Spaces:
Sleeping
Sleeping
File size: 5,223 Bytes
32b2aaa |
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 |
import logging
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from ..melspec import MelSpectrogram
from .hparams import HParams
from .unet import UNet
logger = logging.getLogger(__name__)
def _normalize(x: Tensor) -> Tensor:
return x / (x.abs().max(dim=-1, keepdim=True).values + 1e-7)
class Denoiser(nn.Module):
@property
def stft_cfg(self) -> dict:
hop_size = self.hp.hop_size
return dict(hop_length=hop_size, n_fft=hop_size * 4, win_length=hop_size * 4)
@property
def n_fft(self):
return self.stft_cfg["n_fft"]
@property
def eps(self):
return 1e-7
def __init__(self, hp: HParams):
super().__init__()
self.hp = hp
self.net = UNet(input_dim=3, output_dim=3)
self.mel_fn = MelSpectrogram(hp)
self.dummy: Tensor
self.register_buffer("dummy", torch.zeros(1), persistent=False)
def to_mel(self, x: Tensor, drop_last=True):
"""
Args:
x: (b t), wavs
Returns:
o: (b c t), mels
"""
if drop_last:
return self.mel_fn(x)[..., :-1] # (b d t)
return self.mel_fn(x)
def _stft(self, x):
"""
Args:
x: (b t)
Returns:
mag: (b f t) in [0, inf)
cos: (b f t) in [-1, 1]
sin: (b f t) in [-1, 1]
"""
dtype = x.dtype
device = x.device
if x.is_mps:
x = x.cpu()
window = torch.hann_window(self.stft_cfg["win_length"], device=x.device)
s = torch.stft(x.float(), **self.stft_cfg, window=window, return_complex=True) # (b f t+1)
s = s[..., :-1] # (b f t)
mag = s.abs() # (b f t)
phi = s.angle() # (b f t)
cos = phi.cos() # (b f t)
sin = phi.sin() # (b f t)
mag = mag.to(dtype=dtype, device=device)
cos = cos.to(dtype=dtype, device=device)
sin = sin.to(dtype=dtype, device=device)
return mag, cos, sin
def _istft(self, mag: Tensor, cos: Tensor, sin: Tensor):
"""
Args:
mag: (b f t) in [0, inf)
cos: (b f t) in [-1, 1]
sin: (b f t) in [-1, 1]
Returns:
x: (b t)
"""
device = mag.device
dtype = mag.dtype
if mag.is_mps:
mag = mag.cpu()
cos = cos.cpu()
sin = sin.cpu()
real = mag * cos # (b f t)
imag = mag * sin # (b f t)
s = torch.complex(real, imag) # (b f t)
if s.isnan().any():
logger.warning("NaN detected in ISTFT input.")
s = F.pad(s, (0, 1), "replicate") # (b f t+1)
window = torch.hann_window(self.stft_cfg["win_length"], device=s.device)
x = torch.istft(s, **self.stft_cfg, window=window, return_complex=False)
if x.isnan().any():
logger.warning("NaN detected in ISTFT output, set to zero.")
x = torch.where(x.isnan(), torch.zeros_like(x), x)
x = x.to(dtype=dtype, device=device)
return x
def _magphase(self, real, imag):
mag = (real.pow(2) + imag.pow(2) + self.eps).sqrt()
cos = real / mag
sin = imag / mag
return mag, cos, sin
def _predict(self, mag: Tensor, cos: Tensor, sin: Tensor):
"""
Args:
mag: (b f t)
cos: (b f t)
sin: (b f t)
Returns:
mag_mask: (b f t) in [0, 1], magnitude mask
cos_res: (b f t) in [-1, 1], phase residual
sin_res: (b f t) in [-1, 1], phase residual
"""
x = torch.stack([mag, cos, sin], dim=1) # (b 3 f t)
mag_mask, real, imag = self.net(x).unbind(1) # (b 3 f t)
mag_mask = mag_mask.sigmoid() # (b f t)
real = real.tanh() # (b f t)
imag = imag.tanh() # (b f t)
_, cos_res, sin_res = self._magphase(real, imag) # (b f t)
return mag_mask, sin_res, cos_res
def _separate(self, mag, cos, sin, mag_mask, cos_res, sin_res):
"""Ref: https://audio-agi.github.io/Separate-Anything-You-Describe/AudioSep_arXiv.pdf"""
sep_mag = F.relu(mag * mag_mask)
sep_cos = cos * cos_res - sin * sin_res
sep_sin = sin * cos_res + cos * sin_res
return sep_mag, sep_cos, sep_sin
def forward(self, x: Tensor, y: Tensor | None = None):
"""
Args:
x: (b t), a mixed audio
y: (b t), a fg audio
"""
assert x.dim() == 2, f"Expected (b t), got {x.size()}"
x = x.to(self.dummy)
x = _normalize(x)
if y is not None:
assert y.dim() == 2, f"Expected (b t), got {y.size()}"
y = y.to(self.dummy)
y = _normalize(y)
mag, cos, sin = self._stft(x) # (b 2f t)
mag_mask, sin_res, cos_res = self._predict(mag, cos, sin)
sep_mag, sep_cos, sep_sin = self._separate(mag, cos, sin, mag_mask, cos_res, sin_res)
o = self._istft(sep_mag, sep_cos, sep_sin)
npad = x.shape[-1] - o.shape[-1]
o = F.pad(o, (0, npad))
if y is not None:
self.losses = dict(l1=F.l1_loss(o, y))
return o
|