search

ndey96 / GCNN-Explainability

Implementation of "Explainability Methods for Graph Convolutional Neural Networks" from HRL Laboratories
MIT License
75 stars 3 forks source link

readme

GCNN-Explainability

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.

2

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