Implementation of linear equivariant network layers for Pytorch.
Usage is similar to Pytorch's Linear
layer.
import torch
from equivariant.nn import LinearEquivariant2to2
x = torch.rand(32, 4, 12, 12) # shape (batch, features, dim, dim)
equi = LinearEquivariant2to2(in_features=4, out_features=3)
y = equi(x) # shape (32, 3, 12, 12)
Haggai Maron, Heli Ben-Hamu, Nadav Shamir, Yaron Lipman. Invariant and Equivariant Graph Networks. International Conference on Learning Representations (ICLR) 2019. http://arxiv.org/abs/1812.09902