conv2d function in pytorch

Question

I'm trying to use the function torch.conv2d from Pytorch but can't get a result I understand...

Here is a simple example where the kernel (filt) is the same size as the input (im) to explain what I'm looking for.

import pytorch

filt = torch.rand(3, 3)
im = torch.rand(3, 3)

I want to compute a simple convolution with no padding, so the result should be a scalar (i.e. a 1x1 tensor).

I tried this with conv2d:

# I have to convert image and kernel to 4 dimensions tensors to use conv2d
im_torch = im.reshape((im_height, filt_height, 1, 1))
filt_torch = filt.reshape((filt_height, im_height, 1, 1))
out = torch.nn.functional.conv2d(im_torch, filt_torch, stride=1, padding=0)
print(out)

But the result is not what I expected:

tensor([[[[0.6067]], [[0.3564]], [[0.5397]]],
    [[[0.2557]], [[0.0493]], [[0.2562]]],
    [[[0.6067]], [[0.3564]], [[0.5397]]]])

To give an idea of what I'd like, I want to reproduce scipy convolve2d behavior:

import scipy.signal
out_scipy = scipy.signal.convolve2d(im.detach().numpy(), filt.detach().numpy(), 'valid')
print(out_scipy)

which prints:

array([[1.195723]], dtype=float32)

Answer 1

To get equivalent results :

im_torch = im.reshape((1, 1, 3, 3))
filt_torch = filt.reshape((1, 1, 3, 3))

out = torch.nn.functional.conv2d(filt_torch, im_torch, stride=1, padding=0)
print(out)


import scipy.signal


out_scipy = scipy.signal.convolve2d(im.detach().numpy(), filt.detach().numpy()[::-1, ::-1], 'valid')
print(out_scipy)

There was a shaping issue (https://pytorch.org/docs/stable/nn.functional.html) and the application of the filter in torch differs a bit with how it is handled in scipy : https://discuss.pytorch.org/t/functional-conv2d-produces-different-results-vs-scipy-convolved2d/17762

Answer 2

Ok, I didn't find the exact answer to my question (i.e. how to use conv2d) but I found another way to do it.

First of all, I learned that I'm looking for is called a valid cross-correlation and it is actually the operation implemented by the [Conv2d][1] class.

Hence my solution uses the Conv2d class instead of the conv2d function.

import pytorch

img = torch.rand(3, 3)

model = torch.nn.Conv2d(in_channels=1, out_channels=1, kernel_size=(3, 3), stride=1, padding=0, bias=False)

res = conv_mdl(img)
print(res.shape)

Which prints the scalar I wanted:

torch.Size([1, 1, 1, 1])

PS: I also checked that the result is the right one, not just the dimension.