utils_multiple_choice.py 20.1 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.
Julien Chaumond's avatar
Julien Chaumond committed
16
""" Multiple choice fine-tuning: utilities to work with multiple choice tasks of reading comprehension """
17
18


Aymeric Augustin's avatar
Aymeric Augustin committed
19
20
21
import csv
import glob
import json
22
23
import logging
import os
Julien Chaumond's avatar
Julien Chaumond committed
24
25
26
from dataclasses import dataclass
from enum import Enum
from typing import List, Optional
Aymeric Augustin's avatar
Aymeric Augustin committed
27
28

import tqdm
29
from filelock import FileLock
Aymeric Augustin's avatar
Aymeric Augustin committed
30

31
from transformers import PreTrainedTokenizer, is_tf_available, is_torch_available
32
33
34
35
36


logger = logging.getLogger(__name__)


Julien Chaumond's avatar
Julien Chaumond committed
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
@dataclass(frozen=True)
class InputExample:
    """
    A single training/test example for multiple choice

    Args:
        example_id: Unique id for the example.
        question: string. The untokenized text of the second sequence (question).
        contexts: list of str. The untokenized text of the first sequence (context of corresponding question).
        endings: list of str. multiple choice's options. Its length must be equal to contexts' length.
        label: (Optional) string. The label of the example. This should be
        specified for train and dev examples, but not for test examples.
    """

    example_id: str
    question: str
    contexts: List[str]
    endings: List[str]
    label: Optional[str]


@dataclass(frozen=True)
class InputFeatures:
    """
    A single set of features of data.
    Property names are the same names as the corresponding inputs to a model.
    """

    example_id: str
    input_ids: List[List[int]]
    attention_mask: Optional[List[List[int]]]
    token_type_ids: Optional[List[List[int]]]
    label: Optional[int]


class Split(Enum):
    train = "train"
    dev = "dev"
    test = "test"


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
104
if is_torch_available():
    import torch
    from torch.utils.data.dataset import Dataset

    class MultipleChoiceDataset(Dataset):
        """
        This will be superseded by a framework-agnostic approach
        soon.
        """

        features: List[InputFeatures]

        def __init__(
            self,
            data_dir: str,
            tokenizer: PreTrainedTokenizer,
            task: str,
            max_seq_length: Optional[int] = None,
            overwrite_cache=False,
            mode: Split = Split.train,
        ):
            processor = processors[task]()

            cached_features_file = os.path.join(
                data_dir,
                "cached_{}_{}_{}_{}".format(mode.value, tokenizer.__class__.__name__, str(max_seq_length), task,),
            )
105
106
107
108
109

            # Make sure only the first process in distributed training processes the dataset,
            # and the others will use the cache.
            lock_path = cached_features_file + ".lock"
            with FileLock(lock_path):
Julien Chaumond's avatar
Julien Chaumond committed
110

111
112
113
                if os.path.exists(cached_features_file) and not overwrite_cache:
                    logger.info(f"Loading features from cached file {cached_features_file}")
                    self.features = torch.load(cached_features_file)
Julien Chaumond's avatar
Julien Chaumond committed
114
                else:
115
116
117
118
119
120
121
122
123
                    logger.info(f"Creating features from dataset file at {data_dir}")
                    label_list = processor.get_labels()
                    if mode == Split.dev:
                        examples = processor.get_dev_examples(data_dir)
                    elif mode == Split.test:
                        examples = processor.get_test_examples(data_dir)
                    else:
                        examples = processor.get_train_examples(data_dir)
                    logger.info("Training examples: %s", len(examples))
124
                    self.features = convert_examples_to_features(examples, label_list, max_seq_length, tokenizer,)
125
126
                    logger.info("Saving features into cached file %s", cached_features_file)
                    torch.save(self.features, cached_features_file)
127
128
129
130
131
132
133
134
135
136
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

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

        def __getitem__(self, i) -> InputFeatures:
            return self.features[i]


if is_tf_available():
    import tensorflow as tf

    class TFMultipleChoiceDataset:
        """
        This will be superseded by a framework-agnostic approach
        soon.
        """

        features: List[InputFeatures]

        def __init__(
            self,
            data_dir: str,
            tokenizer: PreTrainedTokenizer,
            task: str,
            max_seq_length: Optional[int] = 128,
            overwrite_cache=False,
            mode: Split = Split.train,
        ):
            processor = processors[task]()

            logger.info(f"Creating features from dataset file at {data_dir}")
            label_list = processor.get_labels()
            if mode == Split.dev:
                examples = processor.get_dev_examples(data_dir)
            elif mode == Split.test:
                examples = processor.get_test_examples(data_dir)
            else:
                examples = processor.get_train_examples(data_dir)
            logger.info("Training examples: %s", len(examples))
166
167

            self.features = convert_examples_to_features(examples, label_list, max_seq_length, tokenizer,)
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206

            def gen():
                for (ex_index, ex) in tqdm.tqdm(enumerate(self.features), desc="convert examples to features"):
                    if ex_index % 10000 == 0:
                        logger.info("Writing example %d of %d" % (ex_index, len(examples)))

                    yield (
                        {
                            "example_id": 0,
                            "input_ids": ex.input_ids,
                            "attention_mask": ex.attention_mask,
                            "token_type_ids": ex.token_type_ids,
                        },
                        ex.label,
                    )

            self.dataset = tf.data.Dataset.from_generator(
                gen,
                (
                    {
                        "example_id": tf.int32,
                        "input_ids": tf.int32,
                        "attention_mask": tf.int32,
                        "token_type_ids": tf.int32,
                    },
                    tf.int64,
                ),
                (
                    {
                        "example_id": tf.TensorShape([]),
                        "input_ids": tf.TensorShape([None, None]),
                        "attention_mask": tf.TensorShape([None, None]),
                        "token_type_ids": tf.TensorShape([None, None]),
                    },
                    tf.TensorShape([]),
                ),
            )

        def get_dataset(self):
Julien Plu's avatar
Julien Plu committed
207
208
            self.dataset = self.dataset.apply(tf.data.experimental.assert_cardinality(len(self.features)))

209
            return self.dataset
210

211
212
        def __len__(self):
            return len(self.features)
Julien Chaumond's avatar
Julien Chaumond committed
213

214
215
        def __getitem__(self, i) -> InputFeatures:
            return self.features[i]
Julien Chaumond's avatar
Julien Chaumond committed
216
217
218


class DataProcessor:
erenup's avatar
erenup committed
219
    """Base class for data converters for multiple choice data sets."""
220
221
222
223
224
225
226
227
228

    def get_train_examples(self, data_dir):
        """Gets a collection of `InputExample`s for the train set."""
        raise NotImplementedError()

    def get_dev_examples(self, data_dir):
        """Gets a collection of `InputExample`s for the dev set."""
        raise NotImplementedError()

erenup's avatar
erenup committed
229
    def get_test_examples(self, data_dir):
erenup's avatar
erenup committed
230
        """Gets a collection of `InputExample`s for the test set."""
erenup's avatar
erenup committed
231
232
        raise NotImplementedError()

233
234
235
236
237
238
    def get_labels(self):
        """Gets the list of labels for this data set."""
        raise NotImplementedError()


class RaceProcessor(DataProcessor):
erenup's avatar
erenup committed
239
    """Processor for the RACE data set."""
240
241
242
243

    def get_train_examples(self, data_dir):
        """See base class."""
        logger.info("LOOKING AT {} train".format(data_dir))
244
245
        high = os.path.join(data_dir, "train/high")
        middle = os.path.join(data_dir, "train/middle")
246
247
        high = self._read_txt(high)
        middle = self._read_txt(middle)
248
        return self._create_examples(high + middle, "train")
249
250
251
252

    def get_dev_examples(self, data_dir):
        """See base class."""
        logger.info("LOOKING AT {} dev".format(data_dir))
253
254
        high = os.path.join(data_dir, "dev/high")
        middle = os.path.join(data_dir, "dev/middle")
255
256
        high = self._read_txt(high)
        middle = self._read_txt(middle)
257
        return self._create_examples(high + middle, "dev")
258

erenup's avatar
erenup committed
259
260
261
    def get_test_examples(self, data_dir):
        """See base class."""
        logger.info("LOOKING AT {} test".format(data_dir))
262
263
        high = os.path.join(data_dir, "test/high")
        middle = os.path.join(data_dir, "test/middle")
erenup's avatar
erenup committed
264
265
        high = self._read_txt(high)
        middle = self._read_txt(middle)
266
        return self._create_examples(high + middle, "test")
erenup's avatar
erenup committed
267

268
269
270
271
272
273
274
275
    def get_labels(self):
        """See base class."""
        return ["0", "1", "2", "3"]

    def _read_txt(self, input_dir):
        lines = []
        files = glob.glob(input_dir + "/*txt")
        for file in tqdm.tqdm(files, desc="read files"):
276
            with open(file, "r", encoding="utf-8") as fin:
277
278
279
280
281
282
283
284
285
286
287
288
                data_raw = json.load(fin)
                data_raw["race_id"] = file
                lines.append(data_raw)
        return lines

    def _create_examples(self, lines, set_type):
        """Creates examples for the training and dev sets."""
        examples = []
        for (_, data_raw) in enumerate(lines):
            race_id = "%s-%s" % (set_type, data_raw["race_id"])
            article = data_raw["article"]
            for i in range(len(data_raw["answers"])):
289
290
291
                truth = str(ord(data_raw["answers"][i]) - ord("A"))
                question = data_raw["questions"][i]
                options = data_raw["options"][i]
292
293
294
295
296

                examples.append(
                    InputExample(
                        example_id=race_id,
                        question=question,
297
                        contexts=[article, article, article, article],  # this is not efficient but convenient
298
                        endings=[options[0], options[1], options[2], options[3]],
299
300
301
                        label=truth,
                    )
                )
302
303
        return examples

304

305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
class SynonymProcessor(DataProcessor):
    """Processor for the Synonym data set."""

    def get_train_examples(self, data_dir):
        """See base class."""
        logger.info("LOOKING AT {} train".format(data_dir))
        return self._create_examples(self._read_csv(os.path.join(data_dir, "mctrain.csv")), "train")

    def get_dev_examples(self, data_dir):
        """See base class."""
        logger.info("LOOKING AT {} dev".format(data_dir))
        return self._create_examples(self._read_csv(os.path.join(data_dir, "mchp.csv")), "dev")

    def get_test_examples(self, data_dir):
        """See base class."""
        logger.info("LOOKING AT {} dev".format(data_dir))

        return self._create_examples(self._read_csv(os.path.join(data_dir, "mctest.csv")), "test")

    def get_labels(self):
        """See base class."""
        return ["0", "1", "2", "3", "4"]

    def _read_csv(self, input_file):
        with open(input_file, "r", encoding="utf-8") as f:
            return list(csv.reader(f))

    def _create_examples(self, lines: List[List[str]], type: str):
        """Creates examples for the training and dev sets."""

        examples = [
            InputExample(
                example_id=line[0],
                question="",  # in the swag dataset, the
                # common beginning of each
                # choice is stored in "sent2".
                contexts=[line[1], line[1], line[1], line[1], line[1]],
                endings=[line[2], line[3], line[4], line[5], line[6]],
                label=line[7],
            )
            for line in lines  # we skip the line with the column names
        ]

        return examples


351
class SwagProcessor(DataProcessor):
erenup's avatar
erenup committed
352
    """Processor for the SWAG data set."""
353
354
355
356
357
358
359
360
361
362
363

    def get_train_examples(self, data_dir):
        """See base class."""
        logger.info("LOOKING AT {} train".format(data_dir))
        return self._create_examples(self._read_csv(os.path.join(data_dir, "train.csv")), "train")

    def get_dev_examples(self, data_dir):
        """See base class."""
        logger.info("LOOKING AT {} dev".format(data_dir))
        return self._create_examples(self._read_csv(os.path.join(data_dir, "val.csv")), "dev")

erenup's avatar
erenup committed
364
365
    def get_test_examples(self, data_dir):
        """See base class."""
erenup's avatar
erenup committed
366
367
368
369
370
        logger.info("LOOKING AT {} dev".format(data_dir))
        raise ValueError(
            "For swag testing, the input file does not contain a label column. It can not be tested in current code"
            "setting!"
        )
erenup's avatar
erenup committed
371
        return self._create_examples(self._read_csv(os.path.join(data_dir, "test.csv")), "test")
372

373
374
375
376
377
    def get_labels(self):
        """See base class."""
        return ["0", "1", "2", "3"]

    def _read_csv(self, input_file):
378
        with open(input_file, "r", encoding="utf-8") as f:
379
            return list(csv.reader(f))
380

381
    def _create_examples(self, lines: List[List[str]], type: str):
382
        """Creates examples for the training and dev sets."""
383
384
        if type == "train" and lines[0][-1] != "label":
            raise ValueError("For training, the input file must contain a label column.")
385
386
387
388
389
390
391

        examples = [
            InputExample(
                example_id=line[2],
                question=line[5],  # in the swag dataset, the
                # common beginning of each
                # choice is stored in "sent2".
392
393
394
395
396
                contexts=[line[4], line[4], line[4], line[4]],
                endings=[line[7], line[8], line[9], line[10]],
                label=line[11],
            )
            for line in lines[1:]  # we skip the line with the column names
397
398
399
400
401
402
        ]

        return examples


class ArcProcessor(DataProcessor):
erenup's avatar
erenup committed
403
    """Processor for the ARC data set (request from allennlp)."""
404
405
406
407
408
409
410
411
412
413
414

    def get_train_examples(self, data_dir):
        """See base class."""
        logger.info("LOOKING AT {} train".format(data_dir))
        return self._create_examples(self._read_json(os.path.join(data_dir, "train.jsonl")), "train")

    def get_dev_examples(self, data_dir):
        """See base class."""
        logger.info("LOOKING AT {} dev".format(data_dir))
        return self._create_examples(self._read_json(os.path.join(data_dir, "dev.jsonl")), "dev")

erenup's avatar
erenup committed
415
416
417
418
    def get_test_examples(self, data_dir):
        logger.info("LOOKING AT {} test".format(data_dir))
        return self._create_examples(self._read_json(os.path.join(data_dir, "test.jsonl")), "test")

419
420
421
422
423
    def get_labels(self):
        """See base class."""
        return ["0", "1", "2", "3"]

    def _read_json(self, input_file):
424
        with open(input_file, "r", encoding="utf-8") as fin:
425
426
427
428
429
430
            lines = fin.readlines()
            return lines

    def _create_examples(self, lines, type):
        """Creates examples for the training and dev sets."""

431
        # There are two types of labels. They should be normalized
432
433
434
435
436
437
        def normalize(truth):
            if truth in "ABCD":
                return ord(truth) - ord("A")
            elif truth in "1234":
                return int(truth) - 1
            else:
erenup's avatar
erenup committed
438
439
                logger.info("truth ERROR! %s", str(truth))
                return None
erenup's avatar
erenup committed
440

441
442
443
444
445
        examples = []
        three_choice = 0
        four_choice = 0
        five_choice = 0
        other_choices = 0
erenup's avatar
erenup committed
446
        # we deleted example which has more than or less than four choices
447
448
449
450
451
452
453
454
455
456
457
458
459
        for line in tqdm.tqdm(lines, desc="read arc data"):
            data_raw = json.loads(line.strip("\n"))
            if len(data_raw["question"]["choices"]) == 3:
                three_choice += 1
                continue
            elif len(data_raw["question"]["choices"]) == 5:
                five_choice += 1
                continue
            elif len(data_raw["question"]["choices"]) != 4:
                other_choices += 1
                continue
            four_choice += 1
            truth = str(normalize(data_raw["answerKey"]))
erenup's avatar
erenup committed
460
            assert truth != "None"
461
462
463
464
465
466
467
            question_choices = data_raw["question"]
            question = question_choices["stem"]
            id = data_raw["id"]
            options = question_choices["choices"]
            if len(options) == 4:
                examples.append(
                    InputExample(
468
                        example_id=id,
469
                        question=question,
470
471
472
473
474
475
                        contexts=[
                            options[0]["para"].replace("_", ""),
                            options[1]["para"].replace("_", ""),
                            options[2]["para"].replace("_", ""),
                            options[3]["para"].replace("_", ""),
                        ],
476
                        endings=[options[0]["text"], options[1]["text"], options[2]["text"], options[3]["text"]],
477
478
479
                        label=truth,
                    )
                )
480
481
482
483
484
485
486
487
488
489
490
491
492

        if type == "train":
            assert len(examples) > 1
            assert examples[0].label is not None
        logger.info("len examples: %s}", str(len(examples)))
        logger.info("Three choices: %s", str(three_choice))
        logger.info("Five choices: %s", str(five_choice))
        logger.info("Other choices: %s", str(other_choices))
        logger.info("four choices: %s", str(four_choice))

        return examples


493
def convert_examples_to_features(
494
    examples: List[InputExample], label_list: List[str], max_length: int, tokenizer: PreTrainedTokenizer,
495
496
497
) -> List[InputFeatures]:
    """
    Loads a data file into a list of `InputFeatures`
498
499
    """

500
    label_map = {label: i for i, label in enumerate(label_list)}
501
502
503
504
505

    features = []
    for (ex_index, example) in tqdm.tqdm(enumerate(examples), desc="convert examples to features"):
        if ex_index % 10000 == 0:
            logger.info("Writing example %d of %d" % (ex_index, len(examples)))
Julien Chaumond's avatar
Julien Chaumond committed
506
        choices_inputs = []
507
        for ending_idx, (context, ending) in enumerate(zip(example.contexts, example.endings)):
508
            text_a = context
509
            if example.question.find("_") != -1:
510
511
                # this is for cloze question
                text_b = example.question.replace("_", ending)
512
            else:
513
514
                text_b = example.question + " " + ending

515
            inputs = tokenizer(
516
517
518
519
                text_a,
                text_b,
                add_special_tokens=True,
                max_length=max_length,
520
521
                padding="max_length",
                truncation=True,
522
                return_overflowing_tokens=True,
523
            )
524
525
526
527
528
529
            if "num_truncated_tokens" in inputs and inputs["num_truncated_tokens"] > 0:
                logger.info(
                    "Attention! you are cropping tokens (swag task is ok). "
                    "If you are training ARC and RACE and you are poping question + options,"
                    "you need to try to use a bigger max seq length!"
                )
530

Julien Chaumond's avatar
Julien Chaumond committed
531
            choices_inputs.append(inputs)
532

Julien Chaumond's avatar
Julien Chaumond committed
533
        label = label_map[example.label]
534

Julien Chaumond's avatar
Julien Chaumond committed
535
536
537
538
539
540
541
        input_ids = [x["input_ids"] for x in choices_inputs]
        attention_mask = (
            [x["attention_mask"] for x in choices_inputs] if "attention_mask" in choices_inputs[0] else None
        )
        token_type_ids = (
            [x["token_type_ids"] for x in choices_inputs] if "token_type_ids" in choices_inputs[0] else None
        )
542

Julien Chaumond's avatar
Julien Chaumond committed
543
544
545
546
547
548
549
550
551
        features.append(
            InputFeatures(
                example_id=example.example_id,
                input_ids=input_ids,
                attention_mask=attention_mask,
                token_type_ids=token_type_ids,
                label=label,
            )
        )
552

Julien Chaumond's avatar
Julien Chaumond committed
553
554
555
    for f in features[:2]:
        logger.info("*** Example ***")
        logger.info("feature: %s" % f)
556
557
558
559

    return features


560
561
processors = {"race": RaceProcessor, "swag": SwagProcessor, "arc": ArcProcessor, "syn": SynonymProcessor}
MULTIPLE_CHOICE_TASKS_NUM_LABELS = {"race", 4, "swag", 4, "arc", 4, "syn", 5}