File size: 10,837 Bytes
85ba398 |
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 |
# 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
import numpy as np
from omegaconf import II, MISSING, OmegaConf
from fairseq import utils
from fairseq.data import (
Dictionary,
IdDataset,
MaskTokensDataset,
NestedDictionaryDataset,
NumelDataset,
NumSamplesDataset,
PrependTokenDataset,
RightPadDataset,
RightPaddingMaskDataset,
SortDataset,
TokenBlockDataset,
data_utils,
)
from fairseq.data.encoders.utils import get_whole_word_mask
from fairseq.data.shorten_dataset import maybe_shorten_dataset
from fairseq.dataclass import FairseqDataclass
from fairseq.tasks import FairseqTask, register_task
from .language_modeling import SAMPLE_BREAK_MODE_CHOICES, SHORTEN_METHOD_CHOICES
logger = logging.getLogger(__name__)
@dataclass
class MaskedLMConfig(FairseqDataclass):
data: str = field(
default=MISSING,
metadata={
"help": "colon separated path to data directories list, \
will be iterated upon during epochs in round-robin manner"
},
)
sample_break_mode: SAMPLE_BREAK_MODE_CHOICES = field(
default="none",
metadata={
"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.'
},
)
tokens_per_sample: int = field(
default=1024,
metadata={"help": "max number of tokens per sample for LM dataset"},
)
mask_prob: float = field(
default=0.15,
metadata={"help": "probability of replacing a token with mask"},
)
leave_unmasked_prob: float = field(
default=0.1,
metadata={"help": "probability that a masked token is unmasked"},
)
random_token_prob: float = field(
default=0.1,
metadata={"help": "probability of replacing a token with a random token"},
)
freq_weighted_replacement: bool = field(
default=False,
metadata={"help": "sample random replacement words based on word frequencies"},
)
mask_whole_words: bool = field(
default=False,
metadata={"help": "mask whole words; you may also want to set --bpe"},
)
mask_multiple_length: int = field(
default=1,
metadata={"help": "repeat the mask indices multiple times"},
)
mask_stdev: float = field(
default=0.0,
metadata={"help": "stdev of the mask length"},
)
shorten_method: SHORTEN_METHOD_CHOICES = field(
default="none",
metadata={
"help": "if not none, shorten sequences that exceed --tokens-per-sample"
},
)
shorten_data_split_list: str = field(
default="",
metadata={
"help": "comma-separated list of dataset splits to apply shortening to, "
'e.g., "train,valid" (default: all dataset splits)'
},
)
seed: int = II("common.seed")
include_target_tokens: bool = field(
default=False,
metadata={
"help": "include target tokens in model input. this is used for data2vec"
},
)
include_index: bool = field(
default=True,
metadata={"help": "include index in model input. this is used for data2vec"},
)
skip_masking: bool = field(
default=False,
metadata={"help": "skip masking at dataset"},
)
# subsample_train: float = field(
# default=1,
# metadata={"help": "shorten training set for debugging"},
# )
d2v2_multi: bool = field(
default=False,
metadata={"help": "prepare dataset for data2vec_multi"},
)
@register_task("masked_lm", dataclass=MaskedLMConfig)
class MaskedLMTask(FairseqTask):
cfg: MaskedLMConfig
"""Task for training masked language models (e.g., BERT, RoBERTa)."""
def __init__(self, cfg: MaskedLMConfig, dictionary=None):
super().__init__(cfg)
self.dictionary = dictionary or self.load_dict(cfg)
# add mask token
self.mask_idx = self.dictionary.add_symbol("<mask>")
@classmethod
def setup_task(cls, cfg: MaskedLMConfig, **kwargs):
dictionary = cls.load_dict(cfg)
return cls(cfg, dictionary)
@classmethod
def load_dict(cls, cfg):
paths = utils.split_paths(cfg.data)
assert len(paths) > 0
dictionary = Dictionary.load(os.path.join(paths[0], "dict.txt"))
logger.info("dictionary: {} types".format(len(dictionary)))
return dictionary
def _load_dataset_split(self, split, epoch, combine):
paths = utils.split_paths(self.cfg.data)
assert len(paths) > 0
data_path = paths[(epoch - 1) % len(paths)]
split_path = os.path.join(data_path, split)
dataset = data_utils.load_indexed_dataset(
split_path,
self.source_dictionary,
combine=combine,
)
if dataset is None:
raise FileNotFoundError(
"Dataset not found: {} ({})".format(split, split_path)
)
dataset = maybe_shorten_dataset(
dataset,
split,
self.cfg.shorten_data_split_list,
self.cfg.shorten_method,
self.cfg.tokens_per_sample,
self.cfg.seed,
)
# create continuous blocks of tokens
dataset = TokenBlockDataset(
dataset,
dataset.sizes,
self.cfg.tokens_per_sample - 1, # one less for <s>
pad=self.source_dictionary.pad(),
eos=self.source_dictionary.eos(),
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)
return PrependTokenDataset(dataset, self.source_dictionary.bos())
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)
"""
dataset = self._load_dataset_split(split, epoch, combine)
# create masked input and targets
mask_whole_words = (
get_whole_word_mask(self.args, self.source_dictionary)
if self.cfg.mask_whole_words
else None
)
src_dataset, tgt_dataset = MaskTokensDataset.apply_mask(
dataset,
self.source_dictionary,
pad_idx=self.source_dictionary.pad(),
mask_idx=self.mask_idx,
seed=self.cfg.seed,
mask_prob=self.cfg.mask_prob,
leave_unmasked_prob=self.cfg.leave_unmasked_prob,
random_token_prob=self.cfg.random_token_prob,
freq_weighted_replacement=self.cfg.freq_weighted_replacement,
mask_whole_words=mask_whole_words,
mask_multiple_length=self.cfg.mask_multiple_length,
mask_stdev=self.cfg.mask_stdev,
skip_masking=self.cfg.skip_masking,
)
with data_utils.numpy_seed(self.cfg.seed):
shuffle = np.random.permutation(len(src_dataset))
target_dataset = RightPadDataset(
tgt_dataset,
pad_idx=self.source_dictionary.pad(),
)
if self.cfg.d2v2_multi:
dataset = self._d2v2_multi_dataset(src_dataset)
else:
dataset = self._regular_dataset(src_dataset, target_dataset)
self.datasets[split] = SortDataset(
dataset, sort_order=[shuffle, src_dataset.sizes]
)
def _regular_dataset(self, src_dataset, target_dataset):
input_dict = {
"src_tokens": RightPadDataset(
src_dataset,
pad_idx=self.source_dictionary.pad(),
),
"src_lengths": NumelDataset(src_dataset, reduce=False),
}
if self.cfg.include_target_tokens:
input_dict["target_tokens"] = target_dataset
if self.cfg.include_index:
input_dict["src_id"] = IdDataset()
dataset = NestedDictionaryDataset(
{
"id": IdDataset(),
"net_input": input_dict,
"target": target_dataset,
"nsentences": NumSamplesDataset(),
"ntokens": NumelDataset(src_dataset, reduce=True),
},
sizes=[src_dataset.sizes],
)
return dataset
def _d2v2_multi_dataset(self, src_dataset):
input_dict = {
"source": RightPadDataset(
src_dataset,
pad_idx=self.source_dictionary.pad(),
),
"id": IdDataset(),
"padding_mask": RightPaddingMaskDataset(src_dataset),
}
dataset = NestedDictionaryDataset(
{
"id": IdDataset(),
"net_input": input_dict,
"nsentences": NumSamplesDataset(),
"ntokens": NumelDataset(src_dataset, reduce=True),
},
sizes=[src_dataset.sizes],
)
return dataset
def build_dataset_for_inference(self, src_tokens, src_lengths, sort=True):
src_dataset = RightPadDataset(
TokenBlockDataset(
src_tokens,
src_lengths,
self.cfg.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(),
)
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
def begin_epoch(self, epoch, model):
model.set_epoch(epoch)
def max_positions(self):
return self.cfg.tokens_per_sample
|