• D²GNAS is an efficient graph neural architecture search method with decoupled differentiable search.

• The framework of D²GNAS is as follows:

• Ensure you have installed CUDA 11.1 before installing other packages

1.Python environment: recommending using Conda package manager to install

conda create -n d2gnas python=3.7
source activate d2gnas

2.Python package:

torch == 1.8.2
torch-geometric == 2.0.2
torch-scatter == 2.0.7
torch-sparse == 0.6.11
hyperopt == 0.1.2

1.Performance test with manual GNN and the optimal GNN designed by D²GNAS

python performance_test.py

2.Search the top promising GNN architectures from scratch with all GNAS methods and test with HPO

step1:python search_from_scratch.py
step2:python performance_test_with_hpo.py

3.Unzip the cs_phy.tar.gz in the dataset file first for CS and Physics testing

tar -xvf cs_phy.tar.gz

If you think D2GNN is useful tool for you, please cite our paper, thank you for your support:

@article{CHEN2024120700,
title = {Decoupled differentiable graph neural architecture search},
journal = {Information Sciences},
volume = {673},
pages = {120700},
year = {2024},
issn = {0020-0255},
doi = {https://doi.org/10.1016/j.ins.2024.120700},
url = {https://www.sciencedirect.com/science/article/pii/S0020025524006133}}