run_summarization.py 30.3 KB
Newer Older
1
#!/usr/bin/env python
2
# coding=utf-8
3
# Copyright 2021 The HuggingFace Team. All rights reserved.
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
#
# 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.
"""
Fine-tuning the library models for sequence to sequence.
"""
# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.

import logging
import os
import sys
from dataclasses import dataclass, field
from typing import Optional

27
import datasets
28
import nltk  # Here to have a nice missing dependency error message early on
29
30
31
32
import numpy as np
from datasets import load_dataset, load_metric

import transformers
33
from filelock import FileLock
34
35
36
37
38
39
from transformers import (
    AutoConfig,
    AutoModelForSeq2SeqLM,
    AutoTokenizer,
    DataCollatorForSeq2Seq,
    HfArgumentParser,
40
41
42
43
    MBart50Tokenizer,
    MBart50TokenizerFast,
    MBartTokenizer,
    MBartTokenizerFast,
44
45
46
47
    Seq2SeqTrainer,
    Seq2SeqTrainingArguments,
    set_seed,
)
48
from transformers.trainer_utils import get_last_checkpoint
49
from transformers.utils import check_min_version, is_offline_mode
50
from transformers.utils.versions import require_version
51
52


53
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
Lysandre Debut's avatar
Lysandre Debut committed
54
check_min_version("4.20.0.dev0")
Sylvain Gugger's avatar
Sylvain Gugger committed
55

56
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/summarization/requirements.txt")
57

58
59
logger = logging.getLogger(__name__)

60
61
try:
    nltk.data.find("tokenizers/punkt")
Stas Bekman's avatar
Stas Bekman committed
62
except (LookupError, OSError):
63
64
65
66
67
68
69
    if is_offline_mode():
        raise LookupError(
            "Offline mode: run this script without TRANSFORMERS_OFFLINE first to download nltk data files"
        )
    with FileLock(".lock") as lock:
        nltk.download("punkt", quiet=True)

70
71
72
# A list of all multilingual tokenizer which require lang attribute.
MULTILINGUAL_TOKENIZERS = [MBartTokenizer, MBartTokenizerFast, MBart50Tokenizer, MBart50TokenizerFast]

73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103

@dataclass
class ModelArguments:
    """
    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
    """

    model_name_or_path: str = field(
        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
    )
    config_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
    )
    tokenizer_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
    )
    cache_dir: Optional[str] = field(
        default=None,
        metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"},
    )
    use_fast_tokenizer: bool = field(
        default=True,
        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
    )
    model_revision: str = field(
        default="main",
        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
    )
    use_auth_token: bool = field(
        default=False,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
104
105
106
107
            "help": (
                "Will use the token generated when running `transformers-cli login` (necessary to use this script "
                "with private models)."
            )
108
109
        },
    )
110
111
112
    resize_position_embeddings: Optional[bool] = field(
        default=None,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
113
114
115
116
            "help": (
                "Whether to automatically resize the position embeddings if `max_source_length` exceeds "
                "the model's position embeddings."
            )
117
118
        },
    )
119
120
121
122
123
124
125
126


@dataclass
class DataTrainingArguments:
    """
    Arguments pertaining to what data we are going to input our model for training and eval.
    """

127
128
    lang: str = field(default=None, metadata={"help": "Language id for summarization."})

129
130
131
132
133
134
135
136
137
138
139
140
141
142
    dataset_name: Optional[str] = field(
        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
    )
    dataset_config_name: Optional[str] = field(
        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
    )
    text_column: Optional[str] = field(
        default=None,
        metadata={"help": "The name of the column in the datasets containing the full texts (for summarization)."},
    )
    summary_column: Optional[str] = field(
        default=None,
        metadata={"help": "The name of the column in the datasets containing the summaries (for summarization)."},
    )
143
144
145
    train_file: Optional[str] = field(
        default=None, metadata={"help": "The input training data file (a jsonlines or csv file)."}
    )
146
147
    validation_file: Optional[str] = field(
        default=None,
148
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
149
150
151
            "help": (
                "An optional input evaluation data file to evaluate the metrics (rouge) on (a jsonlines or csv file)."
            )
152
153
154
155
156
        },
    )
    test_file: Optional[str] = field(
        default=None,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
157
            "help": "An optional input test data file to evaluate the metrics (rouge) on (a jsonlines or csv file)."
158
        },
159
160
161
162
163
164
165
166
167
168
169
    )
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
    )
    max_source_length: Optional[int] = field(
        default=1024,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
170
171
172
173
            "help": (
                "The maximum total input sequence length after tokenization. Sequences longer "
                "than this will be truncated, sequences shorter will be padded."
            )
174
175
176
177
178
        },
    )
    max_target_length: Optional[int] = field(
        default=128,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
179
180
181
182
            "help": (
                "The maximum total sequence length for target text after tokenization. Sequences longer "
                "than this will be truncated, sequences shorter will be padded."
            )
183
184
185
        },
    )
    val_max_target_length: Optional[int] = field(
186
        default=None,
187
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
188
189
190
191
192
193
            "help": (
                "The maximum total sequence length for validation target text after tokenization. Sequences longer "
                "than this will be truncated, sequences shorter will be padded. Will default to `max_target_length`."
                "This argument is also used to override the ``max_length`` param of ``model.generate``, which is used "
                "during ``evaluate`` and ``predict``."
            )
194
195
196
197
198
        },
    )
    pad_to_max_length: bool = field(
        default=False,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
199
200
201
202
203
            "help": (
                "Whether to pad all samples to model maximum sentence length. "
                "If False, will pad the samples dynamically when batching to the maximum length in the batch. More "
                "efficient on GPU but very bad for TPU."
            )
204
205
206
207
208
        },
    )
    max_train_samples: Optional[int] = field(
        default=None,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
209
210
211
212
            "help": (
                "For debugging purposes or quicker training, truncate the number of training examples to this "
                "value if set."
            )
213
214
        },
    )
215
    max_eval_samples: Optional[int] = field(
216
217
        default=None,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
218
219
220
221
            "help": (
                "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
                "value if set."
            )
222
223
        },
    )
224
    max_predict_samples: Optional[int] = field(
225
226
        default=None,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
227
228
229
230
            "help": (
                "For debugging purposes or quicker training, truncate the number of prediction examples to this "
                "value if set."
            )
231
232
233
234
235
        },
    )
    num_beams: Optional[int] = field(
        default=None,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
236
237
238
239
            "help": (
                "Number of beams to use for evaluation. This argument will be passed to ``model.generate``, "
                "which is used during ``evaluate`` and ``predict``."
            )
240
241
        },
    )
242
243
244
245
246
247
    ignore_pad_token_for_loss: bool = field(
        default=True,
        metadata={
            "help": "Whether to ignore the tokens corresponding to padded labels in the loss computation or not."
        },
    )
248
    source_prefix: Optional[str] = field(
249
250
251
252
253
254
        default="", metadata={"help": "A prefix to add before every source text (useful for T5 models)."}
    )

    forced_bos_token: Optional[str] = field(
        default=None,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
255
256
257
258
259
            "help": (
                "The token to force as the first generated token after the decoder_start_token_id."
                "Useful for multilingual models like mBART where the first generated token"
                "needs to be the target language token (Usually it is the target language token)"
            )
260
        },
261
    )
262
263
264
265
266
267
268
269
270
271
272

    def __post_init__(self):
        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
            raise ValueError("Need either a dataset name or a training/validation file.")
        else:
            if self.train_file is not None:
                extension = self.train_file.split(".")[-1]
                assert extension in ["csv", "json"], "`train_file` should be a csv or a json file."
            if self.validation_file is not None:
                extension = self.validation_file.split(".")[-1]
                assert extension in ["csv", "json"], "`validation_file` should be a csv or a json file."
273
274
        if self.val_max_target_length is None:
            self.val_max_target_length = self.max_target_length
275
276
277


summarization_name_mapping = {
278
279
    "amazon_reviews_multi": ("review_body", "review_title"),
    "big_patent": ("description", "abstract"),
280
    "cnn_dailymail": ("article", "highlights"),
281
282
283
284
285
286
    "orange_sum": ("text", "summary"),
    "pn_summary": ("article", "summary"),
    "psc": ("extract_text", "summary_text"),
    "samsum": ("dialogue", "summary"),
    "thaisum": ("body", "summary"),
    "xglue": ("news_body", "news_title"),
287
    "xsum": ("document", "summary"),
288
    "wiki_summary": ("article", "highlights"),
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
}


def main():
    # See all possible arguments in src/transformers/training_args.py
    # or by passing the --help flag to this script.
    # We now keep distinct sets of args, for a cleaner separation of concerns.

    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
        # If we pass only one argument to the script and it's the path to a json file,
        # let's parse it to get our arguments.
        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
    else:
        model_args, data_args, training_args = parser.parse_args_into_dataclasses()

305
306
    # Setup logging
    logging.basicConfig(
307
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
308
309
310
        datefmt="%m/%d/%Y %H:%M:%S",
        handlers=[logging.StreamHandler(sys.stdout)],
    )
311
312
313
314
315
316
    log_level = training_args.get_process_log_level()
    logger.setLevel(log_level)
    datasets.utils.logging.set_verbosity(log_level)
    transformers.utils.logging.set_verbosity(log_level)
    transformers.utils.logging.enable_default_handler()
    transformers.utils.logging.enable_explicit_format()
317
318
319
320
321
322
323
324

    # Log on each process the small summary:
    logger.warning(
        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
    )
    logger.info(f"Training/evaluation parameters {training_args}")

325
326
327
328
329
330
331
332
333
334
335
336
    if data_args.source_prefix is None and model_args.model_name_or_path in [
        "t5-small",
        "t5-base",
        "t5-large",
        "t5-3b",
        "t5-11b",
    ]:
        logger.warning(
            "You're running a t5 model but didn't provide a source prefix, which is the expected, e.g. with "
            "`--source_prefix 'summarize: ' `"
        )

337
338
339
340
341
342
343
344
345
    # Detecting last checkpoint.
    last_checkpoint = None
    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
        last_checkpoint = get_last_checkpoint(training_args.output_dir)
        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
            raise ValueError(
                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                "Use --overwrite_output_dir to overcome."
            )
346
        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
347
348
349
350
            logger.info(
                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
            )
351
352
353
354
355
356
357
358

    # Set seed before initializing model.
    set_seed(training_args.seed)

    # Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub).
    #
359
360
    # For CSV/JSON files this script will use the first column for the full texts and the second column for the
    # summaries (unless you specify column names for this with the `text_column` and `summary_column` arguments).
361
362
363
364
365
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
366
        raw_datasets = load_dataset(
367
368
369
370
            data_args.dataset_name,
            data_args.dataset_config_name,
            cache_dir=model_args.cache_dir,
            use_auth_token=True if model_args.use_auth_token else None,
371
        )
372
373
374
375
376
377
378
379
    else:
        data_files = {}
        if data_args.train_file is not None:
            data_files["train"] = data_args.train_file
            extension = data_args.train_file.split(".")[-1]
        if data_args.validation_file is not None:
            data_files["validation"] = data_args.validation_file
            extension = data_args.validation_file.split(".")[-1]
380
381
382
        if data_args.test_file is not None:
            data_files["test"] = data_args.test_file
            extension = data_args.test_file.split(".")[-1]
383
384
385
386
387
388
        raw_datasets = load_dataset(
            extension,
            data_files=data_files,
            cache_dir=model_args.cache_dir,
            use_auth_token=True if model_args.use_auth_token else None,
        )
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
    config = AutoConfig.from_pretrained(
        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
    )
    tokenizer = AutoTokenizer.from_pretrained(
        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        use_fast=model_args.use_fast_tokenizer,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
    )
    model = AutoModelForSeq2SeqLM.from_pretrained(
        model_args.model_name_or_path,
        from_tf=bool(".ckpt" in model_args.model_name_or_path),
        config=config,
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
    )

Suraj Patil's avatar
Suraj Patil committed
419
420
    model.resize_token_embeddings(len(tokenizer))

421
422
423
424
425
426
    if model.config.decoder_start_token_id is None and isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):
        if isinstance(tokenizer, MBartTokenizer):
            model.config.decoder_start_token_id = tokenizer.lang_code_to_id[data_args.lang]
        else:
            model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(data_args.lang)

427
428
429
    if model.config.decoder_start_token_id is None:
        raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")

430
431
432
433
434
435
    if (
        hasattr(model.config, "max_position_embeddings")
        and model.config.max_position_embeddings < data_args.max_source_length
    ):
        if model_args.resize_position_embeddings is None:
            logger.warning(
Sylvain Gugger's avatar
Sylvain Gugger committed
436
437
                "Increasing the model's number of position embedding vectors from"
                f" {model.config.max_position_embeddings} to {data_args.max_source_length}."
438
439
440
441
442
443
            )
            model.resize_position_embeddings(data_args.max_source_length)
        elif model_args.resize_position_embeddings:
            model.resize_position_embeddings(data_args.max_source_length)
        else:
            raise ValueError(
Sylvain Gugger's avatar
Sylvain Gugger committed
444
445
446
447
                f"`--max_source_length` is set to {data_args.max_source_length}, but the model only has"
                f" {model.config.max_position_embeddings} position encodings. Consider either reducing"
                f" `--max_source_length` to {model.config.max_position_embeddings} or to automatically resize the"
                " model's position encodings by passing `--resize_position_embeddings`."
448
449
            )

450
    prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
451

452
453
454
    # Preprocessing the datasets.
    # We need to tokenize inputs and targets.
    if training_args.do_train:
455
        column_names = raw_datasets["train"].column_names
456
    elif training_args.do_eval:
457
        column_names = raw_datasets["validation"].column_names
458
    elif training_args.do_predict:
459
        column_names = raw_datasets["test"].column_names
460
461
462
    else:
        logger.info("There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`.")
        return
463

464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
    if isinstance(tokenizer, tuple(MULTILINGUAL_TOKENIZERS)):
        assert (
            data_args.lang is not None
        ), f"{tokenizer.__class__.__name__} is a multilingual tokenizer which requires --lang argument"

        tokenizer.src_lang = data_args.lang
        tokenizer.tgt_lang = data_args.lang

        # For multilingual translation models like mBART-50 and M2M100 we need to force the target language token
        # as the first generated token. We ask the user to explicitly provide this as --forced_bos_token argument.
        forced_bos_token_id = (
            tokenizer.lang_code_to_id[data_args.forced_bos_token] if data_args.forced_bos_token is not None else None
        )
        model.config.forced_bos_token_id = forced_bos_token_id

479
480
481
482
    # Get the column names for input/target.
    dataset_columns = summarization_name_mapping.get(data_args.dataset_name, None)
    if data_args.text_column is None:
        text_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
483
    else:
484
485
486
487
488
489
490
491
492
493
494
495
496
        text_column = data_args.text_column
        if text_column not in column_names:
            raise ValueError(
                f"--text_column' value '{data_args.text_column}' needs to be one of: {', '.join(column_names)}"
            )
    if data_args.summary_column is None:
        summary_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
    else:
        summary_column = data_args.summary_column
        if summary_column not in column_names:
            raise ValueError(
                f"--summary_column' value '{data_args.summary_column}' needs to be one of: {', '.join(column_names)}"
            )
497
498
499
500
501

    # Temporarily set max_target_length for training.
    max_target_length = data_args.max_target_length
    padding = "max_length" if data_args.pad_to_max_length else False

502
    if training_args.label_smoothing_factor > 0 and not hasattr(model, "prepare_decoder_input_ids_from_labels"):
503
        logger.warning(
504
505
506
507
            "label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
            f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
        )

508
    def preprocess_function(examples):
509
        # remove pairs where at least one record is None
510

511
512
513
514
515
516
        inputs, targets = [], []
        for i in range(len(examples[text_column])):
            if examples[text_column][i] is not None and examples[summary_column][i] is not None:
                inputs.append(examples[text_column][i])
                targets.append(examples[summary_column][i])

517
        inputs = [prefix + inp for inp in inputs]
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
        model_inputs = tokenizer(inputs, max_length=data_args.max_source_length, padding=padding, truncation=True)

        # Setup the tokenizer for targets
        with tokenizer.as_target_tokenizer():
            labels = tokenizer(targets, max_length=max_target_length, padding=padding, truncation=True)

        # If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
        # padding in the loss.
        if padding == "max_length" and data_args.ignore_pad_token_for_loss:
            labels["input_ids"] = [
                [(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
            ]

        model_inputs["labels"] = labels["input_ids"]
        return model_inputs

    if training_args.do_train:
535
        if "train" not in raw_datasets:
536
            raise ValueError("--do_train requires a train dataset")
537
        train_dataset = raw_datasets["train"]
538
        if data_args.max_train_samples is not None:
539
540
            max_train_samples = min(len(train_dataset), data_args.max_train_samples)
            train_dataset = train_dataset.select(range(max_train_samples))
541
542
543
544
545
546
547
548
549
        with training_args.main_process_first(desc="train dataset map pre-processing"):
            train_dataset = train_dataset.map(
                preprocess_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on train dataset",
            )
550
551
552

    if training_args.do_eval:
        max_target_length = data_args.val_max_target_length
553
        if "validation" not in raw_datasets:
554
            raise ValueError("--do_eval requires a validation dataset")
555
        eval_dataset = raw_datasets["validation"]
556
        if data_args.max_eval_samples is not None:
557
558
            max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
            eval_dataset = eval_dataset.select(range(max_eval_samples))
559
560
561
562
563
564
565
566
567
        with training_args.main_process_first(desc="validation dataset map pre-processing"):
            eval_dataset = eval_dataset.map(
                preprocess_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on validation dataset",
            )
568

569
570
    if training_args.do_predict:
        max_target_length = data_args.val_max_target_length
571
        if "test" not in raw_datasets:
572
            raise ValueError("--do_predict requires a test dataset")
573
        predict_dataset = raw_datasets["test"]
574
        if data_args.max_predict_samples is not None:
575
576
            max_predict_samples = min(len(predict_dataset), data_args.max_predict_samples)
            predict_dataset = predict_dataset.select(range(max_predict_samples))
577
578
579
580
581
582
583
584
585
        with training_args.main_process_first(desc="prediction dataset map pre-processing"):
            predict_dataset = predict_dataset.map(
                preprocess_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on prediction dataset",
            )
586

587
588
    # Data collator
    label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
589
590
591
592
593
594
    data_collator = DataCollatorForSeq2Seq(
        tokenizer,
        model=model,
        label_pad_token_id=label_pad_token_id,
        pad_to_multiple_of=8 if training_args.fp16 else None,
    )
595
596

    # Metric
597
    metric = load_metric("rouge")
598

599
600
601
602
603
    def postprocess_text(preds, labels):
        preds = [pred.strip() for pred in preds]
        labels = [label.strip() for label in labels]

        # rougeLSum expects newline after each sentence
604
605
        preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
        labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
606
607
608

        return preds, labels

609
610
611
612
613
614
615
616
617
618
619
    def compute_metrics(eval_preds):
        preds, labels = eval_preds
        if isinstance(preds, tuple):
            preds = preds[0]
        decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
        if data_args.ignore_pad_token_for_loss:
            # Replace -100 in the labels as we can't decode them.
            labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
        decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)

        # Some simple post-processing
620
        decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
621

622
623
624
        result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True)
        # Extract a few results from ROUGE
        result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
625
626
627

        prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
        result["gen_len"] = np.mean(prediction_lens)
628
        result = {k: round(v, 4) for k, v in result.items()}
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
        return result

    # Initialize our Trainer
    trainer = Seq2SeqTrainer(
        model=model,
        args=training_args,
        train_dataset=train_dataset if training_args.do_train else None,
        eval_dataset=eval_dataset if training_args.do_eval else None,
        tokenizer=tokenizer,
        data_collator=data_collator,
        compute_metrics=compute_metrics if training_args.predict_with_generate else None,
    )

    # Training
    if training_args.do_train:
644
645
646
647
        checkpoint = None
        if training_args.resume_from_checkpoint is not None:
            checkpoint = training_args.resume_from_checkpoint
        elif last_checkpoint is not None:
648
649
            checkpoint = last_checkpoint
        train_result = trainer.train(resume_from_checkpoint=checkpoint)
650
651
        trainer.save_model()  # Saves the tokenizer too for easy upload

652
653
654
655
656
        metrics = train_result.metrics
        max_train_samples = (
            data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
        )
        metrics["train_samples"] = min(max_train_samples, len(train_dataset))
657

658
659
660
        trainer.log_metrics("train", metrics)
        trainer.save_metrics("train", metrics)
        trainer.save_state()
661
662

    # Evaluation
663
    results = {}
664
665
666
667
668
669
    max_length = (
        training_args.generation_max_length
        if training_args.generation_max_length is not None
        else data_args.val_max_target_length
    )
    num_beams = data_args.num_beams if data_args.num_beams is not None else training_args.generation_num_beams
670
671
    if training_args.do_eval:
        logger.info("*** Evaluate ***")
672
        metrics = trainer.evaluate(max_length=max_length, num_beams=num_beams, metric_key_prefix="eval")
673
674
        max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(eval_dataset)
        metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
675

676
677
        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)
678

679
    if training_args.do_predict:
680
        logger.info("*** Predict ***")
681

682
        predict_results = trainer.predict(
683
            predict_dataset, metric_key_prefix="predict", max_length=max_length, num_beams=num_beams
684
        )
685
686
687
688
689
        metrics = predict_results.metrics
        max_predict_samples = (
            data_args.max_predict_samples if data_args.max_predict_samples is not None else len(predict_dataset)
        )
        metrics["predict_samples"] = min(max_predict_samples, len(predict_dataset))
690

691
692
        trainer.log_metrics("predict", metrics)
        trainer.save_metrics("predict", metrics)
693

694
        if trainer.is_world_process_zero():
695
            if training_args.predict_with_generate:
696
697
                predictions = tokenizer.batch_decode(
                    predict_results.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
698
                )
699
700
701
702
                predictions = [pred.strip() for pred in predictions]
                output_prediction_file = os.path.join(training_args.output_dir, "generated_predictions.txt")
                with open(output_prediction_file, "w") as writer:
                    writer.write("\n".join(predictions))
703

704
705
706
707
708
709
710
711
    kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "summarization"}
    if data_args.dataset_name is not None:
        kwargs["dataset_tags"] = data_args.dataset_name
        if data_args.dataset_config_name is not None:
            kwargs["dataset_args"] = data_args.dataset_config_name
            kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
        else:
            kwargs["dataset"] = data_args.dataset_name
Sylvain Gugger's avatar
Sylvain Gugger committed
712

713
714
715
    if data_args.lang is not None:
        kwargs["language"] = data_args.lang

716
    if training_args.push_to_hub:
Sylvain Gugger's avatar
Sylvain Gugger committed
717
        trainer.push_to_hub(**kwargs)
718
719
    else:
        trainer.create_model_card(**kwargs)
Sylvain Gugger's avatar
Sylvain Gugger committed
720

721
722
    return results

723
724
725
726
727
728
729
730

def _mp_fn(index):
    # For xla_spawn (TPUs)
    main()


if __name__ == "__main__":
    main()