FeiGSSS / LGNN

A PyTorch implementation of Supervised community detection with graph neural networks (ICLR 2019)

LGNN: Line Graph Neural Networks

Pytorch re-implementation of Line Graph Neural Networks from ICLR 2019 paper Supervised community detection with graph neural networks. This implementation is based on the offical codes.

• The original implementation is pretty slow cause the data preprocessing in each iteration (about 6s). By utilizing Dataset and DataLoader in Pytorch, the running time of each iteration is reduced to 1.7s.
• Utilizing torch.sparse to perform matrix multiplication in GNN layers.

Requirements

I recommend using an independent environment to run the codes:

conda create -n LGNN
conda activate LGNN
pip install -r requirements.txt

Dataset

Before you run the code, you should make a new folder name data to store the generated SBM graph dataset. The dataset is named as "sbm_{}_{}_{}_{}_{}_{}.pkl".format(p, q, N, n_classes, J, str(train)) .

Train the Model

The key to accelerating the training process is utilizing the multi-processing of DataLoader, therefore, the num_workers should as large as possible. And the defalut arguments will train LGNN on 5-community dissociative SBM graphs with n = 400, C = 5, p = 0, q = 18/n.

Runing Example:

python main_lgnn.py --num_workers 40