I am trying to replicate the Harmonic Convolution introduced in https://openreview.net/pdf?id=rygjHxrYDB.
Basically, I need a convolution that given a frequency bin omega, convolves normally over the time dimension (think image width) while processing the frequency bins (think image rows) given by (omega*k)n, where k and n are indexes that go from 1 to 7. As you can see, the concept is similar to dilated convolutions but spaced according to the harmonics of the frequency bin.
According to the paper, fractional frequencies should be handled by bilinear interpolation. My dumb question is, thus:
how should this be handled? Naively, I know I should use nn.interpolate, but I am not sure how to select the correct fractional bins from there. If it helps, the authors mentioned that their implementation is based on deformable convolutions.