harmonic

Reimplementation of harmonic networks in PyTorch. Original TensorFlow implementation is here. The reimplementation is not one to one. We reimplement harmonic convolutions, generalizing them to have per-radius phase offsets. Nonlinearities use a multiplicative "attentional" model, rather than the C-ReLU proposed by Worrall et al. Additionally, we split the library into two modules: d2 which contains the standard harmonic network implementation for 2d domains and d3, which is a 3d version, arranged to ensure equivaraince to rotations along (x, y) axes and allowing arbitrary kernels along z axis. This is useful for volume data which is anisotropic along one dimension, for instance slice microscopy.

Installation

1. Install torch-localize
2. Install torch-dimcheck
3. Clone this repository
4. Execute python setup.py install. In some case this doesn't work properly and python setup.py develop is necessary instead. I am not sure what is the reason for these issues

Examples

In examples/mnist2d one can find reimplementation of the RotMNIST experiment from Worrall et al., which requires manually downloading the data using their code.