batch_size'] == -1 or hyperparas['batch_size'] < hyperparas['real_batch_size']:
hyperparas['real_batch_size'] = hyperparas['batch_size']
accum_steps = hyperparas['batch_size'] // hyperparas['real_batch_size']
n_steps = hyperparas['n_steps']
torch_dataset = BERTTorchSeqDataset(dataset_train, self.tokenizer, self.hyperparas['max_tokens'],
self.hyperparas['use_crf'], n_data=n_steps * hyperparas['batch_size'])
train_dataloader = DataLoader(torch_dataset, batch_size=hyperparas['real_batch_size'], shuffle=True, collate_fn=collate_fn)
if y_train is None:
y_train = dataset_train.labels
y_train = torch_dataset.prepare_labels(y_train).to(device)
n_class = dataset_train.n_class
model = BertSeqTagger(n_class, **hyperparas).to(device)
self.model = model
param_optimizer = list(model.named_parameters())
crf_param = ['crf.transitions', ]
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if n not in crf_param]},
{'params' : [p for n, p in param_optimizer if n in crf_param], 'lr': hyperparas['lr_crf'],
'weight_decay': hyperparas['l2_crf']},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=hyperparas['lr'], weight_decay=hyperparas['l2'])
# Set up the learning rate scheduler
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=0, num_training_steps=n_steps)
valid_flag = self._init_valid_step(dataset_valid, y_valid, metric, strict, direction, patience, tolerance)
history = {}
last_step_log = {}
try:
with trange(n_steps, desc=f"[FINETUNE] {hyperparas['model_name']} Tagger", unit="steps", disable=not verbose, ncols=150, position=0, leave=True) as pbar:
cnt = 0
step = 0
model.train()
optimizer.zero_grad()
for batch in train_dataloader:
batch_idx = batch['ids'].to(device)
batch_label = y_train[batch_idx]
loss = model.calculate_loss(batch, batch_label)
loss.backward()
cnt += 1
if cnt % accum_steps == 0:
# Clip the norm of the gradients to 1.0.
nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step()
scheduler.step()
optimizer.zero_grad()
step += 1
if valid_flag and step % evaluation_step == 0:
metric_value, early_stop_flag, info = self._valid_step(step)
if early_stop_flag:
logger.info(info)
break
history[step] = {
'loss' : loss.item(),
f'val_{metric}' : metric_value,
f'best_val_{metric}': self.best_metric_value,
'best_step' : self.best_step,
}
last_step_log.update(history[step])
last_step_log['loss'] = loss.item()
pbar.update()
pbar.set_postfix(ordered_dict=last_step_log)
if step >= n_steps:
break
except KeyboardInterrupt:
logger.info(f'KeyboardInterrupt! do not terminate the process in case need to save the best model')
self._finalize()
return history