adam-dziedzic/bandlimited-cnns

Band-limited Training and Inference for Convolutional Neural Networks

bandlimited-cnns's Issues

FFT conv with wrap-around

[Bit offtopic] While researching convolutions via FFT i stumbled upon this library which seems to be very well implemented.
For research I need convolutions via FFT that don't pad the input. It seems trivial but I don't seem to get it right.
I thought I only need to modify:

bandlimited-cnns/cnns/nnlib/pytorch_layers/conv2D_fft.py

Line 264 in 39840c9

if out_size_H:

with:

        if out_size_H:
            out_H = out_size_H
        elif out_size or stride_type is StrideType.SPECTRAL:
            out_H = (H - HH + 2 * pad_H) // stride_H + 1
        elif wrap_around: #new
            out_H = H
        else:
            out_H = H - HH + 1 + 2 * pad_H

(similarly h)
and

bandlimited-cnns/cnns/nnlib/pytorch_layers/conv2D_fft.py

Line 290 in 39840c9

init_H_fft = H + 2 * pad_H + HHH - 1

if wrap_around: #new
   init_H_fft = H + 2 * pad_H
   init_W_fft = W + 2 * pad_W
else:
   (usual init_H_fft, init_W_fft)

but this does not equal:

p_x_torch = torch.nn.functional.pad(x_torch, (1,1,1,1), mode='circular')
expect = torch.nn.functional.conv2d(input=p_x_torch, weight=y_torch,
                                            bias=b_torch, padding=0)

Thanks for your time!