File size: 9,312 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
# 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
from dataclasses import dataclass, field
from typing import Optional

import numpy as np
from omegaconf import II

from fairseq.data import (
    AppendTokenDataset,
    ConcatDataset,
    DenoisingDataset,
    Dictionary,
    PrependTokenDataset,
    ResamplingDataset,
    SortDataset,
    TokenBlockDataset,
    data_utils,
)
from fairseq.data.encoders.utils import get_whole_word_mask
from fairseq.tasks import register_task

from .denoising import DenoisingConfig, DenoisingTask

logger = logging.getLogger(__name__)


@dataclass
class MultilingualDenoisingConfig(DenoisingConfig):
    multilang_sampling_alpha: float = field(
        default=1.0,
        metadata={"help": "smoothing alpha for sample ratios across multiple datasets"},
    )
    add_lang_token: bool = field(
        default=False,
        metadata={"help": ""},
    )
    langs: Optional[str] = field(
        default=None,
        metadata={"help": "language ids we are considering"},
    )
    no_whole_word_mask_langs: str = field(
        default="",
        metadata={
            "help": "languages without spacing between words don't support whole word masking"
        },
    )
    train_subset: str = II("common.train_subset")
    valid_subset: str = II("common.valid_subset")


@register_task("multilingual_denoising", dataclass=MultilingualDenoisingConfig)
class MultilingualDenoisingTask(DenoisingTask):

    cfg: MultilingualDenoisingConfig

    @classmethod
    def setup_task(cls, cfg: MultilingualDenoisingConfig, **kwargs):
        """Setup the task."""
        paths = cfg.data.split(":")
        assert len(paths) > 0
        dictionary = Dictionary.load(os.path.join(paths[0], "dict.txt"))

        data_path = paths[0]
        if cfg.langs is None:
            languages = sorted(
                [
                    name
                    for name in os.listdir(data_path)
                    if os.path.isdir(os.path.join(data_path, name))
                ]
            )
        else:
            languages = cfg.langs.split(",")

        if cfg.add_lang_token:
            for lang in languages:
                dictionary.add_symbol("[{}]".format(lang))

        logger.info("dictionary: {} types".format(len(dictionary)))
        if not hasattr(cfg, "shuffle_instance"):
            cfg.shuffle_instance = False
        return cls(cfg, dictionary)

    def __init__(self, cfg: MultilingualDenoisingConfig, dictionary):
        super().__init__(cfg, dictionary)
        self.dictionary = dictionary

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

    def _get_sample_prob(self, dataset_lens):
        """
        Get smoothed sampling probability by languages. This helps low resource
        languages by upsampling them.
        """
        prob = dataset_lens / dataset_lens.sum()
        smoothed_prob = prob**self.cfg.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 = self.cfg.data.split(":")
        assert len(paths) > 0
        data_path = paths[(epoch - 1) % len(paths)]
        split_path = os.path.join(data_path, split)

        if self.cfg.langs is None:
            languages = sorted(
                [
                    name
                    for name in os.listdir(data_path)
                    if os.path.isdir(os.path.join(data_path, name))
                ]
            )
        else:
            languages = self.cfg.langs.split(",")
            for name in languages:
                p = os.path.join(data_path, name)
                assert os.path.exists(p), "data not found: {}".format(p)

        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 = get_whole_word_mask(self.cfg.bpe, self.dictionary)
        language_without_segmentations = self.cfg.no_whole_word_mask_langs.split(",")
        lang_datasets = []
        for language in languages:
            split_path = os.path.join(data_path, language, split)

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

            end_token = (
                self.source_dictionary.index("[{}]".format(language))
                if self.cfg.add_lang_token
                else self.source_dictionary.eos()
            )

            # create continuous blocks of tokens
            dataset = TokenBlockDataset(
                dataset,
                dataset.sizes,
                self.cfg.tokens_per_sample - 2,  # one less for <s>
                pad=self.source_dictionary.pad(),
                eos=end_token,
                break_mode=self.cfg.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())
            dataset = AppendTokenDataset(dataset, end_token)

            lang_mask_whole_words = (
                mask_whole_words
                if language not in language_without_segmentations
                else None
            )
            lang_dataset = DenoisingDataset(
                dataset,
                dataset.sizes,
                self.dictionary,
                self.mask_idx,
                lang_mask_whole_words,
                shuffle=self.cfg.shuffle_instance,
                seed=self.cfg.seed,
                mask=self.cfg.mask,
                mask_random=self.cfg.mask_random,
                insert=self.cfg.insert,
                rotate=self.cfg.rotate,
                permute_sentences=self.cfg.permute_sentences,
                bpe=self.cfg.bpe,
                replace_length=self.cfg.replace_length,
                mask_length=self.cfg.mask_length,
                poisson_lambda=self.cfg.poisson_lambda,
                eos=None
                if not self.cfg.add_lang_token
                else self.source_dictionary.index("[{}]".format(language)),
            )
            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(
                int(dataset_lengths.sum()),
            )
        )
        if split == self.cfg.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: {}".format(
                    {
                        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: {}".format(
                    {
                        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.cfg.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

            if split in self.cfg.valid_subset:
                self.cfg.valid_subset = self.cfg.valid_subset.replace(
                    split, ",".join(lang_splits)
                )

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

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