EMTF code for pT LUT training (Run II-RunIII)
EMTFPtAssign2017 is designed to be run on cms-lpc in a CMSSW_10_6_1_patch2 environemnt. Since it will also need to pull files from EOS, it also assumes a grid-certificate has been authenticated. If you fork the repository you can setup your environment and new branch using,
cd ~/nobackup/
cmsrel CMSSW_10_6_1_patch2
cd CMSSW_10_6_1_patch2/src
cmsenv
voms-proxy-init --voms cms
git clone git@github.com:<your github username>/EMTFPtAssign2017.git
git checkout -b <your_branch_name>
git push origin <your_branch_name>
EMTFPtAssign2017 has been built with a python wrapper that will submit jobs to condor automatically. This is done through the condor/submitJobs.py script. Using this wrapper you can call,
cd condor
python submitJobs.py <options>
where <options> contains a --isRun2 or --isRun3 flags, --interactiveRun, which set of training variables, which target variable, eta values, training settings, etc. All options can be seen by running python submitJobs.py --help.
Note: Always run python submitJobs.py <options> --inerativeRun before submitting jobs to condor.
EMTFPtAssign2017 is build using ROOT TMVA with a python wrapper for submitting jobs to condor.
The main training program is done using PtRegressionRun3Prep.C. This program allows for labeling different variables as either spectator, training, or target. The spectator variables will not be used in training but will be written to the final ROOT file for evaluation. The training variables will be used by the BDT for training. The target variable is the varialbe the BDT is trying to learn to evaluate. There are also options for loss functions where BDTG_AWB_Sq is the least squares loss function currently being used.
Note: Training variables in PtRegressionRun3Prep.C must match what is written in allowedTrainingVars
Each event will be looped over and each track in an event will be either marked for training or testing. Tracks in the training set are used by the BDT regression algorithm to generate the Look Up Tables (LUTs) which will then evaluate the tracks in the testing set which they have not been trained on. This allows the BDT's performance to be evaluated for each training.
Additional configuration settings are stored in configs/PtRegression2018 directory which control how the NTuples are parsed, input files are assigned, and standard global variables are initialized.
Lastly, the python wrapper is discussed in the pervious section which is stored in condor directory.
Note: macros and macros_Rice2020 are deprecated and plotting is now done using the EMTF_BDT_PerformancePlotter GitHub Repository
The BDT in PtRegressionRun3Prep.C assumes the input file has been prepared in a FlatNTuple generated by EMTFAnalyzer/NTupleMaker GitHub Repository and hadded into one input file stored on EOS. There are multiple modes that are currently implemented in EMTFAnalyzer/NTupleMaker such as FlatNtupleMCRun2, FlatNtupleMCRun2GEM, FlatNtupleMCRun3, and FlatNtupleMCRun3GEM. If you want to switch between NTuples you must change the treeString variable in the PtRegressionRun3Prep.C training program.
The BDT will output a weights.xml LUT and a PtRegression.root file. Th weights.xml file will contain a large XML LUT that can eventually be reformatted and loaded in to CMSSW for evaluation of rate. PtRegression.root will contain the TrainTree of all tracks used in training and a TestTree of all tracks that have been evaulated using the LUT but not been trained on.
The settings currently used for training are to use a least squares loss function, log_2(pt) target variable, and the training variables stored in Run2Variables.py. Additional spectator variables have been added to evaluate performance under specific situations.
The BDT also applies weights based on track characteristics and generator information. For example, a muon with a pT of 5GeV will have more weight than a muon with pT of 200GeV since it is more important to correctly assign 5GeV than to assign 200GeV +- 5GeV.
Lastly, Bit Compression has been added into the training to more accurately represent the performance in firmware where the address space is limited ot 30 bits.
The weights.xml file is not in the correct format to load into CMSSW directly. Once you have run on all emtf modes you can run copyLutsToL1TMuonEndcap.py to generate LUT that are compatible with CMSSW which will be split into 11 directories (1 per mode) with 400 xml files in each directory. These can then be loaded into CMSSW under L1Trigger/L1TMuonEndCap/data/ as outlined in cms-sw/cmssw#36094. You need to download the repository where you keep the lookup tables
git clone https://github.com/dildick/L1Trigger-L1TMuonEndCap
mkdir L1Trigger/L1TMuonEndCap/data
cp -r L1Trigger-L1TMuonEndCap/pt_xmls L1Trigger/L1TMuonEndCap/data
rm -rf L1Trigger-L1TMuonEndCap
In the pt_xmls directory there should be subdirectories, one per version of the BDT lookup tables. In each version there should be 1 subdirectory per valid EMTF track mode. So a directory containing 400 XML files may be in directory, L1Trigger/L1TMuonEndCap/data/pt_xmls/v0p0/15/. Each mode should contain 400 xml files.
Note that the script copyLutsToL1TMuonEndcap.py is not yet Python3 compatible. So that needs to be done since future release won't be Python2 compatible anymore.
When using the lookup tables, please set useCustomLUTs to True. In addition, you need to specify the version number for the EMTF BDT, e.g. v0p0 as xmlLutVersion. Both options are in L1Trigger/L1TMuonEndCap/python/simEmtfDigis_cfi.py.
You may need to modify the EMTF track builder according to the attached screenshots so that the lookup tables are loaded only once per simulation, as opposed to once per event.
Sven: Add here instructions how to run EMTF simulation with custom lookup tables.
cmsrel CMSSW_12_1_0_pre3
cd CMSSW_12_1_0_pre3/src
cmsenv
git cms-init
git cms-addpkg L1Trigger/L1TMuonEndCap
git cms-cherry-pick-pr 36094
git clone https://github.com/dildick/L1Trigger-L1TMuonEndCap
mkdir L1Trigger/L1TMuonEndCap/data
cp -r L1Trigger-L1TMuonEndCap/pt_xmls L1Trigger/L1TMuonEndCap/data
rm -rf L1Trigger-L1TMuonEndCap
# now customize the trackfinder so that lookup tables are only loaded once per event (see screenshots)
# load these repositories
git clone git@github.com:dildick/EMTFAnalyzer.git
git clone git@github.com:dildick/EMTFPtAssign2017.git #put here for completeness, but not needed for running CMSSW
# download a test data file from CMSDAS
xrdcpglobal() {
xrdcp -f -d 1 "root://cms-xrd-global.cern.ch/$1" .
}
xrdcpglobal /store/data/Run2018D/ZeroBias/RAW/v1/000/322/022/00000/F87A285E-87AD-E811-89A7-FA163E0481A2.root
# go to the test directory
cd EMTFAnalyzer/NTupleMaker/test/
cmsRun RunTrackFinder_Run3_cfg.py inputFiles=file:F87A285E-87AD-E811-89A7-FA163E0481A2.root run3=False unpack=True l1=True
The variable TRK_hit_ids has been added as a spectator variable and is generated as an 8 bit number that contains which CSCs and RPCs were used in the building of a particular track. The first 4 bits hold information about ME1, ME2, ME3, ME4 and the last four bits hold information about RPC1, RPC2, RPC3, RPC4, where a 1 indicates that a hit in the detector was used and a 0 indicates there was no hit in the detector used in the track. An all CSC track for mode 15 would be 11110000 and an all RPC track for mode 15 would be 00001111.