File size: 12,144 Bytes
6a62ffb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
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
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
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
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
166
167
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
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
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
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
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
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

import logging
import os

import numpy as np
import torch

from fairseq import utils
from fairseq.data import (
    ConcatDataset,
    Dictionary,
    IdDataset,
    MaskTokensDataset,
    NestedDictionaryDataset,
    NumelDataset,
    NumSamplesDataset,
    PadDataset,
    PrependTokenDataset,
    RawLabelDataset,
    ResamplingDataset,
    SortDataset,
    TokenBlockDataset,
    data_utils,
    encoders,
)
from fairseq.tasks import LegacyFairseqTask, register_task

logger = logging.getLogger(__name__)


@register_task("multilingual_masked_lm")
class MultiLingualMaskedLMTask(LegacyFairseqTask):
    """Task for training masked language models (e.g., BERT, RoBERTa)."""

    @staticmethod
    def add_args(parser):
        """Add task-specific arguments to the parser."""
        parser.add_argument(
            "data",
            help="colon separated path to data directories list, \
                            will be iterated upon during epochs in round-robin manner",
        )
        parser.add_argument(
            "--sample-break-mode",
            default="complete",
            choices=["none", "complete", "complete_doc", "eos"],
            help='If omitted or "none", fills each sample with tokens-per-sample '
            'tokens. If set to "complete", splits samples only at the end '
            "of sentence, but may include multiple sentences per sample. "
            '"complete_doc" is similar but respects doc boundaries. '
            'If set to "eos", includes only one sentence per sample.',
        )
        parser.add_argument(
            "--tokens-per-sample",
            default=512,
            type=int,
            help="max number of total tokens over all segments "
            "per sample for BERT dataset",
        )
        parser.add_argument(
            "--mask-prob",
            default=0.15,
            type=float,
            help="probability of replacing a token with mask",
        )
        parser.add_argument(
            "--leave-unmasked-prob",
            default=0.1,
            type=float,
            help="probability that a masked token is unmasked",
        )
        parser.add_argument(
            "--random-token-prob",
            default=0.1,
            type=float,
            help="probability of replacing a token with a random token",
        )
        parser.add_argument(
            "--freq-weighted-replacement",
            action="store_true",
            help="sample random replacement words based on word frequencies",
        )
        parser.add_argument(
            "--mask-whole-words",
            default=False,
            action="store_true",
            help="mask whole words; you may also want to set --bpe",
        )
        parser.add_argument(
            "--multilang-sampling-alpha",
            type=float,
            default=1.0,
            help="smoothing alpha for sample rations across multiple datasets",
        )

    def __init__(self, args, dictionary):
        super().__init__(args)
        self.dictionary = dictionary
        self.seed = args.seed

        # add mask token
        self.mask_idx = dictionary.add_symbol("<mask>")

    @classmethod
    def setup_task(cls, args, **kwargs):
        paths = utils.split_paths(args.data)
        assert len(paths) > 0
        dictionary = Dictionary.load(os.path.join(paths[0], "dict.txt"))
        logger.info("dictionary: {} types".format(len(dictionary)))
        return cls(args, dictionary)

    def _get_whole_word_mask(self):
        # create masked input and targets
        if self.args.mask_whole_words:
            bpe = encoders.build_bpe(self.args)
            if bpe is not None:

                def is_beginning_of_word(i):
                    if i < self.source_dictionary.nspecial:
                        # special elements are always considered beginnings
                        return True
                    tok = self.source_dictionary[i]
                    if tok.startswith("madeupword"):
                        return True
                    try:
                        return bpe.is_beginning_of_word(tok)
                    except ValueError:
                        return True

                mask_whole_words = torch.ByteTensor(
                    list(map(is_beginning_of_word, range(len(self.source_dictionary))))
                )
        else:
            mask_whole_words = None
        return mask_whole_words

    def _get_sample_prob(self, dataset_lens):
        """
        Get smoothed sampling porbability by languages. This helps low resource
        languages by upsampling them.
        """
        prob = dataset_lens / dataset_lens.sum()
        smoothed_prob = prob**self.args.multilang_sampling_alpha
        smoothed_prob = smoothed_prob / smoothed_prob.sum()
        return smoothed_prob

    def load_dataset(self, split, epoch=1, combine=False, **kwargs):
        """Load a given dataset split.

        Args:
            split (str): name of the split (e.g., train, valid, test)
        """
        paths = utils.split_paths(self.args.data)
        assert len(paths) > 0
        data_path = paths[(epoch - 1) % len(paths)]

        languages = sorted(
            name
            for name in os.listdir(data_path)
            if os.path.isdir(os.path.join(data_path, name))
        )

        logger.info("Training on {0} languages: {1}".format(len(languages), languages))
        logger.info(
            "Language to id mapping: ", {lang: id for id, lang in enumerate(languages)}
        )

        mask_whole_words = self._get_whole_word_mask()
        lang_datasets = []
        for lang_id, language in enumerate(languages):
            split_path = os.path.join(data_path, language, split)

            dataset = data_utils.load_indexed_dataset(
                split_path,
                self.source_dictionary,
                self.args.dataset_impl,
                combine=combine,
            )
            if dataset is None:
                raise FileNotFoundError(
                    "Dataset not found: {} ({})".format(split, split_path)
                )

            # create continuous blocks of tokens
            dataset = TokenBlockDataset(
                dataset,
                dataset.sizes,
                self.args.tokens_per_sample - 1,  # one less for <s>
                pad=self.source_dictionary.pad(),
                eos=self.source_dictionary.eos(),
                break_mode=self.args.sample_break_mode,
            )
            logger.info("loaded {} blocks from: {}".format(len(dataset), split_path))

            # prepend beginning-of-sentence token (<s>, equiv. to [CLS] in BERT)
            dataset = PrependTokenDataset(dataset, self.source_dictionary.bos())

            src_dataset, tgt_dataset = MaskTokensDataset.apply_mask(
                dataset,
                self.source_dictionary,
                pad_idx=self.source_dictionary.pad(),
                mask_idx=self.mask_idx,
                seed=self.args.seed,
                mask_prob=self.args.mask_prob,
                leave_unmasked_prob=self.args.leave_unmasked_prob,
                random_token_prob=self.args.random_token_prob,
                freq_weighted_replacement=self.args.freq_weighted_replacement,
                mask_whole_words=mask_whole_words,
            )

            lang_dataset = NestedDictionaryDataset(
                {
                    "net_input": {
                        "src_tokens": PadDataset(
                            src_dataset,
                            pad_idx=self.source_dictionary.pad(),
                            left_pad=False,
                        ),
                        "src_lengths": NumelDataset(src_dataset, reduce=False),
                    },
                    "target": PadDataset(
                        tgt_dataset,
                        pad_idx=self.source_dictionary.pad(),
                        left_pad=False,
                    ),
                    "nsentences": NumSamplesDataset(),
                    "ntokens": NumelDataset(src_dataset, reduce=True),
                    "lang_id": RawLabelDataset([lang_id] * src_dataset.sizes.shape[0]),
                },
                sizes=[src_dataset.sizes],
            )
            lang_datasets.append(lang_dataset)

        dataset_lengths = np.array(
            [len(d) for d in lang_datasets],
            dtype=float,
        )
        logger.info(
            "loaded total {} blocks for all languages".format(
                dataset_lengths.sum(),
            )
        )
        if split == self.args.train_subset:
            # For train subset, additionally up or down sample languages.
            sample_probs = self._get_sample_prob(dataset_lengths)
            logger.info(
                "Sample probability by language: ",
                {
                    lang: "{0:.4f}".format(sample_probs[id])
                    for id, lang in enumerate(languages)
                },
            )
            size_ratio = (sample_probs * dataset_lengths.sum()) / dataset_lengths
            logger.info(
                "Up/Down Sampling ratio by language: ",
                {
                    lang: "{0:.2f}".format(size_ratio[id])
                    for id, lang in enumerate(languages)
                },
            )

            resampled_lang_datasets = [
                ResamplingDataset(
                    lang_datasets[i],
                    size_ratio=size_ratio[i],
                    seed=self.args.seed,
                    epoch=epoch,
                    replace=size_ratio[i] >= 1.0,
                )
                for i, d in enumerate(lang_datasets)
            ]
            dataset = ConcatDataset(resampled_lang_datasets)
        else:
            dataset = ConcatDataset(lang_datasets)
            lang_splits = [split]
            for lang_id, lang_dataset in enumerate(lang_datasets):
                split_name = split + "_" + languages[lang_id]
                lang_splits.append(split_name)
                self.datasets[split_name] = lang_dataset

            # [TODO]: This is hacky for now to print validation ppl for each
            # language individually. Maybe need task API changes to allow it
            # in more generic ways.
            if split in self.args.valid_subset:
                self.args.valid_subset = self.args.valid_subset.replace(
                    split, ",".join(lang_splits)
                )

        with data_utils.numpy_seed(self.args.seed + epoch):
            shuffle = np.random.permutation(len(dataset))

        self.datasets[split] = SortDataset(
            dataset,
            sort_order=[
                shuffle,
                dataset.sizes,
            ],
        )

    def build_dataset_for_inference(self, src_tokens, src_lengths, sort=True):
        src_dataset = PadDataset(
            TokenBlockDataset(
                src_tokens,
                src_lengths,
                self.args.tokens_per_sample - 1,  # one less for <s>
                pad=self.source_dictionary.pad(),
                eos=self.source_dictionary.eos(),
                break_mode="eos",
            ),
            pad_idx=self.source_dictionary.pad(),
            left_pad=False,
        )
        src_dataset = PrependTokenDataset(src_dataset, self.source_dictionary.bos())
        src_dataset = NestedDictionaryDataset(
            {
                "id": IdDataset(),
                "net_input": {
                    "src_tokens": src_dataset,
                    "src_lengths": NumelDataset(src_dataset, reduce=False),
                },
            },
            sizes=src_lengths,
        )
        if sort:
            src_dataset = SortDataset(src_dataset, sort_order=[src_lengths])
        return src_dataset

    @property
    def source_dictionary(self):
        return self.dictionary

    @property
    def target_dictionary(self):
        return self.dictionary