Differentiable FFT Conv Layer with Dense Color Channels
This is meant to be a drop in replacement for torch.conv
- functional_conv1d_fft replaces torch.nn.functional.conv1d
- Conv1d_fft replaces torch.nn.Conv1d
- supports 1d, 2d and 3d convolution
This is just an alpha POC release.
Written in Python, no optimized cuda.
API is not exactly matching torch yet.
unsupported: stride, dilation, groups, etc