run_glue.py 26.6 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
25
26
27
28
29
30
import logging
import os
import random

import numpy as np
import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)
from torch.utils.data.distributed import DistributedSampler
31
32
33
34
35
36

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

thomwolf's avatar
thomwolf committed
37
from tqdm import tqdm, trange
thomwolf's avatar
thomwolf committed
38

39
from transformers import (WEIGHTS_NAME, BertConfig,
thomwolf's avatar
thomwolf committed
40
                                  BertForSequenceClassification, BertTokenizer,
41
42
43
                                  RobertaConfig,
                                  RobertaForSequenceClassification,
                                  RobertaTokenizer,
thomwolf's avatar
thomwolf committed
44
45
46
                                  XLMConfig, XLMForSequenceClassification,
                                  XLMTokenizer, XLNetConfig,
                                  XLNetForSequenceClassification,
47
48
49
50
                                  XLNetTokenizer,
                                  DistilBertConfig,
                                  DistilBertForSequenceClassification,
                                  DistilBertTokenizer)
thomwolf's avatar
thomwolf committed
51

52
from transformers import AdamW, WarmupLinearSchedule
thomwolf's avatar
thomwolf committed
53

54
55
56
57
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
58
59
60

logger = logging.getLogger(__name__)

Brian Ma's avatar
Brian Ma committed
61
62
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig, 
                                                                                RobertaConfig, DistilBertConfig)), ())
63
64

MODEL_CLASSES = {
thomwolf's avatar
thomwolf committed
65
66
67
    'bert': (BertConfig, BertForSequenceClassification, BertTokenizer),
    'xlnet': (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer),
    'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
68
    'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
69
    'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer)
70
}
thomwolf's avatar
thomwolf committed
71

thomwolf's avatar
thomwolf committed
72
73
74
75
76
77
78
79
80

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
81
def train(args, train_dataset, model, tokenizer):
thomwolf's avatar
thomwolf committed
82
83
84
85
    """ Train the model """
    if args.local_rank in [-1, 0]:
        tb_writer = SummaryWriter()

thomwolf's avatar
thomwolf committed
86
    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
87
88
    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
89

thomwolf's avatar
thomwolf committed
90
    if args.max_steps > 0:
thomwolf's avatar
thomwolf committed
91
        t_total = args.max_steps
thomwolf's avatar
thomwolf committed
92
93
        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
    else:
thomwolf's avatar
thomwolf committed
94
        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
thomwolf's avatar
thomwolf committed
95

thomwolf's avatar
thomwolf committed
96
    # Prepare optimizer and schedule (linear warmup and decay)
thomwolf's avatar
thomwolf committed
97
    no_decay = ['bias', 'LayerNorm.weight']
thomwolf's avatar
thomwolf committed
98
    optimizer_grouped_parameters = [
thomwolf's avatar
thomwolf committed
99
100
        {'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
101
        ]
thomwolf's avatar
thomwolf committed
102
103
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
thomwolf's avatar
thomwolf committed
104
105
    if args.fp16:
        try:
thomwolf's avatar
thomwolf committed
106
            from apex import amp
thomwolf's avatar
thomwolf committed
107
108
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
thomwolf's avatar
thomwolf committed
109
        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
thomwolf's avatar
thomwolf committed
110

111
112
113
114
    # multi-gpu training (should be after apex fp16 initialization)
    if args.n_gpu > 1:
        model = torch.nn.DataParallel(model)

thomwolf's avatar
thomwolf committed
115
116
117
118
119
120
    # 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
121
122
    # Train!
    logger.info("***** Running training *****")
123
124
    logger.info("  Num examples = %d", len(train_dataset))
    logger.info("  Num Epochs = %d", args.num_train_epochs)
thomwolf's avatar
thomwolf committed
125
126
127
    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))
128
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
thomwolf's avatar
thomwolf committed
129
    logger.info("  Total optimization steps = %d", t_total)
thomwolf's avatar
thomwolf committed
130
131

    global_step = 0
thomwolf's avatar
thomwolf committed
132
    tr_loss, logging_loss = 0.0, 0.0
thomwolf's avatar
thomwolf committed
133
    model.zero_grad()
thomwolf's avatar
thomwolf committed
134
    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
thomwolf's avatar
thomwolf committed
135
    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
thomwolf's avatar
thomwolf committed
136
137
138
    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):
thomwolf's avatar
thomwolf committed
139
            model.train()
thomwolf's avatar
thomwolf committed
140
            batch = tuple(t.to(args.device) for t in batch)
141
142
143
            inputs = {'input_ids':      batch[0],
                      'attention_mask': batch[1],
                      'labels':         batch[3]}
144
145
            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
146
            outputs = model(**inputs)
147
            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
thomwolf's avatar
thomwolf committed
148
149
150
151
152
153

            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
154
155
156
            if args.fp16:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
thomwolf's avatar
thomwolf committed
157
                torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
thomwolf's avatar
thomwolf committed
158
159
            else:
                loss.backward()
thomwolf's avatar
thomwolf committed
160
                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
thomwolf's avatar
thomwolf committed
161
162
163
164

            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0:
                optimizer.step()
thomwolf's avatar
thomwolf committed
165
                scheduler.step()  # Update learning rate schedule
thomwolf's avatar
thomwolf committed
166
                model.zero_grad()
thomwolf's avatar
thomwolf committed
167
                global_step += 1
thomwolf's avatar
thomwolf committed
168

thomwolf's avatar
thomwolf committed
169
                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
thomwolf's avatar
thomwolf committed
170
                    # Log metrics
thomwolf's avatar
thomwolf committed
171
                    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
172
                        results = evaluate(args, model, tokenizer)
thomwolf's avatar
thomwolf committed
173
174
                        for key, value in results.items():
                            tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
thomwolf's avatar
thomwolf committed
175
                    tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
thomwolf's avatar
thomwolf committed
176
177
                    tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
                    logging_loss = tr_loss
thomwolf's avatar
thomwolf committed
178
179
180
181
182
183
184
185
186

                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
                    # Save model checkpoint
                    output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
                    if not os.path.exists(output_dir):
                        os.makedirs(output_dir)
                    model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
                    model_to_save.save_pretrained(output_dir)
                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
thomwolf's avatar
thomwolf committed
187
                    logger.info("Saving model checkpoint to %s", output_dir)
thomwolf's avatar
thomwolf committed
188

thomwolf's avatar
thomwolf committed
189
            if args.max_steps > 0 and global_step > args.max_steps:
thomwolf's avatar
thomwolf committed
190
                epoch_iterator.close()
thomwolf's avatar
thomwolf committed
191
192
                break
        if args.max_steps > 0 and global_step > args.max_steps:
thomwolf's avatar
thomwolf committed
193
            train_iterator.close()
thomwolf's avatar
thomwolf committed
194
            break
thomwolf's avatar
thomwolf committed
195

thomwolf's avatar
thomwolf committed
196
197
198
    if args.local_rank in [-1, 0]:
        tb_writer.close()

thomwolf's avatar
thomwolf committed
199
200
201
    return global_step, tr_loss / global_step


thomwolf's avatar
thomwolf committed
202
def evaluate(args, model, tokenizer, prefix=""):
thomwolf's avatar
thomwolf committed
203
204
205
206
207
208
209
210
211
212
213
    # 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
214
        args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
thomwolf's avatar
thomwolf committed
215
216
217
218
219
        # Note that DistributedSampler samples randomly
        eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
        eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)

        # Eval!
thomwolf's avatar
thomwolf committed
220
        logger.info("***** Running evaluation {} *****".format(prefix))
thomwolf's avatar
thomwolf committed
221
222
        logger.info("  Num examples = %d", len(eval_dataset))
        logger.info("  Batch size = %d", args.eval_batch_size)
thomwolf's avatar
thomwolf committed
223
        eval_loss = 0.0
thomwolf's avatar
thomwolf committed
224
225
226
227
        nb_eval_steps = 0
        preds = None
        out_label_ids = None
        for batch in tqdm(eval_dataloader, desc="Evaluating"):
thomwolf's avatar
thomwolf committed
228
            model.eval()
thomwolf's avatar
thomwolf committed
229
230
231
232
233
234
            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]}
235
236
                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
237
238
239
                outputs = model(**inputs)
                tmp_eval_loss, logits = outputs[:2]

thomwolf's avatar
thomwolf committed
240
                eval_loss += tmp_eval_loss.mean().item()
thomwolf's avatar
thomwolf committed
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
            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)

        output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
        with open(output_eval_file, "w") as writer:
thomwolf's avatar
thomwolf committed
259
            logger.info("***** Eval results {} *****".format(prefix))
thomwolf's avatar
thomwolf committed
260
261
262
263
264
265
266
            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
267
def load_and_cache_examples(args, task, tokenizer, evaluate=False):
VictorSanh's avatar
VictorSanh committed
268
    if args.local_rank not in [-1, 0] and not evaluate:
thomwolf's avatar
thomwolf committed
269
270
        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
271
    processor = processors[task]()
272
273
274
275
    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',
276
        list(filter(None, args.model_name_or_path.split('/'))).pop(),
thomwolf's avatar
thomwolf committed
277
278
        str(args.max_seq_length),
        str(task)))
thomwolf's avatar
thomwolf committed
279
    if os.path.exists(cached_features_file):
thomwolf's avatar
thomwolf committed
280
        logger.info("Loading features from cached file %s", cached_features_file)
thomwolf's avatar
thomwolf committed
281
282
        features = torch.load(cached_features_file)
    else:
283
284
        logger.info("Creating features from dataset file at %s", args.data_dir)
        label_list = processor.get_labels()
285
286
287
        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] 
288
        examples = processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
thomwolf's avatar
thomwolf committed
289
290
        features = convert_examples_to_features(examples,
                                                tokenizer,
thomwolf's avatar
thomwolf committed
291
292
293
                                                label_list=label_list,
                                                max_length=args.max_seq_length,
                                                output_mode=output_mode,
thomwolf's avatar
thomwolf committed
294
295
296
                                                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,
297
        )
298
        if args.local_rank in [-1, 0]:
thomwolf's avatar
thomwolf committed
299
            logger.info("Saving features into cached file %s", cached_features_file)
thomwolf's avatar
thomwolf committed
300
301
            torch.save(features, cached_features_file)

VictorSanh's avatar
VictorSanh committed
302
    if args.local_rank == 0 and not evaluate:
thomwolf's avatar
thomwolf committed
303
304
        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache

305
306
    # 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
307
308
    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)
309
    if output_mode == "classification":
thomwolf's avatar
thomwolf committed
310
        all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
311
    elif output_mode == "regression":
thomwolf's avatar
thomwolf committed
312
        all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
313

thomwolf's avatar
thomwolf committed
314
    dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
315
    return dataset
thomwolf's avatar
thomwolf committed
316
317


thomwolf's avatar
thomwolf committed
318
319
320
321
322
323
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.")
324
325
326
327
    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
328
    parser.add_argument("--task_name", default=None, type=str, required=True,
329
                        help="The name of the task to train selected in the list: " + ", ".join(processors.keys()))
thomwolf's avatar
thomwolf committed
330
331
332
333
    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
334
335
336
337
    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
338
339
340
    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,
341
342
                        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
343
344
345
346
    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
347
348
    parser.add_argument("--evaluate_during_training", action='store_true',
                        help="Rul evaluation during training at each logging step.")
thomwolf's avatar
thomwolf committed
349
350
    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
351
352

    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
353
                        help="Batch size per GPU/CPU for training.")
thomwolf's avatar
thomwolf committed
354
    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
355
                        help="Batch size per GPU/CPU for evaluation.")
thomwolf's avatar
thomwolf committed
356
357
358
359
    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.")
thomwolf's avatar
thomwolf committed
360
361
    parser.add_argument("--weight_decay", default=0.0, type=float,
                        help="Weight deay if we apply some.")
thomwolf's avatar
thomwolf committed
362
363
    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
                        help="Epsilon for Adam optimizer.")
thomwolf's avatar
thomwolf committed
364
365
    parser.add_argument("--max_grad_norm", default=1.0, type=float,
                        help="Max gradient norm.")
thomwolf's avatar
thomwolf committed
366
367
    parser.add_argument("--num_train_epochs", default=3.0, type=float,
                        help="Total number of training epochs to perform.")
thomwolf's avatar
thomwolf committed
368
369
    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
370
371
    parser.add_argument("--warmup_steps", default=0, type=int,
                        help="Linear warmup over warmup_steps.")
thomwolf's avatar
thomwolf committed
372

thomwolf's avatar
thomwolf committed
373
    parser.add_argument('--logging_steps', type=int, default=50,
thomwolf's avatar
thomwolf committed
374
                        help="Log every X updates steps.")
thomwolf's avatar
thomwolf committed
375
376
377
378
    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
379
380
381
382
    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
383
384
    parser.add_argument('--overwrite_cache', action='store_true',
                        help="Overwrite the cached training and evaluation sets")
thomwolf's avatar
thomwolf committed
385
386
387
388
    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
389
390
391
392
                        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
393
    parser.add_argument("--local_rank", type=int, default=-1,
thomwolf's avatar
thomwolf committed
394
395
396
                        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
397
398
    args = parser.parse_args()

thomwolf's avatar
thomwolf committed
399
400
401
    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
402
403
404
405
406
407
408
409
410
411
412
    # 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
413
        args.n_gpu = torch.cuda.device_count()
thomwolf's avatar
thomwolf committed
414
415
416
417
    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
418
        args.n_gpu = 1
thomwolf's avatar
thomwolf committed
419
420
421
    args.device = device

    # Setup logging
thomwolf's avatar
thomwolf committed
422
423
424
    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
425
    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
thomwolf's avatar
thomwolf committed
426
                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
thomwolf's avatar
thomwolf committed
427

thomwolf's avatar
thomwolf committed
428
429
    # Set seed
    set_seed(args)
thomwolf's avatar
thomwolf committed
430
431

    # Prepare GLUE task
thomwolf's avatar
thomwolf committed
432
433
434
435
436
    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
437
438
439
440
441
    label_list = processor.get_labels()
    num_labels = len(label_list)

    # Load pretrained model and tokenizer
    if args.local_rank not in [-1, 0]:
442
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
thomwolf's avatar
thomwolf committed
443

444
    args.model_type = args.model_type.lower()
thomwolf's avatar
thomwolf committed
445
    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
446
447
448
    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)
    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)
    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
thomwolf's avatar
thomwolf committed
449
450

    if args.local_rank == 0:
451
        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
thomwolf's avatar
thomwolf committed
452

thomwolf's avatar
thomwolf committed
453
    model.to(args.device)
thomwolf's avatar
thomwolf committed
454

thomwolf's avatar
thomwolf committed
455
456
    logger.info("Training/evaluation parameters %s", args)

457

thomwolf's avatar
thomwolf committed
458
    # Training
thomwolf's avatar
thomwolf committed
459
    if args.do_train:
460
        train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
thomwolf's avatar
thomwolf committed
461
        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
thomwolf's avatar
thomwolf committed
462
        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
thomwolf's avatar
thomwolf committed
463
464


thomwolf's avatar
thomwolf committed
465
    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
thomwolf's avatar
thomwolf committed
466
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
467
468
469
470
        # 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
471
        logger.info("Saving model checkpoint to %s", args.output_dir)
472
473
        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
        # They can then be reloaded using `from_pretrained()`
thomwolf's avatar
thomwolf committed
474
475
        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)
476
        tokenizer.save_pretrained(args.output_dir)
thomwolf's avatar
thomwolf committed
477
478

        # Good practice: save your training arguments together with the trained model
479
        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
thomwolf's avatar
thomwolf committed
480

481
        # Load a trained model and vocabulary that you have fine-tuned
482
        model = model_class.from_pretrained(args.output_dir)
Peng Qi's avatar
Peng Qi committed
483
        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
484
        model.to(args.device)
thomwolf's avatar
thomwolf committed
485

486

thomwolf's avatar
thomwolf committed
487
    # Evaluation
thomwolf's avatar
thomwolf committed
488
    results = {}
thomwolf's avatar
thomwolf committed
489
    if args.do_eval and args.local_rank in [-1, 0]:
490
        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
thomwolf's avatar
thomwolf committed
491
        checkpoints = [args.output_dir]
thomwolf's avatar
thomwolf committed
492
        if args.eval_all_checkpoints:
thomwolf's avatar
thomwolf committed
493
            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
494
            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
thomwolf's avatar
thomwolf committed
495
496
        logger.info("Evaluate the following checkpoints: %s", checkpoints)
        for checkpoint in checkpoints:
497
            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
thomwolf's avatar
thomwolf committed
498
            model = model_class.from_pretrained(checkpoint)
thomwolf's avatar
thomwolf committed
499
500
            model.to(args.device)
            result = evaluate(args, model, tokenizer, prefix=global_step)
thomwolf's avatar
thomwolf committed
501
502
503
            result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
            results.update(result)

thomwolf's avatar
thomwolf committed
504
    return results
thomwolf's avatar
thomwolf committed
505
506
507
508


if __name__ == "__main__":
    main()