run_squad.py 26.2 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.
thomwolf's avatar
thomwolf committed
16
""" Finetuning the library models for question-answering on SQuAD (Bert, XLM, XLNet)."""
17
18
19
20
21
22
23

from __future__ import absolute_import, division, print_function

import argparse
import logging
import os
import random
thomwolf's avatar
thomwolf committed
24
import glob
25
26
27
28
29
30
31
32
33
34

import numpy as np
import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)
from torch.utils.data.distributed import DistributedSampler
from tqdm import tqdm, trange

from tensorboardX import SummaryWriter

thomwolf's avatar
thomwolf committed
35
36
37
38
39
40
41
42
from pytorch_transformers import (WEIGHTS_NAME, BertConfig,
                                  BertForQuestionAnswering, BertTokenizer,
                                  XLMConfig, XLMForQuestionAnswering,
                                  XLMTokenizer, XLNetConfig,
                                  XLNetForQuestionAnswering,
                                  XLNetTokenizer)

from pytorch_transformers import AdamW, WarmupLinearSchedule
43
44
45

from utils_squad import read_squad_examples, convert_examples_to_features, RawResult, write_predictions

thomwolf's avatar
thomwolf committed
46
47
48
# 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)
49
50
from utils_squad_evaluate import EVAL_OPTS, main as evaluate_on_squad

51
52
logger = logging.getLogger(__name__)

thomwolf's avatar
thomwolf committed
53
54
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) \
                  for conf in (BertConfig, XLNetConfig, XLMConfig)), ())
thomwolf's avatar
thomwolf committed
55
56

MODEL_CLASSES = {
thomwolf's avatar
thomwolf committed
57
58
59
    'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
    'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
    'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
thomwolf's avatar
thomwolf committed
60
61
}

thomwolf's avatar
thomwolf committed
62
63
64
65
66
67
68
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)

thomwolf's avatar
thomwolf committed
69

70
def train(args, train_dataset, model, tokenizer):
thomwolf's avatar
thomwolf committed
71
72
73
74
    """ Train the model """
    if args.local_rank in [-1, 0]:
        tb_writer = SummaryWriter()

75
    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
thomwolf's avatar
thomwolf committed
76
77
78
79
    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:
80
        t_total = args.max_steps
thomwolf's avatar
thomwolf committed
81
82
        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
    else:
83
        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
thomwolf's avatar
thomwolf committed
84

85
    # Prepare optimizer and schedule (linear warmup and decay)
thomwolf's avatar
thomwolf committed
86
87
    no_decay = ['bias', 'LayerNorm.weight']
    optimizer_grouped_parameters = [
88
        {'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
89
90
        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
        ]
91
92
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
thomwolf's avatar
thomwolf committed
93
94
95
96
97
98
99
100
101
102
103
    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)

    # Train!
    logger.info("***** Running training *****")
    logger.info("  Num examples = %d", len(train_dataset))
    logger.info("  Num Epochs = %d", args.num_train_epochs)
104
105
106
    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
107
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
108
    logger.info("  Total optimization steps = %d", t_total)
thomwolf's avatar
thomwolf committed
109
110
111

    global_step = 0
    tr_loss, logging_loss = 0.0, 0.0
112
113
114
115
116
117
118
    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
119
            batch = tuple(t.to(args.device) for t in batch)
120
121
122
123
124
            inputs = {'input_ids':       batch[0],
                      'token_type_ids':  batch[1] if args.model_type in ['bert', 'xlnet'] else None,  # XLM don't use segment_ids
                      'attention_mask':  batch[2],
                      'start_positions': batch[3],
                      'end_positions':   batch[4]}
thomwolf's avatar
thomwolf committed
125
            ouputs = model(**inputs)
126
            loss = ouputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)
thomwolf's avatar
thomwolf committed
127

128
            if args.n_gpu > 1:
thomwolf's avatar
thomwolf committed
129
                loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training
130
131
            if args.gradient_accumulation_steps > 1:
                loss = loss / args.gradient_accumulation_steps
thomwolf's avatar
thomwolf committed
132

133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
            if args.fp16:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
                torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
            else:
                loss.backward()
                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)

            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0:
                scheduler.step()  # Update learning rate schedule
                optimizer.step()
                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
175
176
177
    if args.local_rank in [-1, 0]:
        tb_writer.close()

178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
    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)

    # 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 = []
    for batch in tqdm(eval_dataloader, desc="Evaluating"):
        model.eval()
        batch = tuple(t.to(args.device) for t in batch)
        example_indices = batch[3]
        with torch.no_grad():
            inputs = {'input_ids':      batch[0],
                        'token_type_ids': batch[1] if args.model_type in ['bert', 'xlnet'] else None,  # XLM don't use segment_ids
                        'attention_mask': batch[2]}
            outputs = model(**inputs)
            batch_start_logits, batch_end_logits = outputs[:2]

        for i, example_index in enumerate(example_indices):
            start_logits = batch_start_logits[i].detach().cpu().tolist()
            end_logits = batch_end_logits[i].detach().cpu().tolist()
            eval_feature = features[example_index.item()]
            unique_id = int(eval_feature.unique_id)
            all_results.append(RawResult(unique_id=unique_id,
                                         start_logits=start_logits,
                                         end_logits=end_logits))

thomwolf's avatar
thomwolf committed
217
    # Compute predictions
218
219
220
    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))
    output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
thomwolf's avatar
thomwolf committed
221
222
223
224
    write_predictions(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)
225

thomwolf's avatar
thomwolf committed
226
    # Evaluate with the official SQuAD script
227
228
229
230
231
232
233
234
235
236
237
    evaluate_options = EVAL_OPTS(data_file=args.predict_file,
                                 pred_file=output_prediction_file,
                                 na_prob_file=output_null_log_odds_file)
    results = evaluate_on_squad(evaluate_options)
    return results


def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
    # Load data features from cache or dataset file
    input_file = args.predict_file if evaluate else args.train_file
    cached_features_file = os.path.join(os.path.dirname(input_file), 'cached_{}_{}_{}'.format(
thomwolf's avatar
thomwolf committed
238
239
        'dev' if evaluate else 'train',
        list(filter(None, args.model_name.split('/'))).pop(),
240
241
        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
242
243
244
        logger.info("Loading features from cached file %s", cached_features_file)
        features = torch.load(cached_features_file)
    else:
245
246
247
248
249
250
251
252
253
254
        logger.info("Creating features from dataset file at %s", input_file)
        examples = read_squad_examples(input_file=input_file,
                                       is_training=not evaluate,
                                       version_2_with_negative=args.version_2_with_negative)
        features = convert_examples_to_features(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)
thomwolf's avatar
thomwolf committed
255
256
257
258
259
        if args.local_rank in [-1, 0]:
            logger.info("Saving features into cached file %s", cached_features_file)
            torch.save(features, cached_features_file)

    # Convert to Tensors and build dataset
260
261
262
263
    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
    all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
    all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
    if evaluate:
thomwolf's avatar
thomwolf committed
264
265
        all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index)
266
267
268
269
    else:
        all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long)
        all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long)
        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_start_positions, all_end_positions)
thomwolf's avatar
thomwolf committed
270

271
272
    if output_examples:
        return dataset, examples, features
thomwolf's avatar
thomwolf committed
273
274
    return dataset

275
276
277
278
279

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

    ## Required parameters
thomwolf's avatar
thomwolf committed
280
281
282
283
284
285
    parser.add_argument("--train_file", default=None, type=str, required=True,
                        help="SQuAD json for training. E.g., train-v1.1.json")
    parser.add_argument("--predict_file", default=None, type=str, required=True,
                        help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json")
    parser.add_argument("--model_name", default=None, type=str, required=True,
                        help="Bert/XLNet/XLM pre-trained model selected in the list: " + ", ".join(ALL_MODELS))
286
287
288
289
    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
290
291
292
293
294
295
296
    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
297
298
299
300
301
    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.")

302
303
304
305
306
307
308
309
    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
310
311
    parser.add_argument("--do_train", action='store_true',
                        help="Whether to run training.")
312
    parser.add_argument("--do_eval", action='store_true',
thomwolf's avatar
thomwolf committed
313
                        help="Whether to run eval on the dev set.")
314
315
    parser.add_argument("--evaluate_during_training", action='store_true',
                        help="Rul evaluation during training at each logging step.")
thomwolf's avatar
thomwolf committed
316
    parser.add_argument("--do_lower_case", action='store_true',
317
                        help="Set this flag if you are using an uncased model.")
thomwolf's avatar
thomwolf committed
318

319
320
321
322
    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
323
324
325
326
    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.")
327
328
329
330
331
332
    parser.add_argument("--weight_decay", default=0.0, type=float,
                        help="Weight deay if we apply some.")
    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.")
333
334
    parser.add_argument("--num_train_epochs", default=3.0, type=float,
                        help="Total number of training epochs to perform.")
335
336
337
338
    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.")
339
    parser.add_argument("--n_best_size", default=20, type=int,
thomwolf's avatar
thomwolf committed
340
                        help="The total number of n-best predictions to generate in the nbest_predictions.json output file.")
341
342
343
344
345
346
    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
347

348
349
350
351
352
353
    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
354
    parser.add_argument("--no_cuda", action='store_true',
355
                        help="Whether not to use CUDA when available")
356
357
358
359
    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
360
    parser.add_argument('--seed', type=int, default=42,
361
                        help="random seed for initialization")
362

thomwolf's avatar
thomwolf committed
363
    parser.add_argument("--local_rank", type=int, default=-1,
364
                        help="local_rank for distributed training on gpus")
thomwolf's avatar
thomwolf committed
365
366
367
368
369
    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")
370
371
372
373
    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
374
375
376
    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))

377
    # Setup distant debugging if needed
378
379
380
381
382
383
384
    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
385
    # Setup CUDA, GPU & distributed training
386
387
    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
388
389
        args.n_gpu = torch.cuda.device_count()
    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
390
391
392
        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
393
394
        args.n_gpu = 1
    args.device = device
395

thomwolf's avatar
thomwolf committed
396
    # Setup logging
397
398
399
    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
400
    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
401
                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
402

403
404
    # Set seed
    set_seed(args)
405

thomwolf's avatar
thomwolf committed
406
    # Load pretrained model and tokenizer
407
    if args.local_rank not in [-1, 0]:
408
409
410
411
412
413
414
415
416
417
418
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab

    args.model_type = ""
    for key in MODEL_CLASSES:
        if key in args.model_name.lower():
            args.model_type = key  # take the first match in model types
            break
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name)
    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name, do_lower_case=args.do_lower_case)
    model = model_class.from_pretrained(args.model_name, from_tf=bool('.ckpt' in args.model_name), config=config)
419
420

    if args.local_rank == 0:
421
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
422

thomwolf's avatar
thomwolf committed
423
424
    # Distributed and parrallel training
    model.to(args.device)
425
    if args.local_rank != -1:
thomwolf's avatar
thomwolf committed
426
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
427
428
                                                          output_device=args.local_rank,
                                                          find_unused_parameters=True)
thomwolf's avatar
thomwolf committed
429
    elif args.n_gpu > 1:
430
431
        model = torch.nn.DataParallel(model)

432
433
    logger.info("Training/evaluation parameters %s", args)

thomwolf's avatar
thomwolf committed
434
    # Training
435
    if args.do_train:
436
437
438
        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
439

440

thomwolf's avatar
thomwolf committed
441
    # Save the trained model and the tokenizer
442
443
444
445
446
447
448
449
450
451
452
    if 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()`
        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)
453
454

        # Good practice: save your training arguments together with the trained model
455
        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
456

457
458
459
460
461
462
        # Load a trained model and vocabulary that you have fine-tuned
        model = model_class.from_pretrained(args.output_dir)
        tokenizer = tokenizer_class.from_pretrained(args.output_dir)
        model.to(args.device)


thomwolf's avatar
thomwolf committed
463
    # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
464
465
466
467
468
    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)))
thomwolf's avatar
thomwolf committed
469
470
            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce model loading logs

471
        logger.info("Evaluate the following checkpoints: %s", checkpoints)
thomwolf's avatar
thomwolf committed
472

473
        for checkpoint in checkpoints:
thomwolf's avatar
thomwolf committed
474
            # Reload the model
475
476
477
            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
478
479

            # Evaluate
480
            result = evaluate(args, model, tokenizer, prefix=global_step)
thomwolf's avatar
thomwolf committed
481

482
483
            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
484

485
    logger.info("Results: {}".format(results))
thomwolf's avatar
thomwolf committed
486

487
    return results
488
489
490
491


if __name__ == "__main__":
    main()