run_classifier.py 27 KB
Newer Older
1
# coding=utf-8
thomwolf's avatar
thomwolf committed
2
# Copyright 2018 The Google AI Language Team Authors and The HugginFace Inc. team.
3
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
4
5
6
7
8
9
10
11
12
13
14
15
16
17
#
# 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.
"""BERT finetuning runner."""

thomwolf's avatar
thomwolf committed
18
from __future__ import absolute_import, division, print_function
19

thomwolf's avatar
thomwolf committed
20
import argparse
21
import csv
22
import logging
thomwolf's avatar
thomwolf committed
23
import os
VictorSanh's avatar
VictorSanh committed
24
import random
thomwolf's avatar
thomwolf committed
25
import sys
thomwolf's avatar
thomwolf committed
26
27

import numpy as np
VictorSanh's avatar
VictorSanh committed
28
import torch
thomwolf's avatar
thomwolf committed
29
30
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)
31
from torch.utils.data.distributed import DistributedSampler
thomwolf's avatar
thomwolf committed
32
from tqdm import tqdm, trange
33

thomwolf's avatar
thomwolf committed
34
from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE
35
from pytorch_pretrained_bert.modeling import BertForSequenceClassification, BertConfig, WEIGHTS_NAME, CONFIG_NAME
thomwolf's avatar
thomwolf committed
36
from pytorch_pretrained_bert.tokenization import BertTokenizer
37
from pytorch_pretrained_bert.optimization import BertAdam, warmup_linear
38

39
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
40
41
42
                    datefmt = '%m/%d/%Y %H:%M:%S',
                    level = logging.INFO)
logger = logging.getLogger(__name__)
43

44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73

class InputExample(object):
    """A single training/test example for simple sequence classification."""

    def __init__(self, guid, text_a, text_b=None, label=None):
        """Constructs a InputExample.

        Args:
            guid: Unique id for the example.
            text_a: string. The untokenized text of the first sequence. For single
            sequence tasks, only this sequence must be specified.
            text_b: (Optional) string. The untokenized text of the second sequence.
            Only must be specified for sequence pair tasks.
            label: (Optional) string. The label of the example. This should be
            specified for train and dev examples, but not for test examples.
        """
        self.guid = guid
        self.text_a = text_a
        self.text_b = text_b
        self.label = label


class InputFeatures(object):
    """A single set of features of data."""

    def __init__(self, input_ids, input_mask, segment_ids, label_id):
        self.input_ids = input_ids
        self.input_mask = input_mask
        self.segment_ids = segment_ids
        self.label_id = label_id
thomwolf's avatar
thomwolf committed
74

75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93

class DataProcessor(object):
    """Base class for data converters for sequence classification data sets."""

    def get_train_examples(self, data_dir):
        """Gets a collection of `InputExample`s for the train set."""
        raise NotImplementedError()

    def get_dev_examples(self, data_dir):
        """Gets a collection of `InputExample`s for the dev set."""
        raise NotImplementedError()

    def get_labels(self):
        """Gets the list of labels for this data set."""
        raise NotImplementedError()

    @classmethod
    def _read_tsv(cls, input_file, quotechar=None):
        """Reads a tab separated value file."""
94
        with open(input_file, "r") as f:
95
96
97
            reader = csv.reader(f, delimiter="\t", quotechar=quotechar)
            lines = []
            for line in reader:
thomwolf's avatar
thomwolf committed
98
99
                if sys.version_info[0] == 2:
                    line = list(unicode(cell, 'utf-8') for cell in line)
100
101
                lines.append(line)
            return lines
thomwolf's avatar
thomwolf committed
102
103


VictorSanh's avatar
wip  
VictorSanh committed
104
105
106
107
108
class MrpcProcessor(DataProcessor):
    """Processor for the MRPC data set (GLUE version)."""

    def get_train_examples(self, data_dir):
        """See base class."""
thomwolf's avatar
thomwolf committed
109
        logger.info("LOOKING AT {}".format(os.path.join(data_dir, "train.tsv")))
VictorSanh's avatar
wip  
VictorSanh committed
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
        return self._create_examples(
            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")

    def get_dev_examples(self, data_dir):
        """See base class."""
        return self._create_examples(
            self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")

    def get_labels(self):
        """See base class."""
        return ["0", "1"]

    def _create_examples(self, lines, set_type):
        """Creates examples for the training and dev sets."""
        examples = []
        for (i, line) in enumerate(lines):
            if i == 0:
                continue
            guid = "%s-%s" % (set_type, i)
129
130
131
            text_a = line[3]
            text_b = line[4]
            label = line[0]
VictorSanh's avatar
wip  
VictorSanh committed
132
133
134
135
            examples.append(
                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
        return examples

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

class MnliProcessor(DataProcessor):
    """Processor for the MultiNLI data set (GLUE version)."""

    def get_train_examples(self, data_dir):
        """See base class."""
        return self._create_examples(
            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")

    def get_dev_examples(self, data_dir):
        """See base class."""
        return self._create_examples(
            self._read_tsv(os.path.join(data_dir, "dev_matched.tsv")),
            "dev_matched")

    def get_labels(self):
        """See base class."""
        return ["contradiction", "entailment", "neutral"]

    def _create_examples(self, lines, set_type):
        """Creates examples for the training and dev sets."""
        examples = []
        for (i, line) in enumerate(lines):
            if i == 0:
                continue
161
            guid = "%s-%s" % (set_type, line[0])
162
163
            text_a = line[8]
            text_b = line[9]
164
            label = line[-1]
165
166
167
            examples.append(
                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
        return examples
thomwolf's avatar
thomwolf committed
168

169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191

class ColaProcessor(DataProcessor):
    """Processor for the CoLA data set (GLUE version)."""

    def get_train_examples(self, data_dir):
        """See base class."""
        return self._create_examples(
            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")

    def get_dev_examples(self, data_dir):
        """See base class."""
        return self._create_examples(
            self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")

    def get_labels(self):
        """See base class."""
        return ["0", "1"]

    def _create_examples(self, lines, set_type):
        """Creates examples for the training and dev sets."""
        examples = []
        for (i, line) in enumerate(lines):
            guid = "%s-%s" % (set_type, i)
192
193
            text_a = line[3]
            label = line[1]
194
195
196
            examples.append(
                InputExample(guid=guid, text_a=text_a, text_b=None, label=label))
        return examples
thomwolf's avatar
thomwolf committed
197
198
199


def convert_examples_to_features(examples, label_list, max_seq_length, tokenizer):
200
201
    """Loads a data file into a list of `InputBatch`s."""

202
    label_map = {label : i for i, label in enumerate(label_list)}
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217

    features = []
    for (ex_index, example) in enumerate(examples):
        tokens_a = tokenizer.tokenize(example.text_a)

        tokens_b = None
        if example.text_b:
            tokens_b = tokenizer.tokenize(example.text_b)
            # Modifies `tokens_a` and `tokens_b` in place so that the total
            # length is less than the specified length.
            # Account for [CLS], [SEP], [SEP] with "- 3"
            _truncate_seq_pair(tokens_a, tokens_b, max_seq_length - 3)
        else:
            # Account for [CLS] and [SEP] with "- 2"
            if len(tokens_a) > max_seq_length - 2:
218
                tokens_a = tokens_a[:(max_seq_length - 2)]
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237

        # The convention in BERT is:
        # (a) For sequence pairs:
        #  tokens:   [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
        #  type_ids: 0   0  0    0    0     0       0 0    1  1  1  1   1 1
        # (b) For single sequences:
        #  tokens:   [CLS] the dog is hairy . [SEP]
        #  type_ids: 0   0   0   0  0     0 0
        #
        # Where "type_ids" are used to indicate whether this is the first
        # sequence or the second sequence. The embedding vectors for `type=0` and
        # `type=1` were learned during pre-training and are added to the wordpiece
        # embedding vector (and position vector). This is not *strictly* necessary
        # since the [SEP] token unambigiously separates the sequences, but it makes
        # it easier for the model to learn the concept of sequences.
        #
        # For classification tasks, the first vector (corresponding to [CLS]) is
        # used as as the "sentence vector". Note that this only makes sense because
        # the entire model is fine-tuned.
238
239
        tokens = ["[CLS]"] + tokens_a + ["[SEP]"]
        segment_ids = [0] * len(tokens)
240
241

        if tokens_b:
242
243
            tokens += tokens_b + ["[SEP]"]
            segment_ids += [1] * (len(tokens_b) + 1)
244
245
246
247
248
249
250
251

        input_ids = tokenizer.convert_tokens_to_ids(tokens)

        # The mask has 1 for real tokens and 0 for padding tokens. Only real
        # tokens are attended to.
        input_mask = [1] * len(input_ids)

        # Zero-pad up to the sequence length.
252
253
254
255
        padding = [0] * (max_seq_length - len(input_ids))
        input_ids += padding
        input_mask += padding
        segment_ids += padding
256
257
258
259
260
261
262
263
264
265

        assert len(input_ids) == max_seq_length
        assert len(input_mask) == max_seq_length
        assert len(segment_ids) == max_seq_length

        label_id = label_map[example.label]
        if ex_index < 5:
            logger.info("*** Example ***")
            logger.info("guid: %s" % (example.guid))
            logger.info("tokens: %s" % " ".join(
266
                    [str(x) for x in tokens]))
267
268
269
270
271
272
273
            logger.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
            logger.info("input_mask: %s" % " ".join([str(x) for x in input_mask]))
            logger.info(
                    "segment_ids: %s" % " ".join([str(x) for x in segment_ids]))
            logger.info("label: %s (id = %d)" % (example.label, label_id))

        features.append(
thomwolf's avatar
thomwolf committed
274
275
276
277
                InputFeatures(input_ids=input_ids,
                              input_mask=input_mask,
                              segment_ids=segment_ids,
                              label_id=label_id))
278
    return features
thomwolf's avatar
thomwolf committed
279
280


281
282
283
284
285
286
287
288
289
290
291
292
293
294
def _truncate_seq_pair(tokens_a, tokens_b, max_length):
    """Truncates a sequence pair in place to the maximum length."""

    # This is a simple heuristic which will always truncate the longer sequence
    # one token at a time. This makes more sense than truncating an equal percent
    # of tokens from each, since if one sequence is very short then each token
    # that's truncated likely contains more information than a longer sequence.
    while True:
        total_length = len(tokens_a) + len(tokens_b)
        if total_length <= max_length:
            break
        if len(tokens_a) > len(tokens_b):
            tokens_a.pop()
        else:
VictorSanh's avatar
VictorSanh committed
295
296
            tokens_b.pop()

297
298
def accuracy(out, labels):
    outputs = np.argmax(out, axis=1)
thomwolf's avatar
thomwolf committed
299
    return np.sum(outputs == labels)
VictorSanh's avatar
WIP  
VictorSanh committed
300

301
def main():
302
303
304
305
306
307
308
309
    parser = argparse.ArgumentParser()

    ## Required parameters
    parser.add_argument("--data_dir",
                        default=None,
                        type=str,
                        required=True,
                        help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
thomwolf's avatar
thomwolf committed
310
311
    parser.add_argument("--bert_model", default=None, type=str, required=True,
                        help="Bert pre-trained model selected in the list: bert-base-uncased, "
312
313
                        "bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
                        "bert-base-multilingual-cased, bert-base-chinese.")
314
315
316
317
318
319
320
321
322
    parser.add_argument("--task_name",
                        default=None,
                        type=str,
                        required=True,
                        help="The name of the task to train.")
    parser.add_argument("--output_dir",
                        default=None,
                        type=str,
                        required=True,
323
                        help="The output directory where the model predictions and checkpoints will be written.")
324
325

    ## Other parameters
326
327
328
329
    parser.add_argument("--cache_dir",
                        default="",
                        type=str,
                        help="Where do you want to store the pre-trained models downloaded from s3")
330
331
332
333
334
335
336
337
338
339
340
341
    parser.add_argument("--max_seq_length",
                        default=128,
                        type=int,
                        help="The maximum total input sequence length after WordPiece tokenization. \n"
                             "Sequences longer than this will be truncated, and sequences shorter \n"
                             "than this will be padded.")
    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.")
thomwolf's avatar
thomwolf committed
342
343
344
    parser.add_argument("--do_lower_case",
                        action='store_true',
                        help="Set this flag if you are using an uncased model.")
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
    parser.add_argument("--train_batch_size",
                        default=32,
                        type=int,
                        help="Total batch size for training.")
    parser.add_argument("--eval_batch_size",
                        default=8,
                        type=int,
                        help="Total batch size for eval.")
    parser.add_argument("--learning_rate",
                        default=5e-5,
                        type=float,
                        help="The initial learning rate for Adam.")
    parser.add_argument("--num_train_epochs",
                        default=3.0,
                        type=float,
                        help="Total number of training epochs to perform.")
    parser.add_argument("--warmup_proportion",
                        default=0.1,
                        type=float,
                        help="Proportion of training to perform linear learning rate warmup for. "
                             "E.g., 0.1 = 10%% of training.")
    parser.add_argument("--no_cuda",
                        action='store_true',
                        help="Whether not to use CUDA when available")
    parser.add_argument("--local_rank",
                        type=int,
                        default=-1,
                        help="local_rank for distributed training on gpus")
373
374
    parser.add_argument('--seed',
                        type=int,
VictorSanh's avatar
VictorSanh committed
375
376
                        default=42,
                        help="random seed for initialization")
377
378
379
    parser.add_argument('--gradient_accumulation_steps',
                        type=int,
                        default=1,
380
                        help="Number of updates steps to accumulate before performing a backward/update pass.")
thomwolf's avatar
thomwolf committed
381
382
383
384
    parser.add_argument('--fp16',
                        action='store_true',
                        help="Whether to use 16-bit float precision instead of 32-bit")
    parser.add_argument('--loss_scale',
385
386
387
388
                        type=float, default=0,
                        help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
                             "0 (default value): dynamic loss scaling.\n"
                             "Positive power of 2: static loss scaling value.\n")
389
390
    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.")
391
392
    args = parser.parse_args()

393
394
395
396
397
398
399
    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()

VictorSanh's avatar
WIP  
VictorSanh committed
400
401
402
403
404
    processors = {
        "cola": ColaProcessor,
        "mnli": MnliProcessor,
        "mrpc": MrpcProcessor,
    }
thomwolf's avatar
thomwolf committed
405

406
407
408
409
410
411
    num_labels_task = {
        "cola": 2,
        "mnli": 3,
        "mrpc": 2,
    }

thomwolf's avatar
thomwolf committed
412
413
414
415
    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")
        n_gpu = torch.cuda.device_count()
    else:
416
        torch.cuda.set_device(args.local_rank)
thomwolf's avatar
thomwolf committed
417
418
        device = torch.device("cuda", args.local_rank)
        n_gpu = 1
thomwolf's avatar
thomwolf committed
419
420
        # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
        torch.distributed.init_process_group(backend='nccl')
421
422
    logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
        device, n_gpu, bool(args.local_rank != -1), args.fp16))
thomwolf's avatar
thomwolf committed
423

424
425
426
    if args.gradient_accumulation_steps < 1:
        raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
                            args.gradient_accumulation_steps))
thomwolf's avatar
thomwolf committed
427

428
    args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
thomwolf's avatar
thomwolf committed
429

VictorSanh's avatar
VictorSanh committed
430
431
432
    random.seed(args.seed)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
thomwolf's avatar
thomwolf committed
433
434
    if n_gpu > 0:
        torch.cuda.manual_seed_all(args.seed)
thomwolf's avatar
thomwolf committed
435

VictorSanh's avatar
WIP  
VictorSanh committed
436
437
    if not args.do_train and not args.do_eval:
        raise ValueError("At least one of `do_train` or `do_eval` must be True.")
thomwolf's avatar
thomwolf committed
438

439
    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train:
thomwolf's avatar
thomwolf committed
440
        raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
thomwolf's avatar
thomwolf committed
441
442
    if not os.path.exists(args.output_dir):
        os.makedirs(args.output_dir)
VictorSanh's avatar
WIP  
VictorSanh committed
443
444

    task_name = args.task_name.lower()
thomwolf's avatar
thomwolf committed
445

VictorSanh's avatar
WIP  
VictorSanh committed
446
447
448
449
    if task_name not in processors:
        raise ValueError("Task not found: %s" % (task_name))

    processor = processors[task_name]()
450
    num_labels = num_labels_task[task_name]
VictorSanh's avatar
WIP  
VictorSanh committed
451
452
    label_list = processor.get_labels()

thomwolf's avatar
thomwolf committed
453
    tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
thomwolf's avatar
thomwolf committed
454

VictorSanh's avatar
WIP  
VictorSanh committed
455
    train_examples = None
456
    num_train_optimization_steps = None
VictorSanh's avatar
WIP  
VictorSanh committed
457
458
    if args.do_train:
        train_examples = processor.get_train_examples(args.data_dir)
459
460
461
462
        num_train_optimization_steps = int(
            len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs
        if args.local_rank != -1:
            num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
thomwolf's avatar
thomwolf committed
463

thomwolf's avatar
thomwolf committed
464
    # Prepare model
465
    cache_dir = args.cache_dir if args.cache_dir else os.path.join(PYTORCH_PRETRAINED_BERT_CACHE, 'distributed_{}'.format(args.local_rank))
466
    model = BertForSequenceClassification.from_pretrained(args.bert_model,
467
              cache_dir=cache_dir,
468
              num_labels = num_labels)
thomwolf's avatar
thomwolf committed
469
470
    if args.fp16:
        model.half()
thomwolf's avatar
thomwolf committed
471
    model.to(device)
thomwolf's avatar
thomwolf committed
472
    if args.local_rank != -1:
thomwolf's avatar
thomwolf committed
473
474
475
476
477
        try:
            from apex.parallel import DistributedDataParallel as DDP
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")

478
        model = DDP(model)
thomwolf's avatar
thomwolf committed
479
    elif n_gpu > 1:
480
        model = torch.nn.DataParallel(model)
thomwolf's avatar
thomwolf committed
481

thomwolf's avatar
thomwolf committed
482
    # Prepare optimizer
483
484
    param_optimizer = list(model.named_parameters())
    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
thomwolf's avatar
thomwolf committed
485
    optimizer_grouped_parameters = [
486
487
        {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
        {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
thomwolf's avatar
thomwolf committed
488
        ]
489
    if args.fp16:
thomwolf's avatar
thomwolf committed
490
491
492
493
494
495
        try:
            from apex.optimizers import FP16_Optimizer
            from apex.optimizers import FusedAdam
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")

496
497
498
499
500
501
502
503
504
505
506
507
508
        optimizer = FusedAdam(optimizer_grouped_parameters,
                              lr=args.learning_rate,
                              bias_correction=False,
                              max_grad_norm=1.0)
        if args.loss_scale == 0:
            optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
        else:
            optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)

    else:
        optimizer = BertAdam(optimizer_grouped_parameters,
                             lr=args.learning_rate,
                             warmup=args.warmup_proportion,
509
                             t_total=num_train_optimization_steps)
thomwolf's avatar
thomwolf committed
510

thomwolf's avatar
thomwolf committed
511
    global_step = 0
Jade Abbott's avatar
Jade Abbott committed
512
    nb_tr_steps = 0
513
    tr_loss = 0
VictorSanh's avatar
WIP  
VictorSanh committed
514
515
516
517
518
519
    if args.do_train:
        train_features = convert_examples_to_features(
            train_examples, label_list, args.max_seq_length, tokenizer)
        logger.info("***** Running training *****")
        logger.info("  Num examples = %d", len(train_examples))
        logger.info("  Batch size = %d", args.train_batch_size)
520
        logger.info("  Num steps = %d", num_train_optimization_steps)
521
522
523
524
        all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
        all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)
        all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)
        all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long)
525
526
527
528
529
530
531
532
        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
        if args.local_rank == -1:
            train_sampler = RandomSampler(train_data)
        else:
            train_sampler = DistributedSampler(train_data)
        train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)

        model.train()
533
        for _ in trange(int(args.num_train_epochs), desc="Epoch"):
534
535
            tr_loss = 0
            nb_tr_examples, nb_tr_steps = 0, 0
536
537
538
            for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
                batch = tuple(t.to(device) for t in batch)
                input_ids, input_mask, segment_ids, label_ids = batch
539
                loss = model(input_ids, segment_ids, input_mask, label_ids)
thomwolf's avatar
thomwolf committed
540
541
                if n_gpu > 1:
                    loss = loss.mean() # mean() to average on multi-gpu.
542
543
                if args.gradient_accumulation_steps > 1:
                    loss = loss / args.gradient_accumulation_steps
544
545
546
547
548
549

                if args.fp16:
                    optimizer.backward(loss)
                else:
                    loss.backward()

550
                tr_loss += loss.item()
551
                nb_tr_examples += input_ids.size(0)
552
                nb_tr_steps += 1
thomwolf's avatar
thomwolf committed
553
                if (step + 1) % args.gradient_accumulation_steps == 0:
554
555
556
                    if args.fp16:
                        # modify learning rate with special warm up BERT uses
                        # if args.fp16 is False, BertAdam is used that handles this automatically
557
                        lr_this_step = args.learning_rate * warmup_linear(global_step/num_train_optimization_steps, args.warmup_proportion)
558
559
                        for param_group in optimizer.param_groups:
                            param_group['lr'] = lr_this_step
560
561
                    optimizer.step()
                    optimizer.zero_grad()
thomwolf's avatar
thomwolf committed
562
                    global_step += 1
thomwolf's avatar
thomwolf committed
563

564
    if args.do_train:
565
566
567
        # Save a trained model and the associated configuration
        model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self
        output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
568
        torch.save(model_to_save.state_dict(), output_model_file)
569
570
571
572
573
574
575
576
577
578
        output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
        with open(output_config_file, 'w') as f:
            f.write(model_to_save.config.to_json_string())

        # Load a trained model and config that you have fine-tuned
        config = BertConfig(output_config_file)
        model = BertForSequenceClassification(config, num_labels=num_labels)
        model.load_state_dict(torch.load(output_model_file))
    else:
        model = BertForSequenceClassification.from_pretrained(args.bert_model, num_labels=num_labels)
579
    model.to(device)
580

581
    if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
VictorSanh's avatar
WIP  
VictorSanh committed
582
583
584
        eval_examples = processor.get_dev_examples(args.data_dir)
        eval_features = convert_examples_to_features(
            eval_examples, label_list, args.max_seq_length, tokenizer)
VictorSanh's avatar
wip  
VictorSanh committed
585
586
587
        logger.info("***** Running evaluation *****")
        logger.info("  Num examples = %d", len(eval_examples))
        logger.info("  Batch size = %d", args.eval_batch_size)
588
589
590
591
        all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
        all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
        all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
        all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)
592
        eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
593
594
        # Run prediction for full data
        eval_sampler = SequentialSampler(eval_data)
595
596
597
        eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)

        model.eval()
598
        eval_loss, eval_accuracy = 0, 0
VictorSanh's avatar
VictorSanh committed
599
        nb_eval_steps, nb_eval_examples = 0, 0
600
601
 
        for input_ids, input_mask, segment_ids, label_ids in tqdm(eval_dataloader, desc="Evaluating"):
602
            input_ids = input_ids.to(device)
thomwolf's avatar
thomwolf committed
603
            input_mask = input_mask.to(device)
604
            segment_ids = segment_ids.to(device)
605
            label_ids = label_ids.to(device)
606

607
            with torch.no_grad():
608
609
                tmp_eval_loss = model(input_ids, segment_ids, input_mask, label_ids)
                logits = model(input_ids, segment_ids, input_mask)
thomwolf's avatar
thomwolf committed
610
611
612

            logits = logits.detach().cpu().numpy()
            label_ids = label_ids.to('cpu').numpy()
613
614
            tmp_eval_accuracy = accuracy(logits, label_ids)

615
            eval_loss += tmp_eval_loss.mean().item()
616
            eval_accuracy += tmp_eval_accuracy
thomwolf's avatar
thomwolf committed
617

VictorSanh's avatar
VictorSanh committed
618
            nb_eval_examples += input_ids.size(0)
619
            nb_eval_steps += 1
VictorSanh's avatar
WIP  
VictorSanh committed
620

621
622
        eval_loss = eval_loss / nb_eval_steps
        eval_accuracy = eval_accuracy / nb_eval_examples
623
        loss = tr_loss/nb_tr_steps if args.do_train else None
624
625
626
        result = {'eval_loss': eval_loss,
                  'eval_accuracy': eval_accuracy,
                  'global_step': global_step,
627
                  'loss': loss}
VictorSanh's avatar
WIP  
VictorSanh committed
628
629

        output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
VictorSanh's avatar
wip  
VictorSanh committed
630
631
        with open(output_eval_file, "w") as writer:
            logger.info("***** Eval results *****")
VictorSanh's avatar
WIP  
VictorSanh committed
632
            for key in sorted(result.keys()):
VictorSanh's avatar
wip  
VictorSanh committed
633
                logger.info("  %s = %s", key, str(result[key]))
VictorSanh's avatar
WIP  
VictorSanh committed
634
                writer.write("%s = %s\n" % (key, str(result[key])))
635

VictorSanh's avatar
WIP  
VictorSanh committed
636
637
if __name__ == "__main__":
    main()