CbmTrdElectronsTrainAnn.cxx

Go to the documentation of this file.

/* Copyright (C) 2009-2021 UGiessen, JINR-LIT
   SPDX-License-Identifier: GPL-3.0-only
   Authors: Semen Lebedev [committer] */
 
#include "CbmTrdElectronsTrainAnn.h"
 
#include "CbmDrawHist.h"
#include "CbmMCTrack.h"
#include "CbmTrackMatchNew.h"
#include "CbmTrdHit.h"
#include "CbmTrdPoint.h"
#include "CbmTrdTrack.h"
#include "TCanvas.h"
#include "TClonesArray.h"
#include "TCut.h"
#include "TFile.h"
#include "TGraph.h"
#include "TLine.h"
#include "TMVA/Config.h"
#include "TMVA/PDF.h"
#include "TMath.h"
#include "TRandom.h"
#include "TSystem.h"
 
#include <FairRootManager.h>
#include <Logger.h>
 
#include <boost/assign/list_of.hpp>
 
#include <iostream>
#include <sstream>
#include <string>
#include <vector>
 
using boost::assign::list_of;
using std::cout;
using std::endl;
using std::string;
using std::stringstream;
using std::vector;
 
CbmTrdElectronsTrainAnn::CbmTrdElectronsTrainAnn(Int_t nofTrdLayers)
  : FairTask()
  , fMCTracks(NULL)
  , fTrdPoints(NULL)
  , fTrdTracks(NULL)
  , fTrdTrackMatches(NULL)
  , fTrdHits(NULL)
  , fEloss()
  , fHists()
  , fhResults()
  , fhMeanEloss()
  , fhEloss()
  , fhElossSort()
  , fEventNum(0)
  , fOutputDir("results/")
  , fSigmaError(0.0)
  , fIsDoTrain(true)
  , fTransformType(1)
  , fBeamDataFile("")
  , fBeamDataPiHist("")
  , fBeamDataElHist("")
  , fAnnInput()
  , fXOut(-1.)
  , fNofTrdLayers(nofTrdLayers)
  , fMaxEval(1.3)
  , fMinEval(-1.3)
  , fNN(NULL)
  , fReader(NULL)
  , fIdMethod(kANN)
  , fNofAnnEpochs(50)
  , fNofTrainSamples(2500)
  , fElIdEfficiency(0.9)
  , fhOutput()
  , fhCumProbOutput()
  , fhInput()
{
  fEloss.resize(2);
  fhMeanEloss.resize(2);
  fhEloss.resize(2);
  string s;
 
  TH1::SetDefaultBufferSize(30000);
 
  fhResults = new TH1D("fhResults", "fhResults", 2, 0, 2);
  fHists.push_back(fhResults);
 
  for (int i = 0; i < 2; i++) {
    if (i == 0) s = "El";
    if (i == 1) s = "Pi";
    fhMeanEloss[i] = new TH1D(("fhMeanEloss" + s).c_str(), "fhMeanEloss;Mean energy loss [a.u.];Yield", 100, 0., 50e-6);
    fHists.push_back(fhMeanEloss[i]);
    fhEloss[i] = new TH1D(("fhEloss" + s).c_str(), "fhEloss;Energy loss [a.u.];Yield", 100, 0., 50e-6);
    fHists.push_back(fhEloss[i]);
  }
 
  Int_t nofSortBins = 200;
  fhElossSort.resize(2);
  for (int j = 0; j < 2; j++) {
    fhElossSort[j].resize(fNofTrdLayers);
    for (Int_t i = 0; i < fNofTrdLayers; i++) {
      std::stringstream ss1;
      if (j == 0) ss1 << "fhElossSortEl" << i;
      if (j == 1) ss1 << "fhElossSortPi" << i;
      fhElossSort[j][i] =
        new TH1D(ss1.str().c_str(), (ss1.str() + ";Energy loss [a.u.];Counters").c_str(), nofSortBins, 0., 0.);
    }
  }
 
  // Histograms for algorithm testing
  Int_t nofBins = 10000;
  fhOutput.resize(2);
  fhCumProbOutput.resize(2);
  fhInput.resize(2);
  for (Int_t i = 0; i < 2; i++) {
    s                  = (i == 0) ? "El" : "Pi";
    fhOutput[i]        = new TH1D(("fhOutput" + s).c_str(), "fhOutput;Output;Counter", nofBins, fMinEval, fMaxEval);
    fhCumProbOutput[i] = new TH1D(("fhCumProbOutput" + s).c_str(), "fhCumProbOutput;Output;Cumulative probability",
                                  nofBins, fMinEval, fMaxEval);
 
    fhInput[i].resize(fNofTrdLayers);
    for (Int_t j = 0; j < fNofTrdLayers; j++) {
      stringstream ss;
      ss << s << j;
      fhInput[i][j] = new TH1D(("fhInput" + ss.str()).c_str(), "fhInput", 100, 0.0, 0.0);
    }
  }
}
 
CbmTrdElectronsTrainAnn::~CbmTrdElectronsTrainAnn() {}
 
InitStatus CbmTrdElectronsTrainAnn::Init()
{
  FairRootManager* ioman = FairRootManager::Instance();
  if (NULL == ioman) {
    LOG(fatal) << GetName() << "::Init: RootManager not instantised";
  }
 
  fMCTracks = (TClonesArray*) ioman->GetObject("MCTrack");
  if (NULL == fMCTracks) {
    LOG(fatal) << GetName() << "::Init: No MCTrack array";
  }
 
  fTrdPoints = (TClonesArray*) ioman->GetObject("TrdPoint");
  if (NULL == fTrdPoints) {
    LOG(fatal) << GetName() << "::Init: No TRDPoint array";
  }
 
  fTrdTracks = (TClonesArray*) ioman->GetObject("TrdTrack");
  if (NULL == fTrdTracks) {
    LOG(fatal) << GetName() << "::Init: No TrdTrack array";
  }
 
  fTrdTrackMatches = (TClonesArray*) ioman->GetObject("TrdTrackMatch");
  if (NULL == fTrdTrackMatches) {
    LOG(fatal) << GetName() << "::Init: No TrdTrackMatch array";
  }
 
  fTrdHits = (TClonesArray*) ioman->GetObject("TrdHit");
  if (NULL == fTrdHits) {
    LOG(fatal) << GetName() << "::Init: No TrdHit array";
  }
 
  return kSUCCESS;
}
 
void CbmTrdElectronsTrainAnn::Exec(Option_t*)
{
  fEventNum++;
  cout << "CbmTrdElectronsTrainAnn, event " << fEventNum << endl;
 
  FillElossVectorSim();
  FillElossHist();
  SortElossAndFillHist();
 
  cout << "Nof electrons = " << fEloss[0].size() << endl;
  cout << "Nof pions = " << fEloss[1].size() << endl;
 
  // Finish();
}
 
void CbmTrdElectronsTrainAnn::Finish()
{
  Run();
  Draw();
 
  if (fOutputDir != "") {
    gSystem->mkdir(fOutputDir.c_str(), true);
  }
 
  TFile* oldFile     = gFile;
  TDirectory* oldDir = gDirectory;
 
  TFile* f = new TFile(string(fOutputDir + "/trd_elid_hist.root").c_str(), "RECREATE");
  for (unsigned int i = 0; i < fHists.size(); i++) {
    fHists[i]->Write();
  }
 
  gFile      = oldFile;
  gDirectory = oldDir;
 
  f->Close();
}
 
void CbmTrdElectronsTrainAnn::RunBeamData()
{
  FillElossVectorReal();
  FillElossHist();
  SortElossAndFillHist();
 
  cout << "Nof electrons = " << fEloss[0].size() << endl;
  cout << "Nof pions = " << fEloss[1].size() << endl;
 
  Finish();
}
 
void CbmTrdElectronsTrainAnn::FillElossVectorReal()
{
  if (fBeamDataFile == "" || fBeamDataPiHist == "" || fBeamDataElHist == "") {
    LOG(fatal) << GetName() << "::FillElossVectorReal: Set input file for beam data and histogram names!";
  }
 
  TFile* oldFile     = gFile;
  TDirectory* oldDir = gDirectory;
 
  TFile* file = new TFile(fBeamDataFile.c_str(), "READ");
  LOG_IF(fatal, !file) << "Could not open file " << fBeamDataFile;
 
  TH1F* hPion = static_cast<TH1F*>(file->Get<TH1F>(fBeamDataPiHist.c_str())->Clone());
  LOG_IF(fatal, !hPion) << "Histogram " << fBeamDataPiHist << " not found in file " << fBeamDataFile;
 
  TH1F* hElectron = static_cast<TH1F*>(file->Get<TH1F>(fBeamDataElHist.c_str())->Clone());
  LOG_IF(fatal, !hElectron) << "Histogram " << fBeamDataElHist << " not found in file " << fBeamDataFile;
 
  double scaleX =
    fhEloss[0]->GetXaxis()->GetBinUpEdge(fhEloss[0]->GetNbinsX()) / hPion->GetXaxis()->GetBinUpEdge(hPion->GetNbinsX());
 
  int nofSimulatedParticles = 1000000;
  fEloss[0].resize(nofSimulatedParticles);
  fEloss[1].resize(nofSimulatedParticles);
 
  for (int iP = 0; iP < 2; iP++) {
    TH1F* hist = hElectron;
    if (iP == 1) hist = hPion;
    for (int iT = 0; iT < nofSimulatedParticles; iT++) {
      fEloss[iP][iT].resize(fNofTrdLayers);
      for (int iH = 0; iH < fNofTrdLayers; iH++) {
        double eloss = scaleX * hist->GetRandom();
        TrdEloss e(eloss);
        fEloss[iP][iT][iH] = e;
      }
    }
  }
 
  gFile      = oldFile;
  gDirectory = oldDir;
}
 
void CbmTrdElectronsTrainAnn::FillElossVectorSim()
{
  Int_t nofTrdTracks = fTrdTracks->GetEntriesFast();
  for (Int_t iTrdTrack = 0; iTrdTrack < nofTrdTracks; iTrdTrack++) {
    CbmTrdTrack* trdtrack = (CbmTrdTrack*) fTrdTracks->At(iTrdTrack);
    Int_t nHits           = trdtrack->GetNofHits();
 
 
    CbmTrackMatchNew* trdmatch = (CbmTrackMatchNew*) fTrdTrackMatches->At(iTrdTrack);
    Int_t iMC                  = trdmatch->GetMatchedLink().GetIndex();
    if (iMC < 0 || iMC > fMCTracks->GetEntriesFast()) continue;
 
    CbmMCTrack* mctrack = (CbmMCTrack*) fMCTracks->At(iMC);
    Int_t pdg           = TMath::Abs(mctrack->GetPdgCode());
    //      Double_t momMC = mctrack->GetP();
    Double_t motherId = mctrack->GetMotherId();
 
    if (motherId != -1) continue;
 
    int iP = 0;              //[0] = electron
    if (pdg == 211) iP = 1;  //[1] = pion
    vector<TrdEloss> v;
 
    if (nHits < fNofTrdLayers) continue;
 
    for (Int_t iHit = 0; iHit < nHits; iHit++) {
      Int_t hitIndex    = trdtrack->GetHitIndex(iHit);
      CbmTrdHit* trdhit = (CbmTrdHit*) fTrdHits->At(hitIndex);
      TrdEloss e(trdhit->GetELoss(), trdhit->GetELoss(), 0 /*trdhit->GetELossdEdX(), trdhit->GetELossTR()*/);
      if (static_cast<UInt_t>(fNofTrdLayers) == v.size()) break;
      v.push_back(e);
    }  //iHit
    fEloss[iP].push_back(v);
  }  //iTrdTrack
}
 
void CbmTrdElectronsTrainAnn::FillElossHist()
{
  // [0]=electron, [1]=pion
  for (int i = 0; i < 2; i++) {
    for (unsigned int iT = 0; iT < fEloss[i].size(); iT++) {
      Int_t nHits     = fEloss[i][iT].size();
      double sumEloss = 0.;
      for (int iH = 0; iH < nHits; iH++) {
        sumEloss += fEloss[i][iT][iH].fEloss;
        fhEloss[i]->Fill(fEloss[i][iT][iH].fEloss);
      }  // iH
      fhMeanEloss[i]->Fill(sumEloss / nHits);
    }  //iT
  }
}
 
void CbmTrdElectronsTrainAnn::SortElossAndFillHist()
{
  vector<double> v;
  v.resize(fNofTrdLayers);
  for (int iP = 0; iP < 2; iP++) {
    for (unsigned int iT = 0; iT < fEloss[iP].size(); iT++) {
      for (int iH = 0; iH < fNofTrdLayers; iH++) {
        v[iH] = fEloss[iP][iT][iH].fEloss;
      }
      std::sort(v.begin(), v.end());
      for (int iH = 0; iH < fNofTrdLayers; iH++) {
        fhElossSort[iP][iH]->Fill(v[iH]);
      }
    }  //iT
 
    for (int iH = 0; iH < fNofTrdLayers; iH++) {
      fhElossSort[iP][iH]->Scale(1. / fhElossSort[iP][iH]->Integral());
    }
  }
}
 
void CbmTrdElectronsTrainAnn::Run()
{
  if (fIdMethod == kBDT || fIdMethod == kANN) {
    if (fIsDoTrain) DoTrain();
    DoTest();
  }
  else if (fIdMethod == kMEDIAN || fIdMethod == kLIKELIHOOD || fIdMethod == kMeanCut) {
    DoTest();
  }
}
 
void CbmTrdElectronsTrainAnn::Transform()
{
  if (fIdMethod == kBDT || fIdMethod == kANN) {
    if (fTransformType == 1) {
      Transform1();
    }
    else if (fTransformType == 2) {
      Transform2();
    }
  }
}
 
 
void CbmTrdElectronsTrainAnn::Transform1()
{
  Double_t ANNCoef1 = 1.06;
  Double_t ANNCoef2 = 0.57;
  //Double_t ANNCoef1[] = {1.04,1.105,1.154,1.277,1.333,1.394,1.47,1.50,1.54,1.58};
  //Double_t ANNCoef2[] = {0.548,0.567,0.585,0.63,0.645,0.664,0.69,0.705,0.716,0.723};
  for (UInt_t i = 0; i < fAnnInput.size(); i++) {
    fAnnInput[i] = fAnnInput[i] * 1e6;
    fAnnInput[i] = (fAnnInput[i] - ANNCoef1) / ANNCoef2 - 0.225;
  }
  sort(fAnnInput.begin(), fAnnInput.end());
  for (UInt_t i = 0; i < fAnnInput.size(); i++) {
    fAnnInput[i] = TMath::LandauI(fAnnInput[i]);
  }
}
 
void CbmTrdElectronsTrainAnn::Transform2()
{
  sort(fAnnInput.begin(), fAnnInput.end());
 
  Int_t size = fAnnInput.size();
  for (Int_t j = 0; j < size; j++) {
    Int_t binNumEl = fhElossSort[0][j]->FindBin(fAnnInput[j]);
    if (binNumEl > fhElossSort[0][j]->GetNbinsX()) binNumEl = fhElossSort[0][j]->GetNbinsX();
    Double_t probEl = fhElossSort[0][j]->GetBinContent(binNumEl);
 
    Int_t binNumPi = fhElossSort[1][j]->FindBin(fAnnInput[j]);
    if (binNumPi > fhElossSort[1][j]->GetNbinsX()) binNumPi = fhElossSort[1][j]->GetNbinsX();
    Double_t probPi = fhElossSort[1][j]->GetBinContent(binNumPi);
 
    if (TMath::IsNaN(probPi / (probEl + probPi))) {
      fAnnInput[j] = 0.;
    }
    else {
      fAnnInput[j] = probPi / (probEl + probPi);
    }
  }
}
 
Double_t CbmTrdElectronsTrainAnn::Likelihood()
{
  Double_t lPi = 1.;
  Double_t lEl = 1.;
  // sort(fAnnInput.begin(), fAnnInput.end());
  for (UInt_t j = 0; j < fAnnInput.size(); j++) {
    Int_t binNumEl = fhEloss[0]->FindBin(fAnnInput[j]);
    if (binNumEl > fhEloss[0]->GetNbinsX()) binNumEl = fhEloss[0]->GetNbinsX();
    Double_t probEl = fhEloss[0]->GetBinContent(binNumEl);
    lEl             = lEl * probEl;
 
    Int_t binNumPi = fhEloss[1]->FindBin(fAnnInput[j]);
    if (binNumPi > fhEloss[1]->GetNbinsX()) binNumPi = fhEloss[1]->GetNbinsX();
    Double_t probPi = fhEloss[1]->GetBinContent(binNumPi);
    lPi             = lPi * probPi;
  }
  return lEl / (lEl + lPi);
}
 
Double_t CbmTrdElectronsTrainAnn::Median()
{
  sort(fAnnInput.begin(), fAnnInput.end());
  double eval = 0.;
  if (fNofTrdLayers % 2 == 0) {
    eval = (fAnnInput[fNofTrdLayers / 2] + fAnnInput[fNofTrdLayers / 2 + 1]) / 2.;
  }
  else {
    eval = fAnnInput[fNofTrdLayers / 2 + 1];
  }
  double scaleX = 1. / fhEloss[0]->GetXaxis()->GetBinUpEdge(fhEloss[0]->GetNbinsX());
  return eval * scaleX;
}
 
Double_t CbmTrdElectronsTrainAnn::MeanCut()
{
  double eval = 0.;
  for (int i = 0; i < fNofTrdLayers; i++) {
    eval += fAnnInput[i];
  }
  eval          = eval / fNofTrdLayers;
  double scaleX = 1. / fhEloss[0]->GetXaxis()->GetBinUpEdge(fhEloss[0]->GetNbinsX());
  return eval * scaleX;
}
 
Double_t CbmTrdElectronsTrainAnn::Eval(Bool_t)
{
  if (fIdMethod == kBDT) {
    return fReader->EvaluateMVA("BDT");
  }
  else if (fIdMethod == kANN) {
    Double_t par[fNofTrdLayers];
    for (UInt_t i = 0; i < fAnnInput.size(); i++)
      par[i] = fAnnInput[i];
    return fNN->Evaluate(0, par);
  }
  else if (fIdMethod == kMEDIAN) {
    return Median();
  }
  else if (fIdMethod == kLIKELIHOOD) {
    return Likelihood();
  }
  else if (fIdMethod == kMeanCut) {
    return MeanCut();
  }
  return -1.;
}
 
 
void CbmTrdElectronsTrainAnn::DoTrain()
{
  fAnnInput.clear();
  fAnnInput.resize(fNofTrdLayers);
 
  TTree* simu = CreateTree();
  for (Int_t i = 0; i < 2; i++) {
    for (UInt_t iT = 0; iT < fEloss[i].size() - 2; iT++) {
      for (int iH = 0; iH < fNofTrdLayers; iH++) {
        fAnnInput[iH] = fEloss[i][iT][iH].fEloss;
      }
      fXOut = (i == 0) ? 1. : -1.;
      Transform();
      simu->Fill();
      if (Int_t(iT) >= fNofTrainSamples) break;
    }  //iT
  }    //i
 
  if (fIdMethod == kANN) {
    if (NULL != fNN) delete fNN;
    TString mlpString = CreateAnnString();
    cout << "-I- Create ANN: " << mlpString << endl;
    cout << "-I- Number of training epochs = " << fNofAnnEpochs << endl;
    fNN = new TMultiLayerPerceptron(mlpString, simu, "(Entry$+1)");
    fNN->Train(fNofAnnEpochs, "+text,update=10");
    stringstream ss;
    ss << fOutputDir + "/ann_weights_" << fNofTrdLayers << ".txt";
    fNN->DumpWeights(ss.str().c_str());
  }
  else if (fIdMethod == kBDT) {
    CreateFactory(simu);
    (TMVA::gConfig().GetIONames()).fWeightFileDir = fOutputDir;
    TCut mycuts                                   = "";
    TCut mycutb                                   = "";
    //      factory->PrepareTrainingAndTestTree(mycuts, mycutb,"SplitMode=Random:NormMode=NumEvents:!V");
    //factory->BookMethod(TMVA::Types::kTMlpANN, "TMlpANN","!H:!V:NCycles=50:HiddenLayers=N+1");
    stringstream ss;
    ss << "nTrain_Signal=" << fNofTrainSamples - 500 << ":nTrain_Background=" << fNofTrainSamples - 500
       << ":nTest_Signal=0:nTest_Background=0";
    //TMVA_API
    //      factory->PrepareTrainingAndTestTree("", ss.str().c_str());
    //      factory->BookMethod(TMVA::Types::kBDT, "BDT", "!H:!V:NTrees=400:BoostType=AdaBoost:SeparationType=GiniIndex:nCuts=20:PruneMethod=CostComplexity:PruneStrength=4.5");
    //      factory->TrainAllMethods();
  }
}
 
void CbmTrdElectronsTrainAnn::DoPreTest()
{
  cout << "-I- Start pretesting " << endl;
  if (fIdMethod == kBDT) {
    cout << "-I- IdMethod = kBDT " << endl;
  }
  else if (fIdMethod == kANN) {
    cout << "-I- IdMethod = kANN " << endl;
  }
  else if (fIdMethod == kMEDIAN) {
    cout << "-I- IdMethod = kMEDIANA " << endl;
  }
  else if (fIdMethod == kLIKELIHOOD) {
    cout << "-I- IdMethod = kLIKELIHOOD " << endl;
  }
 
  fAnnInput.clear();
  fAnnInput.resize(fNofTrdLayers);
 
  if (fIdMethod == kBDT) {
    fReader = CreateTmvaReader();
    fReader->BookMVA("BDT", fOutputDir + "/TMVAnalysis_BDT.weights.xml");
  }
  else if (fIdMethod == kANN) {
    if (fNN != NULL) delete fNN;
    TTree* simu       = CreateTree();
    TString mlpString = CreateAnnString();
    fNN               = new TMultiLayerPerceptron(mlpString, simu, "(Entry$+1)");
    stringstream ss;
    ss << fOutputDir + "/ann_weights_" << fNofTrdLayers << ".txt";
    fNN->LoadWeights(ss.str().c_str());
  }
 
  for (Int_t i = 0; i < 2; i++) {
    for (UInt_t iT = 0; iT < fEloss[i].size(); iT++) {
      if (Int_t(iT) < fNofTrainSamples) continue;  // exclude training samples
      if (iT % 100000 == 0) cout << "-I- read number: " << iT << endl;
 
      //change electron hit to pion hit and vice versa
      //int randPi = rand()%fElossPi.size();
      //for (int i1 = 0; i1 < 5; i1++){
      //     fElossEl[iE][i1] = fElossPi[randPi][i1];
      // }
      for (int iH = 0; iH < fNofTrdLayers; iH++) {
        if (fSigmaError != 0.) fEloss[i][iT][iH].fEloss += gRandom->Gaus(0., fSigmaError);
        fAnnInput[iH] = fEloss[i][iT][iH].fEloss;
      }
 
      Transform();
      Bool_t isEl = (i == 0) ? true : false;
      FillAnnInputHist(isEl);
      Double_t nnEval = Eval(isEl);
      if (nnEval > fMaxEval) nnEval = fMaxEval - 0.01;
      if (nnEval < fMinEval) nnEval = fMinEval + 0.01;
      fhOutput[i]->Fill(nnEval);
    }
  }
  CreateCumProbOutputHist();
}
 
void CbmTrdElectronsTrainAnn::DoTest()
{
  fAnnInput.clear();
  fAnnInput.resize(fNofTrdLayers);
 
  DoPreTest();
  double cut = FindOptimalCut();
  cout << "-I- Optimal cut = " << cut << " for " << 100 * fElIdEfficiency << "% electron efficiency" << endl;
  cout << "-I- Start testing" << endl;
  fAnnInput.clear();
  fAnnInput.resize(fNofTrdLayers);
 
  if (fIdMethod == kBDT) {
    fReader = CreateTmvaReader();
    //reader->BookMVA("TMlpANN", "weights/TMVAnalysis_TMlpANN.weights.txt");
    fReader->BookMVA("BDT", fOutputDir + "/TMVAnalysis_BDT.weights.xml");
  }
  else if (fIdMethod == kANN) {
    if (fNN != NULL) delete fNN;
    TTree* simu       = CreateTree();
    TString mlpString = CreateAnnString();
    fNN               = new TMultiLayerPerceptron(mlpString, simu, "(Entry$+1)");
    stringstream ss;
    ss << fOutputDir + "/ann_weights_" << fNofTrdLayers << ".txt";
    fNN->LoadWeights(ss.str().c_str());
  }
 
  int nofPiLikeEl = 0;
  int nofElLikePi = 0;
  int nofElTest   = 0;
  int nofPiTest   = 0;
 
  for (Int_t i = 0; i < 2; i++) {
    for (UInt_t iT = 0; iT < fEloss[i].size(); iT++) {
      if (Int_t(iT) < fNofTrainSamples) continue;  //exclude training samples
      if (iT % 100000 == 0) cout << "-I- Read number: " << iT << endl;
 
      for (Int_t iH = 0; iH < fNofTrdLayers; iH++) {
        fAnnInput[iH] = fEloss[i][iT][iH].fEloss;
      }
      Transform();
 
      Bool_t isEl     = (i == 0) ? true : false;
      Double_t nnEval = Eval(isEl);
      if (nnEval > fMaxEval) nnEval = fMaxEval - 0.01;
      if (nnEval < fMinEval) nnEval = fMinEval + 0.01;
      if (i == 0) {
        nofElTest++;
        if (nnEval < cut) nofElLikePi++;
      }
      else if (i == 1) {
        nofPiTest++;
        if (nnEval > cut) nofPiLikeEl++;
      }
    }
  }
  double piSupp = -1;
  if (nofPiLikeEl != 0) piSupp = (double) nofPiTest / nofPiLikeEl;
  double elEff = (double) nofElLikePi / nofElTest * 100.;
 
  cout << "Testing results:" << endl;
  cout << "nof TRD layers " << fNofTrdLayers << endl;
  cout << "cut = " << cut << endl;
  cout << "nof El = " << nofElTest << endl;
  cout << "nof Pi = " << nofPiTest << endl;
  cout << "nof Pi identified as El = " << nofPiLikeEl << endl;
  cout << "nof El identified as Pi = " << nofElLikePi << endl;
  cout << "pion suppression = " << nofPiTest << "/" << nofPiLikeEl << " = " << piSupp << endl;
  cout << "electron efficiency loss in % = " << nofElLikePi << "/" << nofElTest << " = " << elEff << endl;
 
  // write results to histogramm
  fhResults->SetBinContent(1, piSupp);
  fhResults->SetBinContent(2, elEff);
}
 
void CbmTrdElectronsTrainAnn::CreateCumProbOutputHist()
{
  for (Int_t i = 0; i < 2; i++) {
    Double_t cumProb = 0.;
    Double_t nofTr   = fhOutput[i]->GetEntries();
    for (Int_t iH = 1; iH <= fhOutput[i]->GetNbinsX(); iH++) {
      cumProb += fhOutput[i]->GetBinContent(iH);
      Double_t pr = (i == 0) ? 1. - cumProb / nofTr : cumProb / nofTr;
      fhCumProbOutput[i]->SetBinContent(iH, pr);
    }
  }
}
 
TGraph* CbmTrdElectronsTrainAnn::CreateRocDiagramm()
{
  Int_t nBins = fhCumProbOutput[0]->GetNbinsX();
  Double_t x[nBins], y[nBins];
 
  for (Int_t i = 1; i <= nBins; i++) {
    Double_t cpi = 100. * fhCumProbOutput[1]->GetBinContent(i);
    if (cpi == 100.) cpi = 100. - 0.0001;
    y[i - 1] = 100. / (100. - cpi);
    x[i - 1] = 100. * fhCumProbOutput[0]->GetBinContent(i);
  }
  TGraph* rocGraph = new TGraph(nBins, x, y);
  rocGraph->SetTitle("ROC diagram");
  rocGraph->GetXaxis()->SetTitle("Efficiency [%]");
  rocGraph->GetYaxis()->SetTitle("Pion suppression");
  rocGraph->SetLineWidth(3);
  return rocGraph;
}
 
Double_t CbmTrdElectronsTrainAnn::FindOptimalCut()
{
  Double_t optimalCut = -1;
  for (Int_t i = 1; i <= fhCumProbOutput[0]->GetNbinsX(); i++) {
    if (fhCumProbOutput[0]->GetBinContent(i) <= fElIdEfficiency) {
      optimalCut = fhCumProbOutput[0]->GetBinCenter(i);
      return optimalCut;
    }
  }
  return optimalCut;
}
 
TTree* CbmTrdElectronsTrainAnn::CreateTree()
{
  fAnnInput.clear();
  fAnnInput.resize(fNofTrdLayers);
  TTree* simu     = new TTree("MonteCarlo", "MontecarloData");
  size_t buf_size = 100;
  char txt1[buf_size], txt2[buf_size];
  for (Int_t i = 0; i < fNofTrdLayers; i++) {
    snprintf(txt1, buf_size - 1, "x%d", i);
    snprintf(txt2, buf_size - 1, "x%d/F", i);
    simu->Branch(txt1, &fAnnInput[i], txt2);
  }
  simu->Branch("xOut", &fXOut, "xOut/F");
 
  return simu;
}
 
string CbmTrdElectronsTrainAnn::CreateAnnString()
{
  string st       = "";
  size_t buf_size = 50;
  char txt[buf_size];
  for (Int_t i = 0; i < fNofTrdLayers - 1; i++) {
    snprintf(txt, buf_size - 1, "x%d", i);
    st = st + txt + ",";
  }
  snprintf(txt, buf_size - 1, "x%d", fNofTrdLayers - 1);
  st = st + txt + "";
  snprintf(txt, buf_size - 1, "%d", 2 * fNofTrdLayers);
  st = st + ":" + txt + ":xOut";
  return st;
}
 
TMVA::Factory* CbmTrdElectronsTrainAnn::CreateFactory(TTree*)
{
  if (fOutputDir != "") gSystem->mkdir(fOutputDir.c_str(), true);
  TFile* outputFile = TFile::Open(string(fOutputDir + "/tmva_output.root").c_str(), "RECREATE");
 
  TMVA::Factory* factory = new TMVA::Factory("TMVAnalysis", outputFile);
 
  TCut piCut = "xOut<0";
  TCut elCut = "xOut>0";
  //TMVA_API
  //factory->SetInputTrees(simu, elCut, piCut);
  size_t buf_size = 100;
  char txt1[buf_size];
  for (Int_t i = 0; i < fNofTrdLayers; i++) {
    snprintf(txt1, buf_size - 1, "x%d", i);
    //factory->AddVariable(txt1, 'F');
  }
 
  return factory;
}
 
TMVA::Reader* CbmTrdElectronsTrainAnn::CreateTmvaReader()
{
  if (!fReader) delete fReader;
 
  fAnnInput.clear();
  fAnnInput.resize(fNofTrdLayers);
 
  TMVA::Reader* reader = new TMVA::Reader();
 
  size_t buf_size = 100;
  char txt1[buf_size];
  for (Int_t i = 0; i < fNofTrdLayers; i++) {
    snprintf(txt1, buf_size - 1, "x%d", i);
    reader->AddVariable(txt1, &fAnnInput[i]);
  }
  return reader;
}
 
void CbmTrdElectronsTrainAnn::FillAnnInputHist(Bool_t isEl)
{
  for (UInt_t j = 0; j < fAnnInput.size(); j++) {
    if (isEl) {
      fhInput[0][j]->Fill(fAnnInput[j]);
    }
    else {
      fhInput[1][j]->Fill(fAnnInput[j]);
    }
  }
}
 
void CbmTrdElectronsTrainAnn::Draw(Option_t*)
{
  SetDefaultDrawStyle();
  TCanvas* cEloss = new TCanvas("trd_elid_eloss", "trd_elid_eloss", 1200, 600);
  cEloss->Divide(2, 1);
  cEloss->cd(1);
  DrawH1(fhMeanEloss, list_of("e^{#pm}")("#pi^{#pm}"), kLinear, kLog, true, 0.8, 0.8, 0.99, 0.99);
  cEloss->cd(2);
  DrawH1(fhEloss, list_of("e^{#pm}")("#pi^{#pm}"), kLinear, kLog, true, 0.8, 0.8, 0.99, 0.99);
 
 
  TCanvas* cElossSort = new TCanvas("trd_elid_eloss_sort", "trd_elid_eloss_sort", 1200, 900);
  cElossSort->Divide(4, 3);
  for (int iL = 0; iL < fNofTrdLayers; iL++) {
    cElossSort->cd(iL + 1);
    DrawH1(list_of(fhElossSort[0][iL])(fhElossSort[1][iL]), list_of("e^{#pm}")("#pi^{#pm}"), kLinear, kLog, true, 0.8,
           0.8, 0.99, 0.99);
  }
 
  //   TCanvas* cClassifierOutput = new TCanvas("trd_elid_classifier_output","trd_elid_classifier_output", 500, 500);
  TH1D* out0 = (TH1D*) fhOutput[0]->Clone();
  TH1D* out1 = (TH1D*) fhOutput[1]->Clone();
  out0->Rebin(50);
  out1->Rebin(50);
  DrawH1(list_of(out0)(out1), list_of("e^{#pm}")("#pi^{#pm}"), kLinear, kLog, true, 0.8, 0.8, 0.99, 0.99);
  out0->Scale(1. / out0->Integral());
  out1->Scale(1. / out1->Integral());
 
  //    TCanvas* cCumProbOutput = new TCanvas("trd_elid_cum_prob_output","trd_elid_cum_prob_output", 500,500);
  DrawH1(fhCumProbOutput, list_of("e^{#pm}")("#pi^{#pm}"), kLinear, kLinear, true, 0.8, 0.8, 0.99, 0.99);
 
  //    TCanvas* cRoc = new TCanvas("trd_elid_roc","trd_elid_roc", 500, 500);
  TGraph* rocGraph = CreateRocDiagramm();
  DrawGraph(rocGraph);
  gPad->SetLogy();
 
  TCanvas* cInput = new TCanvas("trd_elid_input", "trd_elid_ann_input", 10, 10, 800, 800);
  cInput->Divide(4, 3);
  for (int i = 0; i < fNofTrdLayers; i++) {
    cInput->cd(i + 1);
    DrawH1(list_of(fhInput[0][i])(fhInput[1][i]), list_of("e^{#pm}")("#pi^{#pm}"), kLinear, kLog, true, 0.8, 0.8, 0.99,
           0.99);
  }
}
 
ClassImp(CbmTrdElectronsTrainAnn)