each kernel shares the same offsets？

[ran337287]

#According to test.py, the line 'offset=conv(inputs)' is to learn the offset of neighbor domain of each pixel, so the size of offsets is 118H*W. if the size of input is 1x3xHxW and the size of output is 1x4xHxW, does it means that there has 4 kernels，and each kernel shares the same offsets？

[1zb]

Hi, @ran337287. You can say there is only one kernel with shape (4, 3, kH, kW), or four kernels with shape (1, 3, kH, kW). Both are acceptable. So yes, in every spatial location (i, j) ( 0 < i < H, 0 < j < W), four kernels share the same offsets.

[bkvie]

One could probably change that by modifying conv2d( groups=1) although that would not be what the original paper intended.

[bkvie]

@1zb would it make sense to have individual offsets? So 4 kernels of shape (1,3, kH, kW) with 4 offsets?