Spaces:
Sleeping
Sleeping
File size: 10,118 Bytes
6fc683c |
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 |
# Copyright (c) 2017-present, Facebook, Inc.
# All rights reserved.
#
# This source code is licensed under the license found in the LICENSE file in
# the root directory of this source tree. An additional grant of patent rights
# can be found in the PATENTS file in the same directory.
import os
import sys
import torch
from argparse import Namespace
from dataclasses import dataclass, field
from typing import Optional, Any
from omegaconf import MISSING
from fairseq.data import AddTargetDataset, Dictionary, FileAudioDataset, encoders
from fairseq.dataclass import FairseqDataclass
from fairseq.dataclass.configs import GenerationConfig
from . import FairseqTask, register_task
from .. import utils
from ..logging import metrics
class LabelEncoder(object):
def __init__(self, dictionary):
self.dictionary = dictionary
def __call__(self, label):
return self.dictionary.encode_line(
label, append_eos=False, add_if_not_exist=False
)
@dataclass
class AudioPretrainingConfig(FairseqDataclass):
data: str = field(default=MISSING, metadata={"help": "path to data directory"})
labels: Optional[str] = field(
default=None,
metadata={"help": "extension of the label file to load, used for fine-tuning"},
)
sample_rate: int = field(
default=16_000,
metadata={
"help": "target sample rate. audio files will be up/down sampled to this rate"
},
)
normalize: bool = field(
default=False,
metadata={"help": "if set, normalizes input to have 0 mean and unit variance"},
)
enable_padding: bool = field(
default=False, metadata={"help": "pad shorter samples instead of cropping"}
)
max_sample_size: Optional[int] = field(
default=None, metadata={"help": "max sample size to crop to for batching"}
)
min_sample_size: Optional[int] = field(
default=None, metadata={"help": "min sample size to skip small examples"}
)
# Options for reporting WER metrics during validation. Only applicable to
# Seq2Seq models during fine-tuning
eval_wer: bool = field(
default=False, metadata={"help": "compute WER for Seq2Seq models"}
)
eval_wer_config: GenerationConfig = field(
default_factory=lambda: GenerationConfig(),
metadata={"help": "beam search config for evaluating wer during training"},
)
eval_wer_tokenizer: Any = field(
default=None,
metadata={"help": "tokenizer config for evaluating wer during training"},
)
eval_wer_post_process: str = field(
default="letter",
metadata={
"help": "remove BPE tokens before scoring (can be sentencepiece, letter, and more)"
},
)
autoregressive: bool = field(
default=False,
metadata={
"help": "required for autoregressive decoders (like seq2seq models); "
"adds 'prev_output_tokens' to input and appends eos to target"
},
)
@register_task("audio_pretraining", dataclass=AudioPretrainingConfig)
class AudioPretrainingTask(FairseqTask):
""""""
cfg: AudioPretrainingConfig
def __init__(
self,
cfg: AudioPretrainingConfig,
):
super().__init__(cfg)
if cfg.eval_wer:
assert cfg.labels is not None, "eval_wer can only be set during fine-tuning"
self.blank_symbol = "<s>"
self.state.add_factory("target_dictionary", self.load_target_dictionary)
@classmethod
def setup_task(cls, cfg: AudioPretrainingConfig, **kwargs):
"""Setup the task (e.g., load dictionaries).
Args:
cfg (AudioPretrainingConfig): configuration of this task
"""
return cls(cfg)
def load_target_dictionary(self):
if self.cfg.labels:
dict_path = os.path.join(self.cfg.data, f"dict.{self.cfg.labels}.txt")
return Dictionary.load(dict_path)
return None
def load_dataset(self, split: str, task_cfg: FairseqDataclass = None, **kwargs):
data_path = self.cfg.data
task_cfg = task_cfg or self.cfg
# upgrade old task
if isinstance(task_cfg, Namespace):
if not hasattr(task_cfg, "autoregressive"):
task_cfg.autoregressive = not task_cfg.criterion == 'ctc'
manifest = os.path.join(data_path, "{}.tsv".format(split))
self.datasets[split] = FileAudioDataset(
manifest,
sample_rate=task_cfg.get('sample_rate', self.cfg.sample_rate),
max_sample_size=self.cfg.max_sample_size,
min_sample_size=self.cfg.min_sample_size,
pad=task_cfg.labels is not None or task_cfg.enable_padding,
normalize=task_cfg.normalize,
)
if task_cfg.labels:
label_path = os.path.join(data_path, f"{split}.{task_cfg.labels}")
with open(label_path, "r") as f:
labels = [
line for i, line in enumerate(f)
if i in self.datasets[split].line_inds
]
assert len(labels) == len(self.datasets[split]), (
f"labels length ({len(labels)}) and dataset length "
f"({len(self.datasets[split])}) do not match")
process_label = LabelEncoder(self.target_dictionary)
self.datasets[split] = AddTargetDataset(
self.datasets[split],
labels,
pad=self.target_dictionary.pad(),
eos=self.target_dictionary.eos(),
batch_targets=True,
process_label=process_label,
add_to_input=task_cfg.get('autoregressive', False),
)
@property
def source_dictionary(self):
return None
@property
def target_dictionary(self):
"""Return the :class:`~fairseq.data.Dictionary` for the language
model."""
return self.state.target_dictionary
def max_positions(self):
"""Maximum input length supported by the encoder."""
return (sys.maxsize, sys.maxsize)
def filter_indices_by_size(
self,
indices,
dataset,
max_positions=None,
ignore_invalid_inputs=False,
):
# we do not need to filter by size in this task as dataloaders take care of this
return indices
def valid_step(self, sample, model, criterion):
loss, sample_size, logging_output = super().valid_step(sample, model, criterion)
if self.cfg.eval_wer and self.cfg.autoregressive:
metrics = self._inference_with_wer(self.sequence_generator, sample, model)
logging_output["_num_char_errors"] = metrics["num_char_errors"]
logging_output["_num_chars"] = metrics["num_chars"]
logging_output["_num_word_errors"] = metrics["num_word_errors"]
logging_output["_num_words"] = metrics["num_words"]
return loss, sample_size, logging_output
def build_model(self, model_cfg: FairseqDataclass):
model = super().build_model(model_cfg)
if self.cfg.eval_wer and self.cfg.autoregressive:
self.sequence_generator = self.build_generator(
[model],
self.cfg.eval_wer_config,
)
if self.cfg.eval_wer_tokenizer:
self.tokenizer = encoders.build_tokenizer(self.cfg.eval_wer_tokenizer)
else:
self.tokenizer = None
return model
def _inference_with_wer(self, generator, sample, model):
import editdistance
def decode(toks):
s = self.target_dictionary.string(
toks.int().cpu(),
self.cfg.eval_wer_post_process,
escape_unk=True,
)
if self.tokenizer:
s = self.tokenizer.decode(s)
return s
num_word_errors, num_char_errors = 0, 0
num_chars, num_words = 0, 0
gen_out = self.inference_step(generator, [model], sample, None)
for i in range(len(gen_out)):
hyp = decode(gen_out[i][0]["tokens"])
ref = decode(
utils.strip_pad(sample["target"][i], self.target_dictionary.pad()),
)
num_char_errors += editdistance.eval(hyp, ref)
num_chars += len(ref)
hyp_words = hyp.split()
ref_words = ref.split()
num_word_errors += editdistance.eval(hyp_words, ref_words)
num_words += len(ref_words)
return {
"num_char_errors": num_char_errors,
"num_chars": num_chars,
"num_word_errors": num_word_errors,
"num_words": num_words,
}
def reduce_metrics(self, logging_outputs, criterion):
super().reduce_metrics(logging_outputs, criterion)
zero = torch.scalar_tensor(0.0)
num_char_errors = sum(
log.get("_num_char_errors", zero) for log in logging_outputs
)
num_chars = sum(log.get("_num_chars", zero) for log in logging_outputs)
num_word_errors = sum(
log.get("_num_word_errors", zero) for log in logging_outputs
)
num_words = sum(log.get("_num_words", zero) for log in logging_outputs)
metrics.log_scalar("_num_char_errors", num_char_errors)
metrics.log_scalar("_num_chars", num_chars)
metrics.log_scalar("_num_word_errors", num_word_errors)
metrics.log_scalar("_num_words", num_words)
if num_words > 0:
metrics.log_derived(
"uer",
lambda meters: meters["_num_char_errors"].sum
* 100.0
/ meters["_num_chars"].sum
if meters["_num_chars"].sum > 0
else float("nan"),
)
metrics.log_derived(
"wer",
lambda meters: meters["_num_word_errors"].sum
* 100.0
/ meters["_num_words"].sum
if meters["_num_words"].sum > 0
else float("nan"),
)
|