Recurrently Predicting Hypergraphs

This repository contains the code for reproducing the experiments described in Recurrently Predicting Hypergraphs by David Zhang, Gertjan Burghouts and Cees Snoek.

Dependencies

Install environment with conda and pip:

conda create -n RPH python=3.9
pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install uproot==3.12.0
pip install https://ray-wheels.s3-us-west-2.amazonaws.com/python3.9/a902f2e4ab0a9c27ece8562084aa3fc4be68eeb8/ray-1.2.0.dev0-cp39-cp39-manylinux2014_x86_64.whl
pip install numpy scipy pandas sklearn pytorch-lightning wandb

Experiments

Particle partitioning

Follow the data setup described in https://github.com/hadarser/SetToGraphPaper. Specify the data directory in particle_partitioning_main.py and particle_partitioning_baseline.py. Adapt the TYPE variable to either slot_attention or set_transformer to run the baselines. Both scripts are meant to be run without any additional command line arguments. All hyperparameters are specified in the python scripts directly.

Convex hull finding

Run python convex_hull_main.py for RPH and python convex_hull_baseline.py for the baselines. The ablations can be run by setting the hyperparameters in convex_hull_main.py accordingly.

Delaunay triangulation

Run python delaunay_main.py.