run_lm_finetuning.py 30.2 KB
Newer Older
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
"""
LysandreJik's avatar
LysandreJik committed
17
Fine-tuning the library models for language modeling on a text file (GPT, GPT-2, BERT, RoBERTa).
18
19
20
GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa are fine-tuned
using a masked language modeling (MLM) loss.
"""
21
22
23
24
25
26


import argparse
import glob
import logging
import os
27
import pickle
28
import random
jinoobaek-qz's avatar
jinoobaek-qz committed
29
30
import re
import shutil
31
from typing import Tuple
32
33
34

import numpy as np
import torch
Aymeric Augustin's avatar
Aymeric Augustin committed
35
from torch.utils.data import DataLoader, Dataset, RandomSampler, SequentialSampler
36
37
38
from torch.utils.data.distributed import DistributedSampler
from tqdm import tqdm, trange

39
40
41
42
43
44
from transformers import (
    WEIGHTS_NAME,
    AdamW,
    BertConfig,
    BertForMaskedLM,
    BertTokenizer,
Aymeric Augustin's avatar
Aymeric Augustin committed
45
46
47
48
49
50
    CamembertConfig,
    CamembertForMaskedLM,
    CamembertTokenizer,
    DistilBertConfig,
    DistilBertForMaskedLM,
    DistilBertTokenizer,
51
52
53
54
55
56
    GPT2Config,
    GPT2LMHeadModel,
    GPT2Tokenizer,
    OpenAIGPTConfig,
    OpenAIGPTLMHeadModel,
    OpenAIGPTTokenizer,
57
    PreTrainedTokenizer,
58
59
60
    RobertaConfig,
    RobertaForMaskedLM,
    RobertaTokenizer,
Aymeric Augustin's avatar
Aymeric Augustin committed
61
    get_linear_schedule_with_warmup,
62
)
63

64

Aymeric Augustin's avatar
Aymeric Augustin committed
65
66
try:
    from torch.utils.tensorboard import SummaryWriter
67
except ImportError:
Aymeric Augustin's avatar
Aymeric Augustin committed
68
69
70
    from tensorboardX import SummaryWriter


71
logger = logging.getLogger(__name__)
72
73
74


MODEL_CLASSES = {
75
76
77
78
79
80
    "gpt2": (GPT2Config, GPT2LMHeadModel, GPT2Tokenizer),
    "openai-gpt": (OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
    "bert": (BertConfig, BertForMaskedLM, BertTokenizer),
    "roberta": (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer),
    "distilbert": (DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer),
    "camembert": (CamembertConfig, CamembertForMaskedLM, CamembertTokenizer),
81
82
83
}


84
class TextDataset(Dataset):
85
    def __init__(self, tokenizer, args, file_path="train", block_size=512):
86
87
        assert os.path.isfile(file_path)
        directory, filename = os.path.split(file_path)
88
89
90
        cached_features_file = os.path.join(
            directory, args.model_name_or_path + "_cached_lm_" + str(block_size) + "_" + filename
        )
91

Lysandre's avatar
Lysandre committed
92
        if os.path.exists(cached_features_file) and not args.overwrite_cache:
93
            logger.info("Loading features from cached file %s", cached_features_file)
94
            with open(cached_features_file, "rb") as handle:
95
96
97
98
99
100
101
102
103
                self.examples = pickle.load(handle)
        else:
            logger.info("Creating features from dataset file at %s", directory)

            self.examples = []
            with open(file_path, encoding="utf-8") as f:
                text = f.read()

            tokenized_text = tokenizer.convert_tokens_to_ids(tokenizer.tokenize(text))
104

105
106
            for i in range(0, len(tokenized_text) - block_size + 1, block_size):  # Truncate in block of block_size
                self.examples.append(tokenizer.build_inputs_with_special_tokens(tokenized_text[i : i + block_size]))
107
108
109
110
111
            # Note that we are loosing the last truncated example here for the sake of simplicity (no padding)
            # If your dataset is small, first you should loook for a bigger one :-) and second you
            # can change this behavior by adding (model specific) padding.

            logger.info("Saving features into cached file %s", cached_features_file)
112
            with open(cached_features_file, "wb") as handle:
113
114
115
116
117
118
119
120
121
122
                pickle.dump(self.examples, handle, protocol=pickle.HIGHEST_PROTOCOL)

    def __len__(self):
        return len(self.examples)

    def __getitem__(self, item):
        return torch.tensor(self.examples[item])


def load_and_cache_examples(args, tokenizer, evaluate=False):
123
124
125
126
127
128
    dataset = TextDataset(
        tokenizer,
        args,
        file_path=args.eval_data_file if evaluate else args.train_data_file,
        block_size=args.block_size,
    )
129
130
131
    return dataset


132
133
134
135
136
137
138
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)

139

140
141
142
143
144
145
146
def _rotate_checkpoints(args, checkpoint_prefix, use_mtime=False):
    if not args.save_total_limit:
        return
    if args.save_total_limit <= 0:
        return

    # Check if we should delete older checkpoint(s)
147
    glob_checkpoints = glob.glob(os.path.join(args.output_dir, "{}-*".format(checkpoint_prefix)))
jinoobaek-qz's avatar
jinoobaek-qz committed
148
149
150
    if len(glob_checkpoints) <= args.save_total_limit:
        return

151
    ordering_and_checkpoint_path = []
jinoobaek-qz's avatar
jinoobaek-qz committed
152
    for path in glob_checkpoints:
153
154
155
        if use_mtime:
            ordering_and_checkpoint_path.append((os.path.getmtime(path), path))
        else:
156
            regex_match = re.match(".*{}-([0-9]+)".format(checkpoint_prefix), path)
157
158
159
160
            if regex_match and regex_match.groups():
                ordering_and_checkpoint_path.append((int(regex_match.groups()[0]), path))

    checkpoints_sorted = sorted(ordering_and_checkpoint_path)
jinoobaek-qz's avatar
jinoobaek-qz committed
161
162
163
164
165
166
    checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted]
    number_of_checkpoints_to_delete = max(0, len(checkpoints_sorted) - args.save_total_limit)
    checkpoints_to_be_deleted = checkpoints_sorted[:number_of_checkpoints_to_delete]
    for checkpoint in checkpoints_to_be_deleted:
        logger.info("Deleting older checkpoint [{}] due to args.save_total_limit".format(checkpoint))
        shutil.rmtree(checkpoint)
jinoobaek-qz's avatar
jinoobaek-qz committed
167
168


169
def mask_tokens(inputs: torch.Tensor, tokenizer: PreTrainedTokenizer, args) -> Tuple[torch.Tensor, torch.Tensor]:
170
    """ Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. """
171
    labels = inputs.clone()
172
    # We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
173
    probability_matrix = torch.full(labels.shape, args.mlm_probability)
174
175
176
    special_tokens_mask = [
        tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()
    ]
177
    probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
178
    masked_indices = torch.bernoulli(probability_matrix).bool()
LysandreJik's avatar
LysandreJik committed
179
    labels[~masked_indices] = -100  # We only compute loss on masked tokens
180
181

    # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
thomwolf's avatar
thomwolf committed
182
    indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).bool() & masked_indices
183
184
185
    inputs[indices_replaced] = tokenizer.convert_tokens_to_ids(tokenizer.mask_token)

    # 10% of the time, we replace masked input tokens with random word
thomwolf's avatar
thomwolf committed
186
    indices_random = torch.bernoulli(torch.full(labels.shape, 0.5)).bool() & masked_indices & ~indices_replaced
187
188
    random_words = torch.randint(len(tokenizer), labels.shape, dtype=torch.long)
    inputs[indices_random] = random_words[indices_random]
189

190
    # The rest of the time (10% of the time) we keep the masked input tokens unchanged
191
    return inputs, labels
192

193

194
195
196
197
198
199
def train(args, train_dataset, model, tokenizer):
    """ Train the model """
    if args.local_rank in [-1, 0]:
        tb_writer = SummaryWriter()

    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
thomwolf's avatar
thomwolf committed
200
    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
201
    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
202
203
204
205
206
207
208
209

    if args.max_steps > 0:
        t_total = args.max_steps
        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
    else:
        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs

    # Prepare optimizer and schedule (linear warmup and decay)
210
    no_decay = ["bias", "LayerNorm.weight"]
211
    optimizer_grouped_parameters = [
212
213
214
215
216
217
        {
            "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},
    ]
218
    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
219
220
221
    scheduler = get_linear_schedule_with_warmup(
        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
    )
222
223

    # Check if saved optimizer or scheduler states exist
224
225
226
    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")
    ):
227
        # Load in optimizer and scheduler states
228
229
        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")))
230

231
232
233
234
235
236
237
238
239
240
241
242
243
    if args.fp16:
        try:
            from apex import amp
        except ImportError:
            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)

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

    # Distributed training (should be after apex fp16 initialization)
    if args.local_rank != -1:
244
245
246
        model = torch.nn.parallel.DistributedDataParallel(
            model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
        )
247
248
249
250
251
252

    # Train!
    logger.info("***** Running training *****")
    logger.info("  Num examples = %d", len(train_dataset))
    logger.info("  Num Epochs = %d", args.num_train_epochs)
    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
253
254
255
256
257
258
    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),
    )
259
260
261
262
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", t_total)

    global_step = 0
263
264
265
266
267
    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
268
        global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
269
270
271
272
273
274
275
276
        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)

277
    tr_loss, logging_loss = 0.0, 0.0
thomwolf's avatar
thomwolf committed
278

279
    model_to_resize = model.module if hasattr(model, "module") else model  # Take care of distributed/parallel training
thomwolf's avatar
thomwolf committed
280
281
    model_to_resize.resize_token_embeddings(len(tokenizer))

282
    model.zero_grad()
283
284
285
    train_iterator = trange(
        epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]
    )
286
    set_seed(args)  # Added here for reproducibility
Bilal Khan's avatar
Bilal Khan committed
287
    for _ in train_iterator:
288
289
        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
        for step, batch in enumerate(epoch_iterator):
290

291
292
293
294
295
            # Skip past any already trained steps if resuming training
            if steps_trained_in_current_epoch > 0:
                steps_trained_in_current_epoch -= 1
                continue

296
            inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
297
298
299
            inputs = inputs.to(args.device)
            labels = labels.to(args.device)
            model.train()
300
            outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
301
            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
302
303

            if args.n_gpu > 1:
304
                loss = loss.mean()  # mean() to average on multi-gpu parallel training
305
306
307
308
309
310
311
312
313
314
315
            if args.gradient_accumulation_steps > 1:
                loss = loss / args.gradient_accumulation_steps

            if args.fp16:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
            else:
                loss.backward()

            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0:
316
317
318
319
                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)
320
                optimizer.step()
321
                scheduler.step()  # Update learning rate schedule
322
323
324
325
326
                model.zero_grad()
                global_step += 1

                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
                    # Log metrics
327
328
329
                    if (
                        args.local_rank == -1 and args.evaluate_during_training
                    ):  # Only evaluate when single GPU otherwise metrics may not average well
330
331
                        results = evaluate(args, model, tokenizer)
                        for key, value in results.items():
332
333
334
                            tb_writer.add_scalar("eval_{}".format(key), value, global_step)
                    tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
                    tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
335
336
337
                    logging_loss = tr_loss

                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
338
                    checkpoint_prefix = "checkpoint"
339
                    # Save model checkpoint
340
                    output_dir = os.path.join(args.output_dir, "{}-{}".format(checkpoint_prefix, global_step))
341
342
                    if not os.path.exists(output_dir):
                        os.makedirs(output_dir)
343
344
345
                    model_to_save = (
                        model.module if hasattr(model, "module") else model
                    )  # Take care of distributed/parallel training
346
                    model_to_save.save_pretrained(output_dir)
347
348
                    tokenizer.save_pretrained(output_dir)

349
                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
350
351
                    logger.info("Saving model checkpoint to %s", output_dir)

352
                    _rotate_checkpoints(args, checkpoint_prefix)
jinoobaek-qz's avatar
jinoobaek-qz committed
353

354
355
                    torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
                    torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
Bilal Khan's avatar
Bilal Khan committed
356
                    logger.info("Saving optimizer and scheduler states to %s", output_dir)
357

358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
            if args.max_steps > 0 and global_step > args.max_steps:
                epoch_iterator.close()
                break
        if args.max_steps > 0 and global_step > args.max_steps:
            train_iterator.close()
            break

    if args.local_rank in [-1, 0]:
        tb_writer.close()

    return global_step, tr_loss / global_step


def evaluate(args, model, tokenizer, prefix=""):
    # Loop to handle MNLI double evaluation (matched, mis-matched)
    eval_output_dir = args.output_dir

    eval_dataset = load_and_cache_examples(args, tokenizer, evaluate=True)

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

    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
    # Note that DistributedSampler samples randomly
382
    eval_sampler = SequentialSampler(eval_dataset)
383
    eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
384

ronakice's avatar
ronakice committed
385
386
387
388
    # multi-gpu evaluate
    if args.n_gpu > 1:
        model = torch.nn.DataParallel(model)

389
390
391
392
393
394
    # Eval!
    logger.info("***** Running evaluation {} *****".format(prefix))
    logger.info("  Num examples = %d", len(eval_dataset))
    logger.info("  Batch size = %d", args.eval_batch_size)
    eval_loss = 0.0
    nb_eval_steps = 0
395
396
    model.eval()

397
    for batch in tqdm(eval_dataloader, desc="Evaluating"):
altsoph's avatar
altsoph committed
398
399
400
        inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
        inputs = inputs.to(args.device)
        labels = labels.to(args.device)
401
402

        with torch.no_grad():
altsoph's avatar
altsoph committed
403
            outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
404
405
406
407
408
409
410
            lm_loss = outputs[0]
            eval_loss += lm_loss.mean().item()
        nb_eval_steps += 1

    eval_loss = eval_loss / nb_eval_steps
    perplexity = torch.exp(torch.tensor(eval_loss))

411
    result = {"perplexity": perplexity}
412

413
    output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
414
415
416
417
418
419
    with open(output_eval_file, "w") as writer:
        logger.info("***** Eval results {} *****".format(prefix))
        for key in sorted(result.keys()):
            logger.info("  %s = %s", key, str(result[key]))
            writer.write("%s = %s\n" % (key, str(result[key])))

420
    return result
421
422
423
424
425


def main():
    parser = argparse.ArgumentParser()

426
    # Required parameters
427
428
429
430
431
432
433
434
435
436
    parser.add_argument(
        "--train_data_file", default=None, type=str, required=True, help="The input training data file (a text file)."
    )
    parser.add_argument(
        "--output_dir",
        default=None,
        type=str,
        required=True,
        help="The output directory where the model predictions and checkpoints will be written.",
    )
437

438
    # Other parameters
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
    parser.add_argument(
        "--eval_data_file",
        default=None,
        type=str,
        help="An optional input evaluation data file to evaluate the perplexity on (a text file).",
    )

    parser.add_argument("--model_type", default="bert", type=str, help="The model architecture to be fine-tuned.")
    parser.add_argument(
        "--model_name_or_path",
        default="bert-base-cased",
        type=str,
        help="The model checkpoint for weights initialization.",
    )

    parser.add_argument(
        "--mlm", action="store_true", help="Train with masked-language modeling loss instead of language modeling."
    )
    parser.add_argument(
        "--mlm_probability", type=float, default=0.15, help="Ratio of tokens to mask for masked language modeling loss"
    )

    parser.add_argument(
        "--config_name",
        default="",
        type=str,
        help="Optional pretrained config name or path if not the same as model_name_or_path",
    )
    parser.add_argument(
        "--tokenizer_name",
        default="",
        type=str,
        help="Optional pretrained tokenizer name or path if not the same as model_name_or_path",
    )
    parser.add_argument(
        "--cache_dir",
        default="",
        type=str,
        help="Optional directory to store the pre-trained models downloaded from s3 (instread of the default one)",
    )
    parser.add_argument(
        "--block_size",
        default=-1,
        type=int,
        help="Optional input sequence length after tokenization."
        "The training dataset will be truncated in block of this size for training."
        "Default to the model max input length for single sentence inputs (take into account special tokens).",
    )
    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(
        "--evaluate_during_training", action="store_true", help="Run evaluation during training at each logging step."
    )
    parser.add_argument(
        "--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
    )

    parser.add_argument("--per_gpu_train_batch_size", default=4, type=int, help="Batch size per GPU/CPU for training.")
    parser.add_argument(
        "--per_gpu_eval_batch_size", default=4, type=int, help="Batch size per GPU/CPU for evaluation."
    )
    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("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
    parser.add_argument(
        "--num_train_epochs", default=1.0, type=float, help="Total number of training epochs to perform."
    )
    parser.add_argument(
        "--max_steps",
        default=-1,
        type=int,
        help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
    )
    parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")

    parser.add_argument("--logging_steps", type=int, default=50, help="Log every X updates steps.")
    parser.add_argument("--save_steps", type=int, default=50, help="Save checkpoint every X updates steps.")
    parser.add_argument(
        "--save_total_limit",
        type=int,
        default=None,
        help="Limit the total amount of checkpoints, delete the older checkpoints in the output_dir, does not delete by default",
    )
    parser.add_argument(
        "--eval_all_checkpoints",
        action="store_true",
        help="Evaluate all checkpoints starting with the same prefix as model_name_or_path ending and ending with step number",
    )
    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(
        "--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
    )
    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 (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",
    )
    parser.add_argument("--local_rank", type=int, default=-1, 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.")
558
559
    args = parser.parse_args()

maxvidal's avatar
maxvidal committed
560
    if args.model_type in ["bert", "roberta", "distilbert", "camembert"] and not args.mlm:
561
562
563
564
        raise ValueError(
            "BERT and RoBERTa do not have LM heads but masked LM heads. They must be run using the --mlm "
            "flag (masked language modeling)."
        )
565
    if args.eval_data_file is None and args.do_eval:
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
        raise ValueError(
            "Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
            "or remove the --do_eval argument."
        )

    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
            )
        )
582
583
584
585
586

    # 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
587

588
589
590
591
592
593
594
595
596
597
598
        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")
        args.n_gpu = torch.cuda.device_count()
    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)
599
        torch.distributed.init_process_group(backend="nccl")
600
601
602
603
        args.n_gpu = 1
    args.device = device

    # Setup logging
604
605
606
607
608
609
610
611
612
613
614
615
616
    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,
    )
    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,
    )
617
618
619
620
621
622

    # Set seed
    set_seed(args)

    # Load pretrained model and tokenizer
    if args.local_rank not in [-1, 0]:
623
624
625
        torch.distributed.barrier()  # Barrier to make sure only the first process in distributed training download model & vocab

    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
626
627
628
629
630
631
632
633
634
    config = config_class.from_pretrained(
        args.config_name if args.config_name else args.model_name_or_path,
        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,
    )
635
    if args.block_size <= 0:
636
637
638
        args.block_size = (
            tokenizer.max_len_single_sentence
        )  # Our input block size will be the max possible for the model
thomwolf's avatar
thomwolf committed
639
    args.block_size = min(args.block_size, tokenizer.max_len_single_sentence)
640
641
642
643
644
645
    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,
    )
646
    model.to(args.device)
647
648

    if args.local_rank == 0:
649
        torch.distributed.barrier()  # End of barrier to make sure only the first process in distributed training download model & vocab
650
651
652
653
654

    logger.info("Training/evaluation parameters %s", args)

    # Training
    if args.do_train:
655
656
657
        if args.local_rank not in [-1, 0]:
            torch.distributed.barrier()  # Barrier to make sure only the first process in distributed training process the dataset, and the others will use the cache

658
        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False)
659
660
661
662

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

663
664
665
        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)

666
    # Saving best-practices: if you use save_pretrained for the model and tokenizer, you can reload them using from_pretrained()
667
668
669
670
671
672
673
674
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
        # 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)

        logger.info("Saving model checkpoint to %s", args.output_dir)
        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
        # They can then be reloaded using `from_pretrained()`
675
676
677
        model_to_save = (
            model.module if hasattr(model, "module") else model
        )  # Take care of distributed/parallel training
678
679
680
681
        model_to_save.save_pretrained(args.output_dir)
        tokenizer.save_pretrained(args.output_dir)

        # Good practice: save your training arguments together with the trained model
682
        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
683
684
685

        # Load a trained model and vocabulary that you have fine-tuned
        model = model_class.from_pretrained(args.output_dir)
686
        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
687
688
689
690
691
692
693
        model.to(args.device)

    # Evaluation
    results = {}
    if args.do_eval and args.local_rank in [-1, 0]:
        checkpoints = [args.output_dir]
        if args.eval_all_checkpoints:
694
695
696
            checkpoints = list(
                os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
            )
697
            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
698
699
        logger.info("Evaluate the following checkpoints: %s", checkpoints)
        for checkpoint in checkpoints:
700
701
702
            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
            prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""

703
704
            model = model_class.from_pretrained(checkpoint)
            model.to(args.device)
705
            result = evaluate(args, model, tokenizer, prefix=prefix)
706
            result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
707
708
709
710
711
712
            results.update(result)

    return results


if __name__ == "__main__":
altsoph's avatar
altsoph committed
713
    main()