Awesome

Position-aware Graph Neural Networks

This repository is the official PyTorch implementation of "Position-aware Graph Neural Networks".

Jiaxuan You, Rex Ying, Jure Leskovec, Position-aware Graph Neural Networks, ICML 2019 (long oral).

Installation

• Install PyTorch (tested on 1.0.0), please refer to the offical website for further details

conda install pytorch torchvision cudatoolkit=9.0 -c pytorch

• Install PyTorch Geometric (tested on 1.1.2), please refer to the offical website for further details

pip install --verbose --no-cache-dir torch-scatter
pip install --verbose --no-cache-dir torch-sparse
pip install --verbose --no-cache-dir torch-cluster
pip install --verbose --no-cache-dir torch-spline-conv (optional)
pip install torch-geometric

• Install networkx (tested on 2.3), make sure you are not using networkx 1.x version!

pip install networkx

• Install tensorboardx

pip install tensorboardX

• If you wish to use PPI dataset, unzip data/ppi.zip

Run

• 3-layer GCN, grid

python main.py --model GCN --layer_num 3 --dataset grid

• 2-layer P-GNN, grid

python main.py --model PGNN --layer_num 2 --dataset grid

• 2-layer P-GNN, grid, with 2-hop shortest path distance

python main.py --model GCN --layer_num 2 --approximate 2 --dataset grid

• 3-layer GCN, all datasets

python main.py --model GCN --layer_num 3 --dataset All

• 2-layer PGNN, all datasets

python main.py --model PGNN --layer_num 2 --dataset All

You are highly encouraged to tune all kinds of hyper-parameters to get better performance. We only did very limited hyper-parameter tuning.

We recommend using tensorboard to monitor the training process. To do this, you may run

tensorboard --logdir runs