Spaces:
Running
Running
File size: 14,366 Bytes
e86d760 2c1a5a6 e86d760 c797dfd e86d760 c797dfd e86d760 bbe1979 c797dfd ce6b38e e86d760 b0fda13 e86d760 9b1d5cc e86d760 9b1d5cc e86d760 b0fda13 e86d760 7f9e32d e86d760 7b7acb0 e86d760 4927a3a e86d760 7f9e32d e86d760 cba47e4 e86d760 cba47e4 e86d760 |
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 |
#!/usr/bin/python3
# -*- coding: utf-8 -*-
"""
https://zhuanlan.zhihu.com/p/627039860
https://github.com/facebookresearch/denoiser/blob/main/denoiser/stft_loss.py
"""
from typing import List
import torch
import torch.nn as nn
from torch.nn import functional as F
class LSDLoss(nn.Module):
"""
Log Spectral Distance
Mean square error of power spectrum
"""
def __init__(self,
n_fft: int = 512,
win_size: int = 512,
hop_size: int = 256,
center: bool = True,
eps: float = 1e-8,
reduction: str = "mean",
):
super(LSDLoss, self).__init__()
self.n_fft = n_fft
self.win_size = win_size
self.hop_size = hop_size
self.center = center
self.eps = eps
self.reduction = reduction
if reduction not in ("sum", "mean"):
raise AssertionError(f"param reduction must be sum or mean.")
def forward(self, denoise_power: torch.Tensor, clean_power: torch.Tensor):
"""
:param denoise_power: power spectrum of the estimated signal power spectrum (batch_size, ...)
:param clean_power: power spectrum of the target signal (batch_size, ...)
:return:
"""
denoise_power = denoise_power + self.eps
clean_power = clean_power + self.eps
log_denoise_power = torch.log10(denoise_power)
log_clean_power = torch.log10(clean_power)
# mean_square_error shape: [b, f]
mean_square_error = torch.mean(torch.square(log_denoise_power - log_clean_power), dim=-1)
if self.reduction == "mean":
lsd_loss = torch.mean(mean_square_error)
elif self.reduction == "sum":
lsd_loss = torch.sum(mean_square_error)
else:
raise AssertionError
return lsd_loss
class ComplexSpectralLoss(nn.Module):
def __init__(self,
n_fft: int = 512,
win_size: int = 512,
hop_size: int = 256,
center: bool = True,
eps: float = 1e-8,
reduction: str = "mean",
factor_mag: float = 0.5,
factor_pha: float = 0.3,
factor_gra: float = 0.2,
):
super().__init__()
self.n_fft = n_fft
self.win_size = win_size
self.hop_size = hop_size
self.center = center
self.eps = eps
self.reduction = reduction
self.factor_mag = factor_mag
self.factor_pha = factor_pha
self.factor_gra = factor_gra
if reduction not in ("sum", "mean"):
raise AssertionError(f"param reduction must be sum or mean.")
self.window = nn.Parameter(torch.hann_window(win_size), requires_grad=False)
def forward(self, denoise: torch.Tensor, clean: torch.Tensor):
"""
:param denoise: The estimated signal (batch_size, signal_length)
:param clean: The target signal (batch_size, signal_length)
:return:
"""
if denoise.shape != clean.shape:
raise AssertionError("Input signals must have the same shape")
# denoise_stft, clean_stft shape: [b, f, t]
denoise_stft = torch.stft(
denoise,
n_fft=self.n_fft,
win_length=self.win_size,
hop_length=self.hop_size,
window=self.window,
center=self.center,
pad_mode="reflect",
normalized=False,
return_complex=True
)
clean_stft = torch.stft(
clean,
n_fft=self.n_fft,
win_length=self.win_size,
hop_length=self.hop_size,
window=self.window,
center=self.center,
pad_mode="reflect",
normalized=False,
return_complex=True
)
# complex_diff shape: [b, f, t], dtype: torch.complex64
complex_diff = denoise_stft - clean_stft
# magnitude_diff, phase_diff shape: [b, f, t], dtype: torch.float32
magnitude_diff = torch.abs(complex_diff)
phase_diff = torch.angle(complex_diff)
# magnitude_loss, phase_loss shape: [b,]
magnitude_loss = torch.norm(magnitude_diff, p=2, dim=(-1, -2))
phase_loss = torch.norm(phase_diff, p=1, dim=(-1, -2))
# phase_grad shape: [b, f, t-1], dtype: torch.float32
phase_grad = torch.diff(torch.angle(denoise_stft), dim=-1)
grad_loss = torch.mean(torch.abs(phase_grad), dim=(-1, -2))
# loss, grad_loss shape: [b,]
batch_loss = self.factor_mag * magnitude_loss + self.factor_pha * phase_loss + self.factor_gra * grad_loss
# print(f"magnitude_loss: {magnitude_loss}")
# print(f"phase_loss: {phase_loss}")
# print(f"grad_loss: {grad_loss}")
if self.reduction == "mean":
loss = torch.mean(batch_loss)
elif self.reduction == "sum":
loss = torch.sum(batch_loss)
else:
raise AssertionError
return loss
class SpectralConvergenceLoss(torch.nn.Module):
"""Spectral convergence loss module."""
def __init__(self,
reduction: str = "mean",
eps: float = 1e-8,
):
super(SpectralConvergenceLoss, self).__init__()
self.reduction = reduction
self.eps = eps
if reduction not in ("sum", "mean"):
raise AssertionError(f"param reduction must be sum or mean.")
def forward(self,
denoise_magnitude: torch.Tensor,
clean_magnitude: torch.Tensor,
):
"""
:param denoise_magnitude: Tensor, shape: [batch_size, time_steps, freq_bins]
:param clean_magnitude: Tensor, shape: [batch_size, time_steps, freq_bins]
:return:
"""
error_norm = torch.norm(denoise_magnitude - clean_magnitude, p="fro", dim=(-1, -2))
truth_norm = torch.norm(clean_magnitude, p="fro", dim=(-1, -2))
batch_loss = error_norm / (truth_norm + self.eps)
if self.reduction == "mean":
loss = torch.mean(batch_loss)
elif self.reduction == "sum":
loss = torch.sum(batch_loss)
else:
raise AssertionError
return loss
class LogSTFTMagnitudeLoss(torch.nn.Module):
"""Log STFT magnitude loss module."""
def __init__(self,
reduction: str = "mean",
eps: float = 1e-8,
):
super(LogSTFTMagnitudeLoss, self).__init__()
self.reduction = reduction
self.eps = eps
if reduction not in ("sum", "mean"):
raise AssertionError(f"param reduction must be sum or mean.")
def forward(self,
denoise_magnitude: torch.Tensor,
clean_magnitude: torch.Tensor,
):
"""
:param denoise_magnitude: Tensor, shape: [batch_size, time_steps, freq_bins]
:param clean_magnitude: Tensor, shape: [batch_size, time_steps, freq_bins]
:return:
"""
loss = F.l1_loss(torch.log(denoise_magnitude + self.eps), torch.log(clean_magnitude + self.eps))
if torch.any(torch.isnan(loss)) or torch.any(torch.isinf(loss)):
raise AssertionError("SpectralConvergenceLoss, nan or inf in loss")
return loss
class STFTLoss(torch.nn.Module):
"""STFT loss module."""
def __init__(self,
n_fft: int = 1024,
win_size: int = 600,
hop_size: int = 120,
center: bool = True,
reduction: str = "mean",
):
super(STFTLoss, self).__init__()
self.n_fft = n_fft
self.win_size = win_size
self.hop_size = hop_size
self.center = center
self.reduction = reduction
self.window = nn.Parameter(torch.hann_window(win_size), requires_grad=False)
self.spectral_convergence_loss = SpectralConvergenceLoss(reduction=reduction)
self.log_stft_magnitude_loss = LogSTFTMagnitudeLoss(reduction=reduction)
def forward(self, denoise: torch.Tensor, clean: torch.Tensor):
"""
:param denoise:
:param clean:
:return:
"""
if denoise.shape != clean.shape:
raise AssertionError("Input signals must have the same shape")
# denoise_stft, clean_stft shape: [b, f, t]
denoise_stft = torch.stft(
denoise,
n_fft=self.n_fft,
win_length=self.win_size,
hop_length=self.hop_size,
window=self.window,
center=self.center,
pad_mode="reflect",
normalized=False,
return_complex=True
)
clean_stft = torch.stft(
clean,
n_fft=self.n_fft,
win_length=self.win_size,
hop_length=self.hop_size,
window=self.window,
center=self.center,
pad_mode="reflect",
normalized=False,
return_complex=True
)
denoise_magnitude = torch.abs(denoise_stft)
clean_magnitude = torch.abs(clean_stft)
sc_loss = self.spectral_convergence_loss.forward(denoise_magnitude, clean_magnitude)
mag_loss = self.log_stft_magnitude_loss.forward(denoise_magnitude, clean_magnitude)
return sc_loss, mag_loss
class MultiResolutionSTFTLoss(torch.nn.Module):
"""Multi resolution STFT loss module."""
def __init__(self,
fft_size_list: List[int] = None,
win_size_list: List[int] = None,
hop_size_list: List[int] = None,
factor_sc=0.1,
factor_mag=0.1,
reduction: str = "mean",
):
super(MultiResolutionSTFTLoss, self).__init__()
fft_size_list = fft_size_list or [512, 1024, 2048]
win_size_list = win_size_list or [240, 600, 1200]
hop_size_list = hop_size_list or [50, 120, 240]
if not len(fft_size_list) == len(win_size_list) == len(hop_size_list):
raise AssertionError
loss_fn_list = nn.ModuleList([])
for n_fft, win_size, hop_size in zip(fft_size_list, win_size_list, hop_size_list):
loss_fn_list.append(
STFTLoss(
n_fft=n_fft,
win_size=win_size,
hop_size=hop_size,
reduction=reduction,
)
)
self.loss_fn_list = loss_fn_list
self.factor_sc = factor_sc
self.factor_mag = factor_mag
def forward(self, denoise: torch.Tensor, clean: torch.Tensor):
"""
:param denoise:
:param clean:
:return:
"""
if denoise.shape != clean.shape:
raise AssertionError("Input signals must have the same shape")
sc_loss = 0.0
mag_loss = 0.0
for loss_fn in self.loss_fn_list:
sc_l, mag_l = loss_fn.forward(denoise, clean)
sc_loss += sc_l
mag_loss += mag_l
sc_loss = sc_loss / len(self.loss_fn_list)
mag_loss = mag_loss / len(self.loss_fn_list)
sc_loss = self.factor_sc * sc_loss
mag_loss = self.factor_mag * mag_loss
loss = sc_loss + mag_loss
return loss
class WeightedMagnitudePhaseLoss(nn.Module):
def __init__(self,
n_fft: int = 1024,
win_size: int = 600,
hop_size: int = 120,
center: bool = True,
reduction: str = "mean",
mag_weight: float = 0.9,
pha_weight: float = 0.3,
):
super(WeightedMagnitudePhaseLoss, self).__init__()
self.n_fft = n_fft
self.win_size = win_size
self.hop_size = hop_size
self.center = center
self.reduction = reduction
self.mag_weight = mag_weight
self.pha_weight = pha_weight
self.window = nn.Parameter(torch.hann_window(win_size), requires_grad=False)
def forward(self, denoise: torch.Tensor, clean: torch.Tensor):
"""
:param denoise:
:param clean:
:return:
"""
if denoise.shape != clean.shape:
raise AssertionError("Input signals must have the same shape")
# denoise_stft, clean_stft shape: [b, f, t]
denoise_stft = torch.stft(
denoise,
n_fft=self.n_fft,
win_length=self.win_size,
hop_length=self.hop_size,
window=self.window,
center=self.center,
pad_mode="reflect",
normalized=False,
return_complex=True
)
clean_stft = torch.stft(
clean,
n_fft=self.n_fft,
win_length=self.win_size,
hop_length=self.hop_size,
window=self.window,
center=self.center,
pad_mode="reflect",
normalized=False,
return_complex=True
)
denoise_stft_spec = torch.view_as_real(denoise_stft)
denoise_mag = torch.sqrt(denoise_stft_spec.pow(2).sum(-1) + 1e-9)
denoise_pha = torch.atan2(denoise_stft_spec[:, :, :, 1] + 1e-10, denoise_stft_spec[:, :, :, 0] + 1e-5)
clean_stft_spec = torch.view_as_real(clean_stft)
clean_mag = torch.sqrt(clean_stft_spec.pow(2).sum(-1) + 1e-9)
clean_pha = torch.atan2(clean_stft_spec[:, :, :, 1] + 1e-10, clean_stft_spec[:, :, :, 0] + 1e-5)
mag_loss = F.mse_loss(denoise_mag, clean_mag, reduction=self.reduction)
pha_loss = F.mse_loss(denoise_pha, clean_pha, reduction=self.reduction)
loss = self.mag_weight * mag_loss + self.pha_weight * pha_loss
return loss
def main():
batch_size = 2
signal_length = 16000
estimated_signal = torch.randn(batch_size, signal_length)
target_signal = torch.randn(batch_size, signal_length)
# loss_fn = LSDLoss()
# loss_fn = ComplexSpectralLoss()
# loss_fn = MultiResolutionSTFTLoss()
loss_fn = WeightedMagnitudePhaseLoss()
loss = loss_fn.forward(estimated_signal, target_signal)
print(f"loss: {loss.item()}")
return
if __name__ == "__main__":
main()
|