decoder/loss.py

from typing import List, Tuple

import torch
import torchaudio
from torch import nn

from decoder.modules import safe_log

import torch.nn.functional as F


class MelSpecReconstructionLoss(nn.Module):
    """
    L1 distance between the mel-scaled magnitude spectrograms of the ground truth sample and the generated sample
    """

    def __init__(
        self, sample_rate: int = 24000, n_fft: int = 1024, hop_length: int = 256, n_mels: int = 100,
    ):
        super().__init__()
        self.mel_spec = torchaudio.transforms.MelSpectrogram(
            sample_rate=sample_rate, n_fft=n_fft, hop_length=hop_length, n_mels=n_mels, center=True, power=1,
        )

    def forward(self, y_hat, y) -> torch.Tensor:
        """
        Args:
            y_hat (Tensor): Predicted audio waveform.
            y (Tensor): Ground truth audio waveform.

        Returns:
            Tensor: L1 loss between the mel-scaled magnitude spectrograms.
        """
        mel_hat = safe_log(self.mel_spec(y_hat))
        mel = safe_log(self.mel_spec(y))

        loss = torch.nn.functional.l1_loss(mel, mel_hat)

        return loss


class GeneratorLoss(nn.Module):
    """
    Generator Loss module. Calculates the loss for the generator based on discriminator outputs.
    """

    def forward(self, disc_outputs: List[torch.Tensor]) -> Tuple[torch.Tensor, List[torch.Tensor]]:
        """
        Args:
            disc_outputs (List[Tensor]): List of discriminator outputs.

        Returns:
            Tuple[Tensor, List[Tensor]]: Tuple containing the total loss and a list of loss values from
                                         the sub-discriminators
        """
        loss = 0
        gen_losses = []
        for dg in disc_outputs:
            l = torch.mean(torch.clamp(1 - dg, min=0))
            gen_losses.append(l)
            loss += l

        return loss, gen_losses


class DiscriminatorLoss(nn.Module):
    """
    Discriminator Loss module. Calculates the loss for the discriminator based on real and generated outputs.
    """

    def forward(
        self, disc_real_outputs: List[torch.Tensor], disc_generated_outputs: List[torch.Tensor]
    ) -> Tuple[torch.Tensor, List[torch.Tensor], List[torch.Tensor]]:
        """
        Args:
            disc_real_outputs (List[Tensor]): List of discriminator outputs for real samples.
            disc_generated_outputs (List[Tensor]): List of discriminator outputs for generated samples.

        Returns:
            Tuple[Tensor, List[Tensor], List[Tensor]]: A tuple containing the total loss, a list of loss values from
                                                       the sub-discriminators for real outputs, and a list of
                                                       loss values for generated outputs.
        """
        loss = 0
        r_losses = []
        g_losses = []
        for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
            r_loss = torch.mean(torch.clamp(1 - dr, min=0))
            g_loss = torch.mean(torch.clamp(1 + dg, min=0))
            loss += r_loss + g_loss
            r_losses.append(r_loss.item())
            g_losses.append(g_loss.item())

        return loss, r_losses, g_losses


class FeatureMatchingLoss(nn.Module):
    """
    Feature Matching Loss module. Calculates the feature matching loss between feature maps of the sub-discriminators.
    """

    def forward(self, fmap_r: List[List[torch.Tensor]], fmap_g: List[List[torch.Tensor]]) -> torch.Tensor:
        """
        Args:
            fmap_r (List[List[Tensor]]): List of feature maps from real samples.
            fmap_g (List[List[Tensor]]): List of feature maps from generated samples.

        Returns:
            Tensor: The calculated feature matching loss.
        """
        loss = 0
        for dr, dg in zip(fmap_r, fmap_g):
            for rl, gl in zip(dr, dg):
                loss += torch.mean(torch.abs(rl - gl))

        return loss

class DACGANLoss(nn.Module):
    """
    Computes a discriminator loss, given a discriminator on
    generated waveforms/spectrograms compared to ground truth
    waveforms/spectrograms. Computes the loss for both the
    discriminator and the generator in separate functions.
    """

    def __init__(self, discriminator):
        super().__init__()
        self.discriminator = discriminator

    def forward(self, fake, real):
        # d_fake = self.discriminator(fake.audio_data)
        # d_real = self.discriminator(real.audio_data)
        d_fake = self.discriminator(fake)
        d_real = self.discriminator(real)
        return d_fake, d_real

    def discriminator_loss(self, fake, real):
        d_fake, d_real = self.forward(fake.clone().detach(), real)

        loss_d = 0
        for x_fake, x_real in zip(d_fake, d_real):
            loss_d += torch.mean(x_fake[-1] ** 2)
            loss_d += torch.mean((1 - x_real[-1]) ** 2)
        return loss_d

    def generator_loss(self, fake, real):
        d_fake, d_real = self.forward(fake, real)

        loss_g = 0
        for x_fake in d_fake:
            loss_g += torch.mean((1 - x_fake[-1]) ** 2)

        loss_feature = 0

        for i in range(len(d_fake)):
            for j in range(len(d_fake[i]) - 1):
                loss_feature += F.l1_loss(d_fake[i][j], d_real[i][j].detach())
        return loss_g, loss_feature