WindyMadman/so-vits-svc
1
0
Fork 0
mirror of https://github.com/svc-develop-team/so-vits-svc.git synced 2025-01-09 12:37:30 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add files via upload

Browse Source
This commit is contained in:
Ναρουσέ·μ·γιουμεμί·Χινακάννα 2023-03-10 19:41:05 +08:00 committed by GitHub
parent 120444b293
commit 5b249743cd
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 437 additions and 1298 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

97
configs/config.json
View File

@ -1,103 +1,62 @@
{ {
"train": { "train": {
"log_interval": 50, "log_interval": 200,
"eval_interval": 1000, "eval_interval": 800,
"seed": 1234, "seed": 1234,
"port": 8001,
"epochs": 10000, "epochs": 10000,
"learning_rate": 0.0002, "learning_rate": 0.0001,
"betas": [ "betas": [
0.8, 0.8,
0.99 0.99
], ],
"eps": 1e-09, "eps": 1e-09,
"batch_size": 6, "batch_size": 6,
"accumulation_steps": 1,
"fp16_run": false, "fp16_run": false,
"lr_decay": 0.998, "lr_decay": 0.999875,
"segment_size": 10240, "segment_size": 10240,
"init_lr_ratio": 1, "init_lr_ratio": 1,
"warmup_epochs": 0, "warmup_epochs": 0,
"c_mel": 45, "c_mel": 45,
"keep_ckpts":4 "c_kl": 1.0,
"use_sr": true,
"max_speclen": 512,
"port": "8001",
"keep_ckpts": 3
}, },
"data": { "data": {
"data_dir": "dataset", "training_files": "filelists/train.txt",
"dataset_type": "SingDataset", "validation_files": "filelists/val.txt",
"collate_type": "SingCollate",
"training_filelist": "filelists/train.txt",
"validation_filelist": "filelists/val.txt",
"max_wav_value": 32768.0, "max_wav_value": 32768.0,
"sampling_rate": 44100, "sampling_rate": 44100,
"n_fft": 2048, "filter_length": 2048,
"fmin": 0,
"fmax": 22050,
"hop_length": 512, "hop_length": 512,
"win_size": 2048, "win_length": 2048,
"acoustic_dim": 80, "n_mel_channels": 80,
"c_dim": 256, "mel_fmin": 0.0,
"min_level_db": -115, "mel_fmax": 22050
"ref_level_db": 20,
"min_db": -115,
"max_abs_value": 4.0,
"n_speakers": 200
}, },
"model": { "model": {
"inter_channels": 192,
"hidden_channels": 192, "hidden_channels": 192,
"spk_channels": 192,
"filter_channels": 768, "filter_channels": 768,
"n_heads": 2, "n_heads": 2,
"n_layers": 4, "n_layers": 6,
"kernel_size": 3, "kernel_size": 3,
"p_dropout": 0.1, "p_dropout": 0.1,
"prior_hidden_channels": 192,
"prior_filter_channels": 768,
"prior_n_heads": 2,
"prior_n_layers": 4,
"prior_kernel_size": 3,
"prior_p_dropout": 0.1,
"resblock": "1", "resblock": "1",
"resblock_kernel_sizes": [3,7,11],
"resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
"upsample_rates": [ 8, 8, 2, 2, 2],
"upsample_initial_channel": 512,
"upsample_kernel_sizes": [16,16, 4, 4, 4],
"n_layers_q": 3,
"use_spectral_norm": false, "use_spectral_norm": false,
"resblock_kernel_sizes": [ "gin_channels": 256,
3, "ssl_dim": 256,
7, "n_speakers": 200
11
],
"resblock_dilation_sizes": [
[
1,
3,
5
],
[
1,
3,
5
],
[
1,
3,
5
]
],
"upsample_rates": [
8,
8,
4,
2
],
"upsample_initial_channel": 256,
"upsample_kernel_sizes": [
16,
16,
8,
4
],
"n_harmonic": 64,
"n_bands": 65
}, },
"spk": { "spk": {
"jishuang": 0, "nyaru": 0,
"huiyu": 1, "huiyu": 1,
"nen": 2, "nen": 2,
"paimon": 3, "paimon": 3,

7
inference/infer_tool.py
View File

@ -127,8 +127,9 @@ class Svc(object):
def load_model(self): def load_model(self):
# 获取模型配置 # 获取模型配置
self.net_g_ms = SynthesizerTrn( self.net_g_ms = SynthesizerTrn(
self.hps_ms self.hps_ms.data.filter_length // 2 + 1,
) self.hps_ms.train.segment_size // self.hps_ms.data.hop_length,
**self.hps_ms.model)
_ = utils.load_checkpoint(self.net_g_path, self.net_g_ms, None) _ = utils.load_checkpoint(self.net_g_path, self.net_g_ms, None)
if "half" in self.net_g_path and torch.cuda.is_available(): if "half" in self.net_g_path and torch.cuda.is_available():
_ = self.net_g_ms.half().eval().to(self.dev) _ = self.net_g_ms.half().eval().to(self.dev)
@ -173,7 +174,7 @@ class Svc(object):
c = c.half() c = c.half()
with torch.no_grad(): with torch.no_grad():
start = time.time() start = time.time()
audio = self.net_g_ms.infer(c, f0=f0, g=sid, uv=uv, predict_f0=auto_predict_f0, noice_scale=noice_scale)[0][0,0].data.float() audio = self.net_g_ms.infer(c, f0=f0, g=sid, uv=uv, predict_f0=auto_predict_f0, noice_scale=noice_scale)[0,0].data.float()
use_time = time.time() - start use_time = time.time() - start
print("vits use time:{}".format(use_time)) print("vits use time:{}".format(use_time))
return audio, audio.shape[-1] return audio, audio.shape[-1]

535
modules/modules.py
View File

@ -5,187 +5,182 @@ import scipy
import torch import torch
from torch import nn from torch import nn
from torch.nn import functional as F from torch.nn import functional as F
from torch.autograd import Function
from typing import Any, Optional, Tuple
from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
from torch.nn.utils import weight_norm, remove_weight_norm from torch.nn.utils import weight_norm, remove_weight_norm
import modules.commons as commons import modules.commons as commons
from modules.commons import init_weights, get_padding from modules.commons import init_weights, get_padding
from modules.transforms import piecewise_rational_quadratic_transform
LRELU_SLOPE = 0.1 LRELU_SLOPE = 0.1
class LayerNorm(nn.Module): class LayerNorm(nn.Module):
def __init__(self, channels, eps=1e-5): def __init__(self, channels, eps=1e-5):
super().__init__() super().__init__()
self.channels = channels self.channels = channels
self.eps = eps self.eps = eps
self.gamma = nn.Parameter(torch.ones(channels)) self.gamma = nn.Parameter(torch.ones(channels))
self.beta = nn.Parameter(torch.zeros(channels)) self.beta = nn.Parameter(torch.zeros(channels))
def forward(self, x):
x = x.transpose(1, -1)
x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
return x.transpose(1, -1)
def forward(self, x):
x = x.transpose(1, -1)
x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
return x.transpose(1, -1)
class ConvReluNorm(nn.Module): class ConvReluNorm(nn.Module):
def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout): def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
super().__init__() super().__init__()
self.in_channels = in_channels self.in_channels = in_channels
self.hidden_channels = hidden_channels self.hidden_channels = hidden_channels
self.out_channels = out_channels self.out_channels = out_channels
self.kernel_size = kernel_size self.kernel_size = kernel_size
self.n_layers = n_layers self.n_layers = n_layers
self.p_dropout = p_dropout self.p_dropout = p_dropout
assert n_layers > 1, "Number of layers should be larger than 0." assert n_layers > 1, "Number of layers should be larger than 0."
self.conv_layers = nn.ModuleList() self.conv_layers = nn.ModuleList()
self.norm_layers = nn.ModuleList() self.norm_layers = nn.ModuleList()
self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size // 2)) self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size//2))
self.norm_layers.append(LayerNorm(hidden_channels)) self.norm_layers.append(LayerNorm(hidden_channels))
self.relu_drop = nn.Sequential( self.relu_drop = nn.Sequential(
nn.ReLU(), nn.ReLU(),
nn.Dropout(p_dropout)) nn.Dropout(p_dropout))
for _ in range(n_layers - 1): for _ in range(n_layers-1):
self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size // 2)) self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size//2))
self.norm_layers.append(LayerNorm(hidden_channels)) self.norm_layers.append(LayerNorm(hidden_channels))
self.proj = nn.Conv1d(hidden_channels, out_channels, 1) self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
self.proj.weight.data.zero_() self.proj.weight.data.zero_()
self.proj.bias.data.zero_() self.proj.bias.data.zero_()
def forward(self, x, x_mask): def forward(self, x, x_mask):
x_org = x x_org = x
for i in range(self.n_layers): for i in range(self.n_layers):
x = self.conv_layers[i](x * x_mask) x = self.conv_layers[i](x * x_mask)
x = self.norm_layers[i](x) x = self.norm_layers[i](x)
x = self.relu_drop(x) x = self.relu_drop(x)
x = x_org + self.proj(x) x = x_org + self.proj(x)
return x * x_mask return x * x_mask
class DDSConv(nn.Module): class DDSConv(nn.Module):
""" """
Dialted and Depth-Separable Convolution Dialted and Depth-Separable Convolution
""" """
def __init__(self, channels, kernel_size, n_layers, p_dropout=0.):
super().__init__()
self.channels = channels
self.kernel_size = kernel_size
self.n_layers = n_layers
self.p_dropout = p_dropout
def __init__(self, channels, kernel_size, n_layers, p_dropout=0.): self.drop = nn.Dropout(p_dropout)
super().__init__() self.convs_sep = nn.ModuleList()
self.channels = channels self.convs_1x1 = nn.ModuleList()
self.kernel_size = kernel_size self.norms_1 = nn.ModuleList()
self.n_layers = n_layers self.norms_2 = nn.ModuleList()
self.p_dropout = p_dropout for i in range(n_layers):
dilation = kernel_size ** i
padding = (kernel_size * dilation - dilation) // 2
self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size,
groups=channels, dilation=dilation, padding=padding
))
self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
self.norms_1.append(LayerNorm(channels))
self.norms_2.append(LayerNorm(channels))
self.drop = nn.Dropout(p_dropout) def forward(self, x, x_mask, g=None):
self.convs_sep = nn.ModuleList() if g is not None:
self.convs_1x1 = nn.ModuleList() x = x + g
self.norms_1 = nn.ModuleList() for i in range(self.n_layers):
self.norms_2 = nn.ModuleList() y = self.convs_sep[i](x * x_mask)
for i in range(n_layers): y = self.norms_1[i](y)
dilation = kernel_size ** i y = F.gelu(y)
padding = (kernel_size * dilation - dilation) // 2 y = self.convs_1x1[i](y)
self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size, y = self.norms_2[i](y)
groups=channels, dilation=dilation, padding=padding y = F.gelu(y)
)) y = self.drop(y)
self.convs_1x1.append(nn.Conv1d(channels, channels, 1)) x = x + y
self.norms_1.append(LayerNorm(channels)) return x * x_mask
self.norms_2.append(LayerNorm(channels))
def forward(self, x, x_mask, g=None):
if g is not None:
x = x + g
for i in range(self.n_layers):
y = self.convs_sep[i](x * x_mask)
y = self.norms_1[i](y)
y = F.gelu(y)
y = self.convs_1x1[i](y)
y = self.norms_2[i](y)
y = F.gelu(y)
y = self.drop(y)
x = x + y
return x * x_mask
class WN(torch.nn.Module): class WN(torch.nn.Module):
def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, n_speakers=0, spk_channels=0, def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
p_dropout=0): super(WN, self).__init__()
super(WN, self).__init__() assert(kernel_size % 2 == 1)
assert (kernel_size % 2 == 1) self.hidden_channels =hidden_channels
self.hidden_channels = hidden_channels self.kernel_size = kernel_size,
self.kernel_size = kernel_size, self.dilation_rate = dilation_rate
self.dilation_rate = dilation_rate self.n_layers = n_layers
self.n_layers = n_layers self.gin_channels = gin_channels
self.n_speakers = n_speakers self.p_dropout = p_dropout
self.spk_channels = spk_channels
self.p_dropout = p_dropout
self.in_layers = torch.nn.ModuleList() self.in_layers = torch.nn.ModuleList()
self.res_skip_layers = torch.nn.ModuleList() self.res_skip_layers = torch.nn.ModuleList()
self.drop = nn.Dropout(p_dropout) self.drop = nn.Dropout(p_dropout)
if n_speakers > 0: if gin_channels != 0:
cond_layer = torch.nn.Conv1d(spk_channels, 2 * hidden_channels * n_layers, 1) cond_layer = torch.nn.Conv1d(gin_channels, 2*hidden_channels*n_layers, 1)
self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name='weight') self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name='weight')
for i in range(n_layers): for i in range(n_layers):
dilation = dilation_rate ** i dilation = dilation_rate ** i
padding = int((kernel_size * dilation - dilation) / 2) padding = int((kernel_size * dilation - dilation) / 2)
in_layer = torch.nn.Conv1d(hidden_channels, 2 * hidden_channels, kernel_size, in_layer = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, kernel_size,
dilation=dilation, padding=padding) dilation=dilation, padding=padding)
in_layer = torch.nn.utils.weight_norm(in_layer, name='weight') in_layer = torch.nn.utils.weight_norm(in_layer, name='weight')
self.in_layers.append(in_layer) self.in_layers.append(in_layer)
# last one is not necessary # last one is not necessary
if i < n_layers - 1: if i < n_layers - 1:
res_skip_channels = 2 * hidden_channels res_skip_channels = 2 * hidden_channels
else: else:
res_skip_channels = hidden_channels res_skip_channels = hidden_channels
res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1) res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name='weight') res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name='weight')
self.res_skip_layers.append(res_skip_layer) self.res_skip_layers.append(res_skip_layer)
def forward(self, x, x_mask, g=None, **kwargs): def forward(self, x, x_mask, g=None, **kwargs):
output = torch.zeros_like(x) output = torch.zeros_like(x)
n_channels_tensor = torch.IntTensor([self.hidden_channels]) n_channels_tensor = torch.IntTensor([self.hidden_channels])
if g is not None: if g is not None:
g = self.cond_layer(g) g = self.cond_layer(g)
for i in range(self.n_layers): for i in range(self.n_layers):
x_in = self.in_layers[i](x) x_in = self.in_layers[i](x)
if g is not None: if g is not None:
cond_offset = i * 2 * self.hidden_channels cond_offset = i * 2 * self.hidden_channels
g_l = g[:, cond_offset:cond_offset + 2 * self.hidden_channels, :] g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]
else: else:
g_l = torch.zeros_like(x_in) g_l = torch.zeros_like(x_in)
acts = commons.fused_add_tanh_sigmoid_multiply( acts = commons.fused_add_tanh_sigmoid_multiply(
x_in, x_in,
g_l, g_l,
n_channels_tensor) n_channels_tensor)
acts = self.drop(acts) acts = self.drop(acts)
res_skip_acts = self.res_skip_layers[i](acts) res_skip_acts = self.res_skip_layers[i](acts)
if i < self.n_layers - 1: if i < self.n_layers - 1:
res_acts = res_skip_acts[:, :self.hidden_channels, :] res_acts = res_skip_acts[:,:self.hidden_channels,:]
x = (x + res_acts) * x_mask x = (x + res_acts) * x_mask
output = output + res_skip_acts[:, self.hidden_channels:, :] output = output + res_skip_acts[:,self.hidden_channels:,:]
else: else:
output = output + res_skip_acts output = output + res_skip_acts
return output * x_mask return output * x_mask
def remove_weight_norm(self): def remove_weight_norm(self):
if self.n_speakers > 0: if self.gin_channels != 0:
torch.nn.utils.remove_weight_norm(self.cond_layer) torch.nn.utils.remove_weight_norm(self.cond_layer)
for l in self.in_layers: for l in self.in_layers:
torch.nn.utils.remove_weight_norm(l) torch.nn.utils.remove_weight_norm(l)
for l in self.res_skip_layers: for l in self.res_skip_layers:
torch.nn.utils.remove_weight_norm(l) torch.nn.utils.remove_weight_norm(l)
class ResBlock1(torch.nn.Module): class ResBlock1(torch.nn.Module):
@ -261,193 +256,87 @@ class ResBlock2(torch.nn.Module):
class Log(nn.Module): class Log(nn.Module):
def forward(self, x, x_mask, reverse=False, **kwargs): def forward(self, x, x_mask, reverse=False, **kwargs):
if not reverse: if not reverse:
y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
logdet = torch.sum(-y, [1, 2]) logdet = torch.sum(-y, [1, 2])
return y, logdet return y, logdet
else: else:
x = torch.exp(x) * x_mask x = torch.exp(x) * x_mask
return x return x
class Flip(nn.Module): class Flip(nn.Module):
def forward(self, x, *args, reverse=False, **kwargs): def forward(self, x, *args, reverse=False, **kwargs):
x = torch.flip(x, [1]) x = torch.flip(x, [1])
if not reverse: if not reverse:
logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device) logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
return x, logdet return x, logdet
else: else:
return x return x
class ElementwiseAffine(nn.Module): class ElementwiseAffine(nn.Module):
def __init__(self, channels): def __init__(self, channels):
super().__init__() super().__init__()
self.channels = channels self.channels = channels
self.m = nn.Parameter(torch.zeros(channels, 1)) self.m = nn.Parameter(torch.zeros(channels,1))
self.logs = nn.Parameter(torch.zeros(channels, 1)) self.logs = nn.Parameter(torch.zeros(channels,1))
def forward(self, x, x_mask, reverse=False, **kwargs): def forward(self, x, x_mask, reverse=False, **kwargs):
if not reverse: if not reverse:
y = self.m + torch.exp(self.logs) * x y = self.m + torch.exp(self.logs) * x
y = y * x_mask y = y * x_mask
logdet = torch.sum(self.logs * x_mask, [1, 2]) logdet = torch.sum(self.logs * x_mask, [1,2])
return y, logdet return y, logdet
else: else:
x = (x - self.m) * torch.exp(-self.logs) * x_mask x = (x - self.m) * torch.exp(-self.logs) * x_mask
return x return x
class ResidualCouplingLayer(nn.Module): class ResidualCouplingLayer(nn.Module):
def __init__(self, def __init__(self,
channels, channels,
hidden_channels, hidden_channels,
kernel_size, kernel_size,
dilation_rate, dilation_rate,
n_layers, n_layers,
p_dropout=0, p_dropout=0,
n_speakers=0, gin_channels=0,
spk_channels=0, mean_only=False):
mean_only=False): assert channels % 2 == 0, "channels should be divisible by 2"
assert channels % 2 == 0, "channels should be divisible by 2" super().__init__()
super().__init__() self.channels = channels
self.channels = channels self.hidden_channels = hidden_channels
self.hidden_channels = hidden_channels self.kernel_size = kernel_size
self.kernel_size = kernel_size self.dilation_rate = dilation_rate
self.dilation_rate = dilation_rate self.n_layers = n_layers
self.n_layers = n_layers self.half_channels = channels // 2
self.half_channels = channels // 2 self.mean_only = mean_only
self.mean_only = mean_only
self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1) self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, n_speakers=n_speakers, self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
spk_channels=spk_channels) self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1) self.post.weight.data.zero_()
self.post.weight.data.zero_() self.post.bias.data.zero_()
self.post.bias.data.zero_()
def forward(self, x, x_mask, g=None, reverse=False): def forward(self, x, x_mask, g=None, reverse=False):
x0, x1 = torch.split(x, [self.half_channels] * 2, 1) x0, x1 = torch.split(x, [self.half_channels]*2, 1)
h = self.pre(x0) * x_mask h = self.pre(x0) * x_mask
h = self.enc(h, x_mask, g=g) h = self.enc(h, x_mask, g=g)
stats = self.post(h) * x_mask stats = self.post(h) * x_mask
if not self.mean_only: if not self.mean_only:
m, logs = torch.split(stats, [self.half_channels] * 2, 1) m, logs = torch.split(stats, [self.half_channels]*2, 1)
else: else:
m = stats m = stats
logs = torch.zeros_like(m) logs = torch.zeros_like(m)
if not reverse: if not reverse:
x1 = m + x1 * torch.exp(logs) * x_mask x1 = m + x1 * torch.exp(logs) * x_mask
x = torch.cat([x0, x1], 1) x = torch.cat([x0, x1], 1)
logdet = torch.sum(logs, [1, 2]) logdet = torch.sum(logs, [1,2])
return x, logdet return x, logdet
else: else:
x1 = (x1 - m) * torch.exp(-logs) * x_mask x1 = (x1 - m) * torch.exp(-logs) * x_mask
x = torch.cat([x0, x1], 1) x = torch.cat([x0, x1], 1)
return x return x
class ResidualCouplingBlock(nn.Module):
def __init__(self,
channels,
hidden_channels,
kernel_size,
dilation_rate,
n_layers,
n_flows=4,
n_speakers=0,
gin_channels=0):
super().__init__()
self.channels = channels
self.hidden_channels = hidden_channels
self.kernel_size = kernel_size
self.dilation_rate = dilation_rate
self.n_layers = n_layers
self.n_flows = n_flows
self.gin_channels = gin_channels
self.flows = nn.ModuleList()
for i in range(n_flows):
self.flows.append(ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers,
n_speakers=n_speakers, spk_channels=gin_channels, mean_only=True))
self.flows.append(Flip())
def forward(self, x, x_mask, g=None, reverse=False):
if not reverse:
for flow in self.flows:
x, _ = flow(x, x_mask, g=g, reverse=reverse)
else:
for flow in reversed(self.flows):
x = flow(x, x_mask, g=g, reverse=reverse)
return x
class ConvFlow(nn.Module):
def __init__(self, in_channels, filter_channels, kernel_size, n_layers, num_bins=10, tail_bound=5.0):
super().__init__()
self.in_channels = in_channels
self.filter_channels = filter_channels
self.kernel_size = kernel_size
self.n_layers = n_layers
self.num_bins = num_bins
self.tail_bound = tail_bound
self.half_channels = in_channels // 2
self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.)
self.proj = nn.Conv1d(filter_channels, self.half_channels * (num_bins * 3 - 1), 1)
self.proj.weight.data.zero_()
self.proj.bias.data.zero_()
def forward(self, x, x_mask, g=None, reverse=False):
x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
h = self.pre(x0)
h = self.convs(h, x_mask, g=g)
h = self.proj(h) * x_mask
b, c, t = x0.shape
h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2) # [b, cx?, t] -> [b, c, t, ?]
unnormalized_widths = h[..., :self.num_bins] / math.sqrt(self.filter_channels)
unnormalized_heights = h[..., self.num_bins:2 * self.num_bins] / math.sqrt(self.filter_channels)
unnormalized_derivatives = h[..., 2 * self.num_bins:]
x1, logabsdet = piecewise_rational_quadratic_transform(x1,
unnormalized_widths,
unnormalized_heights,
unnormalized_derivatives,
inverse=reverse,
tails='linear',
tail_bound=self.tail_bound
)
x = torch.cat([x0, x1], 1) * x_mask
logdet = torch.sum(logabsdet * x_mask, [1, 2])
if not reverse:
return x, logdet
else:
return x
class ResStack(nn.Module):
def __init__(self, channel, kernel_size=3, base=3, nums=4):
super(ResStack, self).__init__()
self.layers = nn.ModuleList([
nn.Sequential(
nn.LeakyReLU(),
nn.utils.weight_norm(nn.Conv1d(channel, channel,
kernel_size=kernel_size, dilation=base ** i, padding=base ** i)),
nn.LeakyReLU(),
nn.utils.weight_norm(nn.Conv1d(channel, channel,
kernel_size=kernel_size, dilation=1, padding=1)),
)
for i in range(nums)
])
def forward(self, x):
for layer in self.layers:
x = x + layer(x)
return x

1096
onnxexport/model_onnx.py
View File

File diff suppressed because it is too large Load Diff