By Derek Lim*, Joshua Robinson*, Lingxiao Zhao, Tess Smidt, Suvrit Sra, Haggai Maron, Stefanie Jegelka.
[arXiv] [pdf]

Codebase for neural networks (SignNet and BasisNet) and experiments in the paper.

Alchemy contains the experiments for graph-level regression on Alchemy.

GraphPrediction contains the experiments for graph-level regression on ZINC.

LearningFilters contains the spectral graph convolution experiments.

The intrinsic neural fields experiments use private code from the authors of the original paper, so we do not yet publically release the SignNet codes for these.

PyTorch Geometric SignNet for graph prediction: in Alchemy.

PyTorch Geometric SignNet for graph prediction on ZINC: in GINESignNetPyG.

DGL SignNet for graph prediction: in GraphPrediction.

BasisNet for single graphs: in LearningFilters.

The SignNet architecture is rather simple. Here is an example of pseudo-code for SignNet, as used for graph prediction tasks with a GNN base model:

Coming Soon: More experiments and implementations of our models! This repo and our paper are still a work in progress.