benchmarking-gnns

Why we included this project

Researchers who train graph neural networks and need results that hold up next to published numbers will find a solid reference point here. The repository bundles standard tasks such as molecular property prediction on ZINC and AQSOL plus structural benchmarks like PATTERN and CLUSTER, and pairs them with reproducible training scripts for both message-passing and Weisfeiler-Lehman style models. Its real value is consistency: shared data splits, fixed hyperparameter pipelines, and per-dataset leaderboards let you compare a new architecture against the same baselines everyone else cites. Step-by-step guides cover adding a fresh dataset or a new model class, which helps when you want to extend the benchmark rather than just run it. Just keep in mind this is a benchmarking harness, not a ready-to-serve inference library.