run_squad_w_distillation.py 35.1 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
16
""" This is the exact same script as `examples/run_squad.py` (as of 2020, January 8th) with an additional and optional step of distillation."""
17
18

import argparse
Aymeric Augustin's avatar
Aymeric Augustin committed
19
import glob
20
21
22
import logging
import os
import random
Victor SANH's avatar
Victor SANH committed
23
import timeit
24
25
26

import numpy as np
import torch
Aymeric Augustin's avatar
Aymeric Augustin committed
27
28
import torch.nn as nn
import torch.nn.functional as F
29
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
30
from torch.utils.data.distributed import DistributedSampler
31
from tqdm import tqdm, trange
32

33
34
from transformers import (
    WEIGHTS_NAME,
Aymeric Augustin's avatar
Aymeric Augustin committed
35
    AdamW,
36
37
38
    BertConfig,
    BertForQuestionAnswering,
    BertTokenizer,
Aymeric Augustin's avatar
Aymeric Augustin committed
39
40
41
    DistilBertConfig,
    DistilBertForQuestionAnswering,
    DistilBertTokenizer,
42
43
44
45
46
47
    XLMConfig,
    XLMForQuestionAnswering,
    XLMTokenizer,
    XLNetConfig,
    XLNetForQuestionAnswering,
    XLNetTokenizer,
Aymeric Augustin's avatar
Aymeric Augustin committed
48
    get_linear_schedule_with_warmup,
49
    squad_convert_examples_to_features,
50
)
51
52
53
54
from transformers.data.metrics.squad_metrics import (
    compute_predictions_log_probs,
    compute_predictions_logits,
    squad_evaluate,
55
)
56
from transformers.data.processors.squad import SquadResult, SquadV1Processor, SquadV2Processor
Aymeric Augustin's avatar
Aymeric Augustin committed
57
58
59
60


try:
    from torch.utils.tensorboard import SummaryWriter
61
except ImportError:
Aymeric Augustin's avatar
Aymeric Augustin committed
62
63
    from tensorboardX import SummaryWriter

64

65
66
logger = logging.getLogger(__name__)

67
68
69
ALL_MODELS = sum(
    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig)), ()
)
70
71

MODEL_CLASSES = {
72
73
74
75
    "bert": (BertConfig, BertForQuestionAnswering, BertTokenizer),
    "xlnet": (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
    "xlm": (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
    "distilbert": (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
76
77
}

78

79
80
81
82
83
84
85
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)

86

87
88
89
def to_list(tensor):
    return tensor.detach().cpu().tolist()

90

91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
def train(args, train_dataset, model, tokenizer, teacher=None):
    """ Train the model """
    if args.local_rank in [-1, 0]:
        tb_writer = SummaryWriter()

    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)

    if args.max_steps > 0:
        t_total = args.max_steps
        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
    else:
        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs

    # Prepare optimizer and schedule (linear warmup and decay)
107
    no_decay = ["bias", "LayerNorm.weight"]
108
    optimizer_grouped_parameters = [
109
110
111
112
113
114
        {
            "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
            "weight_decay": args.weight_decay,
        },
        {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
    ]
115
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
116
117
118
    scheduler = get_linear_schedule_with_warmup(
        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
    )
119
120
121
122
123
124
125
126

    # Check if saved optimizer or scheduler states exist
    if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
        os.path.join(args.model_name_or_path, "scheduler.pt")
    ):
        # Load in optimizer and scheduler states
        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
Victor SANH's avatar
Victor SANH committed
127

Victor SANH's avatar
indents  
Victor SANH committed
128
    if args.fp16:
129
130
131
132
        try:
            from apex import amp
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
133

134
135
136
137
138
139
140
141
        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)

    # multi-gpu training (should be after apex fp16 initialization)
    if args.n_gpu > 1:
        model = torch.nn.DataParallel(model)

    # Distributed training (should be after apex fp16 initialization)
    if args.local_rank != -1:
142
143
144
        model = torch.nn.parallel.DistributedDataParallel(
            model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
        )
145
146
147
148
149
150

    # Train!
    logger.info("***** Running training *****")
    logger.info("  Num examples = %d", len(train_dataset))
    logger.info("  Num Epochs = %d", args.num_train_epochs)
    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
151
152
153
154
155
156
    logger.info(
        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
        args.train_batch_size
        * args.gradient_accumulation_steps
        * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
    )
157
158
159
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", t_total)

160
    global_step = 1
Victor SANH's avatar
indents  
Victor SANH committed
161
    epochs_trained = 0
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
    steps_trained_in_current_epoch = 0
    # Check if continuing training from a checkpoint
    if os.path.exists(args.model_name_or_path):
        try:
            # set global_step to gobal_step of last saved checkpoint from model path
            checkpoint_suffix = args.model_name_or_path.split("-")[-1].split("/")[0]
            global_step = int(checkpoint_suffix)
            epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
            steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)

            logger.info("  Continuing training from checkpoint, will skip to saved global_step")
            logger.info("  Continuing training from epoch %d", epochs_trained)
            logger.info("  Continuing training from global step %d", global_step)
            logger.info("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
        except ValueError:
            logger.info("  Starting fine-tuning.")

179
180
    tr_loss, logging_loss = 0.0, 0.0
    model.zero_grad()
181
    train_iterator = trange(
Victor SANH's avatar
indents  
Victor SANH committed
182
183
        epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]
    )
184
    # Added here for reproductibility
Victor SANH's avatar
indents  
Victor SANH committed
185
    set_seed(args)
186

187
188
189
    for _ in train_iterator:
        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
        for step, batch in enumerate(epoch_iterator):
190
191
192
193
194
195

            # Skip past any already trained steps if resuming training
            if steps_trained_in_current_epoch > 0:
                steps_trained_in_current_epoch -= 1
                continue

196
197
198
199
            model.train()
            if teacher is not None:
                teacher.eval()
            batch = tuple(t.to(args.device) for t in batch)
200

201
202
203
204
205
206
207
208
209
210
            inputs = {
                "input_ids": batch[0],
                "attention_mask": batch[1],
                "start_positions": batch[3],
                "end_positions": batch[4],
            }
            if args.model_type != "distilbert":
                inputs["token_type_ids"] = None if args.model_type == "xlm" else batch[2]
            if args.model_type in ["xlnet", "xlm"]:
                inputs.update({"cls_index": batch[5], "p_mask": batch[6]})
Victor SANH's avatar
indents  
Victor SANH committed
211
                if args.version_2_with_negative:
212
                    inputs.update({"is_impossible": batch[7]})
213
214
215
216
217
            outputs = model(**inputs)
            loss, start_logits_stu, end_logits_stu = outputs

            # Distillation loss
            if teacher is not None:
218
219
                if "token_type_ids" not in inputs:
                    inputs["token_type_ids"] = None if args.teacher_type == "xlm" else batch[2]
220
                with torch.no_grad():
221
222
223
224
225
                    start_logits_tea, end_logits_tea = teacher(
                        input_ids=inputs["input_ids"],
                        token_type_ids=inputs["token_type_ids"],
                        attention_mask=inputs["attention_mask"],
                    )
226
227
228
                assert start_logits_tea.size() == start_logits_stu.size()
                assert end_logits_tea.size() == end_logits_stu.size()

229
230
231
232
233
234
235
236
237
238
239
240
                loss_fct = nn.KLDivLoss(reduction="batchmean")
                loss_start = loss_fct(
                    F.log_softmax(start_logits_stu / args.temperature, dim=-1),
                    F.softmax(start_logits_tea / args.temperature, dim=-1),
                ) * (args.temperature ** 2)
                loss_end = loss_fct(
                    F.log_softmax(end_logits_stu / args.temperature, dim=-1),
                    F.softmax(end_logits_tea / args.temperature, dim=-1),
                ) * (args.temperature ** 2)
                loss_ce = (loss_start + loss_end) / 2.0

                loss = args.alpha_ce * loss_ce + args.alpha_squad * loss
241
242

            if args.n_gpu > 1:
243
                loss = loss.mean()  # mean() to average on multi-gpu parallel (not distributed) training
244
245
246
247
248
249
250
251
252
253
254
            if args.gradient_accumulation_steps > 1:
                loss = loss / args.gradient_accumulation_steps

            if args.fp16:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
            else:
                loss.backward()

            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0:
255
256
257
258
259
                if args.fp16:
                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
                else:
                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)

260
261
262
263
264
                optimizer.step()
                scheduler.step()  # Update learning rate schedule
                model.zero_grad()
                global_step += 1

Victor SANH's avatar
indents  
Victor SANH committed
265
                # Log metrics
266
                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
267
268
                    # Only evaluate when single GPU otherwise metrics may not average well
                    if args.local_rank == -1 and args.evaluate_during_training:
269
270
                        results = evaluate(args, model, tokenizer)
                        for key, value in results.items():
271
272
273
                            tb_writer.add_scalar("eval_{}".format(key), value, global_step)
                    tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
                    tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
274
275
276
277
                    logging_loss = tr_loss

                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
                    # Save model checkpoint
278
                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
279
280
                    if not os.path.exists(output_dir):
                        os.makedirs(output_dir)
281
282
283
                    model_to_save = (
                        model.module if hasattr(model, "module") else model
                    )  # Take care of distributed/parallel training
284
                    model_to_save.save_pretrained(output_dir)
Victor SANH's avatar
indents  
Victor SANH committed
285
                    tokenizer.save_pretrained(output_dir)
286

287
                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
288
289
                    logger.info("Saving model checkpoint to %s", output_dir)

290
291
292
293
                    torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
                    torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
                    logger.info("Saving optimizer and scheduler states to %s", output_dir)

294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
            if args.max_steps > 0 and global_step > args.max_steps:
                epoch_iterator.close()
                break
        if args.max_steps > 0 and global_step > args.max_steps:
            train_iterator.close()
            break

    if args.local_rank in [-1, 0]:
        tb_writer.close()

    return global_step, tr_loss / global_step


def evaluate(args, model, tokenizer, prefix=""):
    dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)

    if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
        os.makedirs(args.output_dir)

    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
314

315
    # Note that DistributedSampler samples randomly
316
    eval_sampler = SequentialSampler(dataset)
317
318
    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

319
320
321
322
    # multi-gpu evaluate
    if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
        model = torch.nn.DataParallel(model)

323
324
325
326
    # Eval!
    logger.info("***** Running evaluation {} *****".format(prefix))
    logger.info("  Num examples = %d", len(dataset))
    logger.info("  Batch size = %d", args.eval_batch_size)
327

328
    all_results = []
Victor SANH's avatar
indents  
Victor SANH committed
329
    start_time = timeit.default_timer()
330

331
332
333
    for batch in tqdm(eval_dataloader, desc="Evaluating"):
        model.eval()
        batch = tuple(t.to(args.device) for t in batch)
334

335
        with torch.no_grad():
336
337
338
            inputs = {"input_ids": batch[0], "attention_mask": batch[1]}
            if args.model_type != "distilbert":
                inputs["token_type_ids"] = None if args.model_type == "xlm" else batch[2]  # XLM don't use segment_ids
339
            example_indices = batch[3]
340
341
            if args.model_type in ["xlnet", "xlm"]:
                inputs.update({"cls_index": batch[4], "p_mask": batch[5]})
342

343
344
345
346
347
            outputs = model(**inputs)

        for i, example_index in enumerate(example_indices):
            eval_feature = features[example_index.item()]
            unique_id = int(eval_feature.unique_id)
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366

            output = [to_list(output[i]) for output in outputs]

            # Some models (XLNet, XLM) use 5 arguments for their predictions, while the other "simpler"
            # models only use two.
            if len(output) >= 5:
                start_logits = output[0]
                start_top_index = output[1]
                end_logits = output[2]
                end_top_index = output[3]
                cls_logits = output[4]

                result = SquadResult(
                    unique_id,
                    start_logits,
                    end_logits,
                    start_top_index=start_top_index,
                    end_top_index=end_top_index,
                    cls_logits=cls_logits,
367
                )
368

369
            else:
370
371
372
                start_logits, end_logits = output
                result = SquadResult(unique_id, start_logits, end_logits)

373
374
            all_results.append(result)

375
376
377
    evalTime = timeit.default_timer() - start_time
    logger.info("  Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))

378
379
380
    # Compute predictions
    output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
    output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
381

382
383
384
385
386
    if args.version_2_with_negative:
        output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
    else:
        output_null_log_odds_file = None

387
    if args.model_type in ["xlnet", "xlm"]:
388
        # XLNet uses a more complex post-processing procedure
389
        predictions = compute_predictions_log_probs(
390
391
392
393
394
395
396
397
398
399
400
401
402
403
            examples,
            features,
            all_results,
            args.n_best_size,
            args.max_answer_length,
            output_prediction_file,
            output_nbest_file,
            output_null_log_odds_file,
            model.config.start_n_top,
            model.config.end_n_top,
            args.version_2_with_negative,
            tokenizer,
            args.verbose_logging,
        )
404
    else:
405
        predictions = compute_predictions_logits(
406
407
408
409
410
411
412
413
414
415
416
417
            examples,
            features,
            all_results,
            args.n_best_size,
            args.max_answer_length,
            args.do_lower_case,
            output_prediction_file,
            output_nbest_file,
            output_null_log_odds_file,
            args.verbose_logging,
            args.version_2_with_negative,
            args.null_score_diff_threshold,
418
            tokenizer,
419
        )
420

421
422
    # Compute the F1 and exact scores.
    results = squad_evaluate(examples, predictions)
423
424
425
426
427
    return results


def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
    if args.local_rank not in [-1, 0] and not evaluate:
428
        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
Victor SANH's avatar
indents  
Victor SANH committed
429
        torch.distributed.barrier()
430
431
432

    # Load data features from cache or dataset file
    input_file = args.predict_file if evaluate else args.train_file
433
434
    cached_features_file = os.path.join(
        os.path.dirname(input_file),
435
        "cached_distillation_{}_{}_{}".format(
436
437
438
439
440
            "dev" if evaluate else "train",
            list(filter(None, args.model_name_or_path.split("/"))).pop(),
            str(args.max_seq_length),
        ),
    )
441
    if os.path.exists(cached_features_file) and not args.overwrite_cache:
442
        logger.info("Loading features from cached file %s", cached_features_file)
443
444
445
446
447
448
449
450
451
452
453
454
455
        features_and_dataset = torch.load(cached_features_file)

        try:
            features, dataset, examples = (
                features_and_dataset["features"],
                features_and_dataset["dataset"],
                features_and_dataset["examples"],
            )
        except KeyError:
            raise DeprecationWarning(
                "You seem to be loading features from an older version of this script please delete the "
                "file %s in order for it to be created again" % cached_features_file
            )
456
457
    else:
        logger.info("Creating features from dataset file at %s", input_file)
458
459
        processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
        if evaluate:
460
            examples = processor.get_dev_examples(args.data_dir, filename=args.predict_file)
461
        else:
462
            examples = processor.get_train_examples(args.data_dir, filename=args.train_file)
463
464

        features, dataset = squad_convert_examples_to_features(
465
466
467
468
469
470
            examples=examples,
            tokenizer=tokenizer,
            max_seq_length=args.max_seq_length,
            doc_stride=args.doc_stride,
            max_query_length=args.max_query_length,
            is_training=not evaluate,
471
            return_dataset="pt",
Victor SANH's avatar
indents  
Victor SANH committed
472
            threads=args.threads,
473
        )
474

475
476
        if args.local_rank in [-1, 0]:
            logger.info("Saving features into cached file %s", cached_features_file)
477
            torch.save({"features": features, "dataset": dataset, "examples": examples}, cached_features_file)
478
479

    if args.local_rank == 0 and not evaluate:
Victor SANH's avatar
indents  
Victor SANH committed
480
        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
481
        torch.distributed.barrier()
482
483
484
485
486
487
488
489
490

    if output_examples:
        return dataset, examples, features
    return dataset


def main():
    parser = argparse.ArgumentParser()

491
    # Required parameters
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
    parser.add_argument(
        "--model_type",
        default=None,
        type=str,
        required=True,
        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
    )
    parser.add_argument(
        "--model_name_or_path",
        default=None,
        type=str,
        required=True,
        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
    )
    parser.add_argument(
        "--output_dir",
        default=None,
        type=str,
        required=True,
        help="The output directory where the model checkpoints and predictions will be written.",
    )
513
514

    # Distillation parameters (optional)
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
    parser.add_argument(
        "--teacher_type",
        default=None,
        type=str,
        help="Teacher type. Teacher tokenizer and student (model) tokenizer must output the same tokenization. Only for distillation.",
    )
    parser.add_argument(
        "--teacher_name_or_path",
        default=None,
        type=str,
        help="Path to the already SQuAD fine-tuned teacher model. Only for distillation.",
    )
    parser.add_argument(
        "--alpha_ce", default=0.5, type=float, help="Distillation loss linear weight. Only for distillation."
    )
    parser.add_argument(
        "--alpha_squad", default=0.5, type=float, help="True SQuAD loss linear weight. Only for distillation."
    )
    parser.add_argument(
        "--temperature", default=2.0, type=float, help="Distillation temperature. Only for distillation."
    )
536

537
    # Other parameters
538
539
540
541
542
543
544
    parser.add_argument(
        "--data_dir",
        default=None,
        type=str,
        help="The input data dir. Should contain the .json files for the task."
        + "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
    )
Victor SANH's avatar
indents  
Victor SANH committed
545
    parser.add_argument(
546
        "--train_file",
Victor SANH's avatar
indents  
Victor SANH committed
547
548
549
        default=None,
        type=str,
        help="The input training file. If a data dir is specified, will look for the file there"
550
551
552
553
554
555
556
557
558
        + "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
    )
    parser.add_argument(
        "--predict_file",
        default=None,
        type=str,
        help="The input evaluation file. If a data dir is specified, will look for the file there"
        + "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
    )
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
    parser.add_argument(
        "--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name"
    )
    parser.add_argument(
        "--tokenizer_name",
        default="",
        type=str,
        help="Pretrained tokenizer name or path if not the same as model_name",
    )
    parser.add_argument(
        "--cache_dir",
        default="",
        type=str,
        help="Where do you want to store the pre-trained models downloaded from s3",
    )

    parser.add_argument(
        "--version_2_with_negative",
        action="store_true",
        help="If true, the SQuAD examples contain some that do not have an answer.",
    )
    parser.add_argument(
        "--null_score_diff_threshold",
        type=float,
        default=0.0,
        help="If null_score - best_non_null is greater than the threshold predict null.",
    )

    parser.add_argument(
        "--max_seq_length",
        default=384,
        type=int,
        help="The maximum total input sequence length after WordPiece tokenization. Sequences "
        "longer than this will be truncated, and sequences shorter than this will be padded.",
    )
    parser.add_argument(
        "--doc_stride",
        default=128,
        type=int,
        help="When splitting up a long document into chunks, how much stride to take between chunks.",
    )
    parser.add_argument(
        "--max_query_length",
        default=64,
        type=int,
        help="The maximum number of tokens for the question. Questions longer than this will "
        "be truncated to this length.",
    )
    parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
    parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
    parser.add_argument(
        "--evaluate_during_training", action="store_true", help="Rul evaluation during training at each logging step."
    )
    parser.add_argument(
        "--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
    )

    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.")
    parser.add_argument(
        "--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation."
    )
    parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
    parser.add_argument(
        "--gradient_accumulation_steps",
        type=int,
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
627
    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
    parser.add_argument(
        "--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform."
    )
    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("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
    parser.add_argument(
        "--n_best_size",
        default=20,
        type=int,
        help="The total number of n-best predictions to generate in the nbest_predictions.json output file.",
    )
    parser.add_argument(
        "--max_answer_length",
        default=30,
        type=int,
        help="The maximum length of an answer that can be generated. This is needed because the start "
        "and end predictions are not conditioned on one another.",
    )
    parser.add_argument(
        "--verbose_logging",
        action="store_true",
        help="If true, all of the warnings related to data processing will be printed. "
        "A number of warnings are expected for a normal SQuAD evaluation.",
    )

    parser.add_argument("--logging_steps", type=int, default=50, help="Log every X updates steps.")
    parser.add_argument("--save_steps", type=int, default=50, help="Save checkpoint every X updates steps.")
    parser.add_argument(
        "--eval_all_checkpoints",
        action="store_true",
        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
    )
    parser.add_argument("--no_cuda", action="store_true", help="Whether not to use CUDA when available")
    parser.add_argument(
        "--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory"
    )
    parser.add_argument(
        "--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
    )
    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")

    parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
    parser.add_argument(
        "--fp16",
        action="store_true",
        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
    )
    parser.add_argument(
        "--fp16_opt_level",
        type=str,
        default="O1",
        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
        "See details at https://nvidia.github.io/apex/amp.html",
    )
    parser.add_argument("--server_ip", type=str, default="", help="Can be used for distant debugging.")
    parser.add_argument("--server_port", type=str, default="", help="Can be used for distant debugging.")
691

Victor SANH's avatar
indents  
Victor SANH committed
692
    parser.add_argument("--threads", type=int, default=1, help="multiple threads for converting example to features")
693
694
    args = parser.parse_args()

695
696
697
698
699
700
701
702
703
704
705
    if (
        os.path.exists(args.output_dir)
        and os.listdir(args.output_dir)
        and args.do_train
        and not args.overwrite_output_dir
    ):
        raise ValueError(
            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
                args.output_dir
            )
        )
706
707
708
709
710

    # Setup distant debugging if needed
    if args.server_ip and args.server_port:
        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
        import ptvsd
711

712
713
714
715
716
717
718
719
720
721
722
        print("Waiting for debugger attach")
        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
        ptvsd.wait_for_attach()

    # Setup CUDA, GPU & distributed training
    if args.local_rank == -1 or args.no_cuda:
        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
        args.n_gpu = torch.cuda.device_count()
    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
        torch.cuda.set_device(args.local_rank)
        device = torch.device("cuda", args.local_rank)
723
        torch.distributed.init_process_group(backend="nccl")
724
725
726
727
        args.n_gpu = 1
    args.device = device

    # Setup logging
728
729
730
731
732
733
734
735
736
737
738
739
740
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
    )
    logger.warning(
        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
        args.local_rank,
        device,
        args.n_gpu,
        bool(args.local_rank != -1),
        args.fp16,
    )
741
742
743
744
745
746

    # Set seed
    set_seed(args)

    # Load pretrained model and tokenizer
    if args.local_rank not in [-1, 0]:
Victor SANH's avatar
indents  
Victor SANH committed
747
        # Make sure only the first process in distributed training will download model & vocab
Victor SANH's avatar
Victor SANH committed
748
        torch.distributed.barrier()
749
750
751

    args.model_type = args.model_type.lower()
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
    config = config_class.from_pretrained(
        args.config_name if args.config_name else args.model_name_or_path,
        cache_dir=args.cache_dir if args.cache_dir else None,
    )
    tokenizer = tokenizer_class.from_pretrained(
        args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
        do_lower_case=args.do_lower_case,
        cache_dir=args.cache_dir if args.cache_dir else None,
    )
    model = model_class.from_pretrained(
        args.model_name_or_path,
        from_tf=bool(".ckpt" in args.model_name_or_path),
        config=config,
        cache_dir=args.cache_dir if args.cache_dir else None,
    )
767
768
769

    if args.teacher_type is not None:
        assert args.teacher_name_or_path is not None
770
771
772
        assert args.alpha_ce > 0.0
        assert args.alpha_ce + args.alpha_squad > 0.0
        assert args.teacher_type != "distilbert", "We constraint teachers not to be of type DistilBERT."
773
        teacher_config_class, teacher_model_class, _ = MODEL_CLASSES[args.teacher_type]
774
775
776
777
778
779
        teacher_config = teacher_config_class.from_pretrained(
            args.teacher_name_or_path, cache_dir=args.cache_dir if args.cache_dir else None
        )
        teacher = teacher_model_class.from_pretrained(
            args.teacher_name_or_path, config=teacher_config, cache_dir=args.cache_dir if args.cache_dir else None
        )
780
781
782
783
784
        teacher.to(args.device)
    else:
        teacher = None

    if args.local_rank == 0:
Victor SANH's avatar
indents  
Victor SANH committed
785
        # Make sure only the first process in distributed training will download model & vocab
786
        torch.distributed.barrier()
787
788
789
790
791

    model.to(args.device)

    logger.info("Training/evaluation parameters %s", args)

792
793
794
795
796
797
798
799
800
801
802
    # Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
    # Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
    # remove the need for this code, but it is still valid.
    if args.fp16:
        try:
            import apex

            apex.amp.register_half_function(torch, "einsum")
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")

803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
    # Training
    if args.do_train:
        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
        global_step, tr_loss = train(args, train_dataset, model, tokenizer, teacher=teacher)
        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)

    # Save the trained model and the tokenizer
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
        # Create output directory if needed
        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
            os.makedirs(args.output_dir)

        logger.info("Saving model checkpoint to %s", args.output_dir)
        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
        # They can then be reloaded using `from_pretrained()`
818
819
820
        model_to_save = (
            model.module if hasattr(model, "module") else model
        )  # Take care of distributed/parallel training
821
822
823
824
        model_to_save.save_pretrained(args.output_dir)
        tokenizer.save_pretrained(args.output_dir)

        # Good practice: save your training arguments together with the trained model
825
        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
826
827

        # Load a trained model and vocabulary that you have fine-tuned
828
829
        model = model_class.from_pretrained(args.output_dir)
        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
830
831
832
833
834
        model.to(args.device)

    # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
    results = {}
    if args.do_eval and args.local_rank in [-1, 0]:
835
836
        if args.do_train:
            logger.info("Loading checkpoints saved during training for evaluation")
837
838
        checkpoints = [args.output_dir]
        if args.eval_all_checkpoints:
839
840
841
            checkpoints = list(
                os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
            )
842
843
844
845
846
847
            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce model loading logs

        logger.info("Evaluate the following checkpoints: %s", checkpoints)

        for checkpoint in checkpoints:
            # Reload the model
848
            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
849
            model = model_class.from_pretrained(checkpoint)
850
851
852
853
854
            model.to(args.device)

            # Evaluate
            result = evaluate(args, model, tokenizer, prefix=global_step)

855
            result = dict((k + ("_{}".format(global_step) if global_step else ""), v) for k, v in result.items())
856
857
858
859
860
861
862
863
864
            results.update(result)

    logger.info("Results: {}".format(results))

    return results


if __name__ == "__main__":
    main()