|
import torch |
|
import torchaudio |
|
import typing as T |
|
|
|
|
|
class MelspecDiscriminator(torch.nn.Module): |
|
"""mel spectrogram (frequency domain) discriminator""" |
|
|
|
def __init__(self) -> None: |
|
super().__init__() |
|
self.SAMPLE_RATE = 48000 |
|
|
|
self._melspec = torchaudio.transforms.MelSpectrogram( |
|
sample_rate=self.SAMPLE_RATE, |
|
n_fft=2048, |
|
win_length=int(0.025 * self.SAMPLE_RATE), |
|
hop_length=int(0.010 * self.SAMPLE_RATE), |
|
n_mels=128, |
|
power=1, |
|
) |
|
|
|
|
|
kernel_sizes = [(7, 7), (4, 4), (4, 4), (4, 4)] |
|
strides = [(1, 2), (1, 2), (1, 2), (1, 2)] |
|
self._convs = torch.nn.ModuleList( |
|
[ |
|
torch.nn.Sequential( |
|
torch.nn.Conv2d( |
|
in_channels=1 if i == 0 else 32, |
|
out_channels=64, |
|
kernel_size=k, |
|
stride=s, |
|
padding=(1, 2), |
|
bias=False, |
|
), |
|
torch.nn.BatchNorm2d(num_features=64), |
|
torch.nn.GLU(dim=1), |
|
) |
|
for i, (k, s) in enumerate(zip(kernel_sizes, strides)) |
|
] |
|
) |
|
|
|
|
|
self._postnet = torch.nn.Conv2d( |
|
in_channels=32, |
|
out_channels=1, |
|
kernel_size=(15, 3), |
|
stride=(1, 2), |
|
) |
|
|
|
def forward(self, x: torch.Tensor) -> T.Tuple[torch.Tensor, T.List[torch.Tensor]]: |
|
|
|
x = torch.log(self._melspec(x) + 1e-5) |
|
|
|
|
|
f = [] |
|
for c in self._convs: |
|
x = c(x) |
|
f.append(x) |
|
|
|
|
|
x = self._postnet(x) |
|
x = x.mean(dim=[-2, -1]) |
|
|
|
return [(f, x)] |
|
|