File size: 12,888 Bytes
0b32ad6 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 |
import logging
import os
import random
from dataclasses import dataclass
from typing import Callable, List
import torch
import torchaudio
from s3prl.dataio.encoder.category import CategoryEncoder
from s3prl.dataio.encoder.g2p import G2P
from s3prl.dataio.encoder.tokenizer import (
Tokenizer,
default_phoneme_tokenizer,
load_tokenizer,
)
from s3prl.dataio.encoder.vocabulary import generate_vocab
from .base import AugmentedDynamicItemDataset, DataPipe
logger = logging.getLogger(__name__)
class SetOutputKeys(DataPipe):
def __init__(self, output_keys: dict = None) -> None:
super().__init__()
self.output_keys = output_keys
def forward(self, dataset: AugmentedDynamicItemDataset):
dataset.update_output_keys(self.output_keys)
return dataset
@dataclass
class LoadAudio(DataPipe):
audio_sample_rate: int = 16000
audio_channel_reduction: str = "first"
sox_effects: list = None
wav_path_name: str = "wav_path"
wav_name: str = "wav"
start_sec_name: str = "start_sec"
end_sec_name: str = "end_sec"
def load_audio(
self,
wav_path,
start_sec: float = None,
end_sec: float = None,
):
crop_segment = start_sec is not None and end_sec is not None
torchaudio.set_audio_backend("sox_io")
wav, sr = torchaudio.load(
wav_path,
frame_offset=round(start_sec * self.audio_sample_rate)
if crop_segment
else 0,
num_frames=round((end_sec - start_sec) * self.audio_sample_rate)
if crop_segment
else -1,
)
if self.sox_effects is not None:
wav, sr = torchaudio.sox_effects.apply_effects_tensor(
wav, sr, effects=self.sox_effects
)
if sr != self.audio_sample_rate:
resampler = torchaudio.transforms.Resample(sr, self.audio_sample_rate)
wav = resampler(wav)
if self.audio_channel_reduction == "first":
wav = wav[0]
elif self.audio_channel_reduction == "mean":
wav = wav.mean(dim=0)
wav = wav.view(-1, 1)
return wav
def compute_length(self, wav):
return len(wav)
def forward(self, dataset: AugmentedDynamicItemDataset):
item = dataset[0]
if self.start_sec_name in item and self.end_sec_name in item:
crop_segment = True
else:
crop_segment = False
if not crop_segment:
dataset.add_dynamic_item(
self.load_audio, takes=self.wav_path_name, provides=self.wav_name
)
else:
dataset.add_dynamic_item(
self.load_audio,
takes=[self.wav_path_name, self.start_sec_name, self.end_sec_name],
provides=self.wav_name,
)
dataset.add_dynamic_item(
self.compute_length,
takes=self.wav_name,
provides=f"{self.wav_name}_len",
)
return dataset
@dataclass
class EncodeCategory(DataPipe):
train_category_encoder: bool = False
label_name: str = "label"
category_encoder_name: str = "category"
encoded_target_name: str = "class_id"
def prepare_category(self, labels):
return CategoryEncoder(sorted(list(set(labels))))
def encode_label(self, category, label):
return category.encode(label)
def forward(self, dataset: AugmentedDynamicItemDataset):
if self.train_category_encoder:
with dataset.output_keys_as([self.label_name]):
labels = [item[self.label_name] for item in dataset]
category = self.prepare_category(labels)
dataset.add_tool(self.category_encoder_name, category)
category = dataset.get_tool(self.category_encoder_name)
dataset.add_tool("output_size", len(category))
dataset.add_dynamic_item(
self.encode_label,
takes=[self.category_encoder_name, self.label_name],
provides=self.encoded_target_name,
)
return dataset
@dataclass
class EncodeMultipleCategory(EncodeCategory):
train_category_encoder: bool = False
label_name: str = "labels"
category_encoder_name: str = "categories"
encoded_target_name: str = "class_ids"
def encode_label(self, categories, labels):
return torch.LongTensor(
[category.encode(label) for category, label in zip(categories, labels)]
)
def forward(self, dataset: AugmentedDynamicItemDataset):
if self.train_category_encoder:
with dataset.output_keys_as([self.label_name]):
labels = [item[self.label_name] for item in dataset]
label_types = list(zip(*labels))
categories = [
self.prepare_category(label_type) for label_type in label_types
]
dataset.add_tool(self.category_encoder_name, categories)
dataset.add_tool("output_size", sum([len(c) for c in categories]))
dataset.add_dynamic_item(
self.encode_label,
takes=[self.category_encoder_name, self.label_name],
provides=self.encoded_target_name,
)
return dataset
@dataclass
class EncodeMultiLabel(DataPipe):
label_name: str = "labels"
category_encoder_name: str = "category"
encoded_target_name: str = "binary_labels"
@staticmethod
def label_to_binary_vector(label: List, num_labels: int) -> torch.Tensor:
# Lame special case for multilabel with no labels
if len(label) == 0:
# BCEWithLogitsLoss wants float not long targets
binary_labels = torch.zeros((num_labels,), dtype=torch.float)
else:
binary_labels = torch.zeros((num_labels,)).scatter(
0, torch.tensor(label), 1.0
)
# Validate the binary vector we just created
assert set(torch.where(binary_labels == 1.0)[0].numpy()) == set(label)
return binary_labels
def encode_label(self, category, labels):
labels = [category.encode(label) for label in labels]
binary_labels = self.label_to_binary_vector(labels, len(category))
return binary_labels
def forward(self, dataset: AugmentedDynamicItemDataset):
if not dataset.has_tool(self.category_encoder_name):
with dataset.output_keys_as([self.label_name]):
all_labels = []
for item in dataset:
all_labels.extend(item[self.label_name])
all_labels.sort()
all_labels = set(all_labels)
category = CategoryEncoder(all_labels)
dataset.add_tool(self.category_encoder_name, category)
category = dataset.get_tool(self.category_encoder_name)
dataset.add_tool("output_size", len(category))
dataset.add_dynamic_item(
self.encode_label,
takes=[self.category_encoder_name, self.label_name],
provides=self.encoded_target_name,
)
return dataset
@dataclass
class GenerateTokenizer(DataPipe):
generate: bool = True
tokenizer_name: str = "tokenizer"
text_name: str = "transcription"
vocab_type: str = "character"
text_file: str = None
vocab_file: str = None
slots_file: str = None
vocab_args: dict = None
def prepare_tokenizer(self, text_list: str = None) -> Tokenizer:
"""Generates tokenizer from text data.
Args:
text_list (str, optional): List of text. Defaults to None.
Returns:
Tokenizer: Generated tokenizer
"""
vocab_args = self.vocab_args or {}
assert isinstance(vocab_args, dict)
if text_list is not None:
vocab_result = generate_vocab(
self.vocab_type, text_list=text_list, **vocab_args
)
else:
vocab_result = generate_vocab(
self.vocab_type, text_file=self.text_file, **vocab_args
)
vocab_list = vocab_result if isinstance(vocab_result, list) else None
vocab_file = vocab_result if isinstance(vocab_result, str) else None
tokenizer = load_tokenizer(
self.vocab_type,
vocab_file=vocab_file,
vocab_list=vocab_list,
slots_file=self.slots_file,
)
return tokenizer
def forward(self, dataset: AugmentedDynamicItemDataset):
try:
tokenizer = dataset.get_tool(self.tokenizer_name)
logger.info(
f"Tokenizer (name = {self.tokenizer_name}) exists in dataset, skip generation."
)
except KeyError:
if self.generate:
if self.vocab_file is not None and os.path.exists(self.vocab_file):
tokenizer = load_tokenizer(
self.vocab_type,
vocab_file=self.vocab_file,
slots_file=self.slots_file,
)
else:
text_list = None
if self.text_file is None:
with dataset.output_keys_as([self.text_name]):
text_list = [item[self.text_name] for item in dataset]
tokenizer = self.prepare_tokenizer(text_list)
dataset.add_tool(self.tokenizer_name, tokenizer)
else:
logger.warning(
"No tokenizer is found or generated. No-op for this DataPipe"
)
return dataset
@dataclass
class EncodeText(DataPipe):
text_name: str = "transcription"
output_text_name: str = "tokenized_text"
tokenizer_name: str = "tokenizer"
def encode_text(self, tokenizer: Tokenizer, text: str) -> torch.LongTensor:
return torch.LongTensor(tokenizer.encode(text))
def forward(self, dataset: AugmentedDynamicItemDataset):
try:
tokenizer = dataset.get_tool(self.tokenizer_name)
except KeyError:
raise KeyError(f"Tokenizer (name = {self.tokenizer_name}) not found!")
dataset.add_dynamic_item(
self.encode_text,
takes=[self.tokenizer_name, self.text_name],
provides=self.output_text_name,
)
dataset.add_tool("output_size", tokenizer.vocab_size)
return dataset
@dataclass
class Phonemize(DataPipe):
text_name: str = "transcription"
phonemized_text_name: str = "phonemized_text"
output_text_name: str = "tokenized_text"
g2p_name: str = "g2p"
tokenizer_name: str = "tokenizer"
def grapheme2phoneme(self, g2p: G2P, text: str) -> str:
return g2p.encode(text)
def encode_text(self, tokenizer: Tokenizer, text: str) -> torch.LongTensor:
return torch.LongTensor(tokenizer.encode(text))
def forward(self, dataset: AugmentedDynamicItemDataset):
if not dataset.has_tool(self.g2p_name):
logger.warning(
f"Cannot find {self.g2p_name} in dataset, use default G2P instead."
)
dataset.add_tool(self.g2p_name, G2P())
if not dataset.has_tool(self.tokenizer_name):
logger.warning(
f"Cannot find {self.tokenizer_name} in dataset, use default tokenizer instead."
)
dataset.add_tool(self.tokenizer_name, default_phoneme_tokenizer())
dataset.add_dynamic_item(
self.grapheme2phoneme,
takes=[self.g2p_name, self.text_name],
provides=self.phonemized_text_name,
)
dataset.add_dynamic_item(
self.encode_text,
takes=[self.tokenizer_name, self.phonemized_text_name],
provides=self.output_text_name,
)
tokenizer = dataset.get_tool(self.tokenizer_name)
dataset.add_tool("output_size", tokenizer.vocab_size)
return dataset
@dataclass
class RandomCrop(DataPipe):
"""
Completely randomized for every batch even with the same datapoint id.
Only suitable for training.
"""
sample_rate: int = 16000
max_secs: float = None
wav_name: str = "wav"
crop_name: str = "wav_crop"
def crop_wav(self, wav):
if self.max_secs is not None and wav.size(0) > self.max_secs * self.sample_rate:
start = random.randint(0, wav.size(0) - self.max_secs * self.sample_rate)
end = start + self.max_secs * self.sample_rate
wav = wav[round(start) : round(end)]
return wav, wav.size(0)
def forward(
self, dataset: AugmentedDynamicItemDataset
) -> AugmentedDynamicItemDataset:
dataset.add_dynamic_item(
self.crop_wav,
takes=[self.wav_name],
provides=[self.crop_name, f"{self.crop_name}_len"],
)
return dataset
|