THU-numbda/HGNN-Part ? reverse-engineered prompt

Build me a Python app for hypergraph partitioning that can train a deep learning model on .hgr files and then test it on benchmark circuits like the IBM datasets. I want it to learn partition assignments with a variational autoencoder style model, use hypergraph convolutions, and optimize for a mix of low cut size, balanced partitions, and good reconstruction. Please include a training script, a testing script, shared config and utility code, and a simple way to save and load checkpoints.

The app should be able to read hypergraph data from a data folder, run on GPU if available, and print useful metrics like cutsize, imbalance, runtime, and loss breakdowns. It should support multiple partitions, batch testing over many files, and a refinement style evaluation flow that can improve the initial result. If you need to check current docs for PyTorch or any related libraries, go ahead and look them up online.