Chainer implementation of GCN/GAT

This project implements GCN (Kipf and Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. ICLR.) and GAT (Veličković et al. 2018. Graph Attention Networks. ICLR.) with Chainer. The project includes codes to reproduce the experiments on multiple graph classification datasets. This is NOT an official implementation by the authors.

I referenced @takyamamoto's implementation of GCN to implement this project.

How to Run

Prerequisite

I have only tested the code on Python 3.6.4. Install dependent library by running:

pip install -r requirements.txt

You need to install cupy to enable GPU.

Running training and evaluation

Run:

python train.py

Refer to python train.py -h for the options. Note that you can enable early stopping by --early-stopping flag, but the overhead for saving intermediate models is quite large.

Reproducing the paper

GCN

Run training in with same parameters as the original paper.

python train.oy --model gcn --lr 0.01 --unit 16 --dataset $DATASET

Following table shows the average (min/max) over 10 runs.

My implementation is comparable with the reported result.

GAT

python train.oy --dataset $DATASET --early-stopping

Average test accuracy was 83.7 (from 82.6 to 84.7) over 10 runs. The test accuracy reported in the paper was 83.0, so my implementation is comparable with the reported result.

My implementation took 0.3214s per epoch on K80 GPU (Google Colabolatory), which is slightly slower than 0.3002s (non-sparse) and 0.2043 (sparse) of the tensorflow implementation by the authors.

Licensing

load_data module and all files under data/ directory have been derived from Dr. Kipf's repository, so refer to the original repository for licensing. Other files are distributed under CC0.