akhaliq3
spaces demo
2b7bf83
raw
history blame
25.9 kB
# -*- coding: utf-8 -*-
"""HiFi-GAN Modules.
This code is based on https://github.com/jik876/hifi-gan.
"""
import copy
import logging
import numpy as np
import torch
import torch.nn.functional as F
from parallel_wavegan.layers import HiFiGANResidualBlock as ResidualBlock
from parallel_wavegan.utils import read_hdf5
class HiFiGANGenerator(torch.nn.Module):
"""HiFiGAN generator module."""
def __init__(
self,
in_channels=80,
out_channels=1,
channels=512,
kernel_size=7,
upsample_scales=(8, 8, 2, 2),
upsample_kernel_sizes=(16, 16, 4, 4),
resblock_kernel_sizes=(3, 7, 11),
resblock_dilations=[(1, 3, 5), (1, 3, 5), (1, 3, 5)],
use_additional_convs=True,
bias=True,
nonlinear_activation="LeakyReLU",
nonlinear_activation_params={"negative_slope": 0.1},
use_weight_norm=True,
):
"""Initialize HiFiGANGenerator module.
Args:
in_channels (int): Number of input channels.
out_channels (int): Number of output channels.
channels (int): Number of hidden representation channels.
kernel_size (int): Kernel size of initial and final conv layer.
upsample_scales (list): List of upsampling scales.
upsample_kernel_sizes (list): List of kernel sizes for upsampling layers.
resblock_kernel_sizes (list): List of kernel sizes for residual blocks.
resblock_dilations (list): List of dilation list for residual blocks.
use_additional_convs (bool): Whether to use additional conv layers in residual blocks.
bias (bool): Whether to add bias parameter in convolution layers.
nonlinear_activation (str): Activation function module name.
nonlinear_activation_params (dict): Hyperparameters for activation function.
use_weight_norm (bool): Whether to use weight norm.
If set to true, it will be applied to all of the conv layers.
"""
super().__init__()
# check hyperparameters are valid
assert kernel_size % 2 == 1, "Kernel size must be odd number."
assert len(upsample_scales) == len(upsample_kernel_sizes)
assert len(resblock_dilations) == len(resblock_kernel_sizes)
# define modules
self.num_upsamples = len(upsample_kernel_sizes)
self.num_blocks = len(resblock_kernel_sizes)
self.input_conv = torch.nn.Conv1d(
in_channels,
channels,
kernel_size,
1,
padding=(kernel_size - 1) // 2,
)
self.upsamples = torch.nn.ModuleList()
self.blocks = torch.nn.ModuleList()
for i in range(len(upsample_kernel_sizes)):
assert upsample_kernel_sizes[i] == 2 * upsample_scales[i]
self.upsamples += [
torch.nn.Sequential(
getattr(torch.nn, nonlinear_activation)(
**nonlinear_activation_params
),
torch.nn.ConvTranspose1d(
channels // (2 ** i),
channels // (2 ** (i + 1)),
upsample_kernel_sizes[i],
upsample_scales[i],
padding=upsample_scales[i] // 2 + upsample_scales[i] % 2,
output_padding=upsample_scales[i] % 2,
),
)
]
for j in range(len(resblock_kernel_sizes)):
self.blocks += [
ResidualBlock(
kernel_size=resblock_kernel_sizes[j],
channels=channels // (2 ** (i + 1)),
dilations=resblock_dilations[j],
bias=bias,
use_additional_convs=use_additional_convs,
nonlinear_activation=nonlinear_activation,
nonlinear_activation_params=nonlinear_activation_params,
)
]
self.output_conv = torch.nn.Sequential(
# NOTE(kan-bayashi): follow official implementation but why
# using different slope parameter here? (0.1 vs. 0.01)
torch.nn.LeakyReLU(),
torch.nn.Conv1d(
channels // (2 ** (i + 1)),
out_channels,
kernel_size,
1,
padding=(kernel_size - 1) // 2,
),
torch.nn.Tanh(),
)
# apply weight norm
if use_weight_norm:
self.apply_weight_norm()
# reset parameters
self.reset_parameters()
def forward(self, c):
"""Calculate forward propagation.
Args:
c (Tensor): Input tensor (B, in_channels, T).
Returns:
Tensor: Output tensor (B, out_channels, T).
"""
c = self.input_conv(c)
for i in range(self.num_upsamples):
c = self.upsamples[i](c)
cs = 0.0 # initialize
for j in range(self.num_blocks):
cs += self.blocks[i * self.num_blocks + j](c)
c = cs / self.num_blocks
c = self.output_conv(c)
return c
def reset_parameters(self):
"""Reset parameters.
This initialization follows the official implementation manner.
https://github.com/jik876/hifi-gan/blob/master/models.py
"""
def _reset_parameters(m):
if isinstance(m, (torch.nn.Conv1d, torch.nn.ConvTranspose1d)):
m.weight.data.normal_(0.0, 0.01)
logging.debug(f"Reset parameters in {m}.")
self.apply(_reset_parameters)
def remove_weight_norm(self):
"""Remove weight normalization module from all of the layers."""
def _remove_weight_norm(m):
try:
logging.debug(f"Weight norm is removed from {m}.")
torch.nn.utils.remove_weight_norm(m)
except ValueError: # this module didn't have weight norm
return
self.apply(_remove_weight_norm)
def apply_weight_norm(self):
"""Apply weight normalization module from all of the layers."""
def _apply_weight_norm(m):
if isinstance(m, torch.nn.Conv1d) or isinstance(
m, torch.nn.ConvTranspose1d
):
torch.nn.utils.weight_norm(m)
logging.debug(f"Weight norm is applied to {m}.")
self.apply(_apply_weight_norm)
def register_stats(self, stats):
"""Register stats for de-normalization as buffer.
Args:
stats (str): Path of statistics file (".npy" or ".h5").
"""
assert stats.endswith(".h5") or stats.endswith(".npy")
if stats.endswith(".h5"):
mean = read_hdf5(stats, "mean").reshape(-1)
scale = read_hdf5(stats, "scale").reshape(-1)
else:
mean = np.load(stats)[0].reshape(-1)
scale = np.load(stats)[1].reshape(-1)
self.register_buffer("mean", torch.from_numpy(mean).float())
self.register_buffer("scale", torch.from_numpy(scale).float())
logging.info("Successfully registered stats as buffer.")
def inference(self, c, normalize_before=False):
"""Perform inference.
Args:
c (Union[Tensor, ndarray]): Input tensor (T, in_channels).
normalize_before (bool): Whether to perform normalization.
Returns:
Tensor: Output tensor (T ** prod(upsample_scales), out_channels).
"""
if not isinstance(c, torch.Tensor):
c = torch.tensor(c, dtype=torch.float).to(next(self.parameters()).device)
if normalize_before:
c = (c - self.mean) / self.scale
c = self.forward(c.transpose(1, 0).unsqueeze(0))
return c.squeeze(0).transpose(1, 0)
class HiFiGANPeriodDiscriminator(torch.nn.Module):
"""HiFiGAN period discriminator module."""
def __init__(
self,
in_channels=1,
out_channels=1,
period=3,
kernel_sizes=[5, 3],
channels=32,
downsample_scales=[3, 3, 3, 3, 1],
max_downsample_channels=1024,
bias=True,
nonlinear_activation="LeakyReLU",
nonlinear_activation_params={"negative_slope": 0.1},
use_weight_norm=True,
use_spectral_norm=False,
):
"""Initialize HiFiGANPeriodDiscriminator module.
Args:
in_channels (int): Number of input channels.
out_channels (int): Number of output channels.
period (int): Period.
kernel_sizes (list): Kernel sizes of initial conv layers and the final conv layer.
channels (int): Number of initial channels.
downsample_scales (list): List of downsampling scales.
max_downsample_channels (int): Number of maximum downsampling channels.
use_additional_convs (bool): Whether to use additional conv layers in residual blocks.
bias (bool): Whether to add bias parameter in convolution layers.
nonlinear_activation (str): Activation function module name.
nonlinear_activation_params (dict): Hyperparameters for activation function.
use_weight_norm (bool): Whether to use weight norm.
If set to true, it will be applied to all of the conv layers.
use_spectral_norm (bool): Whether to use spectral norm.
If set to true, it will be applied to all of the conv layers.
"""
super().__init__()
assert len(kernel_sizes) == 2
assert kernel_sizes[0] % 2 == 1, "Kernel size must be odd number."
assert kernel_sizes[1] % 2 == 1, "Kernel size must be odd number."
self.period = period
self.convs = torch.nn.ModuleList()
in_chs = in_channels
out_chs = channels
for downsample_scale in downsample_scales:
self.convs += [
torch.nn.Sequential(
torch.nn.Conv2d(
in_chs,
out_chs,
(kernel_sizes[0], 1),
(downsample_scale, 1),
padding=((kernel_sizes[0] - 1) // 2, 0),
),
getattr(torch.nn, nonlinear_activation)(
**nonlinear_activation_params
),
)
]
in_chs = out_chs
# NOTE(kan-bayashi): Use downsample_scale + 1?
out_chs = min(out_chs * 4, max_downsample_channels)
self.output_conv = torch.nn.Conv2d(
out_chs,
out_channels,
(kernel_sizes[1] - 1, 1),
1,
padding=((kernel_sizes[1] - 1) // 2, 0),
)
if use_weight_norm and use_spectral_norm:
raise ValueError("Either use use_weight_norm or use_spectral_norm.")
# apply weight norm
if use_weight_norm:
self.apply_weight_norm()
# apply spectral norm
if use_spectral_norm:
self.apply_spectral_norm()
def forward(self, x):
"""Calculate forward propagation.
Args:
c (Tensor): Input tensor (B, in_channels, T).
Returns:
list: List of each layer's tensors.
"""
# transform 1d to 2d -> (B, C, T/P, P)
b, c, t = x.shape
if t % self.period != 0:
n_pad = self.period - (t % self.period)
x = F.pad(x, (0, n_pad), "reflect")
t += n_pad
x = x.view(b, c, t // self.period, self.period)
# forward conv
outs = []
for layer in self.convs:
x = layer(x)
outs += [x]
x = self.output_conv(x)
x = torch.flatten(x, 1, -1)
outs += [x]
return outs
def apply_weight_norm(self):
"""Apply weight normalization module from all of the layers."""
def _apply_weight_norm(m):
if isinstance(m, torch.nn.Conv2d):
torch.nn.utils.weight_norm(m)
logging.debug(f"Weight norm is applied to {m}.")
self.apply(_apply_weight_norm)
def apply_spectral_norm(self):
"""Apply spectral normalization module from all of the layers."""
def _apply_spectral_norm(m):
if isinstance(m, torch.nn.Conv2d):
torch.nn.utils.spectral_norm(m)
logging.debug(f"Spectral norm is applied to {m}.")
self.apply(_apply_spectral_norm)
class HiFiGANMultiPeriodDiscriminator(torch.nn.Module):
"""HiFiGAN multi-period discriminator module."""
def __init__(
self,
periods=[2, 3, 5, 7, 11],
discriminator_params={
"in_channels": 1,
"out_channels": 1,
"kernel_sizes": [5, 3],
"channels": 32,
"downsample_scales": [3, 3, 3, 3, 1],
"max_downsample_channels": 1024,
"bias": True,
"nonlinear_activation": "LeakyReLU",
"nonlinear_activation_params": {"negative_slope": 0.1},
"use_weight_norm": True,
"use_spectral_norm": False,
},
):
"""Initialize HiFiGANMultiPeriodDiscriminator module.
Args:
periods (list): List of periods.
discriminator_params (dict): Parameters for hifi-gan period discriminator module.
The period parameter will be overwritten.
"""
super().__init__()
self.discriminators = torch.nn.ModuleList()
for period in periods:
params = copy.deepcopy(discriminator_params)
params["period"] = period
self.discriminators += [HiFiGANPeriodDiscriminator(**params)]
def forward(self, x):
"""Calculate forward propagation.
Args:
x (Tensor): Input noise signal (B, 1, T).
Returns:
List: List of list of each discriminator outputs, which consists of each layer output tensors.
"""
outs = []
for f in self.discriminators:
outs += [f(x)]
return outs
class HiFiGANScaleDiscriminator(torch.nn.Module):
"""HiFi-GAN scale discriminator module."""
def __init__(
self,
in_channels=1,
out_channels=1,
kernel_sizes=[15, 41, 5, 3],
channels=128,
max_downsample_channels=1024,
max_groups=16,
bias=True,
downsample_scales=[2, 2, 4, 4, 1],
nonlinear_activation="LeakyReLU",
nonlinear_activation_params={"negative_slope": 0.1},
use_weight_norm=True,
use_spectral_norm=False,
):
"""Initilize HiFiGAN scale discriminator module.
Args:
in_channels (int): Number of input channels.
out_channels (int): Number of output channels.
kernel_sizes (list): List of four kernel sizes. The first will be used for the first conv layer,
and the second is for downsampling part, and the remaining two are for output layers.
channels (int): Initial number of channels for conv layer.
max_downsample_channels (int): Maximum number of channels for downsampling layers.
bias (bool): Whether to add bias parameter in convolution layers.
downsample_scales (list): List of downsampling scales.
nonlinear_activation (str): Activation function module name.
nonlinear_activation_params (dict): Hyperparameters for activation function.
use_weight_norm (bool): Whether to use weight norm.
If set to true, it will be applied to all of the conv layers.
use_spectral_norm (bool): Whether to use spectral norm.
If set to true, it will be applied to all of the conv layers.
"""
super().__init__()
self.layers = torch.nn.ModuleList()
# check kernel size is valid
assert len(kernel_sizes) == 4
for ks in kernel_sizes:
assert ks % 2 == 1
# add first layer
self.layers += [
torch.nn.Sequential(
torch.nn.Conv1d(
in_channels,
channels,
# NOTE(kan-bayashi): Use always the same kernel size
kernel_sizes[0],
bias=bias,
padding=(kernel_sizes[0] - 1) // 2,
),
getattr(torch.nn, nonlinear_activation)(**nonlinear_activation_params),
)
]
# add downsample layers
in_chs = channels
out_chs = channels
# NOTE(kan-bayashi): Remove hard coding?
groups = 4
for downsample_scale in downsample_scales:
self.layers += [
torch.nn.Sequential(
torch.nn.Conv1d(
in_chs,
out_chs,
kernel_size=kernel_sizes[1],
stride=downsample_scale,
padding=(kernel_sizes[1] - 1) // 2,
groups=groups,
bias=bias,
),
getattr(torch.nn, nonlinear_activation)(
**nonlinear_activation_params
),
)
]
in_chs = out_chs
# NOTE(kan-bayashi): Remove hard coding?
out_chs = min(in_chs * 2, max_downsample_channels)
# NOTE(kan-bayashi): Remove hard coding?
groups = min(groups * 4, max_groups)
# add final layers
out_chs = min(in_chs * 2, max_downsample_channels)
self.layers += [
torch.nn.Sequential(
torch.nn.Conv1d(
in_chs,
out_chs,
kernel_size=kernel_sizes[2],
stride=1,
padding=(kernel_sizes[2] - 1) // 2,
bias=bias,
),
getattr(torch.nn, nonlinear_activation)(**nonlinear_activation_params),
)
]
self.layers += [
torch.nn.Conv1d(
out_chs,
out_channels,
kernel_size=kernel_sizes[3],
stride=1,
padding=(kernel_sizes[3] - 1) // 2,
bias=bias,
),
]
if use_weight_norm and use_spectral_norm:
raise ValueError("Either use use_weight_norm or use_spectral_norm.")
# apply weight norm
if use_weight_norm:
self.apply_weight_norm()
# apply spectral norm
if use_spectral_norm:
self.apply_spectral_norm()
def forward(self, x):
"""Calculate forward propagation.
Args:
x (Tensor): Input noise signal (B, 1, T).
Returns:
List: List of output tensors of each layer.
"""
outs = []
for f in self.layers:
x = f(x)
outs += [x]
return outs
def apply_weight_norm(self):
"""Apply weight normalization module from all of the layers."""
def _apply_weight_norm(m):
if isinstance(m, torch.nn.Conv2d):
torch.nn.utils.weight_norm(m)
logging.debug(f"Weight norm is applied to {m}.")
self.apply(_apply_weight_norm)
def apply_spectral_norm(self):
"""Apply spectral normalization module from all of the layers."""
def _apply_spectral_norm(m):
if isinstance(m, torch.nn.Conv2d):
torch.nn.utils.spectral_norm(m)
logging.debug(f"Spectral norm is applied to {m}.")
self.apply(_apply_spectral_norm)
class HiFiGANMultiScaleDiscriminator(torch.nn.Module):
"""HiFi-GAN multi-scale discriminator module."""
def __init__(
self,
scales=3,
downsample_pooling="AvgPool1d",
# follow the official implementation setting
downsample_pooling_params={
"kernel_size": 4,
"stride": 2,
"padding": 2,
},
discriminator_params={
"in_channels": 1,
"out_channels": 1,
"kernel_sizes": [15, 41, 5, 3],
"channels": 128,
"max_downsample_channels": 1024,
"max_groups": 16,
"bias": True,
"downsample_scales": [2, 2, 4, 4, 1],
"nonlinear_activation": "LeakyReLU",
"nonlinear_activation_params": {"negative_slope": 0.1},
},
follow_official_norm=False,
):
"""Initilize HiFiGAN multi-scale discriminator module.
Args:
scales (int): Number of multi-scales.
downsample_pooling (str): Pooling module name for downsampling of the inputs.
downsample_pooling_params (dict): Parameters for the above pooling module.
discriminator_params (dict): Parameters for hifi-gan scale discriminator module.
follow_official_norm (bool): Whether to follow the norm setting of the official
implementaion. The first discriminator uses spectral norm and the other
discriminators use weight norm.
"""
super().__init__()
self.discriminators = torch.nn.ModuleList()
# add discriminators
for i in range(scales):
params = copy.deepcopy(discriminator_params)
if follow_official_norm:
if i == 0:
params["use_weight_norm"] = False
params["use_spectral_norm"] = True
else:
params["use_weight_norm"] = True
params["use_spectral_norm"] = False
self.discriminators += [HiFiGANScaleDiscriminator(**params)]
self.pooling = getattr(torch.nn, downsample_pooling)(
**downsample_pooling_params
)
def forward(self, x):
"""Calculate forward propagation.
Args:
x (Tensor): Input noise signal (B, 1, T).
Returns:
List: List of list of each discriminator outputs, which consists of each layer output tensors.
"""
outs = []
for f in self.discriminators:
outs += [f(x)]
x = self.pooling(x)
return outs
class HiFiGANMultiScaleMultiPeriodDiscriminator(torch.nn.Module):
"""HiFi-GAN multi-scale + multi-period discriminator module."""
def __init__(
self,
# Multi-scale discriminator related
scales=3,
scale_downsample_pooling="AvgPool1d",
scale_downsample_pooling_params={
"kernel_size": 4,
"stride": 2,
"padding": 2,
},
scale_discriminator_params={
"in_channels": 1,
"out_channels": 1,
"kernel_sizes": [15, 41, 5, 3],
"channels": 128,
"max_downsample_channels": 1024,
"max_groups": 16,
"bias": True,
"downsample_scales": [2, 2, 4, 4, 1],
"nonlinear_activation": "LeakyReLU",
"nonlinear_activation_params": {"negative_slope": 0.1},
},
follow_official_norm=True,
# Multi-period discriminator related
periods=[2, 3, 5, 7, 11],
period_discriminator_params={
"in_channels": 1,
"out_channels": 1,
"kernel_sizes": [5, 3],
"channels": 32,
"downsample_scales": [3, 3, 3, 3, 1],
"max_downsample_channels": 1024,
"bias": True,
"nonlinear_activation": "LeakyReLU",
"nonlinear_activation_params": {"negative_slope": 0.1},
"use_weight_norm": True,
"use_spectral_norm": False,
},
):
"""Initilize HiFiGAN multi-scale + multi-period discriminator module.
Args:
scales (int): Number of multi-scales.
scale_downsample_pooling (str): Pooling module name for downsampling of the inputs.
scale_downsample_pooling_params (dict): Parameters for the above pooling module.
scale_discriminator_params (dict): Parameters for hifi-gan scale discriminator module.
follow_official_norm (bool): Whether to follow the norm setting of the official
implementaion. The first discriminator uses spectral norm and the other
discriminators use weight norm.
periods (list): List of periods.
period_discriminator_params (dict): Parameters for hifi-gan period discriminator module.
The period parameter will be overwritten.
"""
super().__init__()
self.msd = HiFiGANMultiScaleDiscriminator(
scales=scales,
downsample_pooling=scale_downsample_pooling,
downsample_pooling_params=scale_downsample_pooling_params,
discriminator_params=scale_discriminator_params,
follow_official_norm=follow_official_norm,
)
self.mpd = HiFiGANMultiPeriodDiscriminator(
periods=periods,
discriminator_params=period_discriminator_params,
)
def forward(self, x):
"""Calculate forward propagation.
Args:
x (Tensor): Input noise signal (B, 1, T).
Returns:
List: List of list of each discriminator outputs,
which consists of each layer output tensors.
Multi scale and multi period ones are concatenated.
"""
msd_outs = self.msd(x)
mpd_outs = self.mpd(x)
return msd_outs + mpd_outs