GraphNetJetTaggerCalo

Graph Network aiming at jet tagging for binary classification (signal vs background) for searches for long lived particles with CMS detector.

Currently implemented in keras.

Input and output root files are designed to be compatible with LLP repository.

Preliminary naf gpu setup

See: https://github.com/lbenato/ParticleNet-LLP-fork

Once in your environment folder, install the additional packages:

pip install GPUtil
pip install xgboost

Prepare folders

mkdir dataframes
mkdir dataframes_graphnet
mkdir model_weights
mkdir model_weights_graphnet
mkdir root_files

write_pd_graphnet.py

It reads input root files and transforms them into h5 files.

samples*.py

Same as LLP repo; used to load sample names.

dnn_functions.py

Simple function to draw training and validation losses and accuracies.

tf_keras_model.py

Original ParticleNet functions and architectures, added FCN jet tagger architecture (https://github.com/lbenato/LEADER). Added some modifications to take into account angular variables for matrix distance.

prepare_dataset_graphnet.py

• Convert per-event into per-jet dataframes
• Split training/test samples for both signal and background
• Transform them into train, validation, test h5

graphnet_tagger.py

• Load FCN or ParticleNet models
• Training function
• Calculate performances
• Write the output scores of test samples
• Convert h5 back to root files, compatible with any macro of LLP repository

training_script.py

• Define parameters and submits routines defined in graphnet_tagger.py