mirrored_conv.py

import torch
import torch.nn as nn


class MirroredConv1d(nn.Module):
    def __init__(self, **kwargs):
        super(MirroredConv1d, self).__init__()

        if kwargs['out_channels'] % 2 == 1:
            raise ValueError('Number of output channels must be even')
        kwargs['out_channels'] //= 2

        self.conv = nn.Conv1d(**kwargs)

    def forward(self, x):
        x1 = self.conv(x)
        x2 = torch.flip(
            self.conv(
                torch.flip(x, dims=(2,))
            ),
            dims=(2,)
        )
        return torch.cat((x1, x2), dim=1)


if __name__ == '__main__':
    layer = MirroredConv1d(
        in_channels=3,
        out_channels=10,
        kernel_size=3
    )

    x = torch.zeros(7, 3, 128)
    y = layer(x)
    print(y.shape)  # (7, 10, 126)