Unofficial implementation of "Explainability Methods for Graph Convolutional Neural Networks" from HRL Laboratories. I also added a new method called unsigned Grad-CAM (UGrad-CAM) which shows both positive and negative contributions from nodes. Implemented using PyTorch Geometric and RDKit.
To train a GCNN on the BBBP dataset and save the model weights: python train.py.
You can download pretrained weights here.
To load the weights of a trained GCNN and generate explanations: python explain.py