run_squad_w_distillation.py 35.4 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/question-answering/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
import transformers
34
35
from transformers import (
    WEIGHTS_NAME,
Aymeric Augustin's avatar
Aymeric Augustin committed
36
    AdamW,
37
38
39
    BertConfig,
    BertForQuestionAnswering,
    BertTokenizer,
Aymeric Augustin's avatar
Aymeric Augustin committed
40
41
42
    DistilBertConfig,
    DistilBertForQuestionAnswering,
    DistilBertTokenizer,
Julien Chaumond's avatar
Julien Chaumond committed
43
44
45
    RobertaConfig,
    RobertaForQuestionAnswering,
    RobertaTokenizer,
46
47
48
49
50
51
    XLMConfig,
    XLMForQuestionAnswering,
    XLMTokenizer,
    XLNetConfig,
    XLNetForQuestionAnswering,
    XLNetTokenizer,
Aymeric Augustin's avatar
Aymeric Augustin committed
52
    get_linear_schedule_with_warmup,
53
    squad_convert_examples_to_features,
54
)
55
56
57
58
from transformers.data.metrics.squad_metrics import (
    compute_predictions_log_probs,
    compute_predictions_logits,
    squad_evaluate,
59
)
60
from transformers.data.processors.squad import SquadResult, SquadV1Processor, SquadV2Processor
61
from transformers.trainer_utils import is_main_process
Aymeric Augustin's avatar
Aymeric Augustin committed
62
63
64
65


try:
    from torch.utils.tensorboard import SummaryWriter
66
except ImportError:
Aymeric Augustin's avatar
Aymeric Augustin committed
67
68
    from tensorboardX import SummaryWriter

69

70
71
72
73
logger = logging.getLogger(__name__)


MODEL_CLASSES = {
74
75
76
77
    "bert": (BertConfig, BertForQuestionAnswering, BertTokenizer),
    "xlnet": (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
    "xlm": (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
    "distilbert": (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
78
    "roberta": (RobertaConfig, RobertaForQuestionAnswering, RobertaTokenizer),
79
80
}

81

82
83
84
85
86
87
88
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)

89

90
91
92
def to_list(tensor):
    return tensor.detach().cpu().tolist()

93

94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
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)
110
    no_decay = ["bias", "LayerNorm.weight"]
111
    optimizer_grouped_parameters = [
112
113
114
115
116
117
        {
            "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},
    ]
118
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
119
120
121
    scheduler = get_linear_schedule_with_warmup(
        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
    )
122
123
124
125
126
127
128
129

    # 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
130

Victor SANH's avatar
indents  
Victor SANH committed
131
    if args.fp16:
132
133
134
135
        try:
            from apex import amp
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
136

137
138
139
140
141
142
143
144
        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:
145
146
147
        model = torch.nn.parallel.DistributedDataParallel(
            model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
        )
148
149
150
151
152
153

    # 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)
154
155
156
157
158
159
    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),
    )
160
161
162
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", t_total)

163
    global_step = 1
Victor SANH's avatar
indents  
Victor SANH committed
164
    epochs_trained = 0
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
    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.")

182
183
    tr_loss, logging_loss = 0.0, 0.0
    model.zero_grad()
184
    train_iterator = trange(
Victor SANH's avatar
indents  
Victor SANH committed
185
186
        epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]
    )
187
    # Added here for reproductibility
Victor SANH's avatar
indents  
Victor SANH committed
188
    set_seed(args)
189

190
191
192
    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):
193
194
195
196
197
198

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

199
200
201
202
            model.train()
            if teacher is not None:
                teacher.eval()
            batch = tuple(t.to(args.device) for t in batch)
203

204
205
206
207
208
209
210
211
212
213
            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
214
                if args.version_2_with_negative:
215
                    inputs.update({"is_impossible": batch[7]})
216
217
218
219
220
            outputs = model(**inputs)
            loss, start_logits_stu, end_logits_stu = outputs

            # Distillation loss
            if teacher is not None:
221
222
                if "token_type_ids" not in inputs:
                    inputs["token_type_ids"] = None if args.teacher_type == "xlm" else batch[2]
223
                with torch.no_grad():
224
225
226
227
228
                    start_logits_tea, end_logits_tea = teacher(
                        input_ids=inputs["input_ids"],
                        token_type_ids=inputs["token_type_ids"],
                        attention_mask=inputs["attention_mask"],
                    )
229
230
231
                assert start_logits_tea.size() == start_logits_stu.size()
                assert end_logits_tea.size() == end_logits_stu.size()

232
                loss_fct = nn.KLDivLoss(reduction="batchmean")
Lysandre's avatar
Lysandre committed
233
234
235
236
237
238
239
240
241
242
243
244
245
246
                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)
                )
247
248
249
                loss_ce = (loss_start + loss_end) / 2.0

                loss = args.alpha_ce * loss_ce + args.alpha_squad * loss
250
251

            if args.n_gpu > 1:
252
                loss = loss.mean()  # mean() to average on multi-gpu parallel (not distributed) training
253
254
255
256
257
258
259
260
261
262
263
            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:
264
265
266
267
268
                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)

269
270
271
272
273
                optimizer.step()
                scheduler.step()  # Update learning rate schedule
                model.zero_grad()
                global_step += 1

Victor SANH's avatar
indents  
Victor SANH committed
274
                # Log metrics
275
                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
276
277
                    # Only evaluate when single GPU otherwise metrics may not average well
                    if args.local_rank == -1 and args.evaluate_during_training:
278
279
                        results = evaluate(args, model, tokenizer)
                        for key, value in results.items():
280
281
282
                            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)
283
284
285
286
                    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
287
                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
288
289
                    if not os.path.exists(output_dir):
                        os.makedirs(output_dir)
290
291
292
                    model_to_save = (
                        model.module if hasattr(model, "module") else model
                    )  # Take care of distributed/parallel training
293
                    model_to_save.save_pretrained(output_dir)
Victor SANH's avatar
indents  
Victor SANH committed
294
                    tokenizer.save_pretrained(output_dir)
295

296
                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
297
298
                    logger.info("Saving model checkpoint to %s", output_dir)

299
300
301
302
                    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)

303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
            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)
323

324
    # Note that DistributedSampler samples randomly
325
    eval_sampler = SequentialSampler(dataset)
326
327
    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

328
329
330
331
    # multi-gpu evaluate
    if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
        model = torch.nn.DataParallel(model)

332
333
334
335
    # Eval!
    logger.info("***** Running evaluation {} *****".format(prefix))
    logger.info("  Num examples = %d", len(dataset))
    logger.info("  Batch size = %d", args.eval_batch_size)
336

337
    all_results = []
Victor SANH's avatar
indents  
Victor SANH committed
338
    start_time = timeit.default_timer()
339

340
341
342
    for batch in tqdm(eval_dataloader, desc="Evaluating"):
        model.eval()
        batch = tuple(t.to(args.device) for t in batch)
343

344
        with torch.no_grad():
345
346
347
            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
348
            example_indices = batch[3]
349
350
            if args.model_type in ["xlnet", "xlm"]:
                inputs.update({"cls_index": batch[4], "p_mask": batch[5]})
351

352
353
354
355
356
            outputs = model(**inputs)

        for i, example_index in enumerate(example_indices):
            eval_feature = features[example_index.item()]
            unique_id = int(eval_feature.unique_id)
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375

            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,
376
                )
377

378
            else:
379
380
381
                start_logits, end_logits = output
                result = SquadResult(unique_id, start_logits, end_logits)

382
383
            all_results.append(result)

384
385
386
    evalTime = timeit.default_timer() - start_time
    logger.info("  Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))

387
388
389
    # 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))
390

391
392
393
394
395
    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

396
    if args.model_type in ["xlnet", "xlm"]:
397
        # XLNet uses a more complex post-processing procedure
398
        predictions = compute_predictions_log_probs(
399
400
401
402
403
404
405
406
407
408
409
410
411
412
            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,
        )
413
    else:
414
        predictions = compute_predictions_logits(
415
416
417
418
419
420
421
422
423
424
425
426
            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,
427
            tokenizer,
428
        )
429

430
431
    # Compute the F1 and exact scores.
    results = squad_evaluate(examples, predictions)
432
433
434
435
436
    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:
437
        # 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
438
        torch.distributed.barrier()
439
440
441

    # Load data features from cache or dataset file
    input_file = args.predict_file if evaluate else args.train_file
442
443
    cached_features_file = os.path.join(
        os.path.dirname(input_file),
444
        "cached_distillation_{}_{}_{}".format(
445
446
447
448
449
            "dev" if evaluate else "train",
            list(filter(None, args.model_name_or_path.split("/"))).pop(),
            str(args.max_seq_length),
        ),
    )
450
    if os.path.exists(cached_features_file) and not args.overwrite_cache:
451
        logger.info("Loading features from cached file %s", cached_features_file)
452
453
454
455
456
457
458
459
460
461
462
463
464
        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
            )
465
466
    else:
        logger.info("Creating features from dataset file at %s", input_file)
467
468
        processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
        if evaluate:
469
            examples = processor.get_dev_examples(args.data_dir, filename=args.predict_file)
470
        else:
471
            examples = processor.get_train_examples(args.data_dir, filename=args.train_file)
472
473

        features, dataset = squad_convert_examples_to_features(
474
475
476
477
478
479
            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,
480
            return_dataset="pt",
Victor SANH's avatar
indents  
Victor SANH committed
481
            threads=args.threads,
482
        )
483

484
485
        if args.local_rank in [-1, 0]:
            logger.info("Saving features into cached file %s", cached_features_file)
486
            torch.save({"features": features, "dataset": dataset, "examples": examples}, cached_features_file)
487
488

    if args.local_rank == 0 and not evaluate:
Victor SANH's avatar
indents  
Victor SANH committed
489
        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
490
        torch.distributed.barrier()
491
492
493
494
495
496
497
498
499

    if output_examples:
        return dataset, examples, features
    return dataset


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

500
    # Required parameters
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,
513
        help="Path to pretrained model or model identifier from huggingface.co/models",
514
515
516
517
518
519
520
521
    )
    parser.add_argument(
        "--output_dir",
        default=None,
        type=str,
        required=True,
        help="The output directory where the model checkpoints and predictions will be written.",
    )
522
523

    # Distillation parameters (optional)
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
    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."
    )
545

546
    # Other parameters
547
548
549
550
551
552
553
    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
554
    parser.add_argument(
555
        "--train_file",
Victor SANH's avatar
indents  
Victor SANH committed
556
557
558
        default=None,
        type=str,
        help="The input training file. If a data dir is specified, will look for the file there"
559
560
561
562
563
564
565
566
567
        + "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.",
    )
568
569
570
571
572
573
574
575
576
577
578
579
580
    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,
581
        help="Where do you want to store the pre-trained models downloaded from huggingface.co",
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
627
628
629
630
631
632
633
634
635
    )

    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.",
    )
636
    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
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
691
692
693
694
695
696
697
698
699
    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.")
700

Victor SANH's avatar
indents  
Victor SANH committed
701
    parser.add_argument("--threads", type=int, default=1, help="multiple threads for converting example to features")
702
703
    args = parser.parse_args()

704
705
706
707
708
709
710
711
712
713
714
    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
            )
        )
715
716
717
718
719

    # 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
720

721
722
723
724
725
726
727
        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")
728
        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
729
730
731
    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)
732
        torch.distributed.init_process_group(backend="nccl")
733
734
735
736
        args.n_gpu = 1
    args.device = device

    # Setup logging
737
738
739
740
741
742
743
744
745
746
747
748
749
    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,
    )
750
751
752
753
754
    # Set the verbosity to info of the Transformers logger (on main process only):
    if is_main_process(args.local_rank):
        transformers.utils.logging.set_verbosity_info()
        transformers.utils.logging.enable_default_handler()
        transformers.utils.logging.enable_explicit_format()
755
756
757
758
759
    # 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
760
        # Make sure only the first process in distributed training will download model & vocab
Victor SANH's avatar
Victor SANH committed
761
        torch.distributed.barrier()
762
763
764

    args.model_type = args.model_type.lower()
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
    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,
    )
780
781
782

    if args.teacher_type is not None:
        assert args.teacher_name_or_path is not None
783
784
785
        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."
786
        teacher_config_class, teacher_model_class, _ = MODEL_CLASSES[args.teacher_type]
787
788
789
790
791
792
        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
        )
793
794
795
796
797
        teacher.to(args.device)
    else:
        teacher = None

    if args.local_rank == 0:
Victor SANH's avatar
indents  
Victor SANH committed
798
        # Make sure only the first process in distributed training will download model & vocab
799
        torch.distributed.barrier()
800
801
802
803
804

    model.to(args.device)

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

805
806
807
808
809
810
811
812
813
814
815
    # 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.")

816
817
818
819
820
821
822
823
824
825
826
    # 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):
        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()`
827
828
829
        model_to_save = (
            model.module if hasattr(model, "module") else model
        )  # Take care of distributed/parallel training
830
831
832
833
        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
834
        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
835
836

        # Load a trained model and vocabulary that you have fine-tuned
837
838
        model = model_class.from_pretrained(args.output_dir)
        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
839
840
841
842
843
        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]:
844
845
        if args.do_train:
            logger.info("Loading checkpoints saved during training for evaluation")
846
847
        checkpoints = [args.output_dir]
        if args.eval_all_checkpoints:
848
849
850
            checkpoints = list(
                os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
            )
851
852
853
854
855

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

        for checkpoint in checkpoints:
            # Reload the model
856
            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
857
            model = model_class.from_pretrained(checkpoint)
858
859
860
861
862
            model.to(args.device)

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

863
            result = dict((k + ("_{}".format(global_step) if global_step else ""), v) for k, v in result.items())
864
865
866
867
868
869
870
871
872
            results.update(result)

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

    return results


if __name__ == "__main__":
    main()