Home

Awesome

Ala-GCN

Dependencies

(pytorch-lightning has gone through some breaking changes since we wrote this code. If you try using the latest version of pytorch-lightning, you will run into errors like train_dataloader not defined.)

Datasets

Hyperparameters

Cora

python train.py --model=adagnn --dataset=cora --lr=0.01 --percentage=0.05 --n-layers=5 --dropout=0.5 --weight-decay=5e-6 --n-hidden=16 --n-epochs=200 --self-loop

Use 5 layers for 5% and 3%, 9 layers for 1%.

python train.py --percentage=0.05 --n-layers=3  --dataset=cora --model=adagat --lr=0.005 --dropout=0.6 --weight-decay=5.00E-04 --n-hidden=8 --n-epochs=300 --self-loop

Use 3 layers for 5%, 3% and 1%. Notice that due to the size of the dataset, the variance of performance with 1% seed labels is relatively large.

Citeseer

python train.py  --percentage=0.05 --n-layers=3 --dataset=citeseer --model=adagnn --lr=0.01 --dropout=0.5 --weight-decay=5.00E-06 --n-hidden=16 --n-epochs=200 --self-loop

Use 3 layers for 5%, 4 layers for 3% and 7 layers for 1%.

python train.py --percentage=0.05 --n-layers=3 --dataset=citeseer --model=adagat --lr=0.005 --dropout=0.8 --weight-decay=5.00E-04 --n-hidden=8 --n-epochs=300 --self-loop

Use 3 layers for 5%, 3 layers for 3% and 7 layers for 1%.

Pubmed

python train.py --percentage=0.003 --n-layers=9 --dataset=pubmed --model=adagnn --lr=0.01 --dropout=0.5 --weight-decay=5.00E-06 --n-hidden=16 --n-epochs=500 --self-loop

Use 9 layers for 0.3%, 0.15% and 0.05%.

python train.py --percentage=0.003 --n-layers=5 --dataset=pubmed --model=adagat --lr=0.005 --dropout=0.5 --weight-decay=5.00E-04 --n-hidden=16 --n-epochs=500' --self-loop

Use 5 layers for 0.3%, 5 layers for 0.15% and 7 layers for 0.05%.

Models