Paper: Automatic Symmetry Discovery with Lie Algebra Convolutional Network Nima Dehmamy, Robin Walters, Yanchen Liu, Dashun Wang, Rose Yu NeurIPS 2021
(find updated versions on arxiv)
A simple implementation of the L-conv layer in PyTorch (>=1.8) can be found in src/lconv.py.
The L-conv layer acts similar to a graph convlutional layer (GCN), so prepare your input in a similar fashion (e.g. flatten the spatial dimensions).
The input should have shape (batch, channels, #nodes) (e.g. on an image, # nodes = # pixels)
This repository also contains code and notebooks for the experiemnts in the paper (appendix C and D) under paper-code.
Most experiments in appendix D use an older (but identical) implementation in Tensoflow (>=2.1).
Comparison with LieConv in appendix D requires the LieConv packages.
Exmaples of uses will be added soon.