wehub-resource-sync
105 lines
4.1 KiB
Python
105 lines
4.1 KiB
Python
import math
|
|
from math import sqrt
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
|
|
from modules.commons.common_layers import SinusoidalPosEmb, AdamWConv1d
|
|
from modules.commons.common_layers import KaimingNormalConv1d as Conv1d
|
|
from utils.hparams import hparams
|
|
|
|
|
|
class ResidualBlock(nn.Module):
|
|
def __init__(self, encoder_hidden, residual_channels, dilation):
|
|
super().__init__()
|
|
self.residual_channels = residual_channels
|
|
self.dilated_conv = nn.Conv1d(
|
|
residual_channels,
|
|
2 * residual_channels,
|
|
kernel_size=3,
|
|
padding=dilation,
|
|
dilation=dilation
|
|
)
|
|
self.diffusion_projection = nn.Linear(residual_channels, residual_channels)
|
|
self.conditioner_projection = nn.Conv1d(encoder_hidden, 2 * residual_channels, 1)
|
|
self.output_projection = nn.Conv1d(residual_channels, 2 * residual_channels, 1)
|
|
|
|
def forward(self, x, conditioner, diffusion_step):
|
|
diffusion_step = self.diffusion_projection(diffusion_step).unsqueeze(-1)
|
|
conditioner = self.conditioner_projection(conditioner)
|
|
y = x + diffusion_step
|
|
|
|
y = self.dilated_conv(y) + conditioner
|
|
|
|
# Using torch.split instead of torch.chunk to avoid using onnx::Slice
|
|
gate, filter = torch.split(y, [self.residual_channels, self.residual_channels], dim=1)
|
|
y = torch.sigmoid(gate) * torch.tanh(filter)
|
|
|
|
y = self.output_projection(y)
|
|
|
|
# Using torch.split instead of torch.chunk to avoid using onnx::Slice
|
|
residual, skip = torch.split(y, [self.residual_channels, self.residual_channels], dim=1)
|
|
return (x + residual) / math.sqrt(2.0), skip
|
|
|
|
|
|
class WaveNet(nn.Module):
|
|
def __init__(self, in_dims, n_feats, *, num_layers=20, num_channels=256, dilation_cycle_length=4):
|
|
super().__init__()
|
|
self.in_dims = in_dims
|
|
self.n_feats = n_feats
|
|
self.input_projection = Conv1d(in_dims * n_feats, num_channels, 1)
|
|
self.diffusion_embedding = SinusoidalPosEmb(num_channels)
|
|
self.mlp = nn.Sequential(
|
|
nn.Linear(num_channels, num_channels * 4),
|
|
nn.Mish(),
|
|
nn.Linear(num_channels * 4, num_channels)
|
|
)
|
|
self.residual_layers = nn.ModuleList([
|
|
ResidualBlock(
|
|
encoder_hidden=hparams['hidden_size'],
|
|
residual_channels=num_channels,
|
|
dilation=2 ** (i % dilation_cycle_length)
|
|
)
|
|
for i in range(num_layers)
|
|
])
|
|
self.skip_projection = Conv1d(num_channels, num_channels, 1)
|
|
self.output_projection = AdamWConv1d(num_channels, in_dims * n_feats, 1)
|
|
nn.init.zeros_(self.output_projection.weight)
|
|
|
|
def forward(self, spec, diffusion_step, cond):
|
|
"""
|
|
:param spec: [B, F, M, T]
|
|
:param diffusion_step: [B, 1]
|
|
:param cond: [B, H, T]
|
|
:return:
|
|
"""
|
|
if self.n_feats == 1:
|
|
# Use indexing instead of squeeze to avoid emitting an onnx::If
|
|
# whose branches have different rank, which breaks shape inference
|
|
# for the downstream Conv on PyTorch >= 2.0.
|
|
x = spec[:, 0] # [B, M, T]
|
|
else:
|
|
x = spec.flatten(start_dim=1, end_dim=2) # [B, F x M, T]
|
|
x = self.input_projection(x) # [B, C, T]
|
|
|
|
x = F.relu(x)
|
|
diffusion_step = self.diffusion_embedding(diffusion_step)
|
|
diffusion_step = self.mlp(diffusion_step)
|
|
skip = []
|
|
for layer in self.residual_layers:
|
|
x, skip_connection = layer(x, cond, diffusion_step)
|
|
skip.append(skip_connection)
|
|
|
|
x = torch.sum(torch.stack(skip), dim=0) / sqrt(len(self.residual_layers))
|
|
x = self.skip_projection(x)
|
|
x = F.relu(x)
|
|
x = self.output_projection(x) # [B, M, T]
|
|
if self.n_feats == 1:
|
|
x = x[:, None, :, :]
|
|
else:
|
|
# Using reshape instead of unflatten for ONNX export compatibility
|
|
# x = x.unflatten(dim=1, sizes=(self.n_feats, self.in_dims))
|
|
x = x.reshape(-1, self.n_feats, self.in_dims, x.shape[2])
|
|
return x
|