🔍 Research Series on Classic GNNs

Benchmarking Series: Reassessing Classic GNNs	Paper
Classic GNNs are Strong Baselines: Reassessing GNNs for Node Classification (NeurIPS 2024)	Link
GNN+: Can Classic GNNs Be Strong Baselines for Graph-level Tasks? (ICML 2025)	Link

GNN+: Can Classic GNNs Be Strong Baselines for Graph-level Tasks? (ICML 2025)

Based on the GPS codebase: https://github.com/rampasek/GraphGPS

Python environment setup with Conda

Tested with Python 3.9/3.10, PyTorch 2.2.0, and PyTorch Geometric 2.3.1.

To set up the environment, run the following commands:

conda create -n GNNPlus python=3.10
conda activate GNNPlus

pip install torch==2.2.0 torchvision==0.17.0 torchaudio==2.2.0 --index-url https://download.pytorch.org/whl/cu118
pip install torch_geometric==2.3.1
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.2.0+cu118.html

pip install scikit-learn==1.4.0
pip install fsspec rdkit
pip install pytorch-lightning yacs torchmetrics
pip install networkx
pip install tensorboardX
pip install ogb
pip install wandb

Running Training

To execute training, activate the environment and run the following commands:

conda activate GNNPlus

sh run.sh 0 cifar10 2 > cifar10.txt 2>&1 &
sh run.sh 1 cluster 2 > cluster.txt 2>&1 &
sh run.sh 2 coco 2 > coco.txt 2>&1 &
sh run.sh 3 code2 1 > code2.txt 2>&1 &
sh run.sh 4 hiv 2 > hiv.txt 2>&1 &
sh run.sh 5 mal 5 > mal.txt 2>&1 &
sh run.sh 6 zinc 2 > zinc.txt 2>&1 &
sh run.sh 7 pattern 4 > pattern.txt 2>&1 &

sh run.sh 2 pcba 2 > pcba.txt 2>&1 &
sh run.sh 3 peptides-func 4 > peptides-func.txt 2>&1 &
sh run.sh 4 peptides-struct 4 > peptides-struct.txt 2>&1 &
sh run.sh 5 voc 2 > voc.txt 2>&1 &
sh run.sh 6 ppa 2 > ppa.txt 2>&1 &
sh run.sh 7 mnist 2 > mnist.txt 2>&1 &

Alternatively, use the following format for executing training runs:

conda activate GNNPlus

python main.py --cfg configs/gcn/peptides-func.yaml --repeat 2 seed 0

python main.py --cfg configs/gatedgcn/ppa.yaml --repeat 2 seed 0 

Reference

If you find our codes useful, please consider citing our work

@inproceedings{
luo2025can,
title={Can Classic {GNN}s Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence},
author={Yuankai Luo and Lei Shi and Xiao-Ming Wu},
booktitle={Forty-second International Conference on Machine Learning},
year={2025},
url={https://openreview.net/forum?id=ZH7YgIZ3DF}
}