cms-sw · cmsbuild · Jun 3, 2024 · May 21, 2024
diff --git a/DataFormats/L1TrackTrigger/interface/TTTrack.h b/DataFormats/L1TrackTrigger/interface/TTTrack.h
@@ -372,6 +372,7 @@ double TTTrack<T>::chi2ZRed() const {
   return theChi2_Z_ / (theStubRefs.size() - 2.);
 }
 
+/// prompt track quality MVA
 template <typename T>
 double TTTrack<T>::trkMVA1() const {
   return theTrkMVA1_;

diff --git a/DataFormats/L1TrackTrigger/interface/TTTrack_TrackWord.h b/DataFormats/L1TrackTrigger/interface/TTTrack_TrackWord.h
@@ -103,8 +103,10 @@ class TTTrack_TrackWord {
       {0.0, 0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5, 5.0, 6.0, 8.0, 10.0, 20.0, 50.0}};
   static constexpr std::array<double, 1 << TrackBitWidths::kBendChi2Size> bendChi2Bins = {
       {0.0, 0.75, 1.0, 1.5, 2.25, 3.5, 5.0, 20.0}};
-  static constexpr std::array<double, 1 << TrackBitWidths::kMVAQualitySize> mvaQualityBins = {
-      {0.0, 0.125, 0.25, 0.375, 0.5, 0.625, 0.750, 0.875}};
+
+  // Bin edges for TQ MVA
+  static constexpr std::array<double, 1 << TrackBitWidths::kMVAQualitySize> tqMVABins = {
+      {0, 0.125, 0.25, 0.375, 0.5, 0.625, 0.750, 0.875}};
 
   // Sector constants
   static constexpr unsigned int nSectors = 9;
@@ -144,7 +146,7 @@ class TTTrack_TrackWord {
                     double bendChi2,
                     unsigned int hitPattern,
                     double mvaQuality,
-                    unsigned int mvaOther,
+                    double mvaOther,
                     unsigned int sector);
   TTTrack_TrackWord(unsigned int valid,
                     unsigned int rInv,
@@ -226,8 +228,8 @@ class TTTrack_TrackWord {
   double getBendChi2() const { return bendChi2Bins[getBendChi2Bits()]; }
   unsigned int getHitPattern() const { return getHitPatternBits(); }
   unsigned int getNStubs() const { return countSetBits(getHitPatternBits()); }
-  double getMVAQuality() const { return mvaQualityBins[getMVAQualityBits()]; }
-  unsigned int getMVAOther() const { return getMVAOtherBits(); }
+  double getMVAQuality() const { return tqMVABins[getMVAQualityBits()]; }
+  double getMVAOther() const { return getMVAOtherBits(); }
 
   // ----------member functions (setters) ------------
   void setTrackWord(unsigned int valid,
@@ -239,7 +241,7 @@ class TTTrack_TrackWord {
                     double bendChi2,
                     unsigned int hitPattern,
                     double mvaQuality,
-                    unsigned int mvaOther,
+                    double mvaOther,
                     unsigned int sector);
 
   void setTrackWord(unsigned int valid,

diff --git a/DataFormats/L1TrackTrigger/src/TTTrack_TrackWord.cc b/DataFormats/L1TrackTrigger/src/TTTrack_TrackWord.cc
@@ -36,7 +36,7 @@ TTTrack_TrackWord::TTTrack_TrackWord(unsigned int valid,
                                      double bendChi2,
                                      unsigned int hitPattern,
                                      double mvaQuality,
-                                     unsigned int mvaOther,
+                                     double mvaOther,
                                      unsigned int sector) {
   setTrackWord(valid, momentum, POCA, rInv, chi2RPhi, chi2RZ, bendChi2, hitPattern, mvaQuality, mvaOther, sector);
 }
@@ -66,7 +66,7 @@ void TTTrack_TrackWord::setTrackWord(unsigned int valid,
                                      double bendChi2,
                                      unsigned int hitPattern,
                                      double mvaQuality,
-                                     unsigned int mvaOther,
+                                     double mvaOther,
                                      unsigned int sector) {
   // first, derive quantities to be packed
   float rPhi = localPhi(momentum.phi(), sector);  // this needs to be phi relative to the center of the sector
@@ -85,7 +85,7 @@ void TTTrack_TrackWord::setTrackWord(unsigned int valid,
   chi2rz_t chi2RZ_ = getBin(chi2RZ, chi2RZBins);
   bendChi2_t bendChi2_ = getBin(bendChi2, bendChi2Bins);
   hit_t hitPattern_ = hitPattern;
-  qualityMVA_t mvaQuality_ = getBin(mvaQuality, mvaQualityBins);
+  qualityMVA_t mvaQuality_ = getBin(mvaQuality, tqMVABins);
   otherMVA_t mvaOther_ = mvaOther;
 
   // pack the track word

diff --git a/L1Trigger/TrackFindingTracklet/plugins/L1FPGATrackProducer.cc b/L1Trigger/TrackFindingTracklet/plugins/L1FPGATrackProducer.cc
@@ -739,7 +739,7 @@ void L1FPGATrackProducer::produce(edm::Event& iEvent, const edm::EventSetup& iSe
     aTrack.setStubPtConsistency(
         StubPtConsistency::getConsistency(aTrack, theTrackerGeom, tTopo, settings_.bfield(), settings_.nHelixPar()));
 
-    // set TTTrack word first to allow Track Quality MVA to read variables:
+    // set track word before TQ MVA calculated which uses track word variables
     aTrack.setTrackWordBits();
 
     if (trackQuality_) {
@@ -751,7 +751,10 @@ void L1FPGATrackProducer::produce(edm::Event& iEvent, const edm::EventSetup& iSe
     //      trackQualityModel_->setBonusFeatures(hph.bonusFeatures());
     //    }
 
-    // set TTTrack word
+    // test track word
+    //aTrack.testTrackWordBits();
+
+    // set track word again to set MVA variable from TTTrack into track word
     aTrack.setTrackWordBits();
     // test track word
     //aTrack.testTrackWordBits();

diff --git a/L1Trigger/TrackFindingTracklet/plugins/ProducerKFout.cc b/L1Trigger/TrackFindingTracklet/plugins/ProducerKFout.cc
@@ -78,13 +78,14 @@ namespace trklet {
 
     int partialTrackWordBits_;
 
-    // Helper function to convert floating chi2 to chi2 bin
+    // Helper function to convert floating value to bin
     template <typename T>
     unsigned int digitise(const T& bins, double value, double factor) {
       unsigned int bin = 0;
       for (unsigned int i = 0; i < bins.size() - 1; i++) {
         if (value * factor > bins[i] && value * factor <= bins[i + 1])
-          bin = i;
+          break;
+        bin++;
       }
       return bin;
     }
@@ -200,7 +201,7 @@ namespace trklet {
       StreamsTrack outputStreamsTracks(setup_->numRegions() * setup_->tfpNumChannel());
 
       // Setup containers for track quality
-      float tempTQMVA = 0.0;
+      float tempTQMVAPreSig = 0.0;
       // Due to ap_fixed implementation in CMSSW this 10,5 must be specified at compile time, TODO make this a changeable parameter
       std::vector<ap_fixed<10, 5>> trackQuality_inputs = {0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0};
 
@@ -295,10 +296,11 @@ namespace trklet {
               digitise(TTTrack_TrackWord::chi2RZBins, tempchi2rz, (double)setup_->kfoutchi2rzConv())};
 
           // Run BDT emulation and package output into 3 bits
-
-          tempTQMVA = trackQualityModel_->runEmulatedTQ(trackQuality_inputs);
-          tempTQMVA = std::trunc(tempTQMVA * ap_fixed_rescale);
-          TTBV tqMVA(digitise(tqBins_, tempTQMVA, 1.0), TTTrack_TrackWord::TrackBitWidths::kMVAQualitySize, false);
+          // output needs sigmoid transformation applied
+          tempTQMVAPreSig = trackQualityModel_->runEmulatedTQ(trackQuality_inputs);
+          TTBV tqMVA(digitise(L1TrackQuality::getTqMVAPreSigBins(), tempTQMVAPreSig, 1.0),
+                     TTTrack_TrackWord::TrackBitWidths::kMVAQualitySize,
+                     false);
 
           // Build 32 bit partial tracks for outputting in 64 bit packets
           //                  12 +  3       +  7         +  3    +  6

diff --git a/L1Trigger/TrackTrigger/interface/L1TrackQuality.h b/L1Trigger/TrackTrigger/interface/L1TrackQuality.h
@@ -20,17 +20,13 @@ C.Brown 28/07/20
 #include "DataFormats/L1TrackTrigger/interface/TTTrack.h"
 #include "DataFormats/L1TrackTrigger/interface/TTTrack_TrackWord.h"
 #include "DataFormats/L1TrackTrigger/interface/TTTypes.h"
-#include "PhysicsTools/ONNXRuntime/interface/ONNXRuntime.h"
 #include <memory>
 
 #include "conifer.h"
 #include "ap_fixed.h"
 
 class L1TrackQuality {
 public:
-  // Enum class used for determining prediction behaviour in setL1TrackQuality
-  enum class QualityAlgorithm { Cut, GBDT, GBDT_cpp, NN, None };
-
   //Default Constructor
   L1TrackQuality();
 
@@ -49,36 +45,29 @@ class L1TrackQuality {
   // and a single output to be returned which is then used to fill the bits in the Track Word for situations
   // where a TTTrack datatype is unavailable to be passed to the track quality
   float runEmulatedTQ(std::vector<ap_fixed<10, 5>> inputFeatures);
-  // To set private member data
-  void setCutParameters(std::string const& AlgorithmString,
-                        float maxZ0,
-                        float maxEta,
-                        float chi2dofMax,
-                        float bendchi2Max,
-                        float minPt,
-                        int nStubmin);
 
-  void setONNXModel(std::string const& AlgorithmString,
-                    edm::FileInPath const& ONNXmodel,
-                    std::string const& ONNXInputName,
-                    std::vector<std::string> const& featureNames);
+  void setModel(edm::FileInPath const& model, std::vector<std::string> const& featureNames);
 
   void setBonusFeatures(std::vector<float> bonusFeatures);
 
+  // TQ MVA bin conversions
+  static constexpr double invSigmoid(double value) { return -log(1. / value - 1.); }
+  static constexpr std::array<double, 1 << TTTrack_TrackWord::TrackBitWidths::kMVAQualitySize> getTqMVAPreSigBins() {
+    return {{-16.,
+             invSigmoid(TTTrack_TrackWord::tqMVABins[1]),
+             invSigmoid(TTTrack_TrackWord::tqMVABins[2]),
+             invSigmoid(TTTrack_TrackWord::tqMVABins[3]),
+             invSigmoid(TTTrack_TrackWord::tqMVABins[4]),
+             invSigmoid(TTTrack_TrackWord::tqMVABins[5]),
+             invSigmoid(TTTrack_TrackWord::tqMVABins[6]),
+             invSigmoid(TTTrack_TrackWord::tqMVABins[7])}};
+  }
+
 private:
   // Private Member Data
-  QualityAlgorithm qualityAlgorithm_ = QualityAlgorithm::None;
-  edm::FileInPath ONNXmodel_;
-  std::string ONNXInputName_;
+  edm::FileInPath model_;
   std::vector<std::string> featureNames_;
-  float maxZ0_;
-  float maxEta_;
-  float chi2dofMax_;
-  float bendchi2Max_;
-  float minPt_;
-  int nStubsmin_;
   bool useHPH_;
   std::vector<float> bonusFeatures_;
-  std::unique_ptr<cms::Ort::ONNXRuntime> runTime_;
 };
 #endif
diff --git a/L1Trigger/TrackTrigger/python/TrackQualityParams_cfi.py b/L1Trigger/TrackTrigger/python/TrackQualityParams_cfi.py
@@ -1,24 +1,11 @@
 import FWCore.ParameterSet.Config as cms
 
-TrackQualityParams = cms.PSet(qualityAlgorithm = cms.string("GBDT_cpp"), #None, Cut, NN, GBDT, GBDT_cpp
-                              # This emulation GBDT is optimised for the HYBRID_NEWKF emulation and works with the emulation of the KF out module
+TrackQualityParams = cms.PSet(# This emulation GBDT is optimised for the HYBRID_NEWKF emulation and works with the emulation of the KF out module
                               # It is compatible with the HYBRID simulation and will give equivilant performance with this workflow
-                              ONNXmodel = cms.FileInPath("L1Trigger/TrackTrigger/data/L1_TrackQuality_GBDT_emulation_digitized.json"),
-                              # The ONNX model should be found at this path, if you want a local version of the model:
-                              # git clone https://github.com/cms-data/L1Trigger-TrackTrigger.git L1Trigger/TrackTrigger/data
-                              ONNXInputName = cms.string("feature_input"),
+                              model = cms.FileInPath("L1Trigger/TrackTrigger/data/L1_TrackQuality_GBDT_emulation_digitized.json"),
                               #Vector of strings of training features, in the order that the model was trained with
                               featureNames = cms.vstring(["tanl", "z0_scaled", "bendchi2_bin", "nstub",
                                                           "nlaymiss_interior", "chi2rphi_bin", "chi2rz_bin"]),
-                              # Parameters for cut based classifier, optimized for L1 Track MET
-                              # (Table 3.7  The Phase-2 Upgrade of the CMS Level-1 Trigger http://cds.cern.ch/record/2714892) 
-                              maxZ0 = cms.double ( 15. ) ,    # in cm
-                              maxEta = cms.double ( 2.4 ) ,
-                              chi2dofMax = cms.double( 40. ),
-                              bendchi2Max = cms.double( 2.4 ),
-                              minPt = cms.double( 2. ),       # in GeV
-                              nStubsmin = cms.int32( 4 ),
-                              tqemu_bins = cms.vint32( [-480, -62, -35, -16, 0, 16, 35, 62, 480] ),
                               tqemu_TanlScale = cms.double( 128.0),
                               tqemu_Z0Scale = cms.double( 64.0 ),
                               )
diff --git a/L1Trigger/TrackTrigger/src/L1TrackQuality.cc b/L1Trigger/TrackTrigger/src/L1TrackQuality.cc
@@ -10,26 +10,9 @@ C.Brown & C.Savard 07/2020
 L1TrackQuality::L1TrackQuality() {}
 
 L1TrackQuality::L1TrackQuality(const edm::ParameterSet& qualityParams) : useHPH_(false), bonusFeatures_() {
-  std::string AlgorithmString = qualityParams.getParameter<std::string>("qualityAlgorithm");
   // Unpacks EDM parameter set itself to save unecessary processing within TrackProducers
-  if (AlgorithmString == "Cut") {
-    setCutParameters(AlgorithmString,
-                     (float)qualityParams.getParameter<double>("maxZ0"),
-                     (float)qualityParams.getParameter<double>("maxEta"),
-                     (float)qualityParams.getParameter<double>("chi2dofMax"),
-                     (float)qualityParams.getParameter<double>("bendchi2Max"),
-                     (float)qualityParams.getParameter<double>("minPt"),
-                     qualityParams.getParameter<int>("nStubsmin"));
-  }
-
-  else {
-    setONNXModel(AlgorithmString,
-                 qualityParams.getParameter<edm::FileInPath>("ONNXmodel"),
-                 qualityParams.getParameter<std::string>("ONNXInputName"),
-                 qualityParams.getParameter<std::vector<std::string>>("featureNames"));
-    if ((AlgorithmString == "GBDT") || (AlgorithmString == "NN"))
-      runTime_ = std::make_unique<cms::Ort::ONNXRuntime>(this->ONNXmodel_.fullPath());
-  }
+  setModel(qualityParams.getParameter<edm::FileInPath>("model"),
+           qualityParams.getParameter<std::vector<std::string>>("featureNames"));
 }
 
 std::vector<float> L1TrackQuality::featureTransform(TTTrack<Ref_Phase2TrackerDigi_>& aTrack,
@@ -101,117 +84,28 @@ std::vector<float> L1TrackQuality::featureTransform(TTTrack<Ref_Phase2TrackerDig
 }
 
 void L1TrackQuality::setL1TrackQuality(TTTrack<Ref_Phase2TrackerDigi_>& aTrack) {
-  if (this->qualityAlgorithm_ == QualityAlgorithm::Cut) {
-    // Get Track parameters
-    float trk_pt = aTrack.momentum().perp();
-    float trk_bend_chi2 = aTrack.stubPtConsistency();
-    float trk_z0 = aTrack.z0();
-    float trk_eta = aTrack.momentum().eta();
-    float trk_chi2 = aTrack.chi2();
-    const auto& stubRefs = aTrack.getStubRefs();
-    int nStubs = stubRefs.size();
-
-    float classification = 0.0;  // Default classification is 0
-
-    if (trk_pt >= this->minPt_ && abs(trk_z0) < this->maxZ0_ && abs(trk_eta) < this->maxEta_ &&
-        trk_chi2 < this->chi2dofMax_ && trk_bend_chi2 < this->bendchi2Max_ && nStubs >= this->nStubsmin_)
-      classification = 1.0;
-    // Classification updated to 1 if conditions are met
-
-    aTrack.settrkMVA1(classification);
-  }
-
-  else if (this->qualityAlgorithm_ == QualityAlgorithm::GBDT_cpp) {
-    // load in bdt
-    conifer::BDT<float, float> bdt(this->ONNXmodel_.fullPath());
-
-    // collect features and classify using bdt
-    std::vector<float> inputs = featureTransform(aTrack, this->featureNames_);
-    std::vector<float> output = bdt.decision_function(inputs);
-    aTrack.settrkMVA1(1. / (1. + exp(-output.at(0))));  // need logistic sigmoid fcn applied to xgb output
-  }
-
-  else if ((this->qualityAlgorithm_ == QualityAlgorithm::NN) || (this->qualityAlgorithm_ == QualityAlgorithm::GBDT)) {
-    // Setup ONNX input and output names and arrays
-    std::vector<std::string> ortinput_names;
-    std::vector<std::string> ortoutput_names;
-
-    cms::Ort::FloatArrays ortinput;
-    cms::Ort::FloatArrays ortoutputs;
-
-    std::vector<float> Transformed_features = featureTransform(aTrack, this->featureNames_);
-    //    cms::Ort::ONNXRuntime runTime(this->ONNXmodel_.fullPath());  //Setup ONNX runtime
-
-    ortinput_names.push_back(this->ONNXInputName_);
-    ortoutput_names = runTime_->getOutputNames();
-
-    //ONNX runtime recieves a vector of vectors of floats so push back the input
-    // vector of float to create a 1,1,21 ortinput
-    ortinput.push_back(Transformed_features);
-
-    // batch_size 1 as only one set of transformed features is being processed
-    int batch_size = 1;
-    // Run classification
-    ortoutputs = runTime_->run(ortinput_names, ortinput, {}, ortoutput_names, batch_size);
-
-    if (this->qualityAlgorithm_ == QualityAlgorithm::NN) {
-      aTrack.settrkMVA1(ortoutputs[0][0]);
-    }
-
-    else if (this->qualityAlgorithm_ == QualityAlgorithm::GBDT) {
-      aTrack.settrkMVA1(ortoutputs[1][1]);
-    }
-    // Slight differences in the ONNX models of the GBDTs and NNs mean different
-    // indices of the ortoutput need to be accessed
-  }
+  // load in bdt
+  conifer::BDT<float, float> bdt(this->model_.fullPath());
 
-  else {
-    aTrack.settrkMVA1(-999);
-  }
+  // collect features and classify using bdt
+  std::vector<float> inputs = featureTransform(aTrack, this->featureNames_);
+  std::vector<float> output = bdt.decision_function(inputs);
+  aTrack.settrkMVA1(1. / (1. + exp(-output.at(0))));
 }
 
 float L1TrackQuality::runEmulatedTQ(std::vector<ap_fixed<10, 5>> inputFeatures) {
   // load in bdt
 
-  conifer::BDT<ap_fixed<10, 5>, ap_fixed<10, 5>> bdt(this->ONNXmodel_.fullPath());
+  conifer::BDT<ap_fixed<10, 5>, ap_fixed<10, 5>> bdt(this->model_.fullPath());
 
   // collect features and classify using bdt
   std::vector<ap_fixed<10, 5>> output = bdt.decision_function(inputFeatures);
   return output.at(0).to_float();  // need logistic sigmoid fcn applied to xgb output
 }
 
-void L1TrackQuality::setCutParameters(std::string const& AlgorithmString,
-                                      float maxZ0,
-                                      float maxEta,
-                                      float chi2dofMax,
-                                      float bendchi2Max,
-                                      float minPt,
-                                      int nStubmin) {
-  qualityAlgorithm_ = QualityAlgorithm::Cut;
-  maxZ0_ = maxZ0;
-  maxEta_ = maxEta;
-  chi2dofMax_ = chi2dofMax;
-  bendchi2Max_ = bendchi2Max;
-  minPt_ = minPt;
-  nStubsmin_ = nStubmin;
-}
-
-void L1TrackQuality::setONNXModel(std::string const& AlgorithmString,
-                                  edm::FileInPath const& ONNXmodel,
-                                  std::string const& ONNXInputName,
-                                  std::vector<std::string> const& featureNames) {
+void L1TrackQuality::setModel(edm::FileInPath const& model, std::vector<std::string> const& featureNames) {
   //Convert algorithm string to Enum class for track by track comparison
-  if (AlgorithmString == "NN") {
-    qualityAlgorithm_ = QualityAlgorithm::NN;
-  } else if (AlgorithmString == "GBDT") {
-    qualityAlgorithm_ = QualityAlgorithm::GBDT;
-  } else if (AlgorithmString == "GBDT_cpp") {
-    qualityAlgorithm_ = QualityAlgorithm::GBDT_cpp;
-  } else {
-    qualityAlgorithm_ = QualityAlgorithm::None;
-  }
-  ONNXmodel_ = ONNXmodel;
-  ONNXInputName_ = ONNXInputName;
+  model_ = model;
   featureNames_ = featureNames;
 }