With PyTorch, how is my Conv1d dimension reducing when I have padding?

Question

My conv module is:

        return torch.nn.Sequential(
            torch.nn.Conv1d(
                in_channels=in_channels,
                out_channels=in_channels,
                kernel_size=2,
                stride=1,
                dilation=1,
                padding=1
            ),
            torch.nn.ReLU(),
            torch.nn.Conv1d(
                in_channels=in_channels,
                out_channels=in_channels,
                kernel_size=2,
                stride=1,
                dilation=2,
                padding=1
            ),
            torch.nn.ReLU(),
            torch.nn.Conv1d(
                in_channels=in_channels,
                out_channels=in_channels,
                kernel_size=2,
                stride=1,
                dilation=4,
                padding=1
            ),
            torch.nn.ReLU()
        )

And in forward , I have:

down_out = self.downscale_time_conv(inputs)

inputs has a .size of torch.Size([8, 161, 24]) . I'd expect down_out to have the same size, but instead it has: torch.Size([8, 161, 23])

Where did that last element go?

Answer 1

The answer can be found on Pytorch documentation online ( here ). For every operation the output shape is expressed with respect to the input parameters:

For each conv1D:

- L1 = 25 → int((24 + 2*1 - 1*(2 - 1) - 1) / 1 + 1)
- L2 = 25 → int((25 + 2*1 - 2*(2 - 1) - 1) / 1 + 1)
- L3 = 23 → int((25 + 2*1 - 4*(2 - 1) - 1) / 1 + 1)

Do not forget that Lin is the previous size.