JetPointNet

This repository contains the code used to develop particle-flow models based on pointcloud data for jets data.

Minimal Installation

For a minimal installation you can first create a conda environment with conda create --name pointcloud python=3.10. Then make use of the requirements.txt file and run:

git clone ssh://git@gitlab.cern.ch:7999/atlas-jetetmiss/pflow/commontools/jetpointnet.git
conda activate pointcloud
pip install -r requirements.txt # --no-cache-dir

Please note that --no-cache-dir option is suggested. If you plan on developing the code, you should install LFS to download the test cases.

Usage

All configuration is done in the prod/configs/USER_config.toml. A config file can also be set in the environment variable JETPOINTNET_CONFIG_FILE. Once a run, either training or data processing is defined, it can be run with:

make run

There are 3 core steps, data processing, chunking, and training. All are defined in the prod/configs/USER_config.toml file. Chunking is optional and can be skipped if the data is already chunked or sufficiently randomized.

Next steps:

• Apply garbage collection to the augmented data processing scripts to ensure clusters still meet cluster min significance
• Try Mask formers instead of PointNet
• Change batching algorithm for training
• Add CI/CD pipeline for the data processing scripts

Josh's next steps:

• Investigate normalization and visualizations for it
• Remove poetry
• Create sample event with only 1 event for testing, reduce git lfs size

Known Issues

• Many... we should probably start a list here
• The early stopping of the train loop can cause the job to crash when enabled
• The augmented data processing scripts work on JZ4 but seem to fail when I run on the server...?

Authors:

• Dr. Maximilian Swiatlowski (TRIUMF)
• Dr. Luca Clissa (UNIBO)
• Joshua Himmens (TRIUMF/UBC)
• Marko Jovanovic (TRIUMF)
• Jessica Bohm (TRIUMF)