SPA-Net

This repository is now deprecated in favor of the general SPANet library

https://github.com/Alexanders101/SPANet

Permutationless Many-Jet Event Reconstruction with Symmetry Preserving Attention Networks.
https://arxiv.org/abs/2010.09206

Requirements

• python >= 3.7
• h5py
• pytorch >= 1.6
• pytorch-lightning >= 1.0
• parameter-sherpa
• nvidia-apex for optimizers

Training Instructions

1. Get the data release from the following link: Data and pretrained model will be released soon.
2. Modify tbar/options.py to match your system and data location.
3. Run python train.py
4. During trianing, metrics will be published into lightning_logs.
5. After training, weights will be available in lightning_logs.

Code Structure

• ttbar/options.py Contains all of the hyperparameters and options used during training.
• ttbar/dataset.py Is reponsible for loading the HDF5 files that we extracted from madgraph root files.
• ttbar/network/quark_triplet_network.py Describes the main network architecture and training procedure.

Citation

The current citation is to the arXiv preprint. This may be updated in the future.

@misc{fenton_2020_spatter,
      title={Permutationless Many-Jet Event Reconstruction with Symmetry Preserving Attention Networks}, 
      author={Michael James Fenton and Alexander Shmakov and Ta-Wei Ho and Shih-Chieh Hsu and Daniel Whiteson and Pierre Baldi},
      year={2020},
      eprint={2010.09206},
      archivePrefix={arXiv},
      primaryClass={hep-ex}
}