File size: 18,736 Bytes
75decff |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 |
import os
import traceback
import logging
logger = logging.getLogger(__name__)
import numpy as np
import torch
import torch.utils.data
from infer.lib.train.mel_processing import spectrogram_torch
from infer.lib.train.utils import load_filepaths_and_text, load_wav_to_torch
class TextAudioLoaderMultiNSFsid(torch.utils.data.Dataset):
"""
1) loads audio, text pairs
2) normalizes text and converts them to sequences of integers
3) computes spectrograms from audio files.
"""
def __init__(self, audiopaths_and_text, hparams):
self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
self.max_wav_value = hparams.max_wav_value
self.sampling_rate = hparams.sampling_rate
self.filter_length = hparams.filter_length
self.hop_length = hparams.hop_length
self.win_length = hparams.win_length
self.sampling_rate = hparams.sampling_rate
self.min_text_len = getattr(hparams, "min_text_len", 1)
self.max_text_len = getattr(hparams, "max_text_len", 5000)
self._filter()
def _filter(self):
"""
Filter text & store spec lengths
"""
# Store spectrogram lengths for Bucketing
# wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
# spec_length = wav_length // hop_length
audiopaths_and_text_new = []
lengths = []
for audiopath, text, pitch, pitchf, dv in self.audiopaths_and_text:
if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
audiopaths_and_text_new.append([audiopath, text, pitch, pitchf, dv])
lengths.append(os.path.getsize(audiopath) // (3 * self.hop_length))
self.audiopaths_and_text = audiopaths_and_text_new
self.lengths = lengths
def get_sid(self, sid):
sid = torch.LongTensor([int(sid)])
return sid
def get_audio_text_pair(self, audiopath_and_text):
# separate filename and text
file = audiopath_and_text[0]
phone = audiopath_and_text[1]
pitch = audiopath_and_text[2]
pitchf = audiopath_and_text[3]
dv = audiopath_and_text[4]
phone, pitch, pitchf = self.get_labels(phone, pitch, pitchf)
spec, wav = self.get_audio(file)
dv = self.get_sid(dv)
len_phone = phone.size()[0]
len_spec = spec.size()[-1]
# print(123,phone.shape,pitch.shape,spec.shape)
if len_phone != len_spec:
len_min = min(len_phone, len_spec)
# amor
len_wav = len_min * self.hop_length
spec = spec[:, :len_min]
wav = wav[:, :len_wav]
phone = phone[:len_min, :]
pitch = pitch[:len_min]
pitchf = pitchf[:len_min]
return (spec, wav, phone, pitch, pitchf, dv)
def get_labels(self, phone, pitch, pitchf):
phone = np.load(phone)
phone = np.repeat(phone, 2, axis=0)
pitch = np.load(pitch)
pitchf = np.load(pitchf)
n_num = min(phone.shape[0], 900) # DistributedBucketSampler
# print(234,phone.shape,pitch.shape)
phone = phone[:n_num, :]
pitch = pitch[:n_num]
pitchf = pitchf[:n_num]
phone = torch.FloatTensor(phone)
pitch = torch.LongTensor(pitch)
pitchf = torch.FloatTensor(pitchf)
return phone, pitch, pitchf
def get_audio(self, filename):
audio, sampling_rate = load_wav_to_torch(filename)
if sampling_rate != self.sampling_rate:
raise ValueError(
"{} SR doesn't match target {} SR".format(
sampling_rate, self.sampling_rate
)
)
audio_norm = audio
# audio_norm = audio / self.max_wav_value
# audio_norm = audio / np.abs(audio).max()
audio_norm = audio_norm.unsqueeze(0)
spec_filename = filename.replace(".wav", ".spec.pt")
if os.path.exists(spec_filename):
try:
spec = torch.load(spec_filename)
except:
logger.warning("%s %s", spec_filename, traceback.format_exc())
spec = spectrogram_torch(
audio_norm,
self.filter_length,
self.sampling_rate,
self.hop_length,
self.win_length,
center=False,
)
spec = torch.squeeze(spec, 0)
torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
else:
spec = spectrogram_torch(
audio_norm,
self.filter_length,
self.sampling_rate,
self.hop_length,
self.win_length,
center=False,
)
spec = torch.squeeze(spec, 0)
torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
return spec, audio_norm
def __getitem__(self, index):
return self.get_audio_text_pair(self.audiopaths_and_text[index])
def __len__(self):
return len(self.audiopaths_and_text)
class TextAudioCollateMultiNSFsid:
"""Zero-pads model inputs and targets"""
def __init__(self, return_ids=False):
self.return_ids = return_ids
def __call__(self, batch):
"""Collate's training batch from normalized text and aduio
PARAMS
------
batch: [text_normalized, spec_normalized, wav_normalized]
"""
# Right zero-pad all one-hot text sequences to max input length
_, ids_sorted_decreasing = torch.sort(
torch.LongTensor([x[0].size(1) for x in batch]), dim=0, descending=True
)
max_spec_len = max([x[0].size(1) for x in batch])
max_wave_len = max([x[1].size(1) for x in batch])
spec_lengths = torch.LongTensor(len(batch))
wave_lengths = torch.LongTensor(len(batch))
spec_padded = torch.FloatTensor(len(batch), batch[0][0].size(0), max_spec_len)
wave_padded = torch.FloatTensor(len(batch), 1, max_wave_len)
spec_padded.zero_()
wave_padded.zero_()
max_phone_len = max([x[2].size(0) for x in batch])
phone_lengths = torch.LongTensor(len(batch))
phone_padded = torch.FloatTensor(
len(batch), max_phone_len, batch[0][2].shape[1]
) # (spec, wav, phone, pitch)
pitch_padded = torch.LongTensor(len(batch), max_phone_len)
pitchf_padded = torch.FloatTensor(len(batch), max_phone_len)
phone_padded.zero_()
pitch_padded.zero_()
pitchf_padded.zero_()
# dv = torch.FloatTensor(len(batch), 256)#gin=256
sid = torch.LongTensor(len(batch))
for i in range(len(ids_sorted_decreasing)):
row = batch[ids_sorted_decreasing[i]]
spec = row[0]
spec_padded[i, :, : spec.size(1)] = spec
spec_lengths[i] = spec.size(1)
wave = row[1]
wave_padded[i, :, : wave.size(1)] = wave
wave_lengths[i] = wave.size(1)
phone = row[2]
phone_padded[i, : phone.size(0), :] = phone
phone_lengths[i] = phone.size(0)
pitch = row[3]
pitch_padded[i, : pitch.size(0)] = pitch
pitchf = row[4]
pitchf_padded[i, : pitchf.size(0)] = pitchf
# dv[i] = row[5]
sid[i] = row[5]
return (
phone_padded,
phone_lengths,
pitch_padded,
pitchf_padded,
spec_padded,
spec_lengths,
wave_padded,
wave_lengths,
# dv
sid,
)
class TextAudioLoader(torch.utils.data.Dataset):
"""
1) loads audio, text pairs
2) normalizes text and converts them to sequences of integers
3) computes spectrograms from audio files.
"""
def __init__(self, audiopaths_and_text, hparams):
self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
self.max_wav_value = hparams.max_wav_value
self.sampling_rate = hparams.sampling_rate
self.filter_length = hparams.filter_length
self.hop_length = hparams.hop_length
self.win_length = hparams.win_length
self.sampling_rate = hparams.sampling_rate
self.min_text_len = getattr(hparams, "min_text_len", 1)
self.max_text_len = getattr(hparams, "max_text_len", 5000)
self._filter()
def _filter(self):
"""
Filter text & store spec lengths
"""
# Store spectrogram lengths for Bucketing
# wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
# spec_length = wav_length // hop_length
audiopaths_and_text_new = []
lengths = []
for audiopath, text, dv in self.audiopaths_and_text:
if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
audiopaths_and_text_new.append([audiopath, text, dv])
lengths.append(os.path.getsize(audiopath) // (3 * self.hop_length))
self.audiopaths_and_text = audiopaths_and_text_new
self.lengths = lengths
def get_sid(self, sid):
sid = torch.LongTensor([int(sid)])
return sid
def get_audio_text_pair(self, audiopath_and_text):
# separate filename and text
file = audiopath_and_text[0]
phone = audiopath_and_text[1]
dv = audiopath_and_text[2]
phone = self.get_labels(phone)
spec, wav = self.get_audio(file)
dv = self.get_sid(dv)
len_phone = phone.size()[0]
len_spec = spec.size()[-1]
if len_phone != len_spec:
len_min = min(len_phone, len_spec)
len_wav = len_min * self.hop_length
spec = spec[:, :len_min]
wav = wav[:, :len_wav]
phone = phone[:len_min, :]
return (spec, wav, phone, dv)
def get_labels(self, phone):
phone = np.load(phone)
phone = np.repeat(phone, 2, axis=0)
n_num = min(phone.shape[0], 900) # DistributedBucketSampler
phone = phone[:n_num, :]
phone = torch.FloatTensor(phone)
return phone
def get_audio(self, filename):
audio, sampling_rate = load_wav_to_torch(filename)
if sampling_rate != self.sampling_rate:
raise ValueError(
"{} SR doesn't match target {} SR".format(
sampling_rate, self.sampling_rate
)
)
audio_norm = audio
# audio_norm = audio / self.max_wav_value
# audio_norm = audio / np.abs(audio).max()
audio_norm = audio_norm.unsqueeze(0)
spec_filename = filename.replace(".wav", ".spec.pt")
if os.path.exists(spec_filename):
try:
spec = torch.load(spec_filename)
except:
logger.warning("%s %s", spec_filename, traceback.format_exc())
spec = spectrogram_torch(
audio_norm,
self.filter_length,
self.sampling_rate,
self.hop_length,
self.win_length,
center=False,
)
spec = torch.squeeze(spec, 0)
torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
else:
spec = spectrogram_torch(
audio_norm,
self.filter_length,
self.sampling_rate,
self.hop_length,
self.win_length,
center=False,
)
spec = torch.squeeze(spec, 0)
torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
return spec, audio_norm
def __getitem__(self, index):
return self.get_audio_text_pair(self.audiopaths_and_text[index])
def __len__(self):
return len(self.audiopaths_and_text)
class TextAudioCollate:
"""Zero-pads model inputs and targets"""
def __init__(self, return_ids=False):
self.return_ids = return_ids
def __call__(self, batch):
"""Collate's training batch from normalized text and aduio
PARAMS
------
batch: [text_normalized, spec_normalized, wav_normalized]
"""
# Right zero-pad all one-hot text sequences to max input length
_, ids_sorted_decreasing = torch.sort(
torch.LongTensor([x[0].size(1) for x in batch]), dim=0, descending=True
)
max_spec_len = max([x[0].size(1) for x in batch])
max_wave_len = max([x[1].size(1) for x in batch])
spec_lengths = torch.LongTensor(len(batch))
wave_lengths = torch.LongTensor(len(batch))
spec_padded = torch.FloatTensor(len(batch), batch[0][0].size(0), max_spec_len)
wave_padded = torch.FloatTensor(len(batch), 1, max_wave_len)
spec_padded.zero_()
wave_padded.zero_()
max_phone_len = max([x[2].size(0) for x in batch])
phone_lengths = torch.LongTensor(len(batch))
phone_padded = torch.FloatTensor(
len(batch), max_phone_len, batch[0][2].shape[1]
)
phone_padded.zero_()
sid = torch.LongTensor(len(batch))
for i in range(len(ids_sorted_decreasing)):
row = batch[ids_sorted_decreasing[i]]
spec = row[0]
spec_padded[i, :, : spec.size(1)] = spec
spec_lengths[i] = spec.size(1)
wave = row[1]
wave_padded[i, :, : wave.size(1)] = wave
wave_lengths[i] = wave.size(1)
phone = row[2]
phone_padded[i, : phone.size(0), :] = phone
phone_lengths[i] = phone.size(0)
sid[i] = row[3]
return (
phone_padded,
phone_lengths,
spec_padded,
spec_lengths,
wave_padded,
wave_lengths,
sid,
)
class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
"""
Maintain similar input lengths in a batch.
Length groups are specified by boundaries.
Ex) boundaries = [b1, b2, b3] -> any batch is included either {x | b1 < length(x) <=b2} or {x | b2 < length(x) <= b3}.
It removes samples which are not included in the boundaries.
Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1 or length(x) > b3 are discarded.
"""
def __init__(
self,
dataset,
batch_size,
boundaries,
num_replicas=None,
rank=None,
shuffle=True,
):
super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
self.lengths = dataset.lengths
self.batch_size = batch_size
self.boundaries = boundaries
self.buckets, self.num_samples_per_bucket = self._create_buckets()
self.total_size = sum(self.num_samples_per_bucket)
self.num_samples = self.total_size // self.num_replicas
def _create_buckets(self):
buckets = [[] for _ in range(len(self.boundaries) - 1)]
for i in range(len(self.lengths)):
length = self.lengths[i]
idx_bucket = self._bisect(length)
if idx_bucket != -1:
buckets[idx_bucket].append(i)
for i in range(len(buckets) - 1, -1, -1): #
if len(buckets[i]) == 0:
buckets.pop(i)
self.boundaries.pop(i + 1)
num_samples_per_bucket = []
for i in range(len(buckets)):
len_bucket = len(buckets[i])
total_batch_size = self.num_replicas * self.batch_size
rem = (
total_batch_size - (len_bucket % total_batch_size)
) % total_batch_size
num_samples_per_bucket.append(len_bucket + rem)
return buckets, num_samples_per_bucket
def __iter__(self):
# deterministically shuffle based on epoch
g = torch.Generator()
g.manual_seed(self.epoch)
indices = []
if self.shuffle:
for bucket in self.buckets:
indices.append(torch.randperm(len(bucket), generator=g).tolist())
else:
for bucket in self.buckets:
indices.append(list(range(len(bucket))))
batches = []
for i in range(len(self.buckets)):
bucket = self.buckets[i]
len_bucket = len(bucket)
ids_bucket = indices[i]
num_samples_bucket = self.num_samples_per_bucket[i]
# add extra samples to make it evenly divisible
rem = num_samples_bucket - len_bucket
ids_bucket = (
ids_bucket
+ ids_bucket * (rem // len_bucket)
+ ids_bucket[: (rem % len_bucket)]
)
# subsample
ids_bucket = ids_bucket[self.rank :: self.num_replicas]
# batching
for j in range(len(ids_bucket) // self.batch_size):
batch = [
bucket[idx]
for idx in ids_bucket[
j * self.batch_size : (j + 1) * self.batch_size
]
]
batches.append(batch)
if self.shuffle:
batch_ids = torch.randperm(len(batches), generator=g).tolist()
batches = [batches[i] for i in batch_ids]
self.batches = batches
assert len(self.batches) * self.batch_size == self.num_samples
return iter(self.batches)
def _bisect(self, x, lo=0, hi=None):
if hi is None:
hi = len(self.boundaries) - 1
if hi > lo:
mid = (hi + lo) // 2
if self.boundaries[mid] < x and x <= self.boundaries[mid + 1]:
return mid
elif x <= self.boundaries[mid]:
return self._bisect(x, lo, mid)
else:
return self._bisect(x, mid + 1, hi)
else:
return -1
def __len__(self):
return self.num_samples // self.batch_size
|