run_classifier.py 27.2 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
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
#
# 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."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

21
22
import csv
import os
23
24
import logging
import argparse
VictorSanh's avatar
VictorSanh committed
25
import random
thomwolf's avatar
thomwolf committed
26
from tqdm import tqdm, trange
thomwolf's avatar
thomwolf committed
27
28

import numpy as np
VictorSanh's avatar
VictorSanh committed
29
import torch
30
31
32
from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler
from torch.utils.data.distributed import DistributedSampler

33
from pytorch_pretrained_bert.tokenization import BertTokenizer
thomwolf's avatar
thomwolf committed
34
35
from pytorch_pretrained_bert.modeling import BertForSequenceClassification
from pytorch_pretrained_bert.optimization import BertAdam
36
from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE
37

38
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
39
40
41
                    datefmt = '%m/%d/%Y %H:%M:%S',
                    level = logging.INFO)
logger = logging.getLogger(__name__)
42

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

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
73

74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98

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."""
        with open(input_file, "r") as f:
            reader = csv.reader(f, delimiter="\t", quotechar=quotechar)
            lines = []
            for line in reader:
                lines.append(line)
            return lines
thomwolf's avatar
thomwolf committed
99
100


VictorSanh's avatar
wip  
VictorSanh committed
101
102
103
104
105
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
106
        logger.info("LOOKING AT {}".format(os.path.join(data_dir, "train.tsv")))
VictorSanh's avatar
wip  
VictorSanh committed
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
        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)
126
127
128
            text_a = line[3]
            text_b = line[4]
            label = line[0]
VictorSanh's avatar
wip  
VictorSanh committed
129
130
131
132
            examples.append(
                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
        return examples

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

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
158
            guid = "%s-%s" % (set_type, line[0])
159
160
            text_a = line[8]
            text_b = line[9]
161
            label = line[-1]
162
163
164
            examples.append(
                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
        return examples
thomwolf's avatar
thomwolf committed
165

166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188

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)
189
190
            text_a = line[3]
            label = line[1]
191
192
193
            examples.append(
                InputExample(guid=guid, text_a=text_a, text_b=None, label=label))
        return examples
thomwolf's avatar
thomwolf committed
194
195
196


def convert_examples_to_features(examples, label_list, max_seq_length, tokenizer):
197
198
    """Loads a data file into a list of `InputBatch`s."""

199
    label_map = {label : i for i, label in enumerate(label_list)}
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214

    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:
215
                tokens_a = tokens_a[:(max_seq_length - 2)]
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234

        # 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.
235
236
        tokens = ["[CLS]"] + tokens_a + ["[SEP]"]
        segment_ids = [0] * len(tokens)
237
238

        if tokens_b:
239
240
            tokens += tokens_b + ["[SEP]"]
            segment_ids += [1] * (len(tokens_b) + 1)
241
242
243
244
245
246
247
248

        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.
249
250
251
252
        padding = [0] * (max_seq_length - len(input_ids))
        input_ids += padding
        input_mask += padding
        segment_ids += padding
253
254
255
256
257
258
259
260
261
262

        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(
263
                    [str(x) for x in tokens]))
264
265
266
267
268
269
270
            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
271
272
273
274
                InputFeatures(input_ids=input_ids,
                              input_mask=input_mask,
                              segment_ids=segment_ids,
                              label_id=label_id))
275
    return features
thomwolf's avatar
thomwolf committed
276
277


278
279
280
281
282
283
284
285
286
287
288
289
290
291
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
292
293
            tokens_b.pop()

294
295
def accuracy(out, labels):
    outputs = np.argmax(out, axis=1)
thomwolf's avatar
thomwolf committed
296
    return np.sum(outputs == labels)
VictorSanh's avatar
WIP  
VictorSanh committed
297

thomwolf's avatar
thomwolf committed
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
def copy_optimizer_params_to_model(named_params_model, named_params_optimizer):
    """ Utility function for optimize_on_cpu and 16-bits training.
        Copy the parameters optimized on CPU/RAM back to the model on GPU
    """
    for (name_opti, param_opti), (name_model, param_model) in zip(named_params_optimizer, named_params_model):
        if name_opti != name_model:
            logger.error("name_opti != name_model: {} {}".format(name_opti, name_model))
            raise ValueError
        param_model.data.copy_(param_opti.data)

def set_optimizer_params_grad(named_params_optimizer, named_params_model, test_nan=False):
    """ Utility function for optimize_on_cpu and 16-bits training.
        Copy the gradient of the GPU parameters to the CPU/RAMM copy of the model
    """
    is_nan = False
    for (name_opti, param_opti), (name_model, param_model) in zip(named_params_optimizer, named_params_model):
        if name_opti != name_model:
            logger.error("name_opti != name_model: {} {}".format(name_opti, name_model))
            raise ValueError
thomwolf's avatar
thomwolf committed
317
318
319
320
321
322
323
324
        if param_model.grad is not None:
            if test_nan and torch.isnan(param_model.grad).sum() > 0:
                is_nan = True
            if param_opti.grad is None:
                param_opti.grad = torch.nn.Parameter(param_opti.data.new().resize_(*param_opti.data.size()))
            param_opti.grad.data.copy_(param_model.grad.data)
        else:
            param_opti.grad = None
thomwolf's avatar
thomwolf committed
325
326
    return is_nan

327
def main():
328
329
330
331
332
333
334
335
    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
336
337
338
    parser.add_argument("--bert_model", default=None, type=str, required=True,
                        help="Bert pre-trained model selected in the list: bert-base-uncased, "
                             "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.")
339
340
341
342
343
344
345
346
347
    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,
348
                        help="The output directory where the model predictions and checkpoints will be written.")
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364

    ## Other parameters
    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",
                        default=False,
                        action='store_true',
                        help="Whether to run training.")
    parser.add_argument("--do_eval",
                        default=False,
                        action='store_true',
                        help="Whether to run eval on the dev set.")
thomwolf's avatar
thomwolf committed
365
366
367
368
    parser.add_argument("--do_lower_case",
                        default=False,
                        action='store_true',
                        help="Set this flag if you are using an uncased model.")
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
    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",
                        default=False,
                        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")
398
399
    parser.add_argument('--seed',
                        type=int,
VictorSanh's avatar
VictorSanh committed
400
401
                        default=42,
                        help="random seed for initialization")
402
403
404
    parser.add_argument('--gradient_accumulation_steps',
                        type=int,
                        default=1,
405
                        help="Number of updates steps to accumulate before performing a backward/update pass.")
thomwolf's avatar
thomwolf committed
406
407
408
409
410
411
412
413
414
415
416
417
    parser.add_argument('--optimize_on_cpu',
                        default=False,
                        action='store_true',
                        help="Whether to perform optimization and keep the optimizer averages on CPU")
    parser.add_argument('--fp16',
                        default=False,
                        action='store_true',
                        help="Whether to use 16-bit float precision instead of 32-bit")
    parser.add_argument('--loss_scale',
                        type=float, default=128,
                        help='Loss scaling, positive power of 2 values can improve fp16 convergence.')

418
419
    args = parser.parse_args()

VictorSanh's avatar
WIP  
VictorSanh committed
420
421
422
423
424
    processors = {
        "cola": ColaProcessor,
        "mnli": MnliProcessor,
        "mrpc": MrpcProcessor,
    }
thomwolf's avatar
thomwolf committed
425

426
427
428
429
430
431
    num_labels_task = {
        "cola": 2,
        "mnli": 3,
        "mrpc": 2,
    }

thomwolf's avatar
thomwolf committed
432
433
434
435
436
437
    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:
        device = torch.device("cuda", args.local_rank)
        n_gpu = 1
thomwolf's avatar
thomwolf committed
438
439
        # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
        torch.distributed.init_process_group(backend='nccl')
thomwolf's avatar
thomwolf committed
440
441
442
        if args.fp16:
            logger.info("16-bits training currently not supported in distributed training")
            args.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
443
    logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1))
thomwolf's avatar
thomwolf committed
444

445
446
447
    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
448

449
    args.train_batch_size = int(args.train_batch_size / args.gradient_accumulation_steps)
thomwolf's avatar
thomwolf committed
450

VictorSanh's avatar
VictorSanh committed
451
452
453
    random.seed(args.seed)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
thomwolf's avatar
thomwolf committed
454
455
    if n_gpu > 0:
        torch.cuda.manual_seed_all(args.seed)
thomwolf's avatar
thomwolf committed
456

VictorSanh's avatar
WIP  
VictorSanh committed
457
458
    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
459

VictorSanh's avatar
WIP  
VictorSanh committed
460
    if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
thomwolf's avatar
thomwolf committed
461
        raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
VictorSanh's avatar
WIP  
VictorSanh committed
462
463
464
    os.makedirs(args.output_dir, exist_ok=True)

    task_name = args.task_name.lower()
thomwolf's avatar
thomwolf committed
465

VictorSanh's avatar
WIP  
VictorSanh committed
466
467
468
469
    if task_name not in processors:
        raise ValueError("Task not found: %s" % (task_name))

    processor = processors[task_name]()
470
    num_labels = num_labels_task[task_name]
VictorSanh's avatar
WIP  
VictorSanh committed
471
472
    label_list = processor.get_labels()

thomwolf's avatar
thomwolf committed
473
    tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
thomwolf's avatar
thomwolf committed
474

VictorSanh's avatar
WIP  
VictorSanh committed
475
476
477
478
479
    train_examples = None
    num_train_steps = None
    if args.do_train:
        train_examples = processor.get_train_examples(args.data_dir)
        num_train_steps = int(
480
            len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs)
thomwolf's avatar
thomwolf committed
481

thomwolf's avatar
thomwolf committed
482
    # Prepare model
483
    model = BertForSequenceClassification.from_pretrained(args.bert_model,
484
485
              cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(args.local_rank),
              num_labels = num_labels)
thomwolf's avatar
thomwolf committed
486
487
    if args.fp16:
        model.half()
thomwolf's avatar
thomwolf committed
488
    model.to(device)
thomwolf's avatar
thomwolf committed
489
490
491
492
    if args.local_rank != -1:
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
                                                          output_device=args.local_rank)
    elif n_gpu > 1:
493
        model = torch.nn.DataParallel(model)
thomwolf's avatar
thomwolf committed
494

thomwolf's avatar
thomwolf committed
495
496
497
498
499
500
501
502
503
    # Prepare optimizer
    if args.fp16:
        param_optimizer = [(n, param.clone().detach().to('cpu').float().requires_grad_()) \
                            for n, param in model.named_parameters()]
    elif args.optimize_on_cpu:
        param_optimizer = [(n, param.clone().detach().to('cpu').requires_grad_()) \
                            for n, param in model.named_parameters()]
    else:
        param_optimizer = list(model.named_parameters())
thomwolf's avatar
thomwolf committed
504
    no_decay = ['bias', 'gamma', 'beta']
thomwolf's avatar
thomwolf committed
505
    optimizer_grouped_parameters = [
xiaoda99's avatar
xiaoda99 committed
506
507
        {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.01},
        {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.0}
thomwolf's avatar
thomwolf committed
508
        ]
509
510
511
    t_total = num_train_steps
    if args.local_rank != -1:
        t_total = t_total // torch.distributed.get_world_size()
thomwolf's avatar
thomwolf committed
512
    optimizer = BertAdam(optimizer_grouped_parameters,
thomwolf's avatar
thomwolf committed
513
                         lr=args.learning_rate,
thomwolf's avatar
thomwolf committed
514
                         warmup=args.warmup_proportion,
515
                         t_total=t_total)
thomwolf's avatar
thomwolf committed
516

thomwolf's avatar
thomwolf committed
517
    global_step = 0
VictorSanh's avatar
WIP  
VictorSanh committed
518
519
520
521
522
523
524
    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)
        logger.info("  Num steps = %d", num_train_steps)
525
526
527
528
        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)
529
530
531
532
533
534
535
536
        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()
537
        for _ in trange(int(args.num_train_epochs), desc="Epoch"):
538
539
            tr_loss = 0
            nb_tr_examples, nb_tr_steps = 0, 0
540
541
542
            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
543
                loss = model(input_ids, segment_ids, input_mask, label_ids)
thomwolf's avatar
thomwolf committed
544
545
                if n_gpu > 1:
                    loss = loss.mean() # mean() to average on multi-gpu.
thomwolf's avatar
thomwolf committed
546
547
548
549
                if args.fp16 and args.loss_scale != 1.0:
                    # rescale loss for fp16 training
                    # see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
                    loss = loss * args.loss_scale
550
551
552
                if args.gradient_accumulation_steps > 1:
                    loss = loss / args.gradient_accumulation_steps
                loss.backward()
553
                tr_loss += loss.item()
554
                nb_tr_examples += input_ids.size(0)
555
                nb_tr_steps += 1
thomwolf's avatar
thomwolf committed
556
                if (step + 1) % args.gradient_accumulation_steps == 0:
thomwolf's avatar
thomwolf committed
557
558
559
560
                    if args.fp16 or args.optimize_on_cpu:
                        if args.fp16 and args.loss_scale != 1.0:
                            # scale down gradients for fp16 training
                            for param in model.parameters():
561
562
                                if param.grad is not None:
                                    param.grad.data = param.grad.data / args.loss_scale
thomwolf's avatar
thomwolf committed
563
564
565
566
567
568
569
570
571
572
                        is_nan = set_optimizer_params_grad(param_optimizer, model.named_parameters(), test_nan=True)
                        if is_nan:
                            logger.info("FP16 TRAINING: Nan in gradients, reducing loss scaling")
                            args.loss_scale = args.loss_scale / 2
                            model.zero_grad()
                            continue
                        optimizer.step()
                        copy_optimizer_params_to_model(model.named_parameters(), param_optimizer)
                    else:
                        optimizer.step()
thomwolf's avatar
thomwolf committed
573
574
                    model.zero_grad()
                    global_step += 1
thomwolf's avatar
thomwolf committed
575

576
577
578
579
580
581
582
583
584
    # Save a trained model
    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, "pytorch_model.bin")
    torch.save(model_to_save.state_dict(), output_model_file)

    # Load a trained model that you have fine-tuned
    model_state_dict = torch.load(output_model_file)
    model = BertForSequenceClassification.from_pretrained(args.bert_model, state_dict=model_state_dict)

585
    if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
VictorSanh's avatar
WIP  
VictorSanh committed
586
587
588
        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
589
590
591
        logger.info("***** Running evaluation *****")
        logger.info("  Num examples = %d", len(eval_examples))
        logger.info("  Batch size = %d", args.eval_batch_size)
592
593
594
595
        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)
596
        eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
597
598
        # Run prediction for full data
        eval_sampler = SequentialSampler(eval_data)
599
600
601
        eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)

        model.eval()
602
        eval_loss, eval_accuracy = 0, 0
VictorSanh's avatar
VictorSanh committed
603
        nb_eval_steps, nb_eval_examples = 0, 0
604
        for input_ids, input_mask, segment_ids, label_ids in eval_dataloader:
605
            input_ids = input_ids.to(device)
thomwolf's avatar
thomwolf committed
606
            input_mask = input_mask.to(device)
607
            segment_ids = segment_ids.to(device)
608
            label_ids = label_ids.to(device)
609

610
            with torch.no_grad():
611
612
                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
613
614
615

            logits = logits.detach().cpu().numpy()
            label_ids = label_ids.to('cpu').numpy()
616
617
            tmp_eval_accuracy = accuracy(logits, label_ids)

618
            eval_loss += tmp_eval_loss.mean().item()
619
            eval_accuracy += tmp_eval_accuracy
thomwolf's avatar
thomwolf committed
620

VictorSanh's avatar
VictorSanh committed
621
            nb_eval_examples += input_ids.size(0)
622
            nb_eval_steps += 1
VictorSanh's avatar
WIP  
VictorSanh committed
623

624
625
        eval_loss = eval_loss / nb_eval_steps
        eval_accuracy = eval_accuracy / nb_eval_examples
VictorSanh's avatar
WIP  
VictorSanh committed
626

627
628
629
        result = {'eval_loss': eval_loss,
                  'eval_accuracy': eval_accuracy,
                  'global_step': global_step,
630
                  'loss': tr_loss/nb_tr_steps}
VictorSanh's avatar
WIP  
VictorSanh committed
631
632

        output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
VictorSanh's avatar
wip  
VictorSanh committed
633
634
        with open(output_eval_file, "w") as writer:
            logger.info("***** Eval results *****")
VictorSanh's avatar
WIP  
VictorSanh committed
635
            for key in sorted(result.keys()):
VictorSanh's avatar
wip  
VictorSanh committed
636
                logger.info("  %s = %s", key, str(result[key]))
VictorSanh's avatar
WIP  
VictorSanh committed
637
                writer.write("%s = %s\n" % (key, str(result[key])))
638

VictorSanh's avatar
WIP  
VictorSanh committed
639
640
if __name__ == "__main__":
    main()