run_squad.py 29.6 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
""" Finetuning the library models for question-answering on SQuAD (DistilBERT, Bert, XLM, XLNet)."""
17
18

from __future__ import absolute_import, division, print_function
LysandreJik's avatar
LysandreJik committed
19
from transformers.data.processors.squad import SquadV1Processor, SquadV2Processor, SquadResult
20
from transformers.data.metrics.squad_metrics import compute_predictions_logits, compute_predictions_log_probs, squad_evaluate
21
22
23
24
25

import argparse
import logging
import os
import random
thomwolf's avatar
thomwolf committed
26
import glob
27
import timeit
28
29
30
31
32
33
import numpy as np
import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)
from torch.utils.data.distributed import DistributedSampler

34
35
36
37
38
39
try:
    from torch.utils.tensorboard import SummaryWriter
except:
    from tensorboardX import SummaryWriter

from tqdm import tqdm, trange
40

41
from transformers import (WEIGHTS_NAME, BertConfig,
thomwolf's avatar
thomwolf committed
42
43
44
45
                                  BertForQuestionAnswering, BertTokenizer,
                                  XLMConfig, XLMForQuestionAnswering,
                                  XLMTokenizer, XLNetConfig,
                                  XLNetForQuestionAnswering,
46
47
                                  XLNetTokenizer,
                                  DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
thomwolf's avatar
thomwolf committed
48

Lysandre's avatar
Lysandre committed
49
from transformers import AdamW, get_linear_schedule_with_warmup, squad_convert_examples_to_features
50

Lysandre's avatar
Lysandre committed
51
from utils_squad import (convert_examples_to_features as old_convert, read_squad_examples as old_read, RawResult, write_predictions,
52
                         RawResultExtended, write_predictions_extended)
53

thomwolf's avatar
thomwolf committed
54
55
56
# The follwing import is the official SQuAD evaluation script (2.0).
# You can remove it from the dependencies if you are using this script outside of the library
# We've added it here for automated tests (see examples/test_examples.py file)
57
58
from utils_squad_evaluate import EVAL_OPTS, main as evaluate_on_squad

59
60
logger = logging.getLogger(__name__)

thomwolf's avatar
thomwolf committed
61
62
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) \
                  for conf in (BertConfig, XLNetConfig, XLMConfig)), ())
thomwolf's avatar
thomwolf committed
63
64

MODEL_CLASSES = {
thomwolf's avatar
thomwolf committed
65
66
67
    'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
    'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
    'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
68
    'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
thomwolf's avatar
thomwolf committed
69
70
}

thomwolf's avatar
thomwolf committed
71
72
73
74
75
76
77
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)

78
79
def to_list(tensor):
    return tensor.detach().cpu().tolist()
thomwolf's avatar
thomwolf committed
80

81
def train(args, train_dataset, model, tokenizer):
thomwolf's avatar
thomwolf committed
82
83
84
85
    """ Train the model """
    if args.local_rank in [-1, 0]:
        tb_writer = SummaryWriter()

86
    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
thomwolf's avatar
thomwolf committed
87
88
89
90
    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:
91
        t_total = args.max_steps
thomwolf's avatar
thomwolf committed
92
93
        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
    else:
94
        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
thomwolf's avatar
thomwolf committed
95

96
    # Prepare optimizer and schedule (linear warmup and decay)
thomwolf's avatar
thomwolf committed
97
98
    no_decay = ['bias', 'LayerNorm.weight']
    optimizer_grouped_parameters = [
99
        {'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},
thomwolf's avatar
thomwolf committed
100
101
        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]
102
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
103
    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
thomwolf's avatar
thomwolf committed
104
105
106
107
108
109
110
    if args.fp16:
        try:
            from apex import amp
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)

111
112
113
114
    # multi-gpu training (should be after apex fp16 initialization)
    if args.n_gpu > 1:
        model = torch.nn.DataParallel(model)

thomwolf's avatar
thomwolf committed
115
116
117
118
119
120
    # Distributed training (should be after apex fp16 initialization)
    if args.local_rank != -1:
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
                                                          output_device=args.local_rank,
                                                          find_unused_parameters=True)

thomwolf's avatar
thomwolf committed
121
122
123
124
    # Train!
    logger.info("***** Running training *****")
    logger.info("  Num examples = %d", len(train_dataset))
    logger.info("  Num Epochs = %d", args.num_train_epochs)
125
126
127
    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
    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))
thomwolf's avatar
thomwolf committed
128
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
129
    logger.info("  Total optimization steps = %d", t_total)
thomwolf's avatar
thomwolf committed
130
131
132

    global_step = 0
    tr_loss, logging_loss = 0.0, 0.0
133
134
135
136
137
138
139
    model.zero_grad()
    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
    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):
            model.train()
thomwolf's avatar
thomwolf committed
140
            batch = tuple(t.to(args.device) for t in batch)
141
            inputs = {'input_ids':       batch[0],
Simon Layton's avatar
Simon Layton committed
142
143
                      'attention_mask':  batch[1],
                      'start_positions': batch[3],
144
                      'end_positions':   batch[4]}
145
146
            if args.model_type != 'distilbert':
                inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
147
148
            if args.model_type in ['xlnet', 'xlm']:
                inputs.update({'cls_index': batch[5],
thomwolf's avatar
thomwolf committed
149
                               'p_mask':       batch[6]})
Peiqin Lin's avatar
typos  
Peiqin Lin committed
150
            outputs = model(**inputs)
151
            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
thomwolf's avatar
thomwolf committed
152

153
            if args.n_gpu > 1:
thomwolf's avatar
thomwolf committed
154
                loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training
155
156
            if args.gradient_accumulation_steps > 1:
                loss = loss / args.gradient_accumulation_steps
thomwolf's avatar
thomwolf committed
157

158
159
160
161
162
163
164
165
            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:
166
167
168
169
170
                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)

171
                optimizer.step()
172
                scheduler.step()  # Update learning rate schedule
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
                model.zero_grad()
                global_step += 1

                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
                    # Log metrics
                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
                        results = evaluate(args, model, tokenizer)
                        for key, value in results.items():
                            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)
                    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
                    output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
                    if not os.path.exists(output_dir):
                        os.makedirs(output_dir)
                    model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
                    model_to_save.save_pretrained(output_dir)
                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
                    logger.info("Saving model checkpoint to %s", output_dir)

            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

thomwolf's avatar
thomwolf committed
203
204
205
    if args.local_rank in [-1, 0]:
        tb_writer.close()

206
207
208
209
210
211
212
213
214
215
216
217
218
219
    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)
    # Note that DistributedSampler samples randomly
    eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

ronakice's avatar
ronakice committed
220
221
222
223
    # multi-gpu evaluate
    if args.n_gpu > 1:
        model = torch.nn.DataParallel(model)

224
225
226
227
228
    # Eval!
    logger.info("***** Running evaluation {} *****".format(prefix))
    logger.info("  Num examples = %d", len(dataset))
    logger.info("  Batch size = %d", args.eval_batch_size)
    all_results = []
229
    start_time = timeit.default_timer()
230
231
232
233
    for batch in tqdm(eval_dataloader, desc="Evaluating"):
        model.eval()
        batch = tuple(t.to(args.device) for t in batch)
        with torch.no_grad():
LysandreJik's avatar
LysandreJik committed
234
235
236
237
238
            inputs = {
                'input_ids':      batch[0],
                'attention_mask': batch[1]
            }
            
239
240
            if args.model_type != 'distilbert':
                inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]  # XLM don't use segment_ids
241
242
243
244
            example_indices = batch[3]
            if args.model_type in ['xlnet', 'xlm']:
                inputs.update({'cls_index': batch[4],
                               'p_mask':    batch[5]})
245
246
247
248
249
            outputs = model(**inputs)

        for i, example_index in enumerate(example_indices):
            eval_feature = features[example_index.item()]
            unique_id = int(eval_feature.unique_id)
LysandreJik's avatar
LysandreJik committed
250

LysandreJik's avatar
LysandreJik committed
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
            output = [to_list(output[i]) for output in outputs]

            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
                )

            else:
                start_logits, end_logits = output
                result = SquadResult(
                    unique_id, start_logits, end_logits
                )

273
            all_results.append(result)
274

275
276
277
    evalTime = timeit.default_timer() - start_time
    logger.info("  Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))

thomwolf's avatar
thomwolf committed
278
    # Compute predictions
279
280
    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))
281
282
283
284
    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
285
286
287

    if args.model_type in ['xlnet', 'xlm']:
        # XLNet uses a more complex post-processing procedure
288
        predictions = compute_predictions_log_probs(examples, features, all_results, args.n_best_size,
289
290
                        args.max_answer_length, output_prediction_file,
                        output_nbest_file, output_null_log_odds_file, args.predict_file,
291
292
                        model.config.start_n_top, model.config.end_n_top,
                        args.version_2_with_negative, tokenizer, args.verbose_logging)
293
    else:
294
        predictions = compute_predictions_logits(examples, features, all_results, args.n_best_size,
295
296
297
                        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)
298

LysandreJik's avatar
LysandreJik committed
299
    results = squad_evaluate(examples, predictions)
300
301
302
    return results

def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
VictorSanh's avatar
VictorSanh committed
303
    if args.local_rank not in [-1, 0] and not evaluate:
thomwolf's avatar
thomwolf committed
304
305
        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache

306
    # Load data features from cache or dataset file
LysandreJik's avatar
Cleanup  
LysandreJik committed
307
308
    input_dir = args.data_dir if args.data_dir else "."
    cached_features_file = os.path.join(input_dir, 'cached_{}_{}_{}'.format(
thomwolf's avatar
thomwolf committed
309
        'dev' if evaluate else 'train',
310
        list(filter(None, args.model_name_or_path.split('/'))).pop(),
311
312
        str(args.max_seq_length)))
    if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
thomwolf's avatar
thomwolf committed
313
        logger.info("Loading features from cached file %s", cached_features_file)
314
315
        features_and_dataset = torch.load(cached_features_file)
        features, dataset = features_and_dataset["features"], features_and_dataset["dataset"]
thomwolf's avatar
thomwolf committed
316
    else:
LysandreJik's avatar
Cleanup  
LysandreJik committed
317
        logger.info("Creating features from dataset file at %s", input_dir)
Lysandre's avatar
Lysandre committed
318

LysandreJik's avatar
Cleanup  
LysandreJik committed
319
320
321
322
323
324
325
326
327
328
329
330
331
332
        if not args.data_dir:
            try:
                import tensorflow_datasets as tfds
            except ImportError:
                raise ImportError("If not data_dir is specified, tensorflow_datasets needs to be installed.")

            if args.version_2_with_negative:
                logger.warn("tensorflow_datasets does not handle version 2 of SQuAD.")

            tfds_examples = tfds.load("squad")
            examples = SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=evaluate)
        else:
            processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
            examples = processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
LysandreJik's avatar
LysandreJik committed
333

334
        features, dataset = squad_convert_examples_to_features( 
Lysandre's avatar
Lysandre committed
335
336
337
338
339
340
            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,
341
            return_dataset='pt'
Lysandre's avatar
Lysandre committed
342
343
344
        )


thomwolf's avatar
thomwolf committed
345
346
        if args.local_rank in [-1, 0]:
            logger.info("Saving features into cached file %s", cached_features_file)
347
            torch.save({"features": features, "dataset": dataset}, cached_features_file)
thomwolf's avatar
thomwolf committed
348

VictorSanh's avatar
VictorSanh committed
349
    if args.local_rank == 0 and not evaluate:
thomwolf's avatar
thomwolf committed
350
351
        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache

352
353
    if output_examples:
        return dataset, examples, features
thomwolf's avatar
thomwolf committed
354
355
    return dataset

356
357
358
359
360

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

    ## Required parameters
361
362
363
364
    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))
365
366
367
368
    parser.add_argument("--output_dir", default=None, type=str, required=True,
                        help="The output directory where the model checkpoints and predictions will be written.")

    ## Other parameters
LysandreJik's avatar
Cleanup  
LysandreJik committed
369
370
    parser.add_argument("--data_dir", default=None, type=str,
                        help="The input data dir. Should contain the .json files for the task. If not specified, will run with tensorflow_datasets.")
371
372
373
374
375
376
377
    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")

thomwolf's avatar
thomwolf committed
378
379
380
381
382
    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.")

383
384
385
386
387
388
389
390
    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.")
thomwolf's avatar
thomwolf committed
391
392
    parser.add_argument("--do_train", action='store_true',
                        help="Whether to run training.")
393
    parser.add_argument("--do_eval", action='store_true',
thomwolf's avatar
thomwolf committed
394
                        help="Whether to run eval on the dev set.")
395
396
    parser.add_argument("--evaluate_during_training", action='store_true',
                        help="Rul evaluation during training at each logging step.")
thomwolf's avatar
thomwolf committed
397
    parser.add_argument("--do_lower_case", action='store_true',
398
                        help="Set this flag if you are using an uncased model.")
thomwolf's avatar
thomwolf committed
399

400
401
402
403
    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.")
thomwolf's avatar
thomwolf committed
404
405
406
407
    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.")
408
    parser.add_argument("--weight_decay", default=0.0, type=float,
Kazutoshi Shinoda's avatar
Kazutoshi Shinoda committed
409
                        help="Weight decay if we apply some.")
410
411
412
413
    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.")
414
415
    parser.add_argument("--num_train_epochs", default=3.0, type=float,
                        help="Total number of training epochs to perform.")
416
417
418
419
    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.")
420
    parser.add_argument("--n_best_size", default=20, type=int,
thomwolf's avatar
thomwolf committed
421
                        help="The total number of n-best predictions to generate in the nbest_predictions.json output file.")
422
423
424
425
426
427
    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.")
thomwolf's avatar
thomwolf committed
428

429
430
431
432
433
434
    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")
thomwolf's avatar
thomwolf committed
435
    parser.add_argument("--no_cuda", action='store_true',
436
                        help="Whether not to use CUDA when available")
437
438
439
440
    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")
thomwolf's avatar
thomwolf committed
441
    parser.add_argument('--seed', type=int, default=42,
442
                        help="random seed for initialization")
443

thomwolf's avatar
thomwolf committed
444
    parser.add_argument("--local_rank", type=int, default=-1,
445
                        help="local_rank for distributed training on gpus")
thomwolf's avatar
thomwolf committed
446
447
448
449
450
    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")
451
452
453
454
    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.")
    args = parser.parse_args()

thomwolf's avatar
thomwolf committed
455
456
457
    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))

458
    # Setup distant debugging if needed
459
460
461
462
463
464
465
    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
        print("Waiting for debugger attach")
        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
        ptvsd.wait_for_attach()

thomwolf's avatar
thomwolf committed
466
    # Setup CUDA, GPU & distributed training
467
468
    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")
thomwolf's avatar
thomwolf committed
469
470
        args.n_gpu = torch.cuda.device_count()
    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
471
472
473
        torch.cuda.set_device(args.local_rank)
        device = torch.device("cuda", args.local_rank)
        torch.distributed.init_process_group(backend='nccl')
thomwolf's avatar
thomwolf committed
474
475
        args.n_gpu = 1
    args.device = device
476

thomwolf's avatar
thomwolf committed
477
    # Setup logging
478
479
480
    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)
thomwolf's avatar
thomwolf committed
481
    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
482
                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
483

484
485
    # Set seed
    set_seed(args)
486

thomwolf's avatar
thomwolf committed
487
    # Load pretrained model and tokenizer
488
    if args.local_rank not in [-1, 0]:
489
490
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab

491
    args.model_type = args.model_type.lower()
492
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
thomwolf's avatar
thomwolf committed
493
494
495
496
497
498
499
500
501
    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)
502
503

    if args.local_rank == 0:
504
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
505

thomwolf's avatar
thomwolf committed
506
    model.to(args.device)
507

508
509
    logger.info("Training/evaluation parameters %s", args)

Simon Layton's avatar
Simon Layton committed
510
511
512
513
514
515
516
517
518
519
    # 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.")

thomwolf's avatar
thomwolf committed
520
    # Training
521
    if args.do_train:
522
        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
523
        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
524
        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
525

526

thomwolf's avatar
thomwolf committed
527
    # Save the trained model and the tokenizer
Peng Qi's avatar
Peng Qi committed
528
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
529
530
531
532
533
534
535
536
537
538
        # 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()`
        model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
        model_to_save.save_pretrained(args.output_dir)
        tokenizer.save_pretrained(args.output_dir)
539
540

        # Good practice: save your training arguments together with the trained model
541
        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
542

543
544
        # Load a trained model and vocabulary that you have fine-tuned
        model = model_class.from_pretrained(args.output_dir)
Peng Qi's avatar
Peng Qi committed
545
        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
546
547
548
        model.to(args.device)


thomwolf's avatar
thomwolf committed
549
    # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
550
551
552
553
554
    results = {}
    if args.do_eval and args.local_rank in [-1, 0]:
        checkpoints = [args.output_dir]
        if args.eval_all_checkpoints:
            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
555
            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce model loading logs
thomwolf's avatar
thomwolf committed
556

557
        logger.info("Evaluate the following checkpoints: %s", checkpoints)
thomwolf's avatar
thomwolf committed
558

559
        for checkpoint in checkpoints:
thomwolf's avatar
thomwolf committed
560
            # Reload the model
561
562
563
            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
            model = model_class.from_pretrained(checkpoint)
            model.to(args.device)
thomwolf's avatar
thomwolf committed
564
565

            # Evaluate
566
            result = evaluate(args, model, tokenizer, prefix=global_step)
thomwolf's avatar
thomwolf committed
567

568
569
            result = dict((k + ('_{}'.format(global_step) if global_step else ''), v) for k, v in result.items())
            results.update(result)
thomwolf's avatar
thomwolf committed
570

571
    logger.info("Results: {}".format(results))
thomwolf's avatar
thomwolf committed
572

573
    return results
574
575
576
577


if __name__ == "__main__":
    main()