IMS-Toucan / Layers /ResidualStack.py
Florian Lux
initial commit
cea6632
raw
history blame
2.18 kB
# Copyright 2019 Tomoki Hayashi
# MIT License (https://opensource.org/licenses/MIT)
# Adapted by Florian Lux 2021
import torch
class ResidualStack(torch.nn.Module):
def __init__(self, kernel_size=3, channels=32, dilation=1, bias=True, nonlinear_activation="LeakyReLU", nonlinear_activation_params={"negative_slope": 0.2},
pad="ReflectionPad1d", pad_params={}, ):
"""
Initialize ResidualStack module.
Args:
kernel_size (int): Kernel size of dilation convolution layer.
channels (int): Number of channels of convolution layers.
dilation (int): Dilation factor.
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.
pad (str): Padding function module name before dilated convolution layer.
pad_params (dict): Hyperparameters for padding function.
"""
super(ResidualStack, self).__init__()
# defile residual stack part
assert (kernel_size - 1) % 2 == 0, "Not support even number kernel size."
self.stack = torch.nn.Sequential(getattr(torch.nn, nonlinear_activation)(**nonlinear_activation_params),
getattr(torch.nn, pad)((kernel_size - 1) // 2 * dilation, **pad_params),
torch.nn.Conv1d(channels, channels, kernel_size, dilation=dilation, bias=bias),
getattr(torch.nn, nonlinear_activation)(**nonlinear_activation_params),
torch.nn.Conv1d(channels, channels, 1, bias=bias), )
# defile extra layer for skip connection
self.skip_layer = torch.nn.Conv1d(channels, channels, 1, bias=bias)
def forward(self, c):
"""
Calculate forward propagation.
Args:
c (Tensor): Input tensor (B, channels, T).
Returns:
Tensor: Output tensor (B, chennels, T).
"""
return self.stack(c) + self.skip_layer(c)