A lightweight distributed GNN library for full batch node property prediction.

Features/Changelog

• Complete refactoring of CAGNET.
• Distributed utilities such as log, timer, etc.
• Node feature cached training.
• Partitioned graph cache on disk.
• More datasets. Most large graphs from pyg, dgl, ogb supported.
• Training depends on pytorch only.
• Distributed GAT training.
• Latest pytorch version supported.
• CSR graph supported.
• Half precision training supported.

Getting started

1. Setup a clean environment.

   conda create --name gnn
   conda activate gnn

2. Install pytorch (needed for training) and other libraries (needed for downloading datasets).

// Cuda 10:
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch-lts
conda install -c dglteam dgl-cuda10.2
conda install pyg -c pyg -c conda-forge
pip install ogb

// Cuda 11:
conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch-lts -c nvidia
conda install -c dglteam dgl-cuda11.1
conda install pyg -c pyg -c conda-forge
pip install ogb

3. Compile and install spmm. (Optional. CUDA dev environment needed.)

   cd spmm_cpp
   python setup.py install

4. Prepare datasets (edit the code according to your needs).

   //This may take a while.
   python prepare_data.py

5. Train.

   python main.py

Experiments for Sancus: Staleness-Aware Communication-Avoiding Full-Graph Decentralized Training in Large-Scale Graph Neural Networks

1. Check the steps in Getting started .
2. Check dataset, epoch, and num of GPUs in main.py.
3. Check model settings in dist_train.py
4. Check cache methods in models.
5. Run and see the result.

Contact

Contact chenzhao@ust.hk for any problems.