Skip to content

Latest commit

 

History

History
60 lines (52 loc) · 3.33 KB

README.md

File metadata and controls

60 lines (52 loc) · 3.33 KB
Raw

Email Badge Zenodo Badge arxiv Python: version Code style: black

Ultrafast Jet Classification using Geometric Learning for the HL-LHC

Code repository corresponding to the Ultrafast Jet Classification at the HL-LHC paper.

Three machine learning models are used to perform jet origin classification. These models are optimized for deployment on a field-programmable gate array device. In this context, we demonstrate how latency and resource consumption scale with the input size and choice of algorithm. Through quantization-aware training and efficient synthetization for a specific field programmable gate array, we show that O(100) ns inference of geometric learning architectures such as Deep Sets and Interaction Networks is feasible at a relatively low computational resource cost.

Installation

The main dependencies can be installed using conda by running the following command in the terminal while in this repository's directory

conda env create -f fast_jetclass.yml

Alternatively, one can install the following packages manually:

  - python=3.10
  - pyyaml
  - numpy
  - python-wget
  - scikit-learn
  - matplotlib
  - pydot
  - pydotplus
  - tensorflow==2.14
  - qkeras==0.9
  - tensorflow-model-optimization==0.8.0

Then, after installing the dependencies, install this repository using

pip install .

while still in this repository's directory.

Additional Dependencies for Synthesis

The synthesis of the models for the FPGA are done using hls4ml. Hence, a particular version of this package is needed for running *_synth part of the scripts. To install this, execute pip install . inside the hls4ml directory of this repository. Additionally, for the profiling of the synthesized models, one also needs a subpackage of hls4ml, installed with

pip install hls4ml[profiling]
conda install pydot
conda install pydotplus

I honestly have not found a way of making the profiling package work without conda.

If you do not need profiling, omit the --diagnose flag when running the *_synth scripts. The profiling subpackage is a bit buggy, so if you encounter an error, contact the authors of this package; in the current version there are a couple of issues that are fixed by going into the source code of the profiling and changing a few lines. Contact me at the email address above if you need help with this.

Running the Scripts

Examples of running the scripts, from training to synthesis, are given for each model in the README of the scripts directory in this repository. All the experiments executed for the corresponding paper of this code have their configuration in the scripts/configs folder.