run_lm_finetuning.py 27 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
27

from __future__ import absolute_import, division, print_function

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

import numpy as np
import torch
thomwolf's avatar
thomwolf committed
35
from torch.utils.data import DataLoader, Dataset, SequentialSampler, RandomSampler
36
37
38
39
from torch.utils.data.distributed import DistributedSampler
from tensorboardX import SummaryWriter
from tqdm import tqdm, trange

40
from transformers import (WEIGHTS_NAME, AdamW, WarmupLinearSchedule,
41
42
43
                                  BertConfig, BertForMaskedLM, BertTokenizer,
                                  GPT2Config, GPT2LMHeadModel, GPT2Tokenizer,
                                  OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer,
44
45
                                  RobertaConfig, RobertaForMaskedLM, RobertaTokenizer,
                                  DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer)
46

47

48
logger = logging.getLogger(__name__)
49
50
51


MODEL_CLASSES = {
52
    'gpt2': (GPT2Config, GPT2LMHeadModel, GPT2Tokenizer),
53
    'openai-gpt': (OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
54
    'bert': (BertConfig, BertForMaskedLM, BertTokenizer),
55
56
    'roberta': (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer),
    'distilbert': (DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer)
57
58
59
}


60
61
62
63
class TextDataset(Dataset):
    def __init__(self, tokenizer, file_path='train', block_size=512):
        assert os.path.isfile(file_path)
        directory, filename = os.path.split(file_path)
thomwolf's avatar
thomwolf committed
64
        cached_features_file = os.path.join(directory, 'cached_lm_{}_{}'.format(block_size, filename))
65
66
67
68
69
70
71
72
73
74
75
76
77

        if os.path.exists(cached_features_file):
            logger.info("Loading features from cached file %s", cached_features_file)
            with open(cached_features_file, 'rb') as handle:
                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))
78

mgrankin's avatar
mgrankin committed
79
            for i in range(0, len(tokenized_text)-block_size+1, block_size): # Truncate in block of block_size
Denny's avatar
Denny committed
80
                self.examples.append(tokenizer.add_special_tokens_single_sequence(tokenized_text[i:i+block_size]))
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
            # 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)
            with open(cached_features_file, 'wb') as handle:
                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):
    dataset = TextDataset(tokenizer, file_path=args.eval_data_file if evaluate else args.train_data_file, block_size=args.block_size)
    return dataset


101
102
103
104
105
106
107
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)

108

109
def mask_tokens(inputs, tokenizer, args):
110
    """ Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. """
111
    labels = inputs.clone()
112
    # We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
thomwolf's avatar
thomwolf committed
113
    masked_indices = torch.bernoulli(torch.full(labels.shape, args.mlm_probability)).bool()
114
115
116
    labels[~masked_indices] = -1  # We only compute loss on masked tokens

    # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
thomwolf's avatar
thomwolf committed
117
    indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).bool() & masked_indices
118
119
120
    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
121
    indices_random = torch.bernoulli(torch.full(labels.shape, 0.5)).bool() & masked_indices & ~indices_replaced
122
123
    random_words = torch.randint(len(tokenizer), labels.shape, dtype=torch.long)
    inputs[indices_random] = random_words[indices_random]
124

125
    # The rest of the time (10% of the time) we keep the masked input tokens unchanged
126
    return inputs, labels
127

128

129
130
131
132
133
134
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
135
    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
136
    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182

    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)
    no_decay = ['bias', 'LayerNorm.weight']
    optimizer_grouped_parameters = [
        {'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}
        ]
    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)
    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:
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
                                                          output_device=args.local_rank,
                                                          find_unused_parameters=True)

    # 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)
    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))
    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", t_total)

    global_step = 0
    tr_loss, logging_loss = 0.0, 0.0
    model.zero_grad()
    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
183
    set_seed(args)  # Added here for reproducibility (even between python 2 and 3)
184
185
186
    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):
187
            inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
188
189
190
            inputs = inputs.to(args.device)
            labels = labels.to(args.device)
            model.train()
191
            outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
192
            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
193
194

            if args.n_gpu > 1:
195
                loss = loss.mean()  # mean() to average on multi-gpu parallel training
196
197
198
199
200
201
202
203
204
205
206
            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:
207
208
209
210
                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)
211
                optimizer.step()
212
                scheduler.step()  # Update learning rate schedule
213
214
215
216
217
218
219
220
221
222
223
224
225
226
                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
                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
                        results = evaluate(args, model, tokenizer)
                        for key, value in results.items():
                            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)
                    logging_loss = tr_loss

                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
227
228
229
230
231
232
233
234
235
                    # 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'))
                    logger.info("Saving model checkpoint to %s", output_dir)

jinoobaek-qz's avatar
jinoobaek-qz committed
236
237
238
                    if args.save_total_limit and args.save_total_limit > 0:
                        # Check if we should delete older checkpoint(s)
                        glob_checkpoints = glob.glob(os.path.join(args.output_dir, 'checkpoint-*'))
239
                        if len(glob_checkpoints) > args.save_total_limit:
jinoobaek-qz's avatar
jinoobaek-qz committed
240
241
242
243
244
245
246
247
                            checkpoints_sorted = []
                            for path in glob_checkpoints:
                                regex_match = re.match('.*checkpoint-([0-9]+)', path)
                                if regex_match and regex_match.groups():
                                    checkpoints_sorted.append((int(regex_match.groups()[0]), path))

                            checkpoints_sorted = sorted(checkpoints_sorted)
                            checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted]
248
                            number_of_checkpoints_to_delete = max(0, len(checkpoints_sorted) - args.save_total_limit)
jinoobaek-qz's avatar
jinoobaek-qz committed
249
250
251
252
253
                            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)

254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
            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
    eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
279
    eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
280
281
282
283
284
285
286

    # 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
287
288
    model.eval()

289
    for batch in tqdm(eval_dataloader, desc="Evaluating"):
290
        batch = batch.to(args.device)
291
292

        with torch.no_grad():
293
            outputs = model(batch, masked_lm_labels=batch) if args.mlm else model(batch, labels=batch)
294
295
296
297
298
299
300
301
302
303
304
            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))

    result = {
        "perplexity": perplexity
    }

305
    output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
306
307
308
309
310
311
    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])))

312
    return result
313
314
315
316
317
318


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

    ## Required parameters
319
320
    parser.add_argument("--train_data_file", default=None, type=str, required=True,
                        help="The input training data file (a text file).")
321
322
323
324
    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
325
326
327
328
    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,
329
                        help="The model architecture to be fine-tuned.")
330
    parser.add_argument("--model_name_or_path", default="bert-base-cased", type=str,
331
332
333
334
335
336
337
                        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")

338
    parser.add_argument("--config_name", default="", type=str,
339
                        help="Optional pretrained config name or path if not the same as model_name_or_path")
340
    parser.add_argument("--tokenizer_name", default="", type=str,
341
                        help="Optional pretrained tokenizer name or path if not the same as model_name_or_path")
342
    parser.add_argument("--cache_dir", default="", type=str,
343
344
345
346
                        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."
thomwolf's avatar
typo  
thomwolf committed
347
                             "Default to the model max input length for single sentence inputs (take into account special tokens).")
348
349
350
351
352
    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',
353
                        help="Run evaluation during training at each logging step.")
354
355
356
    parser.add_argument("--do_lower_case", action='store_true',
                        help="Set this flag if you are using an uncased model.")

357
    parser.add_argument("--per_gpu_train_batch_size", default=4, type=int,
358
                        help="Batch size per GPU/CPU for training.")
359
    parser.add_argument("--per_gpu_eval_batch_size", default=4, type=int,
360
361
362
363
364
365
366
367
368
369
370
                        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 deay 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.")
371
    parser.add_argument("--num_train_epochs", default=1.0, type=float,
372
373
374
375
376
377
378
379
380
381
                        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.")
jinoobaek-qz's avatar
jinoobaek-qz committed
382
383
    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')
384
    parser.add_argument("--eval_all_checkpoints", action='store_true',
385
                        help="Evaluate all checkpoints starting with the same prefix as model_name_or_path ending and ending with step number")
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
    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.")
    args = parser.parse_args()

406
    if args.model_type in ["bert", "roberta", "distilbert"] and not args.mlm:
407
408
        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).")
409
410
411
    if args.eval_data_file is None and args.do_eval:
        raise ValueError("Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
                         "or remove the --do_eval argument.")
412

413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
    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))

    # 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")
        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)
        torch.distributed.init_process_group(backend='nccl')
        args.n_gpu = 1
    args.device = device

    # Setup logging
    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)

    # Set seed
    set_seed(args)

    # Load pretrained model and tokenizer
    if args.local_rank not in [-1, 0]:
447
448
449
450
451
452
        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]
    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path)
    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)
    if args.block_size <= 0:
thomwolf's avatar
thomwolf committed
453
454
        args.block_size = tokenizer.max_len_single_sentence  # Our input block size will be the max possible for the model
    args.block_size = min(args.block_size, tokenizer.max_len_single_sentence)
455
456
    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
    model.to(args.device)
457
458

    if args.local_rank == 0:
459
        torch.distributed.barrier()  # End of barrier to make sure only the first process in distributed training download model & vocab
460
461
462
463
464

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

    # Training
    if args.do_train:
465
466
467
        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

468
        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False)
469
470
471
472

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

473
474
475
476
        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)


477
    # Saving best-practices: if you use save_pretrained for the model and tokenizer, you can reload them using from_pretrained()
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
    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()`
        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)
        tokenizer.save_pretrained(args.output_dir)

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

        # Load a trained model and vocabulary that you have fine-tuned
        model = model_class.from_pretrained(args.output_dir)
495
        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
496
497
498
499
500
501
502
503
504
        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:
            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
505
            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
506
507
508
        logger.info("Evaluate the following checkpoints: %s", checkpoints)
        for checkpoint in checkpoints:
            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
509
510
            prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
            
511
512
            model = model_class.from_pretrained(checkpoint)
            model.to(args.device)
513
            result = evaluate(args, model, tokenizer, prefix=prefix)
514
515
516
517
518
519
520
521
            result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
            results.update(result)

    return results


if __name__ == "__main__":
    main()