MedPart: Multi-Level Evolutionary Differentiable Hypergraph Partitioning

Repository Structure

MedPart
├── benchmark/             * Directory for input .hgr benchmark files
├── coarsening_results/    * Directory for storing coarsened hypergraph results
├── results/               * Directory for storing final partitioning outputs
├── GA_helper.py           * Evolutionary search utilities implemented in PyTorch
├── MedPart.py             * Main entry point for hypergraph partitioning
├── run_coarsening_auto.sh * Example script that wraps GraphZoom's run_coarsening.sh
└── utils.py               * Utilities for reading .hgr files and coarsening outputs

Notes on Coarsening

The script run_coarsening_auto.sh demonstrates how to invoke GraphZoom for spectral graph coarsening.

Input .hgr files must be converted to .mtx format prior to coarsening.

Optional pruning of high-degree hyperedges can be applied to generate multiple coarsening variants

Usage

To run MedPart on a benchmark:

Update the config in MedPart.py

python MedPart.py --design <DESIGN NAME> --device cuda:0 --tag test

Arguments:

--design : name of the benchmark (expected under benchmark/)

--device : CUDA device or CPU

--tag : experiment tag for organizing output files