Closed Chris-Pedersen closed 2 years ago
Ok have bypassed much of the kymatio
frontend to import the target Torch class at: https://github.com/kymatio/kymatio/blob/776ad5988728117dba73a9a8c7a6cbef0f88593a/kymatio/scattering2d/frontend/torch_frontend.py
Next step is to see if we can import the remainder of the code called by this module. Then do a comparison with the previous network to make sure the output of the network hasn't changed, and we can merge.
Ok dropping kymatio
entirely from the code is a bit ambitious for now due to the number of methods we call from it. I should be able to do most of what we need to do just by modifying the scattering2d
function as this is where the convolutions are performed, using the kymatio
backends. Its still a bit annoying to have these redundant wavelets hanging around but lets see if I can tidy that up at a later stage.
PR to address the first stage of #13 , by dropping our
kymatio
imports