This repo contains the reference source code in PyTorch for the IJCAI 2022 paper Learning Continuous Graph Structure with Bilevel Programming for Graph Neural Networks
The code is built with following libraries:
- python 3.7
- PyTorch 1.7.1
- scipy 1.6
- scikit-learn 0.24.1
- networkx 2.6.3
conda create -n LCGS python=3.7
source activate LCGS
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
pip install scipy
pip install scikit-learn
pip install networkx
Please download the dataset from GCN, then change the filepath in data/utils.py for your path.
The main script is in the file main.py. The options are
--dataset: the evaulation dataset. Available datasets are cora, citeseer and pubmed. Default Cora.
--seed: the random seed. Default 1.
--gpu: the gpu device number (must be a single number). Default 0.
--name: the save name for checkpoint. Default None.
For experiments with normal graph scenario on Cora run
python main.py --dataset cora --name {if if you want to specify the save name}
Our codebase is developed based on the LDS from paper Learning Discrete Structures for Graph Neural Networks.