Introduction#
Welcome to the Salt framework!
Salt is a general-purpose framework for training multi-modal, multi-task models for high energy physics.
Salt supports arbitrary combinations of tasks including object classification and regression, and set-to-set reconstruction via edge classification and segmentation. Salt was developed for state-of-the art jet flavour tagging algorithms such as GN1 and GN2, but can be applied much more widely, as seen below.
The code is hosted on the CERN GitLab: https://gitlab.cern.ch/atlas-flavor-tagging-tools/algorithms/salt
Features#
- Built on Pytorch Lightning.
- Support for multiple YAML-configurable input modalities and output tasks.
- ONNX export support to use trained models in C++ environments like Athena.
- Easily extensible: you can implement your own custom dataloaders and models.
- Documented and tested.
Getting Started#
Below are some helpful links to get you started:
You can find out more about flavour tagging algorithms at the FTAG docs
There is a channel for the framework in the FTAG Mattermost workspace
Contributions are welcome! Check out existing issues for inspiration, or open your own
You can become a Salt expert by checking out the API reference
Current Usage#
Salt is currently used for the following projects:
- Jet flavour tagging
- Boosted X \rightarrow bb tagging
- Tau ID
- b-jet energy calibration
- Primary vertexing
- LLP vertexing
- Prompt lepton veto
- multitop analysis
- Pileup rejection using hits
- e/gamma calibration
- Pileup jet rejection with GNJVT
Statement of Need#
High energy physics research increasingly requires sophisticated machine learning tools to address complex data analysis challenges, for example identifying jets from bottom and charm quarks through their distinctive decay signatures in particle detectors. Salt meets this need by providing a versatile and high-performance machine learning framework that supports various tasks including object classification, regression, and set-to-set reconstruction, enabling physicists to effectively analyse complex particle collision signatures such as charged particle trajectories, decay vertices, and jets.
Meeting Contributions#
| Date | Title | Speakers |
|---|---|---|
| 2024-09-24 | Scaling Salt on Large Machines | Nicholas Luongo |
| 2024-07-09 | Scaling Salt on Large Machines Update | Nicholas Luongo |
| 2024-05-23 | Transformer updates in salt [10+10] | Matthew Leigh |
| 2024-05-21 | Scaling Salt on Large Machines Update | Nicholas Luongo |
| 2024-01-16 | Scaling salt on Large Machines | Nicholas Luongo |
search terms: ['salt']
date cutoff: 2022-12-18, for older meetings please check indico
Created: October 7, 2022