-
Notifications
You must be signed in to change notification settings - Fork 1
/
distillation.py
542 lines (450 loc) · 23.6 KB
/
distillation.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
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
import argparse
from datetime import datetime
from typing import Tuple, Any
import numpy as np
import torch
import torch.nn.functional as F
from dataclasses import dataclass
from torch import nn
from torch.utils.tensorboard import SummaryWriter
from torch.utils.data import DataLoader
from tqdm import tqdm
from transformers import (
BertTokenizer,
AdamW, BertForTokenClassification, BertConfig,
PreTrainedTokenizer
)
from data import get_bc2gm_train_data
from data import get_ner_model_inputs
from eval import evaluate_ner_metrics
from log import setup_logging
from loss import loss as ner_loss
from ner_utils import build_dict
from tags import UTIL_TAGS
import random
from copy import deepcopy
def set_seed(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
def setup_argparser() -> argparse.ArgumentParser:
parser = argparse.ArgumentParser()
parser.add_argument('--seed', default=10, type=int, required=False)
parser.add_argument('--measure_time', default=False, type=lambda x: bool(int(x)), required=False)
parser.add_argument('--do_train', default=False, type=lambda x: bool(int(x)), required=False)
parser.add_argument('--do_eval', default=False, type=lambda x: bool(int(x)), required=False)
parser.add_argument('--local_rank', default=-1, type=int, required=False)
parser.add_argument('--world_size', default=1, type=int, required=False)
parser.add_argument('--n_gpu', default=1, type=int, required=False)
parser.add_argument('--logging_level', default=20, type=int, required=False)
parser.add_argument('--teacher_model_name_or_path', default=None, type=str, required=False,
help="used in model_class.from_pretrained()")
parser.add_argument('--student_model_name_or_path', default=None, type=str, required=False,
help="used in model_class.from_pretrained()")
parser.add_argument('--teacher_checkpoint', default=None, type=str, required=False,
help="checkpoint to load the model from")
parser.add_argument('--student_checkpoint', default=None, type=str, required=False,
help="checkpoint to load the model from")
# data params
parser.add_argument('--path_to_train', default=None, type=str, required=False)
parser.add_argument('--path_to_val', default=None, type=str, required=False)
parser.add_argument('--batch_size', default=16, type=int, required=False)
parser.add_argument('--tokenizer_name', default=None, type=str, required=False)
parser.add_argument('--use_fast', default=True, type=lambda x: bool(int(x)), required=False,
help="whether to use fast tokenizer")
# model params
parser.add_argument("--task_name", default=None, type=str, required=True)
parser.add_argument('--model_name', default=None, type=str, required=True)
parser.add_argument('--weight_decay', default=0., type=float, required=False)
parser.add_argument('--lr_params', default=5e-5, type=float, required=False)
parser.add_argument('--scheduler', default='const', type=str, required=False)
parser.add_argument('--warmup_steps', default=0, type=int, required=False)
parser.add_argument('--gradient_accumulation_steps', default=1, type=int, required=False)
parser.add_argument('--max_grad_norm', default=1., type=float, required=False)
parser.add_argument("--max_steps", default=-1, type=int,
help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
parser.add_argument("--dropout", default=0.5, type=float, help="bilstm dropout")
parser.add_argument("--n_layers", default=2, type=int, help="bilstm number of layers")
parser.add_argument("--hidden_size", default=300, type=int, help="bilstm hidden size")
parser.add_argument("--classifier_size", default=256, type=int, help="bilstm classifier hidden size")
parser.add_argument('--embedding_type', default='train', type=str,
help='embeddings used in bilstm: train (nn.Embedding), bert or word2vec')
parser.add_argument('--embedding_size', default=300, type=int,
help='used when embedding is set to train')
# train params
parser.add_argument('--num_train_epochs', default=5, type=int, required=False)
parser.add_argument('--distillation', default=False, type=lambda x: bool(int(x)), required=True)
parser.add_argument('--alpha', default=0., type=float, required=False)
parser.add_argument('--logging_steps', default=5, type=int, required=False)
parser.add_argument('--eval_steps', default=5, type=int, required=False)
parser.add_argument('--write', default=True, type=lambda x: bool(int(x)), required=False,
help="Write logs to summary writer")
parser.add_argument('--save_steps', type=int, default=10,
help="Save last checkpoint every X update steps")
parser.add_argument('--update_steps_start', type=int, default=0,
help="when using pretrained model enter how many update steps it already underwent")
parser.add_argument('--comment', type=str, default=None, help='additional info to log')
return parser
@dataclass
class BiLSTMConfig:
n_layers: int
embedding_size: int
hidden_size: int
dropout: float
classifier_size: int
class MultiChannelEmbedding(nn.Module):
def __init__(self, vocab_size, embedding_size, out_channels, filters: list):
# filters must each be an odd number otherwise token number is lost
# (might be that convolutions work only for sequence classification)
super().__init__()
self.filters_size = out_channels
self.filters = filters
self.embedding = nn.Embedding(vocab_size, embedding_size)
self.conv = nn.ModuleList([
nn.Conv1d(embedding_size, out_channels, kernel_size=filter, padding=filter // 2)
for filter in filters
])
def init_embedding(self, weight):
self.embedding.weight = nn.Parameter(weight.to(self.embedding.weight.device))
def forward(self, input_ids, **kwargs):
emb = self.embedding(input_ids).transpose(1, 2)
filters = []
for conv1d in self.conv:
filters.append(conv1d(emb).transpose(1, 2))
out = F.relu(torch.cat(filters, dim=2))
return out
class BiLSTMForTokenClassification(nn.Module):
def __init__(self, config: BiLSTMConfig, vocab_size: int, n_classes: int, device: torch.device, embedding,
bert=None):
super().__init__()
self.embedding_type = embedding
if embedding == 'bert' and bert is not None:
self.embedding = deepcopy(bert.bert.embeddings)
elif embedding == 'multichannel':
self.embedding = MultiChannelEmbedding(vocab_size=vocab_size,
embedding_size=config.embedding_size,
out_channels=256,
filters=[1, 3, 5])
config.embedding_size = len(self.embedding.filters) * self.embedding.filters_size
else:
self.embedding = nn.Embedding(vocab_size, config.embedding_size)
self.lstm = nn.LSTM(input_size=config.embedding_size,
hidden_size=config.hidden_size,
num_layers=config.n_layers,
dropout=config.dropout,
batch_first=True,
bidirectional=True)
if not config.classifier_size:
self.linear = nn.Linear(2 * config.hidden_size, n_classes)
else:
self.linear = nn.Sequential(
nn.Linear(2 * config.hidden_size, config.classifier_size),
nn.ReLU(),
nn.Dropout(config.dropout),
nn.Linear(config.classifier_size, n_classes)
)
self.dropout = nn.Dropout(0.5)
self.device = device
def forward(self, input_ids: torch.Tensor, **kwargs):
emb = self.embedding(input_ids)
out, _ = self.lstm(emb)
out = self.linear(self.dropout(out))
return BiLSTMOutput(logits=out)
def forward_(self, input_ids: torch.Tensor, **kwargs):
emb = self.embedding(input_ids)
out, (_, _) = self.lstm(emb)
out = self.linear(out)
return BiLSTMOutput(logits=out)
def write_params(writer, args):
output = f'Teacher model path {args.teacher_model_name_or_path} \n'
output += f'Student model path {args.student_model_name_or_path} \n'
output += f'Task name {args.task_name} \n'
output += f'Use distillation {args.distillation} \n'
output += f'Epochs {args.num_train_epochs} \n'
output += f'Learning rate {args.lr_params} \n'
output += f'Weight decay {args.weight_decay} \n'
output += f'Batch size {args.batch_size} \n'
output += f'Distillation alpha {args.alpha} \n'
output += f'BiLSTM number of layers {args.n_layers} \n'
output += f'BiLSTM hidden size {args.hidden_size} \n'
output += f'BiLSTM embedding type {args.embedding_type} \n'
output += f'BiLSTM embedding size {args.embedding_size} \n'
output += f'BiLSTM classifier size {args.classifier_size} \n'
output += f'BiLSTM dropout rate {args.dropout} \n'
output += f'Start update steps {args.update_steps_start} \n'
if args.comment is not None:
output += args.comment
writer.add_text('Parameters', output, args.update_steps_start)
def init_model(args):
if args.embedding_type == 'bert':
args.embedding_size = 768
label2id = build_dict(UTIL_TAGS + ['GENE'], ['B-', 'I-', 'E-', 'S-'])
label_map = {value: key for key, value in label2id.items()}
num_labels = len(label_map)
args.label2id = label2id
args.label_map = label_map
teacher_model = None
teacher_tokenizer = None
teacher_config = None
if args.distillation and (args.do_train or args.embedding_type == 'bert' or (args.measure_time
and args.teacher_checkpoint)):
teacher_config = BertConfig.from_pretrained(
args.teacher_model_name_or_path,
num_labels=num_labels,
id2label=label_map,
label2id=label2id,
)
teacher_model = BertForTokenClassification.from_pretrained(args.teacher_model_name_or_path,
config=teacher_config).to(
args.device)
if args.teacher_checkpoint is not None:
teacher_model.load_state_dict(torch.load(args.teacher_checkpoint))
teacher_model.eval()
teacher_tokenizer = BertTokenizer.from_pretrained(args.teacher_model_name_or_path)
if args.model_name.lower() == 'bilstm':
config = BiLSTMConfig(
n_layers=args.n_layers,
embedding_size=args.embedding_size,
hidden_size=args.hidden_size,
dropout=args.dropout,
classifier_size=args.classifier_size
)
student_tokenizer = BertTokenizer.from_pretrained(args.teacher_model_name_or_path)
if args.embedding_type == 'bert' and args.distillation:
model = BiLSTMForTokenClassification(config, student_tokenizer.vocab_size, num_labels, args.device,
bert=teacher_model, embedding=args.embedding_type).to(args.device)
else:
model = BiLSTMForTokenClassification(config, student_tokenizer.vocab_size, num_labels, args.device,
bert=None, embedding=args.embedding_type).to(args.device)
else:
config = BertConfig(
attention_probs_dropout_prob=0.1,
cell={},
model_type="bert",
hidden_act="gelu",
hidden_dropout_prob=0.1,
hidden_size=312,
initializer_range=0.02,
intermediate_size=1200,
max_position_embeddings=512,
num_attention_heads=12,
num_hidden_layers=4,
pre_trained="",
structure=[],
type_vocab_size=2,
vocab_size=28996,
num_labels=num_labels,
id2label=label_map,
label2id=label2id,
)
student_tokenizer = BertTokenizer.from_pretrained(args.teacher_model_name_or_path)
model = TinyBertForTokenClassification(config=config, device=args.device).to(args.device)
if args.student_checkpoint is not None:
model.load_state_dict(torch.load(args.student_checkpoint))
return student_tokenizer, model, teacher_tokenizer, teacher_model
class DistillLoss:
def __init__(self, loss_func: callable, teacher: nn.Module, alpha: float):
self.teacher = teacher
self.alpha = alpha
self.loss_func = loss_func
def get_loss(self, output, batch: dict):
with torch.no_grad():
teacher_output = self.teacher(**batch, output_hidden_states=True, output_attentions=True)
loss = self.loss_func(output.logits.cpu(), batch['labels'].cpu(), batch['attention_mask'].cpu())
loss_distill = F.mse_loss(teacher_output.logits, output.logits)
if output.attentions and output.hidden_states:
loss_emb = F.mse_loss(teacher_output.hidden_states[0], output.hidden_states[0])
hidden_aligned = [teacher_output.hidden_states[i] for i in range(1, len(teacher_output.hidden_states), 3)]
loss_hid, loss_attn = 0., 0.
for teacher_hid, student_hid in zip(hidden_aligned, output.hidden_states):
loss_hid += F.mse_loss(teacher_hid, student_hid)
attn_aligned = [teacher_output.attentions[i] for i in range(0, len(teacher_output.attentions), 3)]
for teacher_attn, student_attn in zip(attn_aligned, output.attentions):
loss_attn += F.mse_loss(teacher_attn, student_attn)
loss_distill += loss_emb + loss_hid + loss_attn
return args.alpha * loss + (1 - args.alpha) * loss_distill
class NoDistillLoss:
def __init__(self, loss_func: callable, *args, **kwargs):
self.loss_func = loss_func
def get_loss(self, logits: torch.Tensor, batch: dict):
return self.loss_func(logits.cpu(), batch['labels'].cpu(), batch['attention_mask'].cpu())
class NerData:
def __init__(self, student_tokenizer: PreTrainedTokenizer,
teacher_tokenizer: PreTrainedTokenizer,
tags_vocab: dict):
self.student_tokenizer = student_tokenizer
self.teacher_tokenizer = teacher_tokenizer
self.tags_vocab = tags_vocab
def get_inputs(self, batch, teacher=False):
return get_ner_model_inputs(batch, self.student_tokenizer if not teacher else self.teacher_tokenizer,
self.tags_vocab)
class Data:
def get_inputs(self, batch):
return batch
class Evaluator:
def __init__(self, model: nn.Module, eval_func: callable, val_dataloader: DataLoader,
writer: SummaryWriter = None, saving_name=None):
self.model = model
self.writer = writer
self.eval_func = eval_func
self.val_dataloader = val_dataloader
self.best_f1 = -np.inf
self.saving_name = saving_name
def evaluate_and_write(self, update_steps: int, **kwargs):
self.model.eval()
val_loss, metrics = self.eval_func(self.model, self.val_dataloader, **kwargs)
if self.writer is not None:
self.writer.add_scalar('Losses/val', val_loss, update_steps)
for name, metric in metrics.items():
self.writer.add_scalar(f'Metrics/{name}_dev', metric, update_steps)
if metrics['f1'] > self.best_f1 and self.saving_name is not None:
torch.save(self.model.state_dict(), f'{self.saving_name}_best.pt')
self.best_f1 = metrics['f1']
return val_loss, metrics
class TinyBertForTokenClassification(nn.Module):
def __init__(self, config, fit_size=768, device: torch.device = torch.device("cuda")):
super(TinyBertForTokenClassification, self).__init__()
self.bert = BertForTokenClassification(config=config)
self.W_emb = nn.Linear(config.hidden_size, fit_size)
self.W_hidden = nn.Linear(config.hidden_size, fit_size)
self.device = device
def forward(self, input_ids, **kwargs):
out = self.bert(input_ids, **kwargs, output_hidden_states=True, output_attentions=True)
logits, hidden, attentions = out.logits, out.hidden_states, out.attentions
hidden = tuple([self.W_emb(hidden[0])] + [self.W_hidden(hidden_el) for hidden_el in hidden[1:]])
return TinyBertOutput(logits=logits, hidden_states=hidden, attentions=attentions)
@dataclass
class TinyBertOutput:
logits: torch.Tensor
hidden_states: Tuple[Any]
attentions: Tuple[Any]
@dataclass
class BiLSTMOutput:
logits: torch.Tensor
hidden_states: Tuple[Any] = tuple()
attentions: Tuple[Any] = tuple()
def train(args):
assert args.model_name.lower() in ['bilstm', 'tinybert'], "Model not supported"
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
args.device = device
saving_name = f'{args.model_name}-{datetime.now():%Y%m%d-%H%M-%S}'
student_tokenizer, model, teacher_tokenizer, teacher_model = init_model(args)
optimizer = AdamW(model.parameters(), lr=args.lr_params, weight_decay=args.weight_decay)
# student tokenizer used
train_dataloader, val_dataloader = get_bc2gm_train_data(args, student_tokenizer,
args.label2id, return_train=True, return_val=True)
teacher_train_dataloader = [[] for _ in range(len(train_dataloader))]
if args.student_model_name_or_path:
# teacher tokenizer used
teacher_train_dataloader, _ = get_bc2gm_train_data(args, teacher_tokenizer,
args.label2id, return_train=True, return_val=False)
update_steps = args.update_steps_start
set_seed(args)
writer = None
if args.write:
writer = setup_logging(saving_name)
write_params(writer, args)
tags_vocab = {value: key for key, value in args.label_map.items()}
loss_func = ner_loss if args.task_name == 'ner' else None
eval_func = evaluate_ner_metrics if args.task_name == 'ner' else None
loss_cls = DistillLoss(loss_func, teacher_model, args.alpha) if args.distillation else NoDistillLoss(loss_func)
data_cls = NerData(student_tokenizer, teacher_tokenizer, tags_vocab) if args.task_name == 'ner' else Data()
evaluator = Evaluator(model, eval_func, val_dataloader, writer, saving_name)
# eval
_, metrics = evaluator.evaluate_and_write(update_steps,
label_map=args.label_map, tokenizer=student_tokenizer)
for epoch in range(args.num_train_epochs):
epoch_iterator = tqdm(zip(train_dataloader, teacher_train_dataloader),
desc="Train iteration", position=0, leave=True, total=len(train_dataloader))
for step, (batch, teacher_batch) in enumerate(epoch_iterator):
model.train()
batch = data_cls.get_inputs(batch)
batch = {key: value.to(model.device) for key, value in batch.items()}
output = model(**batch)
if teacher_batch:
teacher_batch = data_cls.get_inputs(teacher_batch, teacher=True)
teacher_batch = {key: value.to(model.device) for key, value in teacher_batch.items()}
loss = loss_cls.get_loss(output, batch if not teacher_batch else teacher_batch)
if args.gradient_accumulation_steps > 1:
loss /= args.gradient_accumulation_steps
loss.backward()
if step % args.gradient_accumulation_steps == 0:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
optimizer.zero_grad()
update_steps += 1
if writer is not None:
writer.add_scalar('Losses/train',
loss.item(), update_steps)
if update_steps % args.eval_steps == 0:
_, metrics = evaluator.evaluate_and_write(update_steps,
label_map=args.label_map, tokenizer=student_tokenizer)
if update_steps % args.save_steps:
torch.save(model.state_dict(), f'{saving_name}_last.pt')
def eval(args):
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
args.device = device
tokenizer, model, _, _ = init_model(args)
model.eval()
writer = None
_, val_dataloader = get_bc2gm_train_data(args, tokenizer, args.label2id, return_train=False, return_val=True)
# eval
eval_func = evaluate_ner_metrics if args.task_name == 'ner' else None
evaluator = Evaluator(model, eval_func, val_dataloader, writer)
_, metrics = evaluator.evaluate_and_write(0,
label_map=args.label_map, tokenizer=tokenizer)
print(metrics)
def measure_time(args):
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
args.device = device
args.batch_size = 1
tokenizer, model, teacher_tokenizer, teacher_model = init_model(args)
model.eval()
if teacher_model is not None:
teacher_model.eval()
_, val_dataloader = get_bc2gm_train_data(args, tokenizer, args.label2id, return_train=False, return_val=True)
if teacher_tokenizer is not None:
_, teacher_val_dataloader = get_bc2gm_train_data(args, teacher_tokenizer, args.label2id,
return_train=False, return_val=True)
else:
teacher_val_dataloader = val_dataloader
tags_vocab = {value: key for key, value in args.label_map.items()}
data_cls = NerData(tokenizer, teacher_tokenizer, tags_vocab) if args.task_name == 'ner' else Data()
starter, ender = torch.cuda.Event(enable_timing=True), torch.cuda.Event(enable_timing=True)
# gpu warm-up
for batch in tqdm(val_dataloader, desc="GPU warm-up"):
batch = data_cls.get_inputs(batch)
_ = model(**{key: value.to(device) for key, value in batch.items()})
student_times = torch.zeros(len(val_dataloader))
teacher_times = torch.zeros(len(val_dataloader))
with torch.no_grad():
for i, batch in enumerate(tqdm(val_dataloader, desc="Evaluation")):
batch = data_cls.get_inputs(batch)
batch = {key: value.to(device) for key, value in batch.items()}
starter.record()
_ = model(**batch)
ender.record()
torch.cuda.synchronize()
student_times[i] = starter.elapsed_time(ender)
if teacher_model is not None:
with torch.no_grad():
for i, batch in enumerate(tqdm(teacher_val_dataloader, desc="Teacher")):
batch = data_cls.get_inputs(batch)
batch = {key: value.to(device) for key, value in batch.items()}
starter.record()
_ = teacher_model(**batch)
ender.record()
torch.cuda.synchronize()
teacher_times[i] = starter.elapsed_time(ender)
print(f"Teacher time: mean = {teacher_times.mean()}, std = {teacher_times.std()}")
print(f"Student time: mean = {student_times.mean()}, std = {student_times.std()}")
if __name__ == "__main__":
args = setup_argparser().parse_args()
if args.measure_time:
measure_time(args)
elif args.do_train:
train(args)
else:
eval(args)