PIPPIN

Particles Into Particles with Permutation Invariant Network

This repository contains the code to reproduce the results described in the paper:

G. Quétant, J. A. Raine, M. Leigh, D. Sengupta and T. Golling
PIPPIN: Generating variable length full events from partons
arXiv:2406.13074 [hep-ph]

Setup

Clone the repository and initialise the necessary submodule¹:

git clone git@github.com:rodem-hep/pippin.git
cd pippin
git submodule update --init

Make sure to also install the required dependencies (See requirements.txt). For instance:

conda create -n pippin python=3.12
conda activate pippin
pip install --upgrade pip
pip install --upgrade -r requirements.txt

---

1. The MLTools submodule should be called mattstools and be pointing at the commit 1cc1d7a80351232804bd4fa79ffa46f21bfc9af2.

Training

Train and test the model with the following command:

python run.py

A debug config is available with less epochs and data:

python run.py +debug=debug

Note that data should be located at ./data. See Zenodo to download it.

Evaluation

Evaluate the model with the following command:

python eval.py --name <network_name> --metrics --plots --inclusive

The arguments are:

• --name, -n: the network name
• --metrics, -m: compute the metrics
• --plots, -p: make the plots
• --inclusive, -i: consider the inclusive dataset (i.e. PIPPIN)

Other arguments can be added:

• --compare, -c: make the comparison plots (i.e. restrained PIPPIN vs retrained Turbo-Sim). Needs both -t and -T
• --leading, -l: consider the leading particles (i.e. PIPPIN)
• --turbolike, -t: consider the Turbo-like outputs (i.e. restrained PIPPIN)
• --turbosim, -T: consider the Turbo-Sim outputs (i.e. retrained Turbo-Sim)

Note that several runs of the same model can be combined. See eval.py line 95 and change the list with the desired job_id. The runs outputs should be located at ./outputs/PIPPIN/ttbar/<network_name>/hdf5/test/<job_id>/outputs.h5.