vocos_pytorch/models/discriminator.py

from typing import List, Tuple

import torch
from torch import nn
from torch import Tensor
from torch.nn import Conv2d
from torch.nn.utils import weight_norm
from torchaudio.transforms import Spectrogram


class MultiPeriodDiscriminator(nn.Module):
    def __init__(self, periods: Tuple[int, ...] = (2, 3, 5, 7, 11)):
        super().__init__()
        self.discriminators = nn.ModuleList([DiscriminatorP(period=p) for p in periods])

    def forward(self, y: Tensor, y_hat: Tensor):
        y_d_rs = []
        y_d_gs = []
        fmap_rs = []
        fmap_gs = []
        for d in self.discriminators:
            y_d_r, fmap_r = d(y)
            y_d_g, fmap_g = d(y_hat)
            y_d_rs.append(y_d_r)
            fmap_rs.append(fmap_r)
            y_d_gs.append(y_d_g)
            fmap_gs.append(fmap_g)

        return y_d_rs, y_d_gs, fmap_rs, fmap_gs


class DiscriminatorP(nn.Module):
    def __init__(
        self,
        period: int,
        in_channels: int = 1,
        kernel_size: int = 5,
        stride: int = 3,
        lrelu_slope: float = 0.1,
    ):
        super().__init__()
        self.period = period
        self.convs = nn.ModuleList(
            [
                weight_norm(Conv2d(in_channels, 32, (kernel_size, 1), (stride, 1), padding=(kernel_size // 2, 0))),
                weight_norm(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(kernel_size // 2, 0))),
                weight_norm(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(kernel_size // 2, 0))),
                weight_norm(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(kernel_size // 2, 0))),
                weight_norm(Conv2d(1024, 1024, (kernel_size, 1), (1, 1), padding=(kernel_size // 2, 0))),
            ]
        )

        self.conv_post = weight_norm(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
        self.lrelu_slope = lrelu_slope

    def forward(self, x: Tensor) -> Tuple[Tensor, List[Tensor]]:
        fmap = []
        # 1d to 2d
        b, c, t = x.shape
        if t % self.period != 0:  # pad first
            n_pad = self.period - (t % self.period)
            x = torch.nn.functional.pad(x, (0, n_pad), "reflect")
            t = t + n_pad
        x = x.view(b, c, t // self.period, self.period)

        for i, l in enumerate(self.convs):
            x = l(x)
            x = torch.nn.functional.leaky_relu(x, self.lrelu_slope)
            if i > 0:
                fmap.append(x)
        x = self.conv_post(x)
        fmap.append(x)
        x = torch.flatten(x, 1, -1)

        return x, fmap


class MultiResolutionDiscriminator(nn.Module):
    def __init__(
        self,
        fft_sizes: Tuple[int, ...] = (2048, 1024, 512),
    ):
        """
        Multi-Resolution Discriminator module adapted from https://github.com/descriptinc/descript-audio-codec.

        Args:
            fft_sizes (tuple[int]): Tuple of window lengths for FFT. Defaults to (2048, 1024, 512).
        """

        super().__init__()
        self.discriminators = nn.ModuleList(
            [DiscriminatorR(window_length=w) for w in fft_sizes]
        )

    def forward(self, y: Tensor, y_hat: Tensor) -> Tuple[List[Tensor], List[Tensor], List[List[Tensor]], List[List[Tensor]]]:
        y_d_rs = []
        y_d_gs = []
        fmap_rs = []
        fmap_gs = []

        for d in self.discriminators:
            y_d_r, fmap_r = d(x=y)
            y_d_g, fmap_g = d(x=y_hat)
            y_d_rs.append(y_d_r)
            fmap_rs.append(fmap_r)
            y_d_gs.append(y_d_g)
            fmap_gs.append(fmap_g)

        return y_d_rs, y_d_gs, fmap_rs, fmap_gs


class DiscriminatorR(nn.Module):
    def __init__(
        self,
        window_length: int,
        channels: int = 32,
        hop_factor: float = 0.25,
        bands: Tuple[Tuple[float, float], ...] = ((0.0, 0.1), (0.1, 0.25), (0.25, 0.5), (0.5, 0.75), (0.75, 1.0)),
    ):
        super().__init__()
        self.window_length = window_length
        self.hop_factor = hop_factor
        self.spec_fn = Spectrogram(
            n_fft=window_length, hop_length=int(window_length * hop_factor), win_length=window_length, power=None
        )
        n_fft = window_length // 2 + 1
        bands = [(int(b[0] * n_fft), int(b[1] * n_fft)) for b in bands]
        self.bands = bands
        convs = lambda: nn.ModuleList(
            [
                weight_norm(nn.Conv2d(2, channels, (3, 9), (1, 1), padding=(1, 4))),
                weight_norm(nn.Conv2d(channels, channels, (3, 9), (1, 2), padding=(1, 4))),
                weight_norm(nn.Conv2d(channels, channels, (3, 9), (1, 2), padding=(1, 4))),
                weight_norm(nn.Conv2d(channels, channels, (3, 9), (1, 2), padding=(1, 4))),
                weight_norm(nn.Conv2d(channels, channels, (3, 3), (1, 1), padding=(1, 1))),
            ]
        )
        self.band_convs = nn.ModuleList([convs() for _ in range(len(self.bands))])

        self.conv_post = weight_norm(nn.Conv2d(channels, 1, (3, 3), (1, 1), padding=(1, 1)))

    def spectrogram(self, x):
        x = x.squeeze(1)
        # Remove DC offset
        x = x - x.mean(dim=-1, keepdims=True)
        # Peak normalize the volume of input audio
        x = 0.8 * x / (x.abs().max(dim=-1, keepdim=True)[0] + 1e-9)
        x = self.spec_fn(x)
        x = torch.view_as_real(x)
        x = x.permute(0, 3, 2, 1) # b f t c -> b c t f
        # Split into bands
        x_bands = [x[..., b[0] : b[1]] for b in self.bands]
        return x_bands

    def forward(self, x: Tensor):
        x_bands = self.spectrogram(x)
        fmap = []
        x = []
        for band, stack in zip(x_bands, self.band_convs):
            for i, layer in enumerate(stack):
                band = layer(band)
                band = torch.nn.functional.leaky_relu(band, 0.1)
                if i > 0:
                    fmap.append(band)
            x.append(band)
        x = torch.cat(x, dim=-1)
        x = self.conv_post(x)
        fmap.append(x)

        return x, fmap