Codes for BDT trainings and working point optimisation to be used in D2H

Requirements

In order to execute the code in this folder, the following python libraries are needed:

• hipe4ml
• hipe4ml_converter
• alive_progress

All configurables needed for sample preparation, ML training and working point optimisation are respectively embedded in configuration files such as:

• config_preparation.json;
• config_training.yml;
• config_optimisation.yml.

Training

Download training samples from Hyperloop

Download the derived AO2D.root (produced via a treeCreator task for example) and save in a dedicated directory (one directory per workflow). These directories will be used as labeled inputs for ML training.

Comment on the configurables

Prepare samples

In order to prepare the samples the following script can be used: root -l PrepareSamples.C+(TString nameCfgFile="./config_preparation.json")

Preselections can be applied at this level along with teh combination of TPC and TOF PID variables. This macro shall produce .root files containing labeled data samples (one file for each label) for each input AO2D file.

Perform training

In order to perform the training and produce the BDT models to be used in D2H, the following script can be used: python3 ml_regression.py config_training.yml Given the output directory set in config_training.yml, a directory is created for each pT bin (i.e. each model trained) and filled with:

• plots at data preparation level: variables distributions and correlations
• plots at training-testing level: BDT output scores, ROC curves, precision recall, features importance
• trained models in files containing ModelHandler prefix (in pickle and onnx formats). The .onnx can be used for ML inference in O2Physics selection task.
• model applied to test set in file containing ModelApplied suffix

Working point optimisation [TO BE RELEASED SOON]

The goal is to perform a scan on BDT output scores while computing relevant quantities such as efficiencies (preselections and BDT), yield fractions, expected signal, expected background, allowing to build expected signal-over-background and expected significance.

Required input

To do so, the needed input gathers the:

• preselections acceptance-times-efficiency (i.e. without cuts on BDT scores): this can be obtained using efficiency.py macro;
• pT-differential cross section predicted by FONLL, both for prompt and non-prompt: see this for instance.

Perform ML output scan

The following script can be used: python3 optimisation.py config_optimisation.yml Given the output directory set in config_optimisation.yml, a single directory is created for all pT bins (i.e. for all models trained) and filled with:

• a .root file containing information on background fits, significance distributions as a function of other quantities of interest, quantities of interest distributions as a function of ML output score(s), with a TDirectory set for each model;
• a canvas containing quantities of interest distributions as a function of ML output score(s), with configurable extensions (.pdf by default).