CompressGNN

This is the KDD'25 repository for submitting paper CompressGNN: Accelerating Graph Neural Network Training via Hierarchical Compression.

Code Structure

The project's code base is organized in the following directory structure.

.
├── benchmark
├── dataset
├── experiment
├── genData
├── install.sh
├── layer
├── LICENSE
├── loader
├── model
├── README.md
├── requirements.txt
├── src
├── test
└── third_party

Installaction

To use this repository, please follow the steps below to install the required dependencies.

Prerequisites

• Python (version 3.8.13)
• pip (version 23.0.1)
• cuda (version 11.6)

Installing Dependencies

1. Clone the repository

2. Navigate to the project directory

3. Install the required dependencies using pip:

pip install -r requirements.txt

This command will install all the necessary libraries and packages, including:

• numpy
• pytorch
• Pytorch Geometric (PyG)
• Deep Graph Library (DGL)
• torch_scatter
• torch_sparse
• pybind
• ...

4. Install CompressGNN

bash install.sh

Data Preparation

Download Dataset

We have uploaded the small dataset Cora and cnr-2000. Users can generate a dataset in a format that meets our data specifications from WebGraph and PyTorch Geometric.

Input Data Format

.
├── csr_elist.npy
├── csr_vlist.npy
├── edge.npy
├── features.npy
├── labels.npy
├── test_mask.npy
├── train_mask.npy
└── val_mask.npy

Generate Torch Format Dataset

cd genData
python createTorchDataset.py <input data folder> <output data folder> coo/csr

Generate Compressed Torch Fromat Dataset

cd genData
python createCompressDataset <input data folder> <output data folder> coo/csr 

Generate datasets with different feature lengths

cd genData
python datagen_feature.py --data=xxx.pt --scale_factor=length --output=xxx.pt

Generate dataset using scripts.

cd genData
bash preprocess.sh
bash preprocess_compress.sh
bash datagen_feature.sh

Run Evaluation

End-to-end Performance

cd benchmark/end2end
bash run.sh

Propagate Performance

• Speedup

cd benchmark/propagate/propagate
bash speedup.sh

• Performance with different feature dimension

cd benchmark/propagate/propagate
bash feature_scale.sh

• Peak memory

cd benchmark/propagate/peak_memory
bash run.sh

Transformation Performance

• Time and accuracy

cd benchmark/transform/time_accu
bash run.sh

• Time breakdown

cd benchmark/transform/time_breakdown
bash run.sh