ABCNet: An attention-based method for particle tagging.

This is the main repository for the ABCNet paper. The implementation uses a modified version of GAPNet to suit the High Energy Physics needs. This repository is divided into two main folders: classification and segmentation, for the quark-gluon tagging and pileup mitigation applications, respectively. The input .h5 files are expected to have the following structure:

data: [N,P,F],
label:[N,P]
pid: [N]
global: [N,G]

N = Number of events

F = Number of features per point

P = Number of points

G = Number of global features

For classification, only the pid is required, while for segmentation only label is required.

The files to be used for the training (train_files.txt), test (test_files.txt) and evaluation (evaluate_files.txt) are required to be listed in the respective text files.

Requirements

Tensorflow

h5py

Classification

To train use:

cd classification
python train.py  --data_dir ../data/QG/  --log_dir qg_test

A logs folder will be created with the training results under the main directory. To evaluate the training use:

python evaluate.py  --data_dir ../data/QG --model_path ../logs/qg_test --batch 500 --name qg_test --modeln 1

Segmentation

To train use:

cd segmentation
python train.py  --data_dir ../data/PU/  --log_dir pu_test

To evaluate the training use:

python evaluate.py  --data_dir ../data/PU --model_path ../logs/ou_test --batch 500 --name pu_test

License

MIT License

Acknowledgements