This library contains a PyTorch implementation of the Concentric Spherical Neural Network (CSNN). The associated paper was published at the International Joint Conference for Neural Networks (IJCNN) 2022, which you can reference here.

For any questions about this work, please contact the primary author (James Fox) at jfox43@gatech.edu.

This codebase was developed using Python 3.8, PyTorch 1.9, DGL 0.6.1, and CUDA 11.1.

The following installs dependencies to Anaconda virtual environment:

conda create --name csgnn python=3.8
conda activate csnn
conda install pytorch torchvision cudatoolkit=11.1 -c pytorch -c nvidia
conda install -c dglteam dgl-cuda11.1
conda install h5py
pip install requests

The follow commands are called from the top level directory of this project.

First, retrieve the dataset:

python -m modelnet40.dataset

This downloads the dataset to the path "./modelnet40_ply_hdf5_2048".

There are two pre-trained models: "csgnn-modelnet-z" is trained on z-axis aligned rotations, and "csgnn-modelnet-SO3" is trained on SO3 rotations. For example, to evaluate the SO3-trained model on SO3-rotated test data, run:

python -m modelnet40.test modelnet40/saved/csgnn-modelnet-SO3.pkl --rotate_test SO3

To train the model according to the paper (with default settings for SO3 training), run

python -m modelnet40.train

Extract the carbon dataset as follows:

tar -xzvf carbon_database.tar.gz

In addition to the dependencies listed earlier, the following packages are required to run the experiment:

pip install pymatgen
pip install scikit-learn

To evaluate pre-trained model for overall error, or error grouped by structure type, run:

python -m dos.test dos/saved/csgnn-dos.pkl --mode [all/group]

To train the model according to the paper, run:

python -m dos.train