run_glue.py 29.8 KB
Newer Older
thomwolf's avatar
thomwolf committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
16
""" Finetuning the library models for sequence classification on GLUE (Bert, XLM, XLNet, RoBERTa)."""
thomwolf's avatar
thomwolf committed
17
18
19
20

from __future__ import absolute_import, division, print_function

import argparse
thomwolf's avatar
thomwolf committed
21
import glob
thomwolf's avatar
thomwolf committed
22
23
24
import logging
import os
import random
25
import json
thomwolf's avatar
thomwolf committed
26
27
28
29
30
31

import numpy as np
import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)
from torch.utils.data.distributed import DistributedSampler
32
33
34
35
36
37

try:
    from torch.utils.tensorboard import SummaryWriter
except:
    from tensorboardX import SummaryWriter

thomwolf's avatar
thomwolf committed
38
from tqdm import tqdm, trange
thomwolf's avatar
thomwolf committed
39

40
from transformers import (WEIGHTS_NAME, BertConfig,
thomwolf's avatar
thomwolf committed
41
                                  BertForSequenceClassification, BertTokenizer,
42
43
44
                                  RobertaConfig,
                                  RobertaForSequenceClassification,
                                  RobertaTokenizer,
thomwolf's avatar
thomwolf committed
45
46
47
                                  XLMConfig, XLMForSequenceClassification,
                                  XLMTokenizer, XLNetConfig,
                                  XLNetForSequenceClassification,
48
49
50
                                  XLNetTokenizer,
                                  DistilBertConfig,
                                  DistilBertForSequenceClassification,
Lysandre's avatar
Lysandre committed
51
52
53
54
55
                                  DistilBertTokenizer,
                                  AlbertConfig,
                                  AlbertForSequenceClassification, 
                                  AlbertTokenizer,
                                )
thomwolf's avatar
thomwolf committed
56

57
from transformers import AdamW, get_linear_schedule_with_warmup
thomwolf's avatar
thomwolf committed
58

59
60
61
62
from transformers import glue_compute_metrics as compute_metrics
from transformers import glue_output_modes as output_modes
from transformers import glue_processors as processors
from transformers import glue_convert_examples_to_features as convert_examples_to_features
thomwolf's avatar
thomwolf committed
63
64
65

logger = logging.getLogger(__name__)

Brian Ma's avatar
Brian Ma committed
66
67
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig, 
                                                                                RobertaConfig, DistilBertConfig)), ())
68
69

MODEL_CLASSES = {
thomwolf's avatar
thomwolf committed
70
71
72
    'bert': (BertConfig, BertForSequenceClassification, BertTokenizer),
    'xlnet': (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer),
    'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
73
    'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
Lysandre's avatar
Lysandre committed
74
75
    'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer),
    'albert': (AlbertConfig, AlbertForSequenceClassification, AlbertTokenizer)
76
}
thomwolf's avatar
thomwolf committed
77

thomwolf's avatar
thomwolf committed
78
79
80
81
82
83
84
85
86

def set_seed(args):
    random.seed(args.seed)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    if args.n_gpu > 0:
        torch.cuda.manual_seed_all(args.seed)


thomwolf's avatar
thomwolf committed
87
def train(args, train_dataset, model, tokenizer):
thomwolf's avatar
thomwolf committed
88
89
90
91
    """ Train the model """
    if args.local_rank in [-1, 0]:
        tb_writer = SummaryWriter()

thomwolf's avatar
thomwolf committed
92
    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
93
94
    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
95

thomwolf's avatar
thomwolf committed
96
    if args.max_steps > 0:
thomwolf's avatar
thomwolf committed
97
        t_total = args.max_steps
thomwolf's avatar
thomwolf committed
98
99
        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
    else:
thomwolf's avatar
thomwolf committed
100
        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
thomwolf's avatar
thomwolf committed
101

thomwolf's avatar
thomwolf committed
102
    # Prepare optimizer and schedule (linear warmup and decay)
thomwolf's avatar
thomwolf committed
103
    no_decay = ['bias', 'LayerNorm.weight']
thomwolf's avatar
thomwolf committed
104
    optimizer_grouped_parameters = [
thomwolf's avatar
thomwolf committed
105
106
        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
thomwolf's avatar
thomwolf committed
107
        ]
Lysandre's avatar
Lysandre committed
108

thomwolf's avatar
thomwolf committed
109
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
110
    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
111
112
113
114
115
116
117

    # Check if saved optimizer or scheduler states exist
    if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
        # Load in optimizer and scheduler states
        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))

thomwolf's avatar
thomwolf committed
118
119
    if args.fp16:
        try:
thomwolf's avatar
thomwolf committed
120
            from apex import amp
thomwolf's avatar
thomwolf committed
121
122
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
thomwolf's avatar
thomwolf committed
123
        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
thomwolf's avatar
thomwolf committed
124

125
126
127
128
    # multi-gpu training (should be after apex fp16 initialization)
    if args.n_gpu > 1:
        model = torch.nn.DataParallel(model)

thomwolf's avatar
thomwolf committed
129
130
131
132
133
134
    # Distributed training (should be after apex fp16 initialization)
    if args.local_rank != -1:
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
                                                          output_device=args.local_rank,
                                                          find_unused_parameters=True)

thomwolf's avatar
thomwolf committed
135
136
    # Train!
    logger.info("***** Running training *****")
137
138
    logger.info("  Num examples = %d", len(train_dataset))
    logger.info("  Num Epochs = %d", args.num_train_epochs)
thomwolf's avatar
thomwolf committed
139
140
141
    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
                   args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
142
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
thomwolf's avatar
thomwolf committed
143
    logger.info("  Total optimization steps = %d", t_total)
thomwolf's avatar
thomwolf committed
144
145

    global_step = 0
146
147
148
149
150
151
152
153
154
155
156
157
158
159
    epochs_trained = 0
    steps_trained_in_current_epoch = 0
    # Check if continuing training from a checkpoint
    if os.path.exists(args.model_name_or_path):
        # set global_step to gobal_step of last saved checkpoint from model path
        global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
        epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
        steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)

        logger.info("  Continuing training from checkpoint, will skip to saved global_step")
        logger.info("  Continuing training from epoch %d", epochs_trained)
        logger.info("  Continuing training from global step %d", global_step)
        logger.info("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)

thomwolf's avatar
thomwolf committed
160
    tr_loss, logging_loss = 0.0, 0.0
thomwolf's avatar
thomwolf committed
161
    model.zero_grad()
162
    train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
thomwolf's avatar
thomwolf committed
163
    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
thomwolf's avatar
thomwolf committed
164
165
166
    for _ in train_iterator:
        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
        for step, batch in enumerate(epoch_iterator):
167
168
169
170
171
172

            # Skip past any already trained steps if resuming training
            if steps_trained_in_current_epoch > 0:
                steps_trained_in_current_epoch -= 1
                continue

thomwolf's avatar
thomwolf committed
173
            model.train()
thomwolf's avatar
thomwolf committed
174
            batch = tuple(t.to(args.device) for t in batch)
175
176
177
            inputs = {'input_ids':      batch[0],
                      'attention_mask': batch[1],
                      'labels':         batch[3]}
178
179
            if args.model_type != 'distilbert':
                inputs['token_type_ids'] = batch[2] if args.model_type in ['bert', 'xlnet'] else None  # XLM, DistilBERT and RoBERTa don't use segment_ids
Peiqin Lin's avatar
typos  
Peiqin Lin committed
180
            outputs = model(**inputs)
181
            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
thomwolf's avatar
thomwolf committed
182
183
184
185
186
187

            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

thomwolf's avatar
thomwolf committed
188
189
190
191
192
            if args.fp16:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
            else:
                loss.backward()
thomwolf's avatar
thomwolf committed
193
194

            tr_loss += loss.item()
195
            if (step + 1) % args.gradient_accumulation_steps == 0:
196
197
198
199
200
                if args.fp16:
                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
                else:
                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)

thomwolf's avatar
thomwolf committed
201
                optimizer.step()
thomwolf's avatar
thomwolf committed
202
                scheduler.step()  # Update learning rate schedule
thomwolf's avatar
thomwolf committed
203
                model.zero_grad()
thomwolf's avatar
thomwolf committed
204
                global_step += 1
thomwolf's avatar
thomwolf committed
205

thomwolf's avatar
thomwolf committed
206
                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
Juha Kiili's avatar
Juha Kiili committed
207
                    logs = {}
thomwolf's avatar
thomwolf committed
208
                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
thomwolf's avatar
thomwolf committed
209
                        results = evaluate(args, model, tokenizer)
thomwolf's avatar
thomwolf committed
210
                        for key, value in results.items():
211
                            eval_key = 'eval_{}'.format(key)
Juha Kiili's avatar
Juha Kiili committed
212
213
                            logs[eval_key] = value

214
215
216
217
                    loss_scalar = (tr_loss - logging_loss) / args.logging_steps
                    learning_rate_scalar = scheduler.get_lr()[0]
                    logs['learning_rate'] = learning_rate_scalar
                    logs['loss'] = loss_scalar
thomwolf's avatar
thomwolf committed
218
                    logging_loss = tr_loss
thomwolf's avatar
thomwolf committed
219

Juha Kiili's avatar
Juha Kiili committed
220
221
222
                    for key, value in logs.items():
                        tb_writer.add_scalar(key, value, global_step)
                    print(json.dumps({**logs, **{'step': global_step}}))
thomwolf's avatar
thomwolf committed
223
224
225
226
227
228
229
230

                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
                    # Save model checkpoint
                    output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
                    if not os.path.exists(output_dir):
                        os.makedirs(output_dir)
                    model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
                    model_to_save.save_pretrained(output_dir)
231
232
                    tokenizer.save_pretrained(output_dir)

thomwolf's avatar
thomwolf committed
233
                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
thomwolf's avatar
thomwolf committed
234
                    logger.info("Saving model checkpoint to %s", output_dir)
thomwolf's avatar
thomwolf committed
235

236
237
238
239
                    torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
                    torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
                    logger.info("Saving optimizer and scheduler states to %s", output_dir)

thomwolf's avatar
thomwolf committed
240
            if args.max_steps > 0 and global_step > args.max_steps:
thomwolf's avatar
thomwolf committed
241
                epoch_iterator.close()
thomwolf's avatar
thomwolf committed
242
243
                break
        if args.max_steps > 0 and global_step > args.max_steps:
thomwolf's avatar
thomwolf committed
244
            train_iterator.close()
thomwolf's avatar
thomwolf committed
245
            break
thomwolf's avatar
thomwolf committed
246

thomwolf's avatar
thomwolf committed
247
248
249
    if args.local_rank in [-1, 0]:
        tb_writer.close()

thomwolf's avatar
thomwolf committed
250
251
252
    return global_step, tr_loss / global_step


thomwolf's avatar
thomwolf committed
253
def evaluate(args, model, tokenizer, prefix=""):
thomwolf's avatar
thomwolf committed
254
255
256
257
258
259
260
261
262
263
264
    # Loop to handle MNLI double evaluation (matched, mis-matched)
    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,)

    results = {}
    for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
        eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True)

        if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
            os.makedirs(eval_output_dir)

thomwolf's avatar
thomwolf committed
265
        args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
thomwolf's avatar
thomwolf committed
266
        # Note that DistributedSampler samples randomly
267
        eval_sampler = SequentialSampler(eval_dataset)
thomwolf's avatar
thomwolf committed
268
269
        eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

ronakice's avatar
ronakice committed
270
271
272
273
        # multi-gpu eval
        if args.n_gpu > 1:
            model = torch.nn.DataParallel(model)

thomwolf's avatar
thomwolf committed
274
        # Eval!
thomwolf's avatar
thomwolf committed
275
        logger.info("***** Running evaluation {} *****".format(prefix))
thomwolf's avatar
thomwolf committed
276
277
        logger.info("  Num examples = %d", len(eval_dataset))
        logger.info("  Batch size = %d", args.eval_batch_size)
thomwolf's avatar
thomwolf committed
278
        eval_loss = 0.0
thomwolf's avatar
thomwolf committed
279
280
281
282
        nb_eval_steps = 0
        preds = None
        out_label_ids = None
        for batch in tqdm(eval_dataloader, desc="Evaluating"):
thomwolf's avatar
thomwolf committed
283
            model.eval()
thomwolf's avatar
thomwolf committed
284
285
286
287
288
289
            batch = tuple(t.to(args.device) for t in batch)

            with torch.no_grad():
                inputs = {'input_ids':      batch[0],
                          'attention_mask': batch[1],
                          'labels':         batch[3]}
290
291
                if args.model_type != 'distilbert':
                    inputs['token_type_ids'] = batch[2] if args.model_type in ['bert', 'xlnet'] else None  # XLM, DistilBERT and RoBERTa don't use segment_ids
thomwolf's avatar
thomwolf committed
292
293
294
                outputs = model(**inputs)
                tmp_eval_loss, logits = outputs[:2]

thomwolf's avatar
thomwolf committed
295
                eval_loss += tmp_eval_loss.mean().item()
thomwolf's avatar
thomwolf committed
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
            nb_eval_steps += 1
            if preds is None:
                preds = logits.detach().cpu().numpy()
                out_label_ids = inputs['labels'].detach().cpu().numpy()
            else:
                preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
                out_label_ids = np.append(out_label_ids, inputs['labels'].detach().cpu().numpy(), axis=0)

        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)
        results.update(result)

312
        output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
thomwolf's avatar
thomwolf committed
313
        with open(output_eval_file, "w") as writer:
thomwolf's avatar
thomwolf committed
314
            logger.info("***** Eval results {} *****".format(prefix))
thomwolf's avatar
thomwolf committed
315
316
317
318
319
320
321
            for key in sorted(result.keys()):
                logger.info("  %s = %s", key, str(result[key]))
                writer.write("%s = %s\n" % (key, str(result[key])))

    return results


thomwolf's avatar
thomwolf committed
322
def load_and_cache_examples(args, task, tokenizer, evaluate=False):
VictorSanh's avatar
VictorSanh committed
323
    if args.local_rank not in [-1, 0] and not evaluate:
thomwolf's avatar
thomwolf committed
324
325
        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache

thomwolf's avatar
thomwolf committed
326
    processor = processors[task]()
327
328
329
330
    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',
331
        list(filter(None, args.model_name_or_path.split('/'))).pop(),
thomwolf's avatar
thomwolf committed
332
333
        str(args.max_seq_length),
        str(task)))
334
    if os.path.exists(cached_features_file) and not args.overwrite_cache:
thomwolf's avatar
thomwolf committed
335
        logger.info("Loading features from cached file %s", cached_features_file)
thomwolf's avatar
thomwolf committed
336
337
        features = torch.load(cached_features_file)
    else:
338
339
        logger.info("Creating features from dataset file at %s", args.data_dir)
        label_list = processor.get_labels()
340
341
342
        if task in ['mnli', 'mnli-mm'] and args.model_type in ['roberta']:
            # HACK(label indices are swapped in RoBERTa pretrained model)
            label_list[1], label_list[2] = label_list[2], label_list[1] 
343
        examples = processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
thomwolf's avatar
thomwolf committed
344
345
        features = convert_examples_to_features(examples,
                                                tokenizer,
thomwolf's avatar
thomwolf committed
346
347
348
                                                label_list=label_list,
                                                max_length=args.max_seq_length,
                                                output_mode=output_mode,
thomwolf's avatar
thomwolf committed
349
350
351
                                                pad_on_left=bool(args.model_type in ['xlnet']),                 # pad on the left for xlnet
                                                pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
                                                pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0,
352
        )
353
        if args.local_rank in [-1, 0]:
thomwolf's avatar
thomwolf committed
354
            logger.info("Saving features into cached file %s", cached_features_file)
thomwolf's avatar
thomwolf committed
355
356
            torch.save(features, cached_features_file)

VictorSanh's avatar
VictorSanh committed
357
    if args.local_rank == 0 and not evaluate:
thomwolf's avatar
thomwolf committed
358
359
        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache

360
361
    # Convert to Tensors and build dataset
    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
thomwolf's avatar
thomwolf committed
362
363
    all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
    all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
364
    if output_mode == "classification":
thomwolf's avatar
thomwolf committed
365
        all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
366
    elif output_mode == "regression":
thomwolf's avatar
thomwolf committed
367
        all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
Lysandre's avatar
Lysandre committed
368
 
thomwolf's avatar
thomwolf committed
369
    dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
370
    return dataset
thomwolf's avatar
thomwolf committed
371
372


thomwolf's avatar
thomwolf committed
373
374
375
376
377
378
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.")
379
380
381
382
    parser.add_argument("--model_type", default=None, type=str, required=True,
                        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
thomwolf's avatar
thomwolf committed
383
    parser.add_argument("--task_name", default=None, type=str, required=True,
384
                        help="The name of the task to train selected in the list: " + ", ".join(processors.keys()))
thomwolf's avatar
thomwolf committed
385
386
387
388
    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
thomwolf's avatar
thomwolf committed
389
390
391
392
    parser.add_argument("--config_name", default="", type=str,
                        help="Pretrained config name or path if not the same as model_name")
    parser.add_argument("--tokenizer_name", default="", type=str,
                        help="Pretrained tokenizer name or path if not the same as model_name")
thomwolf's avatar
thomwolf committed
393
394
395
    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,
396
397
                        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
398
399
400
401
    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
402
403
    parser.add_argument("--evaluate_during_training", action='store_true',
                        help="Rul evaluation during training at each logging step.")
thomwolf's avatar
thomwolf committed
404
405
    parser.add_argument("--do_lower_case", action='store_true',
                        help="Set this flag if you are using an uncased model.")
thomwolf's avatar
thomwolf committed
406
407

    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
408
                        help="Batch size per GPU/CPU for training.")
thomwolf's avatar
thomwolf committed
409
    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
410
                        help="Batch size per GPU/CPU for evaluation.")
thomwolf's avatar
thomwolf committed
411
    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
Lysandre's avatar
Lysandre committed
412
                        help="Number of updates steps to accumulate before performing a backward/update pass.")     
thomwolf's avatar
thomwolf committed
413
414
    parser.add_argument("--learning_rate", default=5e-5, type=float,
                        help="The initial learning rate for Adam.")
thomwolf's avatar
thomwolf committed
415
416
    parser.add_argument("--weight_decay", default=0.0, type=float,
                        help="Weight deay if we apply some.")
thomwolf's avatar
thomwolf committed
417
418
    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
                        help="Epsilon for Adam optimizer.")
thomwolf's avatar
thomwolf committed
419
420
    parser.add_argument("--max_grad_norm", default=1.0, type=float,
                        help="Max gradient norm.")
thomwolf's avatar
thomwolf committed
421
422
    parser.add_argument("--num_train_epochs", default=3.0, type=float,
                        help="Total number of training epochs to perform.")
thomwolf's avatar
thomwolf committed
423
424
    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
425
426
    parser.add_argument("--warmup_steps", default=0, type=int,
                        help="Linear warmup over warmup_steps.")
thomwolf's avatar
thomwolf committed
427

thomwolf's avatar
thomwolf committed
428
    parser.add_argument('--logging_steps', type=int, default=50,
thomwolf's avatar
thomwolf committed
429
                        help="Log every X updates steps.")
thomwolf's avatar
thomwolf committed
430
431
432
433
    parser.add_argument('--save_steps', type=int, default=50,
                        help="Save checkpoint every X updates steps.")
    parser.add_argument("--eval_all_checkpoints", action='store_true',
                        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
thomwolf's avatar
thomwolf committed
434
435
436
437
    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")
thomwolf's avatar
thomwolf committed
438
439
    parser.add_argument('--overwrite_cache', action='store_true',
                        help="Overwrite the cached training and evaluation sets")
thomwolf's avatar
thomwolf committed
440
441
442
443
    parser.add_argument('--seed', type=int, default=42,
                        help="random seed for initialization")

    parser.add_argument('--fp16', action='store_true',
thomwolf's avatar
thomwolf committed
444
445
446
447
                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
    parser.add_argument('--fp16_opt_level', type=str, default='O1',
                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
                             "See details at https://nvidia.github.io/apex/amp.html")
thomwolf's avatar
thomwolf committed
448
    parser.add_argument("--local_rank", type=int, default=-1,
thomwolf's avatar
thomwolf committed
449
450
451
                        help="For distributed training: local_rank")
    parser.add_argument('--server_ip', type=str, default='', help="For distant debugging.")
    parser.add_argument('--server_port', type=str, default='', help="For distant debugging.")
thomwolf's avatar
thomwolf committed
452
453
    args = parser.parse_args()

thomwolf's avatar
thomwolf committed
454
455
456
    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
457
458
459
460
461
462
463
464
465
466
467
    # 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
468
        args.n_gpu = torch.cuda.device_count()
thomwolf's avatar
thomwolf committed
469
470
471
472
    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
473
        args.n_gpu = 1
thomwolf's avatar
thomwolf committed
474
475
476
    args.device = device

    # Setup logging
thomwolf's avatar
thomwolf committed
477
478
479
    logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
                        datefmt = '%m/%d/%Y %H:%M:%S',
                        level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
thomwolf's avatar
thomwolf committed
480
    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
thomwolf's avatar
thomwolf committed
481
                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
thomwolf's avatar
thomwolf committed
482

thomwolf's avatar
thomwolf committed
483
484
    # Set seed
    set_seed(args)
thomwolf's avatar
thomwolf committed
485
486

    # Prepare GLUE task
thomwolf's avatar
thomwolf committed
487
488
489
490
491
    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
492
493
494
495
496
    label_list = processor.get_labels()
    num_labels = len(label_list)

    # Load pretrained model and tokenizer
    if args.local_rank not in [-1, 0]:
497
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
thomwolf's avatar
thomwolf committed
498

499
    args.model_type = args.model_type.lower()
thomwolf's avatar
thomwolf committed
500
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
thomwolf's avatar
thomwolf committed
501
502
503
504
505
506
507
508
509
510
511
    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
                                          num_labels=num_labels,
                                          finetuning_task=args.task_name,
                                          cache_dir=args.cache_dir if args.cache_dir else None)
    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
                                                do_lower_case=args.do_lower_case,
                                                cache_dir=args.cache_dir if args.cache_dir else None)
    model = model_class.from_pretrained(args.model_name_or_path,
                                        from_tf=bool('.ckpt' in args.model_name_or_path),
                                        config=config,
                                        cache_dir=args.cache_dir if args.cache_dir else None)
thomwolf's avatar
thomwolf committed
512
513

    if args.local_rank == 0:
514
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
thomwolf's avatar
thomwolf committed
515

thomwolf's avatar
thomwolf committed
516
    model.to(args.device)
thomwolf's avatar
thomwolf committed
517

thomwolf's avatar
thomwolf committed
518
519
    logger.info("Training/evaluation parameters %s", args)

520

thomwolf's avatar
thomwolf committed
521
    # Training
thomwolf's avatar
thomwolf committed
522
    if args.do_train:
523
        train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
thomwolf's avatar
thomwolf committed
524
        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
thomwolf's avatar
thomwolf committed
525
        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
thomwolf's avatar
thomwolf committed
526
527


thomwolf's avatar
thomwolf committed
528
    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
529
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
530
531
532
533
        # 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)

thomwolf's avatar
thomwolf committed
534
        logger.info("Saving model checkpoint to %s", args.output_dir)
535
536
        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
        # They can then be reloaded using `from_pretrained()`
thomwolf's avatar
thomwolf committed
537
538
        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)
539
        tokenizer.save_pretrained(args.output_dir)
thomwolf's avatar
thomwolf committed
540
541

        # Good practice: save your training arguments together with the trained model
542
        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
thomwolf's avatar
thomwolf committed
543

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

549

thomwolf's avatar
thomwolf committed
550
    # Evaluation
thomwolf's avatar
thomwolf committed
551
    results = {}
thomwolf's avatar
thomwolf committed
552
    if args.do_eval and args.local_rank in [-1, 0]:
553
        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
thomwolf's avatar
thomwolf committed
554
        checkpoints = [args.output_dir]
thomwolf's avatar
thomwolf committed
555
        if args.eval_all_checkpoints:
thomwolf's avatar
thomwolf committed
556
            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
557
            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
thomwolf's avatar
thomwolf committed
558
559
        logger.info("Evaluate the following checkpoints: %s", checkpoints)
        for checkpoint in checkpoints:
560
            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
561
562
            prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
            
thomwolf's avatar
thomwolf committed
563
            model = model_class.from_pretrained(checkpoint)
thomwolf's avatar
thomwolf committed
564
            model.to(args.device)
565
            result = evaluate(args, model, tokenizer, prefix=prefix)
thomwolf's avatar
thomwolf committed
566
567
568
            result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
            results.update(result)

thomwolf's avatar
thomwolf committed
569
    return results
thomwolf's avatar
thomwolf committed
570
571
572
573


if __name__ == "__main__":
    main()