run_glue.py 20.3 KB
Newer Older
thomwolf's avatar
thomwolf committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
# 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.
"""BERT finetuning runner."""

from __future__ import absolute_import, division, print_function

import argparse
import logging
import os
import random
from tqdm import tqdm, trange

import numpy as np

import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)
from torch.utils.data.distributed import DistributedSampler

from tensorboardX import SummaryWriter

35
36
37
38
39
from pytorch_transformers import (BertForSequenceClassification, XLNetForSequenceClassification,
                                  XLMForSequenceClassification, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
                                  XLNET_PRETRAINED_MODEL_ARCHIVE_MAP, XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
from pytorch_transformers import (BertTokenizer, XLNetTokenizer,
                                  XLMTokenizer)
thomwolf's avatar
thomwolf committed
40
41
42
43
44
45
46
from pytorch_transformers.optimization import BertAdam, WarmupLinearSchedule

from utils_glue import processors, output_modes, convert_examples_to_features, compute_metrics


logger = logging.getLogger(__name__)

47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
ALL_MODELS = sum((tuple(m.keys()) for m in (BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
                                            XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
                                            XLM_PRETRAINED_MODEL_ARCHIVE_MAP)), ())

MODEL_CLASSES = {
    'bert': BertForSequenceClassification,
    'xlnet': XLNetForSequenceClassification,
    'xlm': XLMForSequenceClassification,
}

TOKENIZER_CLASSES = {
    'bert': BertTokenizer,
    'xlnet': XLNetTokenizer,
    'xlm': XLMTokenizer,
}
thomwolf's avatar
thomwolf committed
62

63
def train(args, train_dataset, model):
thomwolf's avatar
thomwolf committed
64
65
66
67
68
    """ Train the model """
    if args.local_rank in [-1, 0]:
        tb_writer = SummaryWriter()

    args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
69
70
    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)
thomwolf's avatar
thomwolf committed
71

thomwolf's avatar
thomwolf committed
72
73
74
75
76
    if args.max_steps > 0:
        num_train_optimization_steps = args.max_steps
        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
    else:
        num_train_optimization_steps = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
thomwolf's avatar
thomwolf committed
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97

    # Prepare optimizer
    param_optimizer = list(model.named_parameters())
    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
    optimizer_grouped_parameters = [
        {'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}
        ]
    if args.fp16:
        try:
            from apex.optimizers import FP16_Optimizer, FusedAdam
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
        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)
        warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion, t_total=num_train_optimization_steps)

    else:
thomwolf's avatar
thomwolf committed
98
99
        optimizer = BertAdam(optimizer_grouped_parameters, lr=args.learning_rate, warmup=args.warmup_proportion,
                             t_total=num_train_optimization_steps)
thomwolf's avatar
thomwolf committed
100
101
102

    # Train!
    logger.info("***** Running training *****")
103
104
    logger.info("  Num examples = %d", len(train_dataset))
    logger.info("  Num Epochs = %d", args.num_train_epochs)
thomwolf's avatar
thomwolf committed
105
    logger.info("  Batch size = %d", args.train_batch_size)
106
107
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", num_train_optimization_steps)
thomwolf's avatar
thomwolf committed
108
109
110
111

    global_step = 0
    tr_loss = 0
    model.train()
112
    optimizer.zero_grad()
thomwolf's avatar
thomwolf committed
113
114
115
    for _ in trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]):
        for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])):
            batch = tuple(t.to(args.device) for t in batch)
116
117
            inputs = {'input_ids':      batch[0],
                      'attention_mask': batch[1],
thomwolf's avatar
thomwolf committed
118
                      'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None,  # XLM don't use segment_ids
119
120
                      'labels':         batch[3]}
            ouputs = model(**inputs)
thomwolf's avatar
thomwolf committed
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
            loss = ouputs[0]

            if args.n_gpu > 1:
                loss = loss.mean() # mean() to average on multi-gpu parallel training
            if args.gradient_accumulation_steps > 1:
                loss = loss / args.gradient_accumulation_steps

            loss.backward() if not args.fp16 else optimizer.backward(loss)

            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0:
                if args.fp16:
                    # modify learning rate with special warm up BERT uses
                    # if args.fp16 is False, BertAdam is used that handles this automatically
                    lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step, args.warmup_proportion)
                    for param_group in optimizer.param_groups:
                        param_group['lr'] = lr_this_step
                optimizer.step()
                optimizer.zero_grad()
                global_step += 1
                if args.local_rank in [-1, 0]:
                    if not args.fp16:
                        tb_writer.add_scalar('lr', optimizer.get_lr()[0], global_step)
                    tb_writer.add_scalar('loss', loss.item(), global_step)
thomwolf's avatar
thomwolf committed
145
146
147
148
            if args.max_steps > 0 and global_step > args.max_steps:
                break
        if args.max_steps > 0 and global_step > args.max_steps:
            break
thomwolf's avatar
thomwolf committed
149
150
151
152

    return global_step, tr_loss / global_step


153
def evalutate(args, eval_task, eval_output_dir, dataset, model):
thomwolf's avatar
thomwolf committed
154
155
156
157
158
    """ Evaluate the model """
    if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
        os.makedirs(eval_output_dir)

    # Note that DistributedSampler samples randomly
159
160
    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)
thomwolf's avatar
thomwolf committed
161
162
163

    # Eval!
    logger.info("***** Running evaluation *****")
164
    logger.info("  Num examples = %d", len(dataset))
thomwolf's avatar
thomwolf committed
165
166
167
168
169
170
171
172
173
174
    logger.info("  Batch size = %d", args.eval_batch_size)
    model.eval()
    eval_loss = 0
    nb_eval_steps = 0
    preds = None
    out_label_ids = None
    for batch in tqdm(eval_dataloader, desc="Evaluating"):
        batch = tuple(t.to(args.device) for t in batch)

        with torch.no_grad():
thomwolf's avatar
thomwolf committed
175
176
177
178
179
            inputs = {'input_ids':      batch[0],
                      'attention_mask': batch[1],
                      'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None,  # XLM don't use segment_ids
                      'labels':         batch[3]}
            outputs = model(**inputs)
thomwolf's avatar
thomwolf committed
180
181
182
183
184
185
            tmp_eval_loss, logits = outputs[:2]

        eval_loss += tmp_eval_loss.mean().item()
        nb_eval_steps += 1
        if preds is None:
            preds = logits.detach().cpu().numpy()
thomwolf's avatar
thomwolf committed
186
            out_label_ids = inputs['labels'].detach().cpu().numpy()
thomwolf's avatar
thomwolf committed
187
188
        else:
            preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
thomwolf's avatar
thomwolf committed
189
            out_label_ids = np.append(out_label_ids, inputs['labels'].detach().cpu().numpy(), axis=0)
thomwolf's avatar
thomwolf committed
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204

    eval_loss = eval_loss / nb_eval_steps
    if args.output_mode == "classification":
        preds = np.argmax(preds, axis=1)
    elif args.output_mode == "regression":
        preds = np.squeeze(preds)
    result = compute_metrics(eval_task, preds, out_label_ids)

    output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
    with open(output_eval_file, "w") as writer:
        logger.info("***** Eval results *****")
        for key in sorted(result.keys()):
            logger.info("  %s = %s", key, str(result[key]))
            writer.write("%s = %s\n" % (key, str(result[key])))

205
    return result
thomwolf's avatar
thomwolf committed
206
207


208
def load_and_cache_examples(args, task, tokenizer, evaluate=False):
thomwolf's avatar
thomwolf committed
209
    processor = processors[task]()
210
211
212
213
    output_mode = output_modes[task]
    # Load data features from cache or dataset file
    cached_features_file = os.path.join(args.data_dir, 'cached_{}_{}_{}_{}'.format(
        'dev' if evaluate else 'train',
214
        list(filter(None, args.model_name.split('/'))).pop(),
thomwolf's avatar
thomwolf committed
215
216
217
        str(args.max_seq_length),
        str(task)))
    if os.path.exists(cached_features_file):
thomwolf's avatar
thomwolf committed
218
        logger.info("Loading features from cached file %s", cached_features_file)
thomwolf's avatar
thomwolf committed
219
220
        features = torch.load(cached_features_file)
    else:
221
222
223
        logger.info("Creating features from dataset file at %s", args.data_dir)
        label_list = processor.get_labels()
        examples = processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
224
        features = convert_examples_to_features(examples, label_list, args.max_seq_length, tokenizer, output_mode,
225
            cls_token_at_end=bool(args.model_type in ['xlnet']),            # xlnet has a cls token at the end
226
            cls_token=tokenizer.cls_token,
227
228
229
230
231
            sep_token=tokenizer.sep_token,
            cls_token_segment_id=2 if args.model_type in ['xlnet'] else 1,
            pad_on_left=bool(args.model_type in ['xlnet']),                 # pad on the left for xlnet
            pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0)
        if args.local_rank in [-1, 0]:
thomwolf's avatar
thomwolf committed
232
            logger.info("Saving features into cached file %s", cached_features_file)
thomwolf's avatar
thomwolf committed
233
234
            torch.save(features, cached_features_file)

235
236
237
238
239
240
241
242
243
244
245
    # Convert to Tensors and build dataset
    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 output_mode == "classification":
        all_label_ids = torch.tensor([f.label_id for f in features], dtype=torch.long)
    elif output_mode == "regression":
        all_label_ids = torch.tensor([f.label_id for f in features], dtype=torch.float)

    dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
    return dataset
thomwolf's avatar
thomwolf committed
246
247


thomwolf's avatar
thomwolf committed
248
249
250
251
252
253
def main():
    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.")
254
255
    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))
thomwolf's avatar
thomwolf committed
256
    parser.add_argument("--task_name", default=None, type=str, required=True,
257
                        help="The name of the task to train selected in the list: " + ", ".join(processors.keys()))
thomwolf's avatar
thomwolf committed
258
259
260
261
262
263
264
    parser.add_argument("--output_dir", default=None, type=str, required=True,
                        help="The output directory where the model predictions and checkpoints will be written.")

    ## Other parameters
    parser.add_argument("--cache_dir", default="", type=str,
                        help="Where do you want to store the pre-trained models downloaded from s3")
    parser.add_argument("--max_seq_length", default=128, type=int,
265
266
                        help="The maximum total input sequence length after tokenization. Sequences longer "
                             "than this will be truncated, sequences shorter will be padded.")
thomwolf's avatar
thomwolf committed
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
    parser.add_argument("--do_train", action='store_true',
                        help="Whether to run training.")
    parser.add_argument("--do_eval", action='store_true',
                        help="Whether to run eval on the dev set.")
    parser.add_argument("--do_lower_case", action='store_true',
                        help="Set this flag if you are using an uncased model.")
    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('--gradient_accumulation_steps', type=int, default=1,
                        help="Number of updates steps to accumulate before performing a backward/update pass.")
    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.")
thomwolf's avatar
thomwolf committed
283
284
    parser.add_argument("--max_steps", default=-1, type=int,
                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
thomwolf's avatar
thomwolf committed
285
    parser.add_argument("--warmup_proportion", default=0.1, type=float,
286
                        help="Proportion of training with linear learning rate warmup (0.1 = 10%% of training).")
thomwolf's avatar
thomwolf committed
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
    parser.add_argument("--no_cuda", action='store_true',
                        help="Avoid using CUDA when available")
    parser.add_argument('--overwrite_output_dir', action='store_true',
                        help="Overwrite the content of the output directory")
    parser.add_argument('--seed', type=int, default=42,
                        help="random seed for initialization")

    parser.add_argument('--fp16', action='store_true',
                        help="Whether to use 16-bit float precision instead of 32-bit")
    parser.add_argument('--loss_scale', 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")

    parser.add_argument("--local_rank", type=int, default=-1,
                        help="local_rank for distributed training on gpus")

    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
308
309
310
    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))

thomwolf's avatar
thomwolf committed
311
312
313
314
315
316
317
318
319
320
321
    # Setup distant debugging if needed
    if args.server_ip and args.server_port:
        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
        import ptvsd
        print("Waiting for debugger attach")
        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
        ptvsd.wait_for_attach()

    # Setup CUDA, GPU & distributed training
    if args.local_rank == -1 or args.no_cuda:
        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
thomwolf's avatar
thomwolf committed
322
        args.n_gpu = torch.cuda.device_count()
thomwolf's avatar
thomwolf committed
323
324
325
326
    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
        torch.cuda.set_device(args.local_rank)
        device = torch.device("cuda", args.local_rank)
        torch.distributed.init_process_group(backend='nccl')
thomwolf's avatar
thomwolf committed
327
        args.n_gpu = 1
thomwolf's avatar
thomwolf committed
328
329
330
331
    args.device = device

    # Setup logging
    logging.basicConfig(level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
thomwolf's avatar
thomwolf committed
332
333
    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
                args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
thomwolf's avatar
thomwolf committed
334
335
336
337
338

    # Setup seeds
    random.seed(args.seed)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
thomwolf's avatar
thomwolf committed
339
    if args.n_gpu > 0:
thomwolf's avatar
thomwolf committed
340
341
342
        torch.cuda.manual_seed_all(args.seed)

    # Prepare GLUE task
thomwolf's avatar
thomwolf committed
343
344
345
346
347
    args.task_name = args.task_name.lower()
    if args.task_name not in processors:
        raise ValueError("Task not found: %s" % (args.task_name))
    processor = processors[args.task_name]()
    args.output_mode = output_modes[args.task_name]
thomwolf's avatar
thomwolf committed
348
349
350
351
352
353
354
355
    label_list = processor.get_labels()
    num_labels = len(label_list)

    # Load pretrained model and tokenizer
    if args.local_rank not in [-1, 0]:
        # Make sure only the first process in distributed training will download model & vocab
        torch.distributed.barrier()

356
    args.model_type = args.model_name.lower().split('-')[0]
357
358
359
360
    tokenizer_class = TOKENIZER_CLASSES[args.model_type]
    model_class = MODEL_CLASSES[args.model_type]
    tokenizer = tokenizer_class.from_pretrained(args.model_name, do_lower_case=args.do_lower_case)
    model = model_class.from_pretrained(args.model_name, num_labels=num_labels)
thomwolf's avatar
thomwolf committed
361
362
363
364
365
366
367

    if args.local_rank == 0:
        torch.distributed.barrier()

    # Distributed, parrallel and fp16 model
    if args.fp16:
        model.half()
thomwolf's avatar
thomwolf committed
368
    model.to(args.device)
thomwolf's avatar
thomwolf committed
369
370
371
372
373
    if args.local_rank != -1:
        model = torch.nn.parallel.DistributedDataParallel(model,
                                                          device_ids=[args.local_rank],
                                                          output_device=args.local_rank,
                                                          find_unused_parameters=True)
thomwolf's avatar
thomwolf committed
374
    elif args.n_gpu > 1:
thomwolf's avatar
thomwolf committed
375
376
        model = torch.nn.DataParallel(model)

thomwolf's avatar
thomwolf committed
377
    # Training
thomwolf's avatar
thomwolf committed
378
    if args.do_train:
379
380
        train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
        global_step, tr_loss = train(args, train_dataset, model)
thomwolf's avatar
thomwolf committed
381
        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
thomwolf's avatar
thomwolf committed
382
383


thomwolf's avatar
thomwolf committed
384
    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
thomwolf's avatar
thomwolf committed
385
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
386
387
388
389
390
391
        # 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)

        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
        # They can then be reloaded using `from_pretrained()`
thomwolf's avatar
thomwolf committed
392
393
        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)
394
        tokenizer.save_pretrained(args.output_dir)
thomwolf's avatar
thomwolf committed
395
396

        # Good practice: save your training arguments together with the trained model
397
        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
thomwolf's avatar
thomwolf committed
398

399
        # Load a trained model and vocabulary that you have fine-tuned
400
401
        model = model_class.from_pretrained(args.output_dir)
        tokenizer = tokenizer_class.from_pretrained(args.output_dir)
402
        model.to(args.device)
thomwolf's avatar
thomwolf committed
403

thomwolf's avatar
thomwolf committed
404
    # Evaluation
thomwolf's avatar
thomwolf committed
405
    if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
thomwolf's avatar
thomwolf committed
406
407
408
409
410
        # Handle MNLI double evaluation
        eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
        eval_outputs_dirs = (args.output_dir, args.output_dir + '-MM') if args.task_name == "mnli" else (args.output_dir,)

        for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
411
412
413
            eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True)

            result = evalutate(args, eval_task, eval_output_dir, eval_dataset, model)
thomwolf's avatar
thomwolf committed
414

415
        return result
thomwolf's avatar
thomwolf committed
416
417
418
419


if __name__ == "__main__":
    main()