WIP: Add trackMVA device algorithms to master
This merge will add two new device algorithms to Allen: a oneTrack and a twoTrack particle decay selection algorithm using CatBoost decision tree models to discriminate between interesting and non-interesting particle decays. This work was completed over the summer under the supervision of @thboettc and @dcraik.
The following is an approximate changelog of files modified to incorporate these new device algorithms:
-
Added a custom sequence incorporating trackMVA algorithms into HLT1 in configuration/sequences/trackMVAs.py and configuration/sequences/definitions/trackMVAsSequence.py
-
Added names to the algorithms in configuration/sequences/definitions/HLT1Sequence.py so that they could be incorporated as inputs into the trackMVA algorithms.
-
Added device algorithms in device/trackMVAs, also modified CMakeLists.txt in the device directory to incorporate the new directory.
-
Added trackMVAs in target_link_libraries of stream/CMakeLists.txt
-
Added CatBoost models into input/detector_configuration/down
-
Modified stream/sequence/include/Constants.h, stream/sequence/source/Constants.cpp, and main/src/Allen.cpp
Models are trained using ntuples produced using custom branches of MooreAnalysis (https://gitlab.cern.ch/lhcb/MooreAnalysis/-/tree/thboettc_mva_ntuples) and Analysis (https://gitlab.cern.ch/lhcb/Analysis/-/tree/thboettc_mva_signal_tuple). Tracks are reconstructed using Allen and combined using the Moore HLT2 particle combiners.
The data used is available in the TestfileDB. Tags used:
Signal:
- Upgrade_BsPhiPhi_MD_FTv4_DIGI
- Upgrade_Ds2KKPi_MD_FTv4_DIGI
- Upgrade_KstEE_MD_FTv4_DIGI
- Upgrade_KstMuMu_MD_FTv4_DIGI
- upgrade_DC19_01_Bs2JPsiPhi_MD
- upgrade_DC19_01_Dp2KSPip_MU
- upgrade_DC19_01_Dst2D0piMD
Background:
- MiniBrunel_2018_MinBias_FTv4_DIGI
A more detailed description of training and performance can be found here: https://indico.cern.ch/event/946429/contributions/3976994/attachments/2110462/3550001/Condon_IRIS-HEP.pdf