fish-speech/fish_speech/models/vits_decoder/losses.py

import torch
import torch.nn.functional as F
from torch import nn


def feature_loss(fmap_r: list[torch.Tensor], fmap_g: list[torch.Tensor]):
    loss = 0
    for dr, dg in zip(fmap_r, fmap_g):
        dr = dr.float().detach()
        dg = dg.float()
        loss += torch.mean(torch.abs(dr - dg))

    return loss * 2


def discriminator_loss(
    disc_real_outputs: list[torch.Tensor], disc_generated_outputs: list[torch.Tensor]
):
    loss = 0
    r_losses = []
    g_losses = []
    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
        dr = dr.float()
        dg = dg.float()
        r_loss = torch.mean((1 - dr) ** 2)
        g_loss = torch.mean(dg**2)
        loss += r_loss + g_loss
        r_losses.append(r_loss.item())
        g_losses.append(g_loss.item())

    return loss, r_losses, g_losses


def generator_loss(disc_outputs: list[torch.Tensor]):
    loss = 0
    gen_losses = []
    for dg in disc_outputs:
        dg = dg.float()
        l = torch.mean((1 - dg) ** 2)
        gen_losses.append(l)
        loss += l

    return loss, gen_losses


def kl_loss(
    z_p: torch.Tensor,
    logs_q: torch.Tensor,
    m_p: torch.Tensor,
    logs_p: torch.Tensor,
    z_mask: torch.Tensor,
):
    """
    z_p, logs_q: [b, h, t_t]
    m_p, logs_p: [b, h, t_t]
    """
    z_p = z_p.float()
    logs_q = logs_q.float()
    m_p = m_p.float()
    logs_p = logs_p.float()
    z_mask = z_mask.float()

    kl = logs_p - logs_q - 0.5
    kl += 0.5 * ((z_p - m_p) ** 2) * torch.exp(-2.0 * logs_p)
    kl = torch.sum(kl * z_mask)
    l = kl / torch.sum(z_mask)
    return l