Optimization of Graph Neural Networks with Natural Gradient Descent

This repository contains the implementaion of the Optimization of Graph Neural Networks with Natural Gradient Descent. Most of the code is adapted from github.com/rusty1s/pytorch_geometric and github.com/Thrandis/EKFAC-pytorch. To duplicate the results reported in the paper, follow the subsequent steps in order.

• Clone the repository and change your current directory:

git clone https://github.com/russellizadi/ssp
cd ssp

• Create a new conda environment using the default environment.yml:

conda env create

• Activate the default environment:

conda activate ssp

• Go to the experiments folder:

cd experiments

• Run all the experiments performed in the paper:

./run.sh