Home

Awesome

GraphGPS: General Powerful Scalable Graph Transformers

arXiv PWC

GraphGPS-viz

How to build a graph Transformer? We provide a 3-part recipe on how to build graph Transformers with linear complexity. Our GPS recipe consists of choosing 3 main ingredients:

  1. positional/structural encoding: LapPE, RWSE, SignNet, EquivStableLapPE
  2. local message-passing mechanism: GatedGCN, GINE, PNA
  3. global attention mechanism: Transformer, Performer, BigBird

In this GraphGPS package we provide several positional/structural encodings and model choices, implementing the GPS recipe. GraphGPS is built using PyG and GraphGym from PyG2. Specifically PyG v2.2 is required.

Python environment setup with Conda

conda create -n graphgps python=3.10
conda activate graphgps

conda install pytorch=1.13 torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia
conda install pyg=2.2 -c pyg -c conda-forge
pip install pyg-lib -f https://data.pyg.org/whl/torch-1.13.0+cu117.html

# RDKit is required for OGB-LSC PCQM4Mv2 and datasets derived from it.  
conda install openbabel fsspec rdkit -c conda-forge

pip install pytorch-lightning yacs torchmetrics
pip install performer-pytorch
pip install tensorboardX
pip install ogb
pip install wandb

conda clean --all

Running GraphGPS

conda activate graphgps

# Running GPS with RWSE and tuned hyperparameters for ZINC.
python main.py --cfg configs/GPS/zinc-GPS+RWSE.yaml  wandb.use False

# Running config with tuned SAN hyperparams for ZINC.
python main.py --cfg configs/SAN/zinc-SAN.yaml  wandb.use False

# Running a debug/dev config for ZINC.
python main.py --cfg tests/configs/graph/zinc.yaml  wandb.use False

Running GraphGPS on OGB-LSC PCQM4Mv2

Training

# "small" GPS (GatedGCN+Transformer) with RWSE: 5layers, 304dim, 6152001 params 
python main.py --cfg configs/GPS/pcqm4m-GPS+RWSE.yaml
# "medium" GPS (GatedGCN+Transformer) with RWSE: 10layers, 384dim, 19414641 params
python main.py --cfg configs/GPS/pcqm4m-GPSmedium+RWSE.yaml
# "deep" GPS (GatedGCN+Transformer) with RWSE: 16layers, 256dim, 13807345 params
python main.py --cfg configs/GPS/pcqm4m-GPSdeep+RWSE.yaml

Expected performance

Model configparameterstrain MAEcustom valid MAEofficial valid MAE
GPS-small6,152,0010.06380.08490.0937
GPS-medium19,414,6410.07260.08050.0858
GPS-deep13,807,3450.06410.07960.0852

Inference and submission files for OGB-LSC leaderboard

You need a saved pretrained model from the previous step, then run it with an "inference" script that loads official valid, test-dev, and test-challenge splits, then runs inference, and the official OGB Evaluator.

# You can download our pretrained GPS-deep (151 MB).
wget https://www.dropbox.com/s/aomimvak4gb6et3/pcqm4m-GPS%2BRWSE.deep.zip
unzip pcqm4m-GPS+RWSE.deep.zip -d pretrained/

# Run inference and official OGB Evaluator.
python main.py --cfg configs/GPS/pcqm4m-GPSdeep-inference.yaml 

# Result files for OGB-LSC Leaderboard.
results/pcqm4m-GPSdeep-inference/0/y_pred_pcqm4m-v2_test-challenge.npz
results/pcqm4m-GPSdeep-inference/0/y_pred_pcqm4m-v2_test-dev.npz

Benchmarking GPS on 11 datasets

See run/run_experiments.sh script to run multiple random seeds per each of the 11 datasets. We rely on Slurm job scheduling system.

Alternatively, you can run them in terminal following the example below. Configs for all 11 datasets are in configs/GPS/.

conda activate graphgps
# Run 10 repeats with 10 different random seeds (0..9):
python main.py --cfg configs/GPS/zinc-GPS+RWSE.yaml  --repeat 10  wandb.use False
# Run a particular random seed:
python main.py --cfg configs/GPS/zinc-GPS+RWSE.yaml  --repeat 1  seed 42  wandb.use False

W&B logging

To use W&B logging, set wandb.use True and have a gtransformers entity set-up in your W&B account (or change it to whatever else you like by setting wandb.entity).

Unit tests

To run all unit tests, execute from the project root directory:

python -m unittest -v

Or specify a particular test module, e.g.:

python -m unittest -v unittests.test_eigvecs

Citation

If you find this work useful, please cite our NeurIPS 2022 paper:

@article{rampasek2022GPS,
  title={{Recipe for a General, Powerful, Scalable Graph Transformer}}, 
  author={Ladislav Ramp\'{a}\v{s}ek and Mikhail Galkin and Vijay Prakash Dwivedi and Anh Tuan Luu and Guy Wolf and Dominique Beaini},
  journal={Advances in Neural Information Processing Systems},
  volume={35},
  year={2022}
}