The documentation for the nn mentions it does a cross-correlation, however, my results indicate it does a convolution operator.
import torch
import torch.nn.functional as F
im = torch.Tensor([[0,1],[2,3]]).unsqueeze(0).unsqueeze(0)kernel = torch.Tensor([[0,1,2],[3,4,5], [6,7,8]]).unsqueeze(0).unsqueeze(0)
op = F.conv_transpose2d(im, kernel, stride=2)
print(op)
It outputs :
tensor([[[[ 0., 0., 0., 1., 2.],
[ 0., 0., 3., 4., 5.],
[ 0., 2., 10., 10., 14.],
[ 6., 8., 19., 12., 15.],
[12., 14., 34., 21., 24.]]]])
which would be the result if there was a convolution, I had expected the correlation result to be:
tensor([[[[ 0., 0., 8., 7., 6.],
[ 0., 0., 5., 4., 3.],
[16., 14., 38., 22., 18.],
[10., 8., 21., 12., 9.],
[ 4., 2., 6., 3., 0.]]]])
Did I misunderstand something or is it a bug?