modules/downsampling.py

import torch

from modules.base import BaseModule
from modules.interpolation import InterpolationBlock
from modules.layers import Conv1dWithInitialization


class ConvolutionBlock(BaseModule):
    def __init__(self, in_channels, out_channels, dilation):
        super(ConvolutionBlock, self).__init__()
        self.leaky_relu = torch.nn.LeakyReLU(0.2)
        self.convolution = Conv1dWithInitialization(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=3,
            stride=1,
            padding=dilation,
            dilation=dilation
        )
    
    def forward(self, x):
        outputs = self.leaky_relu(x)
        outputs = self.convolution(outputs)
        return outputs


class DownsamplingBlock(BaseModule):
    def __init__(self, in_channels, out_channels, factor, dilations):
        super(DownsamplingBlock, self).__init__()
        in_sizes = [in_channels] + [out_channels for _ in range(len(dilations) - 1)]
        out_sizes = [out_channels for _ in range(len(in_sizes))]
        self.main_branch = torch.nn.Sequential(*([
            InterpolationBlock(
                scale_factor=factor,
                mode='linear',
                align_corners=False,
                downsample=True
            )
        ] + [
            ConvolutionBlock(in_size, out_size, dilation)
            for in_size, out_size, dilation in zip(in_sizes, out_sizes, dilations)
        ]))
        self.residual_branch = torch.nn.Sequential(*[
            Conv1dWithInitialization(
                in_channels=in_channels,
                out_channels=out_channels,
                kernel_size=1,
                stride=1
            ),
            InterpolationBlock(
                scale_factor=factor,
                mode='linear',
                align_corners=False,
                downsample=True
            )
        ])

    def forward(self, x):
        outputs = self.main_branch(x)
        outputs = outputs + self.residual_branch(x)
        return outputs