Wrong output shape of torch.nn.Conv2d with 2d stride overload

[PointerGuide]

Hi,
according to python's torch documentation (image below),
with W_in: 8, kernelSize: (1, 7), stride: (1, 1), padding: (0, 3) and default dilation (1,1) i should have recived W_out = 8.
I got 2 instead. That could be an issue I think.

There isn't that problem with H_out, when i set
W_in: 8, kernelSize: (7, 1), stride: (1, 1), padding: (3, 0)
Screen from
https://pytorch.org/docs/1.9.1/generated/torch.nn.Conv2d.html

[NiklasGustafsson]

I'm seeing the same thing. Thanks for the issue report.

[PointerGuide]

here is the problem
i want to set padding (0,3) but it sets that enum instead

[NiklasGustafsson]

Yes, it's the encoding of padding='same' and padding='valid' that is incorrectly received in the C++ code. Same for Conv3d.

The correct code should look like this:

    torch::nn::Conv2dOptions::padding_t padd =
        (paddingX == -1) && (paddingY == 0) ? torch::kSame :
        (paddingX == 0) && (paddingY == 0) ? torch::kValid :
        (torch::nn::Conv2dOptions::padding_t)torch::ExpandingArray<2>({ paddingX, paddingY });

[NiklasGustafsson]

@241721 -- since you already looked into the details yourself, maybe you could do a quick review of PR #497?

[NiklasGustafsson]

@241721 -- thanks for the PR review!

[PointerGuide]

Thx for your response, i approved.