A “Periodical cicada” (magicicada septendecim) only emerge every 17 years and are noted for their song.
Calorimeter Image Convolutional Anomaly Detection Algorithm (CICADA) uses low level Compact Muon Solenoid experiment's trigger calorimeter information as the input to convolutional autoencoder to find anomalies produced during Large Hadron Collider proton-proton collisionsis. Quantization Aware Training and Knowledge Distillation are used to compress the model for sub-500ns inference on Field-Programmable Gate Arrays.
- First paper coming soon...
You could email us at [email protected]. Visit our website to find our latest talks:
This work is funded by the Eric and Wendy Schmidt foundation.
We suggest using conda to setup the environment as follows:
conda create -n "cicada" python=3.9.16 numpy==1.26.0
conda activate cicada
python3 -m pip install -r misc/requirements.txt
The data needed for running the scripts is available on eos under /eos/project-c/cicada-project/. The directory contains both root and h5 files. To access these files, please contact admins at [email protected] with the title Data access request for [Your Name].
If you still wish to convert the root files to h5 yourself, you can do it as follows:
python3 convert-root-to-h5.py <source_root_directory> <save_path_prefix>
This script can be run using additional options. --acceptance will store acceptance information, i.e. if an event has at least one jet of minimum pT of 30 GeV in the central region. -–split will produce three additional files to help with train-validation-test split. Mind that the h5 files are prepared and stored already in eos.
You may wish to profile the dataset to spot potential problems, e.g.
python3 dataset-profiling.py
In misc/config.yml you will find a configuration file used by most of the scripts throughout this repository. It contains data sources for background, signal and exposure datasets. The usage of these is controlled by use flag. Insted of editing the file, you could copy it and use a custom one using --config scripts.
We follow the training procedure as outlined in this paper. To run, execute the following:
python3 cicada-training.py
This script can be run using additional --epochs argument controlling the length of the training.
We compile QKeras model to C++ using hls4ml. The script run a few additional checks to check there agreement between hls4ml and tensorflow model respective outputs. To run, execute the following:
python3 compile-to-cpp.py --version <version>
Version argument should be either 1.1.2 or 2.1.2. The models used for compilation are loaded from HuggingFace, not the local directory.
The files generated by hls4ml can be integrated within CTP7cpp code to generate the IP containing both CICADA and WOMBAT.