Spaces:
Running
Running
File size: 6,705 Bytes
96134ee |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 |
import math
import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.utils import remove_weight_norm
from torch.utils.checkpoint import checkpoint
from torch.nn.utils.parametrizations import weight_norm
LRELU_SLOPE = 0.1
class MRFLayer(nn.Module):
def __init__(self, channels, kernel_size, dilation):
super().__init__()
self.conv1 = weight_norm(nn.Conv1d(channels, channels, kernel_size, padding=(kernel_size * dilation - dilation) // 2, dilation=dilation))
self.conv2 = weight_norm(nn.Conv1d(channels, channels, kernel_size, padding=kernel_size // 2, dilation=1))
def forward(self, x):
return x + self.conv2(F.leaky_relu(self.conv1(F.leaky_relu(x, LRELU_SLOPE)), LRELU_SLOPE))
def remove_weight_norm(self):
remove_weight_norm(self.conv1)
remove_weight_norm(self.conv2)
class MRFBlock(nn.Module):
def __init__(self, channels, kernel_size, dilations):
super().__init__()
self.layers = nn.ModuleList()
for dilation in dilations:
self.layers.append(MRFLayer(channels, kernel_size, dilation))
def forward(self, x):
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(nn.Module):
def __init__(self, samp_rate, harmonic_num = 0, sine_amp = 0.1, noise_std = 0.003, voiced_threshold = 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):
return torch.ones_like(f0) * (f0 > self.voiced_threshold)
def _f02sine(self, f0_values):
rad_values = (f0_values / self.sampling_rate) % 1
rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], dtype=f0_values.dtype, 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
return torch.sin(torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi)
def forward(self, f0):
with torch.no_grad():
f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, dtype=f0.dtype, 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)
sine_waves = sine_waves * uv + ((uv * self.noise_std + (1 - uv) * self.sine_amp / 3) * torch.randn_like(sine_waves))
return sine_waves
class SourceModuleHnNSF(nn.Module):
def __init__(self, sampling_rate, harmonic_num = 0, sine_amp = 0.1, add_noise_std = 0.003, voiced_threshold = 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 = nn.Linear(harmonic_num + 1, 1)
self.l_tanh = nn.Tanh()
def forward(self, x):
return self.l_tanh(self.l_linear(self.l_sin_gen(x).to(dtype=self.l_linear.weight.dtype)))
class HiFiGANMRFGenerator(nn.Module):
def __init__(self, in_channel, upsample_initial_channel, upsample_rates, upsample_kernel_sizes, resblock_kernel_sizes, resblock_dilations, gin_channels, sample_rate, harmonic_num, checkpointing = False):
super().__init__()
self.num_kernels = len(resblock_kernel_sizes)
self.checkpointing = checkpointing
self.f0_upsample = nn.Upsample(scale_factor=np.prod(upsample_rates))
self.m_source = SourceModuleHnNSF(sample_rate, harmonic_num)
self.conv_pre = weight_norm(nn.Conv1d(in_channel, upsample_initial_channel, kernel_size=7, stride=1, padding=3))
self.upsamples = nn.ModuleList()
self.noise_convs = 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)):
self.upsamples.append(weight_norm(nn.ConvTranspose1d(upsample_initial_channel // (2**i), upsample_initial_channel // (2 ** (i + 1)), kernel_size=k, stride=u, padding=((k - u) // 2) if u % 2 == 0 else (u // 2 + u % 2), output_padding=u % 2)))
stride = stride_f0s[i]
kernel = 1 if stride == 1 else stride * 2 - stride % 2
self.noise_convs.append(nn.Conv1d(1, upsample_initial_channel // (2 ** (i + 1)), kernel_size=kernel, stride=stride, padding=0 if stride == 1 else (kernel - stride) // 2))
self.mrfs = nn.ModuleList()
for i in range(len(self.upsamples)):
channel = upsample_initial_channel // (2 ** (i + 1))
self.mrfs.append(nn.ModuleList([MRFBlock(channel, kernel_size=k, dilations=d) for k, d in zip(resblock_kernel_sizes, resblock_dilations)]))
self.conv_post = weight_norm(nn.Conv1d(channel, 1, kernel_size=7, stride=1, padding=3))
if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
def forward(self, x, f0, g = None):
har_source = self.m_source(self.f0_upsample(f0[:, None, :]).transpose(-1, -2)).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 = F.leaky_relu(x, LRELU_SLOPE)
if self.training and self.checkpointing:
x = checkpoint(ups, x, use_reentrant=False) + noise_conv(har_source)
xs = sum([checkpoint(layer, x, use_reentrant=False) for layer in mrf])
else:
x = ups(x) + noise_conv(har_source)
xs = sum([layer(x) for layer in mrf])
x = xs / self.num_kernels
return torch.tanh(self.conv_post(F.leaky_relu(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) |