180 lines
6.2 KiB
Python
180 lines
6.2 KiB
Python
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from torch.nn.utils.parametrizations import weight_norm, spectral_norm
|
|
|
|
# -------------------------------------------------------------------
|
|
# 1. Multi-Period Discriminator (MPD)
|
|
# Captures periodic structures (pitch/timbre) by folding audio.
|
|
# -------------------------------------------------------------------
|
|
|
|
class DiscriminatorP(nn.Module):
|
|
def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
|
|
super(DiscriminatorP, self).__init__()
|
|
self.period = period
|
|
self.use_spectral_norm = use_spectral_norm
|
|
|
|
# Use spectral_norm for stability, or weight_norm for performance
|
|
norm_f = spectral_norm if use_spectral_norm else weight_norm
|
|
|
|
# We use 2D convs because we "fold" the 1D audio into 2D (Period x Time)
|
|
self.convs = nn.ModuleList([
|
|
norm_f(nn.Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(2, 0))),
|
|
norm_f(nn.Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(2, 0))),
|
|
norm_f(nn.Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(2, 0))),
|
|
norm_f(nn.Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(2, 0))),
|
|
norm_f(nn.Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(2, 0))),
|
|
])
|
|
|
|
self.conv_post = norm_f(nn.Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
|
|
|
|
def forward(self, x):
|
|
fmap = []
|
|
|
|
# 1d to 2d conversion: [B, C, T] -> [B, C, T/P, P]
|
|
b, c, t = x.shape
|
|
if t % self.period != 0: # Pad if not divisible by period
|
|
n_pad = self.period - (t % self.period)
|
|
x = F.pad(x, (0, n_pad), "reflect")
|
|
t = t + n_pad
|
|
|
|
x = x.view(b, c, t // self.period, self.period)
|
|
|
|
for l in self.convs:
|
|
x = l(x)
|
|
x = F.leaky_relu(x, 0.1)
|
|
fmap.append(x) # Store feature map for Feature Matching Loss
|
|
|
|
x = self.conv_post(x)
|
|
fmap.append(x)
|
|
|
|
# Flatten back to 1D for score
|
|
x = torch.flatten(x, 1, -1)
|
|
|
|
return x, fmap
|
|
|
|
|
|
class MultiPeriodDiscriminator(nn.Module):
|
|
def __init__(self, periods=[2, 3, 5, 7, 11]):
|
|
super(MultiPeriodDiscriminator, self).__init__()
|
|
self.discriminators = nn.ModuleList([
|
|
DiscriminatorP(p) for p in periods
|
|
])
|
|
|
|
def forward(self, y, y_hat):
|
|
y_d_rs = [] # Real scores
|
|
y_d_gs = [] # Generated (Fake) scores
|
|
fmap_rs = [] # Real feature maps
|
|
fmap_gs = [] # Generated (Fake) feature maps
|
|
|
|
for i, d in enumerate(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
|
|
|
|
|
|
# -------------------------------------------------------------------
|
|
# 2. Multi-Scale Discriminator (MSD)
|
|
# Captures structure at different audio resolutions (raw, x0.5, x0.25).
|
|
# -------------------------------------------------------------------
|
|
|
|
class DiscriminatorS(nn.Module):
|
|
def __init__(self, use_spectral_norm=False):
|
|
super(DiscriminatorS, self).__init__()
|
|
norm_f = spectral_norm if use_spectral_norm else weight_norm
|
|
|
|
# Standard 1D Convolutions with large receptive field
|
|
self.convs = nn.ModuleList([
|
|
norm_f(nn.Conv1d(1, 16, 15, 1, padding=7)),
|
|
norm_f(nn.Conv1d(16, 64, 41, 4, groups=4, padding=20)),
|
|
norm_f(nn.Conv1d(64, 256, 41, 4, groups=16, padding=20)),
|
|
norm_f(nn.Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
|
|
norm_f(nn.Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
|
|
norm_f(nn.Conv1d(1024, 1024, 5, 1, padding=2)),
|
|
])
|
|
self.conv_post = norm_f(nn.Conv1d(1024, 1, 3, 1, padding=1))
|
|
|
|
def forward(self, x):
|
|
fmap = []
|
|
for l in self.convs:
|
|
x = l(x)
|
|
x = F.leaky_relu(x, 0.1)
|
|
fmap.append(x)
|
|
x = self.conv_post(x)
|
|
fmap.append(x)
|
|
x = torch.flatten(x, 1, -1)
|
|
return x, fmap
|
|
|
|
|
|
class MultiScaleDiscriminator(nn.Module):
|
|
def __init__(self):
|
|
super(MultiScaleDiscriminator, self).__init__()
|
|
# 3 Scales: Original, Downsampled x2, Downsampled x4
|
|
self.discriminators = nn.ModuleList([
|
|
DiscriminatorS(use_spectral_norm=True),
|
|
DiscriminatorS(),
|
|
DiscriminatorS(),
|
|
])
|
|
self.meanpools = nn.ModuleList([
|
|
nn.AvgPool1d(4, 2, padding=2),
|
|
nn.AvgPool1d(4, 2, padding=2)
|
|
])
|
|
|
|
def forward(self, y, y_hat):
|
|
y_d_rs = []
|
|
y_d_gs = []
|
|
fmap_rs = []
|
|
fmap_gs = []
|
|
|
|
for i, d in enumerate(self.discriminators):
|
|
if i != 0:
|
|
# Downsample input for subsequent discriminators
|
|
y = self.meanpools[i-1](y)
|
|
y_hat = self.meanpools[i-1](y_hat)
|
|
|
|
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
|
|
|
|
|
|
# -------------------------------------------------------------------
|
|
# 3. Master Wrapper
|
|
# Combines MPD and MSD into one class to fit your training script.
|
|
# -------------------------------------------------------------------
|
|
|
|
class SISUDiscriminator(nn.Module):
|
|
def __init__(self):
|
|
super(SISUDiscriminator, self).__init__()
|
|
self.mpd = MultiPeriodDiscriminator()
|
|
self.msd = MultiScaleDiscriminator()
|
|
|
|
def forward(self, y, y_hat):
|
|
# Return format:
|
|
# scores_real, scores_fake, features_real, features_fake
|
|
|
|
# Run Multi-Period
|
|
mpd_y_d_rs, mpd_y_d_gs, mpd_fmap_rs, mpd_fmap_gs = self.mpd(y, y_hat)
|
|
|
|
# Run Multi-Scale
|
|
msd_y_d_rs, msd_y_d_gs, msd_fmap_rs, msd_fmap_gs = self.msd(y, y_hat)
|
|
|
|
# Combine all results
|
|
return (
|
|
mpd_y_d_rs + msd_y_d_rs, # All real scores
|
|
mpd_y_d_gs + msd_y_d_gs, # All fake scores
|
|
mpd_fmap_rs + msd_fmap_rs, # All real feature maps
|
|
mpd_fmap_gs + msd_fmap_gs # All fake feature maps
|
|
)
|