说明：该系列文章从本人知乎账号迁入，主要原因是知乎图片附件过于模糊。

知乎专栏地址：
语音生成专栏

系列文章地址：
【GPT-SOVITS-01】源码梳理
【GPT-SOVITS-02】GPT模块解析
【GPT-SOVITS-03】SOVITS 模块-生成模型解析
【GPT-SOVITS-04】SOVITS 模块-鉴别模型解析
【GPT-SOVITS-05】SOVITS 模块-残差量化解析
【GPT-SOVITS-06】特征工程-HuBert原理

1.SOVITS 鉴别器

1.1、概述

GPT-SOVITS 在鉴别器这块在SOVITS原始版本上做了简化，先回顾下SOVITS的鉴别器。主要包含三类：

各个鉴别器的输出都包括两类，即各层中间输出和最终结果输出，分别用来计算特征损失和生成损失。如下：

1.2、MRD举例

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.utils import weight_norm, spectral_norm

class DiscriminatorR(torch.nn.Module):
    def __init__(self, hp, resolution):
        super(DiscriminatorR, self).__init__()

        self.resolution = resolution
        self.LRELU_SLOPE = hp.mpd.lReLU_slope

        norm_f = weight_norm if hp.mrd.use_spectral_norm == False else spectral_norm

        self.convs = nn.ModuleList([
            norm_f(nn.Conv2d(1, 32, (3, 9), padding=(1, 4))),
            norm_f(nn.Conv2d(32, 32, (3, 9), stride=(1, 2), padding=(1, 4))),
            norm_f(nn.Conv2d(32, 32, (3, 9), stride=(1, 2), padding=(1, 4))),
            norm_f(nn.Conv2d(32, 32, (3, 9), stride=(1, 2), padding=(1, 4))),
            norm_f(nn.Conv2d(32, 32, (3, 3), padding=(1, 1))),
        ])
        self.conv_post = norm_f(nn.Conv2d(32, 1, (3, 3), padding=(1, 1)))

    def forward(self, x):
        fmap = []

        # 获取频谱，这里是做了窗口傅里叶变换
        # 傅里叶变换时,频谱数量、窗口的移动、窗口大小由参数 resolution 决定
        x = self.spectrogram(x)
        x = x.unsqueeze(1)
        for l in self.convs:

            # 与其他鉴别器一样经过conv-1d 和 leak-relue 形成中间层特征
            x = l(x)
            x = F.leaky_relu(x, self.LRELU_SLOPE)

            # 中间层特征被保存在 fmap 中
            fmap.append(x)
        x = self.conv_post(x)
        fmap.append(x)
        x = torch.flatten(x, 1, -1)

        # 返回各层的中间层特征 fmap  和 最终输出 x
        return fmap, x

    def spectrogram(self, x):
        n_fft, hop_length, win_length = self.resolution
        x = F.pad(x, (int((n_fft - hop_length) / 2), int((n_fft - hop_length) / 2)), mode='reflect')
        x = x.squeeze(1)
        x = torch.stft(x, n_fft=n_fft, hop_length=hop_length, win_length=win_length, center=False, return_complex=False) #[B, F, TT, 2]
        mag = torch.norm(x, p=2, dim =-1) #[B, F, TT]

        return mag


class MultiResolutionDiscriminator(torch.nn.Module):
    def __init__(self, hp):
        super(MultiResolutionDiscriminator, self).__init__()
        self.resolutions = eval(hp.mrd.resolutions)
        self.discriminators = nn.ModuleList(
            [DiscriminatorR(hp, resolution) for resolution in self.resolutions]
        )

    def forward(self, x):
        ret = list()

        # 这里做了一个不同尺度的 DiscriminatorR

        """
        在 base.yml 中 mrd 的参数如下，有四个不同的尺度：
        mrd:
        resolutions: "[(1024, 120, 600), (2048, 240, 1200), (4096, 480, 2400), (512, 50, 240)]" # (filter_length, hop_length, win_length)
        use_spectral_norm: False
        lReLU_slope: 0.2
        """
        for disc in self.discriminators:
            ret.append(disc(x))

        return ret  # [(feat, score), (feat, score), (feat, score)]

2.GPT-SOVITS 鉴别器

2.1、主要更改

GPT-SOVITS 鉴别器结构与 SOVITS基本类似，只是去除了多分辨率鉴别器，其余基本一样，包括多周期鉴别器的尺度也是 2, 3, 5, 7, 11。其返回结果也包含最终【生成鉴别结果】和各层输出【特征鉴别结果】两类。

class MultiPeriodDiscriminator(torch.nn.Module):
    def __init__(self, use_spectral_norm=False):
        super(MultiPeriodDiscriminator, self).__init__()
        periods = [2, 3, 5, 7, 11]

        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
        discs = discs + [
            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
        ]
        self.discriminators = nn.ModuleList(discs)

    def forward(self, y, y_hat):
        y_d_rs = []
        y_d_gs = []
        fmap_rs = []
        fmap_gs = []
        for i, d in enumerate(self.discriminators):
            y_d_r, fmap_r = d(y)      # 原始音频输入，返回鉴别结果
            y_d_g, fmap_g = d(y_hat)  # 推测音频输入，返回鉴别结果
            y_d_rs.append(y_d_r)
            y_d_gs.append(y_d_g)
            fmap_rs.append(fmap_r)
            fmap_gs.append(fmap_g)

        return y_d_rs, y_d_gs, fmap_rs, fmap_gs

2.2、损失函数

y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
with autocast(enabled=False):
    loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
    loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl

    loss_fm = feature_loss(fmap_r, fmap_g)
    loss_gen, losses_gen = generator_loss(y_d_hat_g)

如前文所述，这里特征损失基于各层输出，计算逻辑在 feature_loss

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

    return loss * 2

最终生成损失判别基于最终结果，计算逻辑在 generator_loss

def generator_loss(disc_outputs):
    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