Spaces:
Runtime error
Runtime error
File size: 9,964 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 |
# 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 Any, Optional
import numpy as np
from omegaconf import II, MISSING
from fairseq import utils
from fairseq.data import (
AppendTokenDataset,
DenoisingDataset,
Dictionary,
IdDataset,
NestedDictionaryDataset,
NumelDataset,
PadDataset,
PrependTokenDataset,
StripTokenDataset,
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 ChoiceEnum, FairseqDataclass
from fairseq.tasks import FairseqTask, register_task
from ..data.indexed_dataset import get_available_dataset_impl
logger = logging.getLogger(__name__)
SAMPLE_BREAK_MODE_CHOICES = ChoiceEnum(["none", "complete", "complete_doc", "eos"])
SHORTEN_METHOD_CHOICES = ChoiceEnum(["none", "truncate", "random_crop"])
MASK_LENGTH_CHOICES = ChoiceEnum(["subword", "word", "span-poisson"])
@dataclass
class DenoisingConfig(FairseqDataclass):
data: str = field(
default=MISSING,
metadata={"help": "path to data directory"},
)
bpe: Optional[str] = field(
default=None,
metadata={"help": "TODO"},
)
tokens_per_sample: int = field(
default=512,
metadata={
"help": "max number of total tokens over all segments "
"per sample for dataset"
},
)
sample_break_mode: SAMPLE_BREAK_MODE_CHOICES = field(
default="complete_doc",
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.'
},
)
replace_length: int = field(
default=0,
metadata={"help": "TODO, should only allow -1, 0 and 1"},
)
mask: float = field(
default=0.0,
metadata={"help": "fraction of words/subwords that will be masked"},
)
mask_random: float = field(
default=0.0,
metadata={"help": "instead of using [MASK], use random token this often"},
)
insert: float = field(
default=0.0,
metadata={"help": "insert this percentage of additional random tokens"},
)
permute: float = field(
default=0.0,
metadata={"help": "take this proportion of subwords and permute them"},
)
rotate: float = field(
default=0.5,
metadata={"help": "rotate this proportion of inputs"},
)
poisson_lambda: float = field(
default=3.0,
metadata={"help": "randomly shuffle sentences for this proportion of inputs"},
)
shuffle_instance: float = field(
default=0.0,
metadata={"help": "shuffle this proportion of sentences in all inputs"},
)
mask_length: MASK_LENGTH_CHOICES = field(
default="subword",
metadata={"help": "mask length to choose"},
)
permute_sentences: int = field(
default=-1,
metadata={
"help": "when masking N tokens, replace with 0, 1, or N tokens (use -1 for N)"
},
)
seed: int = II("common.seed")
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)'
},
)
max_source_positions: int = field(
default=1024,
metadata={"help": "max number of tokens in the source sequence"},
)
max_target_positions: int = field(
default=1024,
metadata={"help": "max number of tokens in the target sequence"},
)
dataset_impl: Optional[ChoiceEnum(get_available_dataset_impl())] = II(
"dataset.dataset_impl"
)
@register_task("denoising", dataclass=DenoisingConfig)
class DenoisingTask(FairseqTask):
"""
Denoising task for applying sequence to sequence denoising. (ie. BART)
"""
cfg: DenoisingConfig
def __init__(self, cfg, dictionary):
super().__init__(cfg)
self.dictionary = dictionary
# add mask token
self.mask_idx = self.dictionary.add_symbol("<mask>")
@classmethod
def setup_task(cls, cfg: DenoisingConfig, **kwargs):
"""Setup the task."""
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)))
if not hasattr(cfg, "shuffle_instance"):
cfg.shuffle_instance = False
return cls(cfg, 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.dictionary,
self.cfg.dataset_impl,
combine=combine,
)
if dataset is None:
raise FileNotFoundError(
"Dataset not found: {} ({})".format(split, split_path)
)
dataset = StripTokenDataset(dataset, self.dictionary.eos())
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 - 2,
# one less for <s> and one for </s>
pad=self.dictionary.pad(),
eos=self.dictionary.eos(),
break_mode=self.cfg.sample_break_mode,
document_sep_len=0,
)
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, self.source_dictionary.eos())
return dataset
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)
mask_whole_words = (
get_whole_word_mask(self.cfg.bpe, self.source_dictionary)
if self.cfg.mask_length != "subword"
else None
)
self.datasets[split] = DenoisingDataset(
dataset,
dataset.sizes,
self.dictionary,
self.mask_idx,
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,
)
logger.info(
"Split: {0}, Loaded {1} samples of denoising_dataset".format(
split,
len(self.datasets[split]),
)
)
def build_dataset_for_inference(self, src_tokens, src_lengths, **kwargs):
"""
Generate batches for inference. We assume that the input begins with a
bos symbol (`<s>`) and ends with an eos symbol (`</s>`).
"""
pad = self.source_dictionary.pad()
eos = self.source_dictionary.eos()
src_dataset = TokenBlockDataset(
src_tokens,
src_lengths,
block_size=self.cfg.tokens_per_sample - 2, # for <s> and </s>
pad=pad,
eos=eos,
break_mode=self.cfg.sample_break_mode,
document_sep_len=0,
)
prev_output_tokens = PrependTokenDataset(
StripTokenDataset(src_dataset, eos), eos
)
src_dataset = PadDataset(src_dataset, pad_idx=pad, left_pad=False)
return NestedDictionaryDataset(
{
"id": IdDataset(),
"net_input": {
"src_tokens": src_dataset,
"src_lengths": NumelDataset(src_dataset, reduce=False),
"prev_output_tokens": PadDataset(
prev_output_tokens, pad_idx=pad, left_pad=False
),
},
"target": src_dataset,
},
sizes=[np.array(src_lengths)],
)
def max_positions(self):
"""Return the max sentence length allowed by the task."""
return (self.cfg.max_source_positions, self.cfg.max_target_positions)
@property
def source_dictionary(self):
"""Return the source :class:`~fairseq.data.Dictionary`."""
return self.dictionary
@property
def target_dictionary(self):
"""Return the target :class:`~fairseq.data.Dictionary`."""
return self.dictionary
|