HyPER (Hypergraph for Particle Event Reconstruction) is Graph Neural Network (GNN) utilising blended graph-hypergraph representation learning to reconstruct short-lived particles from their complex final states.
We have removed graph building precedure following #4.
GraphDatasetnow loads flat data from aHDF5file and computes graph structure on the fly. We no longer recommend nor provide support for release v0.1.
HyPER model is built upon lightning, which uses tensorboard logger to save trained models.
To train HyPER:
python -m HyPER.train --config-name=default [options]
HyPER uses hydra for configuring runs (#10). You can overwrite any option using, for example, all_matched=False at the end, it overwrites the all_matched option provided in your configuration file.
Configuration files must be placed in the
configsfolder. Provide the file name without.yamlextension to--config-name.
To evaluate trained HyPER model on a dataset:
python -m HyPER.predict --config-name=default [options]
Four output variables are saved:
| Variables | Description |
|---|---|
HyPER_HE_IDX |
Indices of the nodes enclosed by a hyperedge |
HyPER_HE_RAW |
Soft probability of a hyperedge |
HyPER_GE_IDX |
Indices of the nodes connected by a graph edge |
HyPER_GE_RAW |
Soft probability of a graph edge |
Based on these RAW outputs, events can then be reconstructed by stating the correct topology in the configuration file. We currently have limited event topologies available in HyPER, see HyPER/topology. If you wish additional ones to be added, you can create a issue here.
Currenlty, conda auto environment solving is only supported by Linux/amd64 machines, due to some ongoing issues with torch_geometric in MacOS. A conda environment file environment_linux.yaml is provided. We recommend using miniforge as your conda package manager due to its lightweightness.
To create a conda environment named "HyPER":
conda env create -f environment_linux.yaml
We have tested the code with CUDA=11.8, HyPER should work with any CUDA versions above.