run_xnli.py 17.2 KB
Newer Older
1
#!/usr/bin/env python
VictorSanh's avatar
VictorSanh committed
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# 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.
17
""" Finetuning multi-lingual models on XNLI (e.g. Bert, DistilBERT, XLM).
18
    Adapted from `examples/text-classification/run_glue.py`"""
VictorSanh's avatar
VictorSanh committed
19
20
21
22

import logging
import os
import random
23
24
25
import sys
from dataclasses import dataclass, field
from typing import Optional
VictorSanh's avatar
VictorSanh committed
26

27
import datasets
VictorSanh's avatar
VictorSanh committed
28
import numpy as np
29
from datasets import load_dataset, load_metric
VictorSanh's avatar
VictorSanh committed
30

31
import transformers
32
from transformers import (
33
34
35
    AutoConfig,
    AutoModelForSequenceClassification,
    AutoTokenizer,
36
37
38
39
40
41
42
    DataCollatorWithPadding,
    EvalPrediction,
    HfArgumentParser,
    Trainer,
    TrainingArguments,
    default_data_collator,
    set_seed,
43
)
44
from transformers.trainer_utils import get_last_checkpoint
45
from transformers.utils import check_min_version
46
from transformers.utils.versions import require_version
Aymeric Augustin's avatar
Aymeric Augustin committed
47

VictorSanh's avatar
VictorSanh committed
48

49
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
Lysandre Debut's avatar
Lysandre Debut committed
50
check_min_version("4.20.0.dev0")
Lysandre's avatar
Lysandre committed
51

52
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
53

VictorSanh's avatar
VictorSanh committed
54
55
56
logger = logging.getLogger(__name__)


57
58
59
60
@dataclass
class DataTrainingArguments:
    """
    Arguments pertaining to what data we are going to input our model for training and eval.
VictorSanh's avatar
VictorSanh committed
61

62
63
64
65
    Using `HfArgumentParser` we can turn this class
    into argparse arguments to be able to specify them on
    the command line.
    """
VictorSanh's avatar
VictorSanh committed
66

67
68
69
    max_seq_length: Optional[int] = field(
        default=128,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
70
71
72
73
            "help": (
                "The maximum total input sequence length after tokenization. Sequences longer "
                "than this will be truncated, sequences shorter will be padded."
            )
74
75
        },
    )
76
77
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
78
    )
79
80
81
    pad_to_max_length: bool = field(
        default=True,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
82
83
84
85
            "help": (
                "Whether to pad all samples to `max_seq_length`. "
                "If False, will pad the samples dynamically when batching to the maximum length in the batch."
            )
86
        },
87
    )
88
89
90
    max_train_samples: Optional[int] = field(
        default=None,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
91
92
93
94
            "help": (
                "For debugging purposes or quicker training, truncate the number of training examples to this "
                "value if set."
            )
95
96
        },
    )
97
    max_eval_samples: Optional[int] = field(
98
99
        default=None,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
100
101
102
103
            "help": (
                "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
                "value if set."
            )
104
105
        },
    )
106
    max_predict_samples: Optional[int] = field(
107
108
        default=None,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
109
110
111
112
            "help": (
                "For debugging purposes or quicker training, truncate the number of prediction examples to this "
                "value if set."
            )
113
114
        },
    )
115
116
    server_ip: Optional[str] = field(default=None, metadata={"help": "For distant debugging."})
    server_port: Optional[str] = field(default=None, metadata={"help": "For distant debugging."})
VictorSanh's avatar
VictorSanh committed
117
118


119
120
121
122
123
@dataclass
class ModelArguments:
    """
    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
    """
VictorSanh's avatar
VictorSanh committed
124

125
126
    model_name_or_path: str = field(
        default=None, metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
127
    )
128
129
    language: str = field(
        default=None, metadata={"help": "Evaluation language. Also train language if `train_language` is set to None."}
130
    )
131
132
    train_language: Optional[str] = field(
        default=None, metadata={"help": "Train language if it is different from the evaluation language."}
133
    )
134
135
    config_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
136
    )
137
138
    tokenizer_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
139
    )
140
    cache_dir: Optional[str] = field(
141
        default=None,
142
        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
143
    )
144
145
146
    do_lower_case: Optional[bool] = field(
        default=False,
        metadata={"help": "arg to indicate if tokenizer should do lower case in AutoTokenizer.from_pretrained()"},
147
    )
148
149
150
    use_fast_tokenizer: bool = field(
        default=True,
        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
151
    )
152
153
154
    model_revision: str = field(
        default="main",
        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
155
    )
156
157
158
    use_auth_token: bool = field(
        default=False,
        metadata={
Sylvain Gugger's avatar
Sylvain Gugger committed
159
160
161
162
            "help": (
                "Will use the token generated when running `transformers-cli login` (necessary to use this script "
                "with private models)."
            )
163
        },
164
    )
165
166
167
168
    ignore_mismatched_sizes: bool = field(
        default=False,
        metadata={"help": "Will enable to load a pretrained model whose head dimensions are different."},
    )
169

170
171
172
173
174
175
176
177
178

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, TrainingArguments))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()

VictorSanh's avatar
VictorSanh committed
179
    # Setup distant debugging if needed
180
    if data_args.server_ip and data_args.server_port:
VictorSanh's avatar
VictorSanh committed
181
182
        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
        import ptvsd
183

VictorSanh's avatar
VictorSanh committed
184
        print("Waiting for debugger attach")
185
        ptvsd.enable_attach(address=(data_args.server_ip, data_args.server_port), redirect_output=True)
VictorSanh's avatar
VictorSanh committed
186
187
188
        ptvsd.wait_for_attach()

    # Setup logging
189
    logging.basicConfig(
190
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
191
        datefmt="%m/%d/%Y %H:%M:%S",
192
        handlers=[logging.StreamHandler(sys.stdout)],
193
    )
194
195
196
197
198
199
200

    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()
201
202

    # Log on each process the small summary:
203
    logger.warning(
204
205
        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}"
206
    )
207
208
    logger.info(f"Training/evaluation parameters {training_args}")

209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
    # 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."
            )
        elif last_checkpoint is not None:
            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."
            )

224
225
226
227
228
229
    # Set seed before initializing model.
    set_seed(training_args.seed)

    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
    # download the dataset.
    # Downloading and loading xnli dataset from the hub.
230
231
    if training_args.do_train:
        if model_args.train_language is None:
232
233
234
235
236
237
238
            train_dataset = load_dataset(
                "xnli",
                model_args.language,
                split="train",
                cache_dir=model_args.cache_dir,
                use_auth_token=True if model_args.use_auth_token else None,
            )
239
        else:
240
            train_dataset = load_dataset(
241
242
243
244
245
                "xnli",
                model_args.train_language,
                split="train",
                cache_dir=model_args.cache_dir,
                use_auth_token=True if model_args.use_auth_token else None,
246
            )
247
248
249
        label_list = train_dataset.features["label"].names

    if training_args.do_eval:
250
251
252
253
254
255
256
        eval_dataset = load_dataset(
            "xnli",
            model_args.language,
            split="validation",
            cache_dir=model_args.cache_dir,
            use_auth_token=True if model_args.use_auth_token else None,
        )
257
258
259
        label_list = eval_dataset.features["label"].names

    if training_args.do_predict:
260
261
262
263
264
265
266
        predict_dataset = load_dataset(
            "xnli",
            model_args.language,
            split="test",
            cache_dir=model_args.cache_dir,
            use_auth_token=True if model_args.use_auth_token else None,
        )
267
        label_list = predict_dataset.features["label"].names
268
269

    # Labels
VictorSanh's avatar
VictorSanh committed
270
271
272
    num_labels = len(label_list)

    # Load pretrained model and tokenizer
273
274
    # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
275
    config = AutoConfig.from_pretrained(
276
        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
277
        num_labels=num_labels,
278
279
280
281
        finetuning_task="xnli",
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
282
    )
283
    tokenizer = AutoTokenizer.from_pretrained(
284
285
286
287
288
289
        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
        do_lower_case=model_args.do_lower_case,
        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,
290
    )
291
    model = AutoModelForSequenceClassification.from_pretrained(
292
293
        model_args.model_name_or_path,
        from_tf=bool(".ckpt" in model_args.model_name_or_path),
294
        config=config,
295
296
297
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
298
        ignore_mismatched_sizes=model_args.ignore_mismatched_sizes,
299
    )
VictorSanh's avatar
VictorSanh committed
300

301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
    # Preprocessing the datasets
    # Padding strategy
    if data_args.pad_to_max_length:
        padding = "max_length"
    else:
        # We will pad later, dynamically at batch creation, to the max sequence length in each batch
        padding = False

    def preprocess_function(examples):
        # Tokenize the texts
        return tokenizer(
            examples["premise"],
            examples["hypothesis"],
            padding=padding,
            max_length=data_args.max_seq_length,
            truncation=True,
        )
VictorSanh's avatar
VictorSanh committed
318

319
320
    if training_args.do_train:
        if data_args.max_train_samples is not None:
321
322
            max_train_samples = min(len(train_dataset), data_args.max_train_samples)
            train_dataset = train_dataset.select(range(max_train_samples))
323
324
325
326
327
328
329
        with training_args.main_process_first(desc="train dataset map pre-processing"):
            train_dataset = train_dataset.map(
                preprocess_function,
                batched=True,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on train dataset",
            )
330
331
332
        # Log a few random samples from the training set:
        for index in random.sample(range(len(train_dataset)), 3):
            logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
333
334

    if training_args.do_eval:
335
        if data_args.max_eval_samples is not None:
336
337
            max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
            eval_dataset = eval_dataset.select(range(max_eval_samples))
338
339
340
341
342
343
344
        with training_args.main_process_first(desc="validation dataset map pre-processing"):
            eval_dataset = eval_dataset.map(
                preprocess_function,
                batched=True,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on validation dataset",
            )
VictorSanh's avatar
VictorSanh committed
345

346
    if training_args.do_predict:
347
        if data_args.max_predict_samples is not None:
348
349
            max_predict_samples = min(len(predict_dataset), data_args.max_predict_samples)
            predict_dataset = predict_dataset.select(range(max_predict_samples))
350
351
352
353
354
355
356
        with training_args.main_process_first(desc="prediction dataset map pre-processing"):
            predict_dataset = predict_dataset.map(
                preprocess_function,
                batched=True,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on prediction dataset",
            )
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379

    # Get the metric function
    metric = load_metric("xnli")

    # You can define your custom compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
    # predictions and label_ids field) and has to return a dictionary string to float.
    def compute_metrics(p: EvalPrediction):
        preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions
        preds = np.argmax(preds, axis=1)
        return metric.compute(predictions=preds, references=p.label_ids)

    # Data collator will default to DataCollatorWithPadding, so we change it if we already did the padding.
    if data_args.pad_to_max_length:
        data_collator = default_data_collator
    elif training_args.fp16:
        data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=8)
    else:
        data_collator = None

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
380
        train_dataset=train_dataset if training_args.do_train else None,
381
382
383
384
385
        eval_dataset=eval_dataset if training_args.do_eval else None,
        compute_metrics=compute_metrics,
        tokenizer=tokenizer,
        data_collator=data_collator,
    )
VictorSanh's avatar
VictorSanh committed
386
387

    # Training
388
    if training_args.do_train:
389
        checkpoint = None
390
391
392
        if training_args.resume_from_checkpoint is not None:
            checkpoint = training_args.resume_from_checkpoint
        elif last_checkpoint is not None:
393
394
            checkpoint = last_checkpoint
        train_result = trainer.train(resume_from_checkpoint=checkpoint)
395
        metrics = train_result.metrics
396
397
398
399
        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))
VictorSanh's avatar
VictorSanh committed
400

401
        trainer.save_model()  # Saves the tokenizer too for easy upload
402

403
404
405
        trainer.log_metrics("train", metrics)
        trainer.save_metrics("train", metrics)
        trainer.save_state()
VictorSanh's avatar
VictorSanh committed
406

407
408
409
    # Evaluation
    if training_args.do_eval:
        logger.info("*** Evaluate ***")
410
411
        metrics = trainer.evaluate(eval_dataset=eval_dataset)

412
413
        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))
414

415
416
        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)
VictorSanh's avatar
VictorSanh committed
417

418
419
420
    # Prediction
    if training_args.do_predict:
        logger.info("*** Predict ***")
421
        predictions, labels, metrics = trainer.predict(predict_dataset, metric_key_prefix="predict")
422

423
424
425
426
        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))
427

428
429
        trainer.log_metrics("predict", metrics)
        trainer.save_metrics("predict", metrics)
430
431

        predictions = np.argmax(predictions, axis=1)
432
        output_predict_file = os.path.join(training_args.output_dir, "predictions.txt")
433
        if trainer.is_world_process_zero():
434
            with open(output_predict_file, "w") as writer:
435
436
437
438
439
                writer.write("index\tprediction\n")
                for index, item in enumerate(predictions):
                    item = label_list[item]
                    writer.write(f"{index}\t{item}\n")

VictorSanh's avatar
VictorSanh committed
440
441
442

if __name__ == "__main__":
    main()