This is an anonymous implementation of DRGAT.
git clone https://github.com/anonymousaabc/DRGCN.git
cd DRGCN
sh drgat_env_install.sh
Two datasets are required to run this code. We have already upload datasets in the directory. You can skip this step.
One is ogbn-arxiv origin data, the directory is ./drgat/dataset/ogbn_arxiv/
.
The other is ogbn-arxiv pretrained node features from GIANT-XRT, the directory is ./drgat/dataset/ogbn-arxiv-pretrain/
.
GIANT+XRT+DRGAT: Run runexp_drgat_ogbnarxiv.sh for reproducing our results for ogbn-arxiv dataset with GIANT-XRT features.
cd drgat
sh runexp_drgat_ogbnarxiv.sh
GIANT+XRT+DRGAT+KD: Run runexp_drgat_ogbnarxiv_kd.sh for reproducing our results for ogbn-arxiv dataset with GIANT-XRT features and KD.
cd drgat
sh runexp_drgat_ogbnarxiv_kd.sh
If execute correctly, you should have the following performance (using pretrained GIANT-XRT features).
Metrics | GIANT-XRT+DRGAT | GIANT-XRT+DRGAT+KD |
---|---|---|
Average val accuracy (%) | 77.16 ± 0.08 | 77.25 ± 0.06 |
Average test accuracy (%) | 76.11 ± 0.09 | 76.33 ± 0.08 |
Number of params: 2685527
Our hardware used for the experiments is Tesla P100-PCIE-16GB.
Remark: We do not fine-tune DRGAT for our GIANT-XRT. It is possible to achieve higher performance by fine-tune it more carefully.