63days / gat

Pytorch Implementations of Graph Attention Network and Graph Convolution Network

Pytorch Implementation of GCN & GAT

Pytorch Implementations of Semi-Supervised Classification with Graph Convolutional Networks and Graph Attention Networks.

Overview of GAT.

Results

<Cora dataset>

	Training Time	Loss	Acc
GCN	2s	0.9790	81.6%
GAT	17m	0.6724	83.8%
spGCN	13s	0.9215	81.4%
spGAT	1m30s	0.6745	84.7%

<Siteseer dataset>

	Training Time	Loss	Acc
GCN	2s	1.2088	60%
GAT	21m	1.1907	59.1%
spGCN	17s	1.602	58.3%
spGAT	1m47s	1.1591	59.2%

GAT achieves better performances compared to GCN.

Usage

python3 train.py --model {gcn, gat, spgcn, spgat}

References

https://github.com/tkipf/pygcn https://github.com/Diego999/pyGAT