Unifews

This is the original code for Unifews: Unified Entry-Wise Sparsification for Efficient Graph Neural Network

Dependencies

Python

Installed env.txt by conda:

conda create --name <env> --file env.txt

C++

• C++ 14
• CMake 3.16
• eigen3

Experiment

Data Preparation

1. Use utils/data_transfer.py to generate processed files under path data/[dataset_name] similar to the example folder data/cora:

• adj.npz: adjacency matrix in scipy.sparse.csr_matrix
• feats.npy: features in .npy array
• labels.npz: node label information
  • 'label': labels (number or one-hot)
  • 'idx_train/idx_val/idx_test': indices of training/validation/test nodes
• adj_el.bin, adj_pl.bin, attribute.txt, degree.npz: graph files for precomputation

Decoupled Model Propagation

1. Compile Cython:

cd precompute
python setup.py build_ext --inplace

Model Training

1. Run full-batch experiment:

python run_fb.py -f [seed] -c [config_file] -v [device]

2. Run mini-batch experiment

python run_mb.py -f [seed] -c [config_file] -v [device]

Reference & Links

Datasets

• cora, citeseer, pubmed: Pytorch Geometric
• arxiv, products, papers100m: OGBl

Baselines

• GLT: A Unified Lottery Ticket Hypothesis for Graph Neural Networks
• GEBT: Early-Bird GCNs: Graph-Network Co-optimization towards More Efficient GCN Training and Inference via Drawing Early-Bird Lottery Tickets
• CGP: Comprehensive Graph Gradual Pruning for Sparse Training in Graph Neural Networks
• DSpar: DSpar: An Embarrassingly Simple Strategy for Efficient GNN Training and Inference via Degree-Based Sparsification
• NDLS: Node Dependent Local Smoothing for Scalable Graph Learning
• NIGCN: Node-wise Diffusion for Scalable Graph Learning