About

This repository is aimed to document the small modifications I made on CGCNN for a set of experiments. For the original implementation I redirect to his own repository: https://github.com/txie-93/cgcnn

Changes

• Changed the property scalers so they take the global mean and variance instead of dynamically scaling based on batch-statistics
• Added option to concatente a property in the representation after the readout function of the graph model.
• Added option for 'fooling' the bond-perception algorithm to remove structure in the way the graph is constructed.

Requirements

This code was tested on Python 3.7.17 and Python 3.8.6. To set up the environment use the requirements in requirements.txt

python -m venv ~/python-envs/cgcnn
source ~/python-envs/cgcnn/bin/activate
python -m pip install -r requirements.txt