egnn-pytorch

Why we included this project

Equivariance matters when your input coordinates can be rotated, translated, or reflected and you still want the same answer. This repo packages E(n)-equivariant graph neural networks in PyTorch, with both individual layers and a complete network ready to use. You can add edge features, limit attention to sparse neighbors through an adjacency matrix, and clamp coordinate weights to keep training stable as the neighborhood grows. It is a building block rather than a full application, so it suits teams that already have a training pipeline and want a solid equivariant layer to drop in. The code is short and readable, which makes adapting it to your own data and loss functions straightforward.