Understanding GNN Computational Graph: A Coordinated Computation, IO, and Memory Perspective (FusedGAT)
- Paper link: https://arxiv.org/abs/2110.09524
- Author's code repo (in PyTorch): https://github.com/dgSPARSE/dgNN.
Dataset | # Nodes | # Edges | # Classes |
---|---|---|---|
Cora | 2,708 | 10,556 | 7 |
Citeseer | 3,327 | 9,228 | 6 |
Pubmed | 19,717 | 88,651 | 3 |
Refer to Planetoid.
TL_BACKEND="torch" python fusedgat_trainer.py --dataset cora --lr 0.01 --l2_coef 0.005 --drop_rate 0.7
TL_BACKEND="torch" python fusedgat_trainer.py --dataset citeseer --lr 0.01 --l2_coef 0.01 --drop_rate 0.6
TL_BACKEND="torch" python fusedgat_trainer.py --dataset pubmed --lr 0.01 --l2_coef 0.001 --drop_rate 0.2
TL_BACKEND="torch" python fusedgat_tester.py --dataset cora --lr 0.01 --l2_coef 0.005 --drop_rate 0.7
TL_BACKEND="torch" python fusedgat_tester.py --dataset citeseer --lr 0.01 --l2_coef 0.01 --drop_rate 0.6
TL_BACKEND="torch" python fusedgat_tester.py --dataset pubmed --lr 0.01 --l2_coef 0.001 --drop_rate 0.2
TL_BACKEND="torch" python gat_tester.py --dataset cora --lr 0.01 --l2_coef 0.005 --drop_rate 0.7
TL_BACKEND="torch" python gat_tester.py --dataset citeseer --lr 0.01 --l2_coef 0.01 --drop_rate 0.6
TL_BACKEND="torch" python gat_tester.py --dataset pubmed --lr 0.01 --l2_coef 0.001 --drop_rate 0.2
Dataset | Our(torch) |
---|---|
cora | 79.68 |
citeseer | 66.20 |
pubmed | 76.96 |
Dataset | Metric | GAT | FusedGAT |
---|---|---|---|
cora | train | 20.37 | 10.09 |
cora | infer | 4.04 | 2.11 |
cora | memory | 359 | 341 |
citeseer | train | 21.16 | 9.99 |
citeseer | infer | 4.28 | 2.19 |
citeseer | memory | 421 | 451 |
pubmed | train | 20.44 | 9.05 |
pubmed | infer | 4.21 | 2.11 |
pubmed | memory | 505 | 427 |
train : ms / epoch
infer : ms / epoch
memory: MB