forked from RUCAIBox/TextBox
-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
571 lines (485 loc) · 24.5 KB
/
trainer.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
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
import collections
import os
from accelerate.logging import get_logger
from typing import Optional, Union, List, Dict
import math
import torch
import torch.optim as optim
import transformers
from accelerate import Accelerator
from accelerate.utils import set_seed
from torch.utils.data import DataLoader
from tqdm import tqdm
from textbox import Config
from textbox.utils.dashboard import get_dashboard, Timestamp, EpochTracker
from .scheduler import (
AbstractScheduler, InverseSquareRootScheduler, CosineScheduler, LinearScheduler, ConstantScheduler
)
from torch.utils.data import DataLoader
from ..evaluator import BaseEvaluator
from ..model.abstract_model import AbstractModel
from ..utils import serialized_save, init_seed
from ..utils.enum_type import PLM_MODELS, RNN_MODELS
class AbstractTrainer:
r"""Trainer Class is used to manage the training and evaluation processes of text generation system models.
AbstractTrainer is an abstract class in which the fit() and evaluate() method should be implemented according
to different training and evaluation strategies.
"""
def __init__(self, config: Config, model: AbstractModel):
self.config = config
self.model = model
self.logger = get_logger(__name__)
def fit(self, train_data: DataLoader):
r"""Train the model based on the train data.
"""
raise NotImplementedError('Method `fit()` should be implemented.')
def evaluate(self, eval_data: DataLoader):
r"""Evaluate the model based on the eval data.
"""
raise NotImplementedError('Method `evaluate()` should be implemented.')
class Trainer(AbstractTrainer):
r"""The basic Trainer for basic training and evaluation strategies in text generation systems.
This class defines common functions for training and evaluation processes of most text generation system models,
including `fit()`, `evaluate()`, `resume_checkpoint()` and some other features helpful for model training and
evaluation.
Generally speaking, this class can serve most text generation system models, If the training process of the model
is to simply optimize a single loss without involving any complex training strategies, such as adversarial learning,
pre-training and so on.
Initializing the Trainer needs two parameters: `config` and `model`. `config` records the parameters' information
for controlling training and evaluation, such as `learning_rate`, `epochs` and so on.
More information can be found in [placeholder]. `model` is the instantiated object of a Model Class.
"""
def __init__(self, config: Config, model: AbstractModel, accelerator: Accelerator):
super(Trainer, self).__init__(config, model)
self.device: torch.device = config['device']
self.filename = config['filename']
self.post_processing = config['post_processing']
self.accelerator = accelerator
# Optimization strategy
self.learning_rate = config['learning_rate']
self.optimizer_kwargs = {'lr': config['learning_rate']}
self.optimizer_kwargs.update(config['optimizer_kwargs'])
self.adafactor_kwargs = config['adafactor_kwargs']
self.scheduler_kwargs = config['scheduler_kwargs']
self.grad_clip = config['grad_clip']
self._trainable_parameters = filter(lambda x: x.requires_grad, self.model.parameters())
self.optimizer = self._build_optimizer(config['optimizer'], config['scheduler'])
self.accumulation_steps = config['accumulation_steps']
# Training strategy
self.quick_test = bool(config['quick_test'])
self.max_steps = config['max_steps'] or 10000000000 # max training batch step
self.start_epoch = 1
r"""Start epoch index. That is, `epoch_idx` iterates through `range(self.start_epoch, self.epochs)`"""
self.epochs = config['epochs'] if self.max_steps == 10000000000 else 10000000000
r"""End epoch index + 1, aka max iteration times. That is, `epoch_idx` iterates through
`range(self.start_epoch, self.epochs)`"""
self.valid_steps = self.config['valid_steps']
self.valid_strategy = self.config['valid_strategy']
self._valid_count = 0
self.train_loss_list: List[float] = list()
self.valid_result_dict: Dict[int, EpochTracker] = dict()
self.stopping_steps = config['stopping_steps']
self.stopped = False
self.stopping_count = 0
# Evaluation strategy
self.metrics_for_best_model = set(self.config["metrics_for_best_model"])
self.evaluator = BaseEvaluator(config, self.config["metrics"])
# Functionality
self.saved_dir = os.path.join(config['saved_dir'], self.filename)
self.saved_model_filename = os.path.join(self.saved_dir, 'checkpoint')
self.saved_text_filename: str = os.path.join(self.saved_dir, 'generation.txt')
self.max_save = config['max_save'] if config['max_save'] is not None else 2
if self.max_save == 0:
# The saved checkpoint will be deleted at the end of experiment
self.logger.warning('max_save has been set to 0. None of the checkpoint will be saved.')
self.max_save = 1
self.disable_tqdm = config['disable_tqdm'] or not self.accelerator.is_local_main_process
self._summary_tracker = get_dashboard()
def _build_optimizer(self, optimizer: str, scheduler: Optional[str])\
-> Union[optim.Optimizer, AbstractScheduler]:
"""Init the optimizer and scheduler.
Returns:
Union[optim.Optimizer, AbstractScheduler]: the optimizer
"""
optimizer_class = collections.defaultdict(
lambda: optim.AdamW, {
'adam': optim.Adam,
'adamw': optim.AdamW,
'sgd': optim.SGD,
'adagrad': optim.Adagrad,
'rmsprop': optim.RMSprop,
'adafactor': transformers.Adafactor,
}
)
scheduler_class = {
'inverse': InverseSquareRootScheduler,
'cosine': CosineScheduler,
'linear': LinearScheduler,
'constant': ConstantScheduler,
}
# dealing with adafactor
if optimizer == 'adafactor':
# using adafactor_kwargs in overall.yaml
if self.grad_clip is not None:
self.grad_clip = None
self.logger.warning(
"Additional optimizer operations like gradient clipping "
"should not be used alongside Adafactor."
)
self.optimizer_kwargs.update(self.adafactor_kwargs)
# get optimizer (use default value of pytorch if self.optimizer_kwargs is empty)
self.logger.debug(f'Using optimizer {optimizer}')
optimizer = optimizer_class[optimizer](params=self._trainable_parameters, **self.optimizer_kwargs)
# scheduling
if scheduler is not None and scheduler in scheduler_class:
assert isinstance(self.scheduler_kwargs, dict), "Please specify scheduler_kwargs"
self.logger.debug(f'Using scheduler {scheduler}.')
self.scheduler_kwargs.setdefault("max_lr", self.learning_rate)
optimizer = scheduler_class[scheduler](base_optimizer=optimizer, **self.scheduler_kwargs)
return optimizer
@property
def timestamp(self) -> Timestamp:
"""Return the timestamp for the moment."""
return self._summary_tracker.axes
@property
def best_valid_result(self) -> EpochTracker:
"""Retrieve best result dict from `self.valid_result_list`."""
return self.valid_result_dict[self.best_valid_timestamp.valid_epoch]
@property
def best_valid_timestamp(self) -> Timestamp:
"""Retrieve timestamp of best valid result."""
return self._summary_tracker.best_valid_timestamp
def _train_epoch(
self,
train_data: DataLoader,
epoch_idx: int,
valid_data: Optional[DataLoader] = None,
) -> dict:
r"""Train the model in an epoch
Args:
train_data:
epoch_idx: the current epoch index.
valid_data: Optional (default = None) the dataloader of validation set
Returns:
dict: Training losses.
"""
self.model.train()
if not self.disable_tqdm:
train_data_len = math.ceil(len(train_data) / self.accumulation_steps)
train_tqdm = tqdm(
range(train_data_len),
desc=f"train {epoch_idx:4}",
dynamic_ncols=True,
postfix={'loss': None},
unit='step'
)
with self._summary_tracker.new_epoch('train'):
for step, data in enumerate(train_data):
with self.accelerator.accumulate(self.model):
if step % self.accumulation_steps == 0:
self._summary_tracker.new_step()
if self.timestamp.train_step == self.max_steps + 1:
self.stopped = True
break
self.optimizer.zero_grad()
loss = self.model(data, epoch_idx=epoch_idx)
# avg_loss = self.accelerator.gather(loss).mean().item()
avg_loss = loss.item()
self._summary_tracker.append_loss(avg_loss)
self.accelerator.backward(loss)
if self.accelerator.sync_gradients and self.grad_clip is not None:
self.accelerator.clip_grad_norm_(self.model.parameters(), self.grad_clip)
self.optimizer.step()
if self.accelerator.sync_gradients:
if not self.disable_tqdm:
train_tqdm.update(1)
train_tqdm.set_postfix(loss=self._summary_tracker.epoch_loss)
if valid_data:
self.stopped |= self._valid(valid_data, 'step')
self.accelerator.gradient_state._set_end_of_dataloader(False)
if self.stopped:
break
if not self.disable_tqdm:
train_tqdm.close()
return self._summary_tracker.epoch_dict()
@torch.no_grad()
def _valid(
self,
valid_data: DataLoader,
valid_mode: str,
) -> bool:
"""Validate every `self.eval_interval` step or epoch if evaluation strategy matches attribute
`self.eval_strategy`. Specifically, if `self.eval_interval` is set to `0`, validation will be skipped.
Early stopping will also be checked if `self.stopping_steps` is positive integer.
Args:
valid_data: The dataloader of validation set.
valid_mode: The evaluation strategy of current call ("epoch" or "step").
Returns:
bool: Early stopping. Return true if `self.stopping_steps` is positive integer and `self._early_stopping()`
is True.
"""
if (self.valid_steps <= 0) or (valid_mode != self.valid_strategy):
return False
self._valid_count += 1
if self._valid_count % self.valid_steps != 0:
return False
self.temp_mode = self._summary_tracker._current_mode
self.temp_epoch = self._summary_tracker._current_epoch
state = torch.default_generator.get_state()
with self._summary_tracker.new_epoch('valid'):
if 'loss' in self.metrics_for_best_model:
self.model.eval()
if not self.disable_tqdm:
valid_tqdm = tqdm(
valid_data,
desc=f"valid {self.timestamp.valid_epoch:4}",
dynamic_ncols=True,
postfix={'loss': None},
unit='step'
)
else:
valid_tqdm = valid_data
for data in valid_tqdm:
self._summary_tracker.new_step()
loss = self.model(data)
avg_loss = self.accelerator.gather(loss).mean().item()
self._summary_tracker.append_loss(avg_loss)
if not self.disable_tqdm:
valid_tqdm.set_postfix(loss=self._summary_tracker.epoch_loss)
valid_results = self._summary_tracker.epoch_dict()
else:
valid_results = self.evaluate(valid_data, is_valid=True)
self._summary_tracker.set_metrics_results(valid_results)
self.valid_result_dict[self.timestamp.valid_epoch] = self._summary_tracker._current_epoch
self._summary_tracker._current_mode = self.temp_mode
self._summary_tracker._current_epoch = self.temp_epoch
self.model.train()
stopped = bool(self.stopping_steps) and self._early_stopping(self._summary_tracker.is_best_valid)
self.accelerator.wait_for_everyone()
if self.accelerator.is_local_main_process:
# os.system('sleep 10s')
self.save_checkpoint()
torch.default_generator.set_state(state)
return stopped
def _early_stopping(self, current_best: bool) -> bool:
r""" Check early stopping with `stopping_steps`, a maximum amount of non-best validation.
Args:
current_best: Whether current epoch is the one with the best score.
Return:
bool: If true, the training process will be stopped, else the `self.stopping_count` will accumulate.
"""
stop_flag = False
if current_best:
self.stopping_count = 0
else:
self.stopping_count += 1
stop_flag = self.stopping_count > self.stopping_steps
return stop_flag
def _get_checkpoint(self) -> Optional[dict]:
if len(self.valid_result_dict) == 0:
self.logger.warning('Get checkpoint failed. No validation has been performed.')
return None
# get optimizer, config and validation summary
checkpoint = {
# parameters that needed to be loaded
'stopping_count': self.stopping_count,
'best_valid_score': self._summary_tracker.best_valid_score,
'epoch': self.timestamp.train_epoch,
'timestamp': self.timestamp,
'config': self.config,
# parameters for recording only
'summary': self.valid_result_dict[self.timestamp.valid_epoch],
}
self.logger.debug(checkpoint)
return checkpoint
def save_checkpoint(self):
serial_idx = self.timestamp.valid_epoch
serial_of_soft_link = self.best_valid_timestamp.valid_epoch
serialized_save(
self.accelerator.unwrap_model(self.model),
self.optimizer,
self._get_checkpoint(),
serial=serial_idx,
serial_of_soft_link=serial_of_soft_link,
path_without_extension=self.saved_model_filename,
tag='epoch',
extension_name='pth',
max_save=self.max_save,
)
def save_generated_text(self, generated_corpus: List[str], is_valid: bool = False):
r"""Store the generated text by our model into `self.saved_text_filename`."""
# path_to_save = self.saved_model_filename + '_epoch-' + str(self.timestamp.valid_epoch)
# saved_text_filename = os.path.join(path_to_save, 'generation.txt')
# os.makedirs(path_to_save, exist_ok=True)
if not is_valid:
saved_text_filename = self.saved_text_filename
with open(saved_text_filename, 'w') as fout:
for text in generated_corpus:
fout.write(text + '\n')
def resume_checkpoint(self, resume_dir: str):
r"""Load training information.
Args:
resume_dir: the checkpoint file (specific by `model_path`).
"""
# check
self.logger.info("Resuming checkpoint from {}...".format(resume_dir))
resume_checkpoint = os.path.join(resume_dir, 'textbox_configuration.pt')
resume_optimizer = os.path.join(resume_dir, 'optimizer.pt')
if os.path.isfile(resume_checkpoint):
checkpoint = torch.load(resume_checkpoint, map_location=self.device)
else:
self.logger.warning('Checkpoint file "{}" not found. Resuming stopped.'.format(resume_dir))
return
if checkpoint['config']['optimizer'].lower() != self.config['optimizer']:
self.logger.warning(
'Optimizer configuration given in config file is different from that of checkpoint. '
'This may yield an exception while state_dict is being loaded.'
)
if os.path.isfile(resume_optimizer):
optim_state_dict = torch.load(resume_optimizer, map_location=self.device)
self.optimizer.load_state_dict(optim_state_dict)
else:
self.logger.warning('Checkpoint file "{}" not found. Resuming stopped.'.format(resume_dir))
return
# load start epoch and early stopping
self.start_epoch = checkpoint['epoch'] + 1 # start from the next step
self._summary_tracker.axes = checkpoint['timestamp']
self.stopping_count = checkpoint['stopping_count']
self._summary_tracker.best_valid_score = checkpoint['best_valid_score']
self._summary_tracker.best_valid_timestamp = self._summary_tracker.axes
self.valid_result_dict = {self._summary_tracker.axes.valid_epoch: checkpoint['summary']}
if checkpoint['config']['seed']:
init_seed(checkpoint['config']['seed'], checkpoint['config']['reproducibility'])
set_seed(checkpoint['config']['seed'])
# load architecture params from checkpoint
if checkpoint['config']['model_name'] != self.config['model_name']:
self.logger.warning(
'Architecture configuration given in config file is different from that of checkpoint. '
'This may yield an exception while state_dict is being loaded.'
)
# load optimizer state from checkpoint only when optimizer type is not changed
self.logger.info(
'Checkpoint loaded. Resume training from epoch {} steps {}'.format(
self.start_epoch, self._summary_tracker.axes.train_step + 1
)
)
def fit(
self,
train_data: DataLoader,
valid_data: Optional[DataLoader] = None,
) -> dict:
r"""Train the model based on the train data.
Args:
train_data: The dataloader of training set.
valid_data: (default = None) The dataloader of training set.
Returns:
dict: the best valid score and best valid result.
"""
self.model = self.accelerator.prepare(self.model)
self.optimizer = self.accelerator.prepare(self.optimizer)
self.logger.info("====== Start training ======")
self.accelerator.wait_for_everyone()
for epoch_idx in range(self.start_epoch, self.epochs + 1):
# train
loss = self._train_epoch(train_data, epoch_idx, valid_data)['loss']
self.train_loss_list.append(loss)
# valid
if valid_data:
self.stopped |= self._valid(valid_data, 'epoch')
if self.stopped:
if self.stopping_steps:
self.logger.info(f'Early stopped at {self.stopping_count} non-best validation.')
elif self.max_steps < 10000000000:
self.logger.info(f'Stopped at max_steps {self.max_steps}.')
break
file = self.saved_model_filename + '_best'
if os.path.exists(file):
self.logger.info(f'Soft link created: {file} -> {os.readlink(file)}')
self.logger.info(
f'====== Finished training, best validation result '
f'at train epoch {self.best_valid_timestamp.train_epoch} ======'
)
self.model = self.accelerator.unwrap_model(self.model)
self.logger.info('Best valid result: {}'.format(self.best_valid_result.as_str()))
return self.best_valid_result.as_dict()
@torch.no_grad()
def evaluate(
self,
eval_data: DataLoader,
load_best_model: bool = True,
is_valid: bool = False,
) -> Optional[dict]:
r"""Evaluate the model based on the `eval_data`.
Args:
eval_data (DataLoader): the eval data
load_best_model (bool, optional): whether load the best model in the training process, default: True.
It should be set True, if users want to test the model after training.
model_file (str, optional): the saved model file, default: None. If users want to test the previously
trained model file, they can set this parameter.
is_valid: (default = False) True if evaluate during validation
Returns:
dict: eval result, key is the eval metric and value in the corresponding metric value
"""
if is_valid:
load_best_model = False
if load_best_model:
checkpoint_dir = self.saved_model_filename + '_best'
self.logger.info('Loading model structure and parameters from {} ...'.format(checkpoint_dir))
self.accelerator.wait_for_everyone()
unwrap_model = self.accelerator.unwrap_model(self.model)
unwrap_model.from_pretrained(checkpoint_dir)
unwrap_model.tokenizer.from_pretrained(checkpoint_dir)
if not is_valid:
self.model = self.accelerator.prepare(self.model)
self.model.eval()
if self.config['dataset'] == 'multiwoz':
self.evaluator.evaluators[0].load_data('valid' if is_valid else 'test')
turn_domains = self.evaluator.evaluators[0].turn_domains
# generate
generate_corpus = []
eval_tqdm = tqdm(eval_data, desc="generating", dynamic_ncols=True) if not self.disable_tqdm else eval_data
state = torch.default_generator.get_state()
for i, batch_data in enumerate(eval_tqdm):
if self.config['dataset'] != 'multiwoz':
generated = self.accelerator.unwrap_model(self.model).generate(batch_data, self.accelerator)
else:
batch_size = batch_data['source_ids'].size(0)
idx_mask = torch.zeros(batch_size).to(self.device).bool()
idx_mask[::3] = 1
bs_batch = {}
bs_batch['source_ids'] = batch_data['source_ids'][idx_mask]
bs_batch['source_mask'] = batch_data['source_mask'][idx_mask]
asrs_batch = {}
asrs_batch['source_ids'] = batch_data['source_ids'][~idx_mask]
asrs_batch['source_mask'] = batch_data['source_mask'][~idx_mask]
bs_outputs = self.accelerator.unwrap_model(self.model).generate(bs_batch, self.accelerator)
batch_size //= 3
db_texts = [
self.evaluator.evaluators[0].span_db(bs, td)
for bs, td in zip(bs_outputs, turn_domains[i * batch_size:(i + 1) * batch_size])
]
db_ids = torch.tensor(self.model.tokenizer.convert_tokens_to_ids(db_texts)).long()
db_ids = db_ids.repeat_interleave(2).to(self.device)
db_idx = torch.eq(asrs_batch['source_ids'], self.model.tokenizer.convert_tokens_to_ids('[db_nores]'))
asrs_batch['source_ids'][db_idx] = db_ids
asrs_outputs = self.accelerator.unwrap_model(self.model).generate(asrs_batch, self.accelerator)
generated = sum([[bs, aspn, rs]
for bs, aspn, rs in zip(bs_outputs, asrs_outputs[::2], asrs_outputs[1::2])], [])
generate_corpus.extend(generated)
torch.default_generator.set_state(state)
corpus_len = len(eval_data.dataset.target_text)
reference_dataset = eval_data.dataset
generate_corpus = generate_corpus[:corpus_len]
if self.post_processing == 'paraphrase':
for i, gen in enumerate(generate_corpus):
if gen.find('[SEP]') >= 0:
gen = gen.split('[SEP]')[1].strip()
else:
last = max(gen.rfind('('), gen.rfind(')'))
last = gen.find(' ', last)
gen = gen[last:].strip()
generate_corpus[i] = gen
if self.accelerator.is_local_main_process:
self.save_generated_text(generate_corpus, is_valid)
result = self.evaluator.evaluate(generate_corpus, reference_dataset)
return result