Introduction

This package is a code supplement to the paper Anomalous Jet Identification via Sequence Modeling by Kahn et al.

The package contains the same model used to produce the results of the paper, with a small dataset to provide a proof-of-concept scenario of the model.

Paper Links

• IOPScience
• arXiv

Contents

• main.py
  • Primary script to run pre-processing, training, and evaluation
• model.py
  • Model definition in PyTorch
• helpers/processing.py
  • Implementation of the alignment algorithm (Algorithm 1) in the paper
• helpers/train.py
  • Functions to train and evaluate the model
• helpers/plotting.py
  • Code to plot ROC curves, other evaluation plots using Matplotlib
• helpers/eval.py
  • Helper functions to handle evaluation data
• helpers/defaults.yaml
  • Definition of default argument values
• unprocessed_data/
  • Small datasets for both 2 and 3 prong signal hypotheses. Contains both contaminated datasets as well as background and signal sets independently
    • Number of events: 1000 Background, 100 Signal for both 2 and 3 prong

Usage

python main.py [arguments]

Example: python main.py -p -t -d

Arguments

Option	Action
-p	Pre-process the datasets in unprocessed_data/ using the alignment algorithm (Algorithm 1)
-t	Train and evaluate the model (requires processing first)
-d	Draw evaluation plots (requires training first)
-j	Maximum number of jets per event to consider, sorted in pT descending (Default 1)
-c	Maximum number of constituents per jet to consider, sorted in pT descending (Default 20)
-k	KL-Divergence weight term in the VRNN loss function (Default 0.1)
-l	Hidden layer dimensionality (Default 16)
-z	Latent layer dimensionality (Default 2)
-s	Signal sample to process/train/evaluate (Default "2Prong")

Output

Processed data in .hdf5 format stored in the Output_h5/ directory
Network weights in .pth format stored in the saves/ directory
Evaluation data in .npy format stored in the eval_data/ directory
Evaluation plots stored in the plots/ directory