|
import math |
|
from typing import Optional |
|
|
|
import numpy as np |
|
import torch |
|
from torch.nn.utils import remove_weight_norm |
|
from torch.nn.utils.parametrizations import weight_norm |
|
from torch.utils.checkpoint import checkpoint |
|
|
|
LRELU_SLOPE = 0.1 |
|
|
|
|
|
class MRFLayer(torch.nn.Module): |
|
""" |
|
A single layer of the Multi-Receptive Field (MRF) block. |
|
|
|
This layer consists of two 1D convolutional layers with weight normalization |
|
and Leaky ReLU activation in between. The first convolution has a dilation, |
|
while the second has a dilation of 1. A skip connection is added from the input |
|
to the output. |
|
|
|
Args: |
|
channels (int): The number of input and output channels. |
|
kernel_size (int): The kernel size of the convolutional layers. |
|
dilation (int): The dilation rate for the first convolutional layer. |
|
""" |
|
|
|
def __init__(self, channels, kernel_size, dilation): |
|
super().__init__() |
|
self.conv1 = weight_norm( |
|
torch.nn.Conv1d( |
|
channels, |
|
channels, |
|
kernel_size, |
|
padding=(kernel_size * dilation - dilation) // 2, |
|
dilation=dilation, |
|
) |
|
) |
|
self.conv2 = weight_norm( |
|
torch.nn.Conv1d( |
|
channels, channels, kernel_size, padding=kernel_size // 2, dilation=1 |
|
) |
|
) |
|
|
|
def forward(self, x: torch.Tensor): |
|
|
|
y = torch.nn.functional.leaky_relu(x, LRELU_SLOPE) |
|
y = self.conv1(y) |
|
|
|
y = torch.nn.functional.leaky_relu_(y, LRELU_SLOPE) |
|
y = self.conv2(y) |
|
return x + y |
|
|
|
def remove_weight_norm(self): |
|
remove_weight_norm(self.conv1) |
|
remove_weight_norm(self.conv2) |
|
|
|
|
|
class MRFBlock(torch.nn.Module): |
|
""" |
|
A Multi-Receptive Field (MRF) block. |
|
|
|
This block consists of multiple MRFLayers with different dilation rates. |
|
It applies each layer sequentially to the input. |
|
|
|
Args: |
|
channels (int): The number of input and output channels for the MRFLayers. |
|
kernel_size (int): The kernel size for the convolutional layers in the MRFLayers. |
|
dilations (list[int]): A list of dilation rates for the MRFLayers. |
|
""" |
|
|
|
def __init__(self, channels, kernel_size, dilations): |
|
super().__init__() |
|
self.layers = torch.nn.ModuleList() |
|
for dilation in dilations: |
|
self.layers.append(MRFLayer(channels, kernel_size, dilation)) |
|
|
|
def forward(self, x: torch.Tensor): |
|
for layer in self.layers: |
|
x = layer(x) |
|
return x |
|
|
|
def remove_weight_norm(self): |
|
for layer in self.layers: |
|
layer.remove_weight_norm() |
|
|
|
|
|
class SineGenerator(torch.nn.Module): |
|
""" |
|
Definition of sine generator |
|
|
|
Generates sine waveforms with optional harmonics and additive noise. |
|
Can be used to create harmonic noise source for neural vocoders. |
|
|
|
Args: |
|
samp_rate (int): Sampling rate in Hz. |
|
harmonic_num (int): Number of harmonic overtones (default 0). |
|
sine_amp (float): Amplitude of sine-waveform (default 0.1). |
|
noise_std (float): Standard deviation of Gaussian noise (default 0.003). |
|
voiced_threshold (float): F0 threshold for voiced/unvoiced classification (default 0). |
|
""" |
|
|
|
def __init__( |
|
self, |
|
samp_rate: int, |
|
harmonic_num: int = 0, |
|
sine_amp: float = 0.1, |
|
noise_std: float = 0.003, |
|
voiced_threshold: float = 0, |
|
): |
|
super(SineGenerator, self).__init__() |
|
self.sine_amp = sine_amp |
|
self.noise_std = noise_std |
|
self.harmonic_num = harmonic_num |
|
self.dim = self.harmonic_num + 1 |
|
self.sampling_rate = samp_rate |
|
self.voiced_threshold = voiced_threshold |
|
|
|
def _f02uv(self, f0: torch.Tensor): |
|
""" |
|
Generates voiced/unvoiced (UV) signal based on the fundamental frequency (F0). |
|
|
|
Args: |
|
f0 (torch.Tensor): Fundamental frequency tensor of shape (batch_size, length, 1). |
|
""" |
|
|
|
uv = torch.ones_like(f0) |
|
uv = uv * (f0 > self.voiced_threshold) |
|
return uv |
|
|
|
def _f02sine(self, f0_values: torch.Tensor): |
|
""" |
|
Generates sine waveforms based on the fundamental frequency (F0) and its harmonics. |
|
|
|
Args: |
|
f0_values (torch.Tensor): Tensor of fundamental frequency and its harmonics, |
|
shape (batch_size, length, dim), where dim indicates |
|
the fundamental tone and overtones. |
|
""" |
|
|
|
|
|
rad_values = (f0_values / self.sampling_rate) % 1 |
|
|
|
|
|
rand_ini = torch.rand( |
|
f0_values.shape[0], f0_values.shape[2], device=f0_values.device |
|
) |
|
rand_ini[:, 0] = 0 |
|
rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini |
|
|
|
|
|
tmp_over_one = torch.cumsum(rad_values, 1) % 1 |
|
tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0 |
|
cumsum_shift = torch.zeros_like(rad_values) |
|
cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0 |
|
|
|
sines = torch.sin(torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi) |
|
|
|
return sines |
|
|
|
def forward(self, f0: torch.Tensor): |
|
with torch.no_grad(): |
|
f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device) |
|
|
|
f0_buf[:, :, 0] = f0[:, :, 0] |
|
for idx in np.arange(self.harmonic_num): |
|
f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (idx + 2) |
|
|
|
sine_waves = self._f02sine(f0_buf) * self.sine_amp |
|
|
|
uv = self._f02uv(f0) |
|
|
|
noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3 |
|
noise = noise_amp * torch.randn_like(sine_waves) |
|
|
|
sine_waves = sine_waves * uv + noise |
|
return sine_waves, uv, noise |
|
|
|
|
|
class SourceModuleHnNSF(torch.nn.Module): |
|
""" |
|
Generates harmonic and noise source features. |
|
|
|
This module uses the SineGenerator to create harmonic signals based on the |
|
fundamental frequency (F0) and merges them into a single excitation signal. |
|
|
|
Args: |
|
sample_rate (int): Sampling rate in Hz. |
|
harmonic_num (int, optional): Number of harmonics above F0. Defaults to 0. |
|
sine_amp (float, optional): Amplitude of sine source signal. Defaults to 0.1. |
|
add_noise_std (float, optional): Standard deviation of additive Gaussian noise. Defaults to 0.003. |
|
voiced_threshod (float, optional): Threshold to set voiced/unvoiced given F0. Defaults to 0. |
|
""" |
|
|
|
def __init__( |
|
self, |
|
sampling_rate: int, |
|
harmonic_num: int = 0, |
|
sine_amp: float = 0.1, |
|
add_noise_std: float = 0.003, |
|
voiced_threshold: float = 0, |
|
): |
|
super(SourceModuleHnNSF, self).__init__() |
|
|
|
self.sine_amp = sine_amp |
|
self.noise_std = add_noise_std |
|
|
|
|
|
self.l_sin_gen = SineGenerator( |
|
sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshold |
|
) |
|
|
|
|
|
self.l_linear = torch.nn.Linear(harmonic_num + 1, 1) |
|
self.l_tanh = torch.nn.Tanh() |
|
|
|
def forward(self, x: torch.Tensor): |
|
sine_wavs, uv, _ = self.l_sin_gen(x) |
|
sine_wavs = sine_wavs.to(dtype=self.l_linear.weight.dtype) |
|
sine_merge = self.l_tanh(self.l_linear(sine_wavs)) |
|
|
|
return sine_merge, None, None |
|
|
|
|
|
class HiFiGANMRFGenerator(torch.nn.Module): |
|
""" |
|
HiFi-GAN generator with Multi-Receptive Field (MRF) blocks. |
|
|
|
This generator takes an input feature sequence and fundamental frequency (F0) |
|
as input and generates an audio waveform. It utilizes transposed convolutions |
|
for upsampling and MRF blocks for feature refinement. It can also condition |
|
on global conditioning features. |
|
|
|
Args: |
|
in_channel (int): Number of input channels. |
|
upsample_initial_channel (int): Number of channels after the initial convolution. |
|
upsample_rates (list[int]): List of upsampling rates for the transposed convolutions. |
|
upsample_kernel_sizes (list[int]): List of kernel sizes for the transposed convolutions. |
|
resblock_kernel_sizes (list[int]): List of kernel sizes for the convolutional layers in the MRF blocks. |
|
resblock_dilations (list[list[int]]): List of lists of dilation rates for the MRF blocks. |
|
gin_channels (int): Number of global conditioning input channels (0 if no global conditioning). |
|
sample_rate (int): Sampling rate of the audio. |
|
harmonic_num (int): Number of harmonics to generate. |
|
checkpointing (bool): Whether to use checkpointing to save memory during training (default: False). |
|
""" |
|
|
|
def __init__( |
|
self, |
|
in_channel: int, |
|
upsample_initial_channel: int, |
|
upsample_rates: list[int], |
|
upsample_kernel_sizes: list[int], |
|
resblock_kernel_sizes: list[int], |
|
resblock_dilations: list[list[int]], |
|
gin_channels: int, |
|
sample_rate: int, |
|
harmonic_num: int, |
|
checkpointing: bool = False, |
|
): |
|
super().__init__() |
|
self.num_kernels = len(resblock_kernel_sizes) |
|
self.checkpointing = checkpointing |
|
|
|
self.f0_upsample = torch.nn.Upsample(scale_factor=np.prod(upsample_rates)) |
|
self.m_source = SourceModuleHnNSF(sample_rate, harmonic_num) |
|
|
|
self.conv_pre = weight_norm( |
|
torch.nn.Conv1d( |
|
in_channel, upsample_initial_channel, kernel_size=7, stride=1, padding=3 |
|
) |
|
) |
|
self.upsamples = torch.nn.ModuleList() |
|
self.noise_convs = torch.nn.ModuleList() |
|
|
|
stride_f0s = [ |
|
math.prod(upsample_rates[i + 1 :]) if i + 1 < len(upsample_rates) else 1 |
|
for i in range(len(upsample_rates)) |
|
] |
|
|
|
for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)): |
|
|
|
if u % 2 == 0: |
|
|
|
padding = (k - u) // 2 |
|
else: |
|
padding = u // 2 + u % 2 |
|
|
|
self.upsamples.append( |
|
weight_norm( |
|
torch.nn.ConvTranspose1d( |
|
upsample_initial_channel // (2**i), |
|
upsample_initial_channel // (2 ** (i + 1)), |
|
kernel_size=k, |
|
stride=u, |
|
padding=padding, |
|
output_padding=u % 2, |
|
) |
|
) |
|
) |
|
""" handling odd upsampling rates |
|
# s k p |
|
# 40 80 20 |
|
# 32 64 16 |
|
# 4 8 2 |
|
# 2 3 1 |
|
# 63 125 31 |
|
# 9 17 4 |
|
# 3 5 1 |
|
# 1 1 0 |
|
""" |
|
stride = stride_f0s[i] |
|
kernel = 1 if stride == 1 else stride * 2 - stride % 2 |
|
padding = 0 if stride == 1 else (kernel - stride) // 2 |
|
|
|
self.noise_convs.append( |
|
torch.nn.Conv1d( |
|
1, |
|
upsample_initial_channel // (2 ** (i + 1)), |
|
kernel_size=kernel, |
|
stride=stride, |
|
padding=padding, |
|
) |
|
) |
|
self.mrfs = torch.nn.ModuleList() |
|
for i in range(len(self.upsamples)): |
|
channel = upsample_initial_channel // (2 ** (i + 1)) |
|
self.mrfs.append( |
|
torch.nn.ModuleList( |
|
[ |
|
MRFBlock(channel, kernel_size=k, dilations=d) |
|
for k, d in zip(resblock_kernel_sizes, resblock_dilations) |
|
] |
|
) |
|
) |
|
self.conv_post = weight_norm( |
|
torch.nn.Conv1d(channel, 1, kernel_size=7, stride=1, padding=3) |
|
) |
|
if gin_channels != 0: |
|
self.cond = torch.nn.Conv1d(gin_channels, upsample_initial_channel, 1) |
|
|
|
def forward( |
|
self, x: torch.Tensor, f0: torch.Tensor, g: Optional[torch.Tensor] = None |
|
): |
|
f0 = self.f0_upsample(f0[:, None, :]).transpose(-1, -2) |
|
har_source, _, _ = self.m_source(f0) |
|
har_source = har_source.transpose(-1, -2) |
|
|
|
x = self.conv_pre(x) |
|
|
|
if g is not None: |
|
|
|
x += self.cond(g) |
|
|
|
for ups, mrf, noise_conv in zip(self.upsamples, self.mrfs, self.noise_convs): |
|
|
|
x = torch.nn.functional.leaky_relu_(x, LRELU_SLOPE) |
|
|
|
if self.training and self.checkpointing: |
|
x = checkpoint(ups, x, use_reentrant=False) |
|
else: |
|
x = ups(x) |
|
|
|
x += noise_conv(har_source) |
|
|
|
def mrf_sum(x, layers): |
|
return sum(layer(x) for layer in layers) / self.num_kernels |
|
|
|
if self.training and self.checkpointing: |
|
x = checkpoint(mrf_sum, x, mrf, use_reentrant=False) |
|
else: |
|
x = mrf_sum(x, mrf) |
|
|
|
x = torch.nn.functional.leaky_relu_(x) |
|
x = self.conv_post(x) |
|
|
|
x = torch.tanh_(x) |
|
return x |
|
|
|
def remove_weight_norm(self): |
|
remove_weight_norm(self.conv_pre) |
|
for up in self.upsamples: |
|
remove_weight_norm(up) |
|
for mrf in self.mrfs: |
|
mrf.remove_weight_norm() |
|
remove_weight_norm(self.conv_post) |
|
|