Incidence Networks for Geometric Deep Learning

This is a pytorch implementation of "Incidence Networks for Geometric Deep Learning" paper. Link to paper: https://arxiv.org/pdf/1905.11460.pdf

Requirements:

• numpy==1.17.4
• torch==1.4.0
• chainer_chemistry
• rdkit
• (optional) torch_geometric==1.4.2 (https://github.com/rusty1s/pytorch_geometric)

Dataset prepration

There are two ways to get the processed dataset before start running:

1- Download the dataset from the following links and save it under data direcotry.

https://figshare.com/articles/dataset/qm9_complete_inhomo_zip/12649757 (dense graph) https://figshare.com/articles/dataset/Processed_QM9_sparse_/12649790 (sparse graph)

2- Run qm9_prep_main.py from data_prep directory. (It takes a bit time to process data)

Training and Evaluation

To train and evaluate the model on a target, run the following command at src directory.

python main.py --mode mode --target_index target_index --data_path data_path --log_path log_path --is_linear is_linear --is_sym is_sym --graph_type graph_type

Arguments:

--mode: 0 for node-node adjacency and 1 for node-edge

--target_index: index to the molecular target [0-11] (see table below)

--data_path: path to input data

--log_path: path to the log (default: ../results/{inhomo/homo}_checkpoints/t{target_index}_{taget_name}/)

--is_linear: 0 for non-linear and 1 for linear

--is_sym: 0 for symmetric adjacency and 1 for non-symmetric adjacency

--graph_type: dense or sparse

Molecular target properties

Target index	Target name
0	mu
1	alpha
2	homo
3	lumo
4	gap
5	R2
6	ZPVE
7	U0
8	U
9	H
10	G
11	Cv