The offical implementation of paper "Neural Scaling Laws on Graphs". In this repo, we provided pipelines to test the data and model scaling behaviors of graph deep learning models.
pip install -r "requirement.txt"
To run the codes, please first create three folders: dataset/, figures/ and ./results under the main folder.
To test the data scaling scaling behaviors, run
bash scale_data.sh
The user could customize the training set sizes by changing the data_array in scale_data.sh.
To test the data scaling scaling behaviors, run
bash scale_model.sh
The user could customize the training set sizes by changing the emb_array in model_data.sh.
The dataset can be defined with the argument --dataset and will be downloaded automatically under the /dataset folder. The scaling results will recorded under the ./results folder.
To visualize the scaling results, run
python curve_draw.py --filename $name
Here $name is the target result file name. The command will generate a scaling curve under /figures folder and calculte the value of R-square of the fitting.
We thank the authors of OGB for their codes and datasets shared!