AdaptiveVertexFitter.cc

Go to the documentation of this file.

#include "RecoVertex/AdaptiveVertexFit/interface/AdaptiveVertexFitter.h"
#include "DataFormats/GeometryCommonDetAlgo/interface/GlobalError.h"
#include "RecoVertex/VertexTools/interface/AnnealingSchedule.h"
#include "RecoVertex/VertexTools/interface/GeometricAnnealing.h"
#include "RecoVertex/VertexTools/interface/VertexTrackFactory.h"
#include "RecoVertex/VertexPrimitives/interface/VertexException.h"
#include "FWCore/MessageLogger/interface/MessageLogger.h"

#include <algorithm>

using namespace edm;

// #define STORE_WEIGHTS
#ifdef STORE_WEIGHTS
#include <dataharvester/Writer.h>
#endif

using namespace std;

namespace {
    void sortTracksByPt(std::vector<reco::TransientTrack> & cont) {
      auto s = cont.size();
      float pt2[s]; int ind[s]; int i=0; 
      for (auto const & tk : cont) { ind[i]=i; pt2[i++] = tk.impactPointState().globalMomentum().perp2();}
      //clang can not handle lambdas with variable length arrays
      auto * p_pt2 = pt2;
      std::sort(ind,ind+s, [p_pt2](int i, int j){return p_pt2[i]>p_pt2[j];} ); 
      std::vector<reco::TransientTrack> tmp; tmp.reserve(s);
      for (auto i=0U; i<s; ++i) tmp.emplace_back(std::move( cont[ind[i]] ) );
      cont.swap(tmp);
    }

  // typedef ReferenceCountingPointer<VertexTrack<5> > RefCountedVertexTrack;
  typedef AdaptiveVertexFitter::RefCountedVertexTrack RefCountedVertexTrack;


  AlgebraicSymMatrix33 initFitError() {
      // that's how we model the lin pt error for the initial seed!
      const float initialError = 10000;
      AlgebraicSymMatrix33 ret;
      ret(0,0)=initialError;
      ret(1,1)=initialError;
      ret(2,2)=initialError;
      return ret;
  }

  GlobalError const fitError = initFitError();

  AlgebraicSymMatrix33 initLinePointError() {
      // that's how we model the error of the linearization point.
      // for track weighting!
      AlgebraicSymMatrix33 ret;
      ret(0,0)=.3; ret(1,1)=.3; ret(2,2)=3.;
      // ret(0,0)=1e-7; ret(1,1)=1e-7; ret(2,2)=1e-7;
      return ret;
  }

  GlobalError const linPointError = initLinePointError();




  void
  sortByDistanceToRefPoint (std::vector<RefCountedVertexTrack> & cont, const GlobalPoint  ref ) {
      auto s = cont.size();
      float d2[s]; int ind[s]; int i=0;
      for (auto const & tk : cont) { ind[i]=i; d2[i++] = (tk->linearizedTrack()->track().initialFreeState().position() - ref ).mag2();}
      //clang can not handle lambdas with variable length arrays
      auto * p_d2=d2;
      std::sort(ind,ind+s, [p_d2](int i, int j){return p_d2[i]<p_d2[j];} );
      std::vector<RefCountedVertexTrack> tmp; tmp.reserve(s);
      for (auto i=0U; i<s; ++i) tmp.emplace_back(std::move( cont[ind[i]] ) );
      cont.swap(tmp);
  }



  #ifdef STORE_WEIGHTS
  //NOTE: This is not thread safe
  map < RefCountedLinearizedTrackState, int > ids;
  int iter=0;

  int getId ( const RefCountedLinearizedTrackState & r )
  {
    static int ctr=1;
    if ( ids.count(r) == 0 )
    {
      ids[r]=ctr++;
    }
    return ids[r];
  }
  #endif
}

AdaptiveVertexFitter::AdaptiveVertexFitter(
      const AnnealingSchedule & ann,
      const LinearizationPointFinder & linP,
      const VertexUpdator<5> & updator,
      const VertexTrackCompatibilityEstimator<5> & crit,
      const VertexSmoother<5> & smoother,
      const AbstractLTSFactory<5> & ltsf ) :
    theNr(0),
    theLinP(linP.clone()), theUpdator( updator.clone()),
    theSmoother ( smoother.clone() ), theAssProbComputer( ann.clone() ),
    theComp ( crit.clone() ), theLinTrkFactory ( ltsf.clone() ),
    gsfIntermediarySmoothing_(false)
{
  setParameters();
}

void AdaptiveVertexFitter::setWeightThreshold ( float w )
{
  theWeightThreshold=w;
}

AdaptiveVertexFitter::AdaptiveVertexFitter
                        (const AdaptiveVertexFitter & o ) :
    theMaxShift ( o.theMaxShift ), theMaxLPShift ( o.theMaxLPShift ),
    theMaxStep ( o.theMaxStep ), theWeightThreshold ( o.theWeightThreshold ),
    theNr ( o.theNr ),
    theLinP ( o.theLinP->clone() ), theUpdator ( o.theUpdator->clone() ),
    theSmoother ( o.theSmoother->clone() ),
    theAssProbComputer ( o.theAssProbComputer->clone() ),
    theComp ( o.theComp->clone() ),
    theLinTrkFactory ( o.theLinTrkFactory->clone() ),
    gsfIntermediarySmoothing_(o.gsfIntermediarySmoothing_)
{}

AdaptiveVertexFitter::~AdaptiveVertexFitter()
{
  delete theLinP;
  delete theUpdator;
  delete theSmoother;
  delete theAssProbComputer;
  delete theComp;
  delete theLinTrkFactory;
}

void AdaptiveVertexFitter::setParameters( double maxshift, double maxlpshift, 
                                          unsigned maxstep, double weightthreshold )
{
  theMaxShift = maxshift;
  theMaxLPShift = maxlpshift;
  theMaxStep = maxstep;
  theWeightThreshold=weightthreshold;
}


void AdaptiveVertexFitter::setParameters ( const edm::ParameterSet & s )
{
    setParameters ( s.getParameter<double>("maxshift"),
                    s.getParameter<double>("maxlpshift"),
                    s.getParameter<int>("maxstep"),
                    s.getParameter<double>("weightthreshold") );
}

CachingVertex<5>
AdaptiveVertexFitter::vertex(const vector<reco::TransientTrack> & unstracks) const
{
  if ( unstracks.size() < 2 )
  {
    LogError("RecoVertex|AdaptiveVertexFitter")
      << "Supplied fewer than two tracks. Vertex is invalid.";
    return CachingVertex<5>(); // return invalid vertex
  };
  vector < reco::TransientTrack > tracks = unstracks;
  sortTracksByPt( tracks);
  // Linearization Point
  GlobalPoint linP = theLinP->getLinearizationPoint(tracks);
  // Initial vertex seed, with a very large error matrix
  VertexState lseed (linP, linPointError );
  vector<RefCountedVertexTrack> vtContainer = linearizeTracks(tracks, lseed);

  VertexState seed (linP, fitError );
  return fit(vtContainer, seed, false);
}

CachingVertex<5>
AdaptiveVertexFitter::vertex(const vector<RefCountedVertexTrack> & tracks) const
{
  if ( tracks.size() < 2 )
  {
    LogError("RecoVertex|AdaptiveVertexFitter")
      << "Supplied fewer than two tracks. Vertex is invalid.";
    return CachingVertex<5>(); // return invalid vertex
  };
  // Initial vertex seed, with a very small weight matrix
  GlobalPoint linP = tracks[0]->linearizedTrack()->linearizationPoint();
  VertexState seed (linP, fitError );
  return fit(tracks, seed, false);
}

CachingVertex<5>
AdaptiveVertexFitter::vertex(const vector<RefCountedVertexTrack> & tracks, const reco::BeamSpot & spot ) const
{
  if ( tracks.empty() )
  {
    LogError("RecoVertex|AdaptiveVertexFitter")
      << "Supplied no tracks. Vertex is invalid.";
    return CachingVertex<5>(); // return invalid vertex
  };
  VertexState beamSpotState(spot);
  return fit(tracks, beamSpotState, true );
}



CachingVertex<5>
AdaptiveVertexFitter::vertex(const vector<reco::TransientTrack> & tracks,
                  const GlobalPoint& linPoint) const
{
  if ( tracks.size() < 2 )
  {
    LogError("RecoVertex|AdaptiveVertexFitter")
      << "Supplied fewer than two tracks. Vertex is invalid.";
    return CachingVertex<5>(); // return invalid vertex
  };
  // Initial vertex seed, with a very large error matrix
  VertexState seed (linPoint, linPointError );
  vector<RefCountedVertexTrack> vtContainer = linearizeTracks(tracks, seed);
  VertexState fitseed (linPoint, fitError );
  return fit(vtContainer, fitseed, false);
}


CachingVertex<5> 
AdaptiveVertexFitter::vertex(const vector<reco::TransientTrack> & unstracks,
                   const reco::BeamSpot& beamSpot) const
{
  if ( unstracks.empty() )
  {
    LogError("RecoVertex|AdaptiveVertexFitter")
      << "Supplied no tracks. Vertex is invalid.";
    return CachingVertex<5>(); // return invalid vertex
  };

  VertexState beamSpotState(beamSpot);
  vector<RefCountedVertexTrack> vtContainer;

  vector < reco::TransientTrack > tracks = unstracks;
  sortTracksByPt(tracks);

  if (tracks.size() > 1) {
    // Linearization Point search if there are more than 1 track
    GlobalPoint linP = theLinP->getLinearizationPoint(tracks);
    VertexState lpState(linP, linPointError );
    vtContainer = linearizeTracks(tracks, lpState);
  } else {
    // otherwise take the beamspot position.
    vtContainer = linearizeTracks(tracks, beamSpotState);
  }

  return fit(vtContainer, beamSpotState, true);
}


CachingVertex<5> AdaptiveVertexFitter::vertex(const vector<reco::TransientTrack> & tracks,
                  const GlobalPoint& priorPos,
                  const GlobalError& priorError) const

{
  if ( tracks.empty() )
  {
    LogError("RecoVertex|AdaptiveVertexFitter")
      << "Supplied no tracks. Vertex is invalid.";
    return CachingVertex<5>(); // return invalid vertex
  };
  VertexState seed (priorPos, priorError);
  vector<RefCountedVertexTrack> vtContainer = linearizeTracks(tracks, seed);
  return fit( vtContainer, seed, true );
}


CachingVertex<5> AdaptiveVertexFitter::vertex(
                const vector<RefCountedVertexTrack> & tracks,
                  const GlobalPoint& priorPos,
                  const GlobalError& priorError) const
{
  if ( tracks.empty() )
  {
    LogError("RecoVertex|AdaptiveVertexFitter")
      << "Supplied no tracks. Vertex is invalid.";
    return CachingVertex<5>(); // return invalid vertex
  };
  VertexState seed (priorPos, priorError);
  return fit(tracks, seed, true);
}


vector<AdaptiveVertexFitter::RefCountedVertexTrack>
AdaptiveVertexFitter::linearizeTracks(const vector<reco::TransientTrack> & tracks,
                                      const VertexState & seed ) const
{
  const GlobalPoint & linP ( seed.position() );
  vector<RefCountedLinearizedTrackState> lTracks;
  for(vector<reco::TransientTrack>::const_iterator i = tracks.begin();
      i != tracks.end(); ++i )
  {
    try {
      RefCountedLinearizedTrackState lTrData
        = theLinTrkFactory->linearizedTrackState(linP, *i);
      lTracks.push_back(lTrData);
    } catch ( exception & e ) {
      LogInfo("RecoVertex/AdaptiveVertexFitter") 
        << "Exception " << e.what() << " in ::linearizeTracks."
        << "Your future vertex has just lost a track.";
    };
  }
  return weightTracks(lTracks, seed );
}

vector<AdaptiveVertexFitter::RefCountedVertexTrack>
AdaptiveVertexFitter::reLinearizeTracks(
                                const vector<RefCountedVertexTrack> & tracks,
                                const CachingVertex<5> & vertex ) const
{
  const VertexState& seed = vertex.vertexState();
  GlobalPoint linP = seed.position();
  vector<RefCountedLinearizedTrackState> lTracks;
  for(vector<RefCountedVertexTrack>::const_iterator i = tracks.begin();
    i != tracks.end(); i++)
  {
    try {
      RefCountedLinearizedTrackState lTrData
        = theLinTrkFactory->linearizedTrackState( linP, (**i).linearizedTrack()->track() );
      /*
      RefCountedLinearizedTrackState lTrData =
              (**i).linearizedTrack()->stateWithNewLinearizationPoint(linP);
              */
      lTracks.push_back(lTrData);
    } catch ( exception & e ) {
      LogInfo("RecoVertex/AdaptiveVertexFitter") 
        << "Exception " << e.what() << " in ::relinearizeTracks. "
        << "Will not relinearize this track.";
      lTracks.push_back ( (**i).linearizedTrack() );
    };
  };
  return reWeightTracks(lTracks, vertex );
}

AdaptiveVertexFitter * AdaptiveVertexFitter::clone() const
{
  return new AdaptiveVertexFitter( * this );
}

double AdaptiveVertexFitter::getWeight ( float chi2 ) const
{
  double weight = theAssProbComputer->weight(chi2);

  if ( weight > 1.0 )
  {
    LogInfo("RecoVertex/AdaptiveVertexFitter") << "Weight " << weight << " > 1.0!";
    weight=1.0;
  };

  if ( weight < 1e-20 )
  {
    // LogInfo("RecoVertex/AdaptiveVertexFitter") << "Weight " << weight << " < 0.0!";
    weight=1e-20;
  };
  return weight;
}

vector<AdaptiveVertexFitter::RefCountedVertexTrack>
AdaptiveVertexFitter::reWeightTracks(
                    const vector<RefCountedLinearizedTrackState> & lTracks,
                    const CachingVertex<5> & vertex ) const
{
  const VertexState& seed = vertex.vertexState();
  // cout << "[AdaptiveVertexFitter] now reweight around " << seed.position() << endl;
  theNr++;
  // GlobalPoint pos = seed.position();

  vector<RefCountedVertexTrack> finalTracks;
  VertexTrackFactory<5> vTrackFactory;
  #ifdef STORE_WEIGHTS
  iter++;
  #endif
  for(vector<RefCountedLinearizedTrackState>::const_iterator i
        = lTracks.begin(); i != lTracks.end(); i++)
  {
    double weight=0.;
    // cout << "[AdaptiveVertexFitter] estimate " << endl;
    pair < bool, double > chi2Res ( false, 0. );
    try {
      chi2Res =  theComp->estimate ( vertex, *i, std::distance(lTracks.begin(),i) );
    } catch ( exception const & e ) {};
    // cout << "[AdaptiveVertexFitter] /estimate " << endl;
    if (!chi2Res.first) {
      // cout << "[AdaptiveVertexFitter] aie... vertex candidate is at  " << vertex.position() << endl;
      LogInfo("AdaptiveVertexFitter" ) << "When reweighting, chi2<0. Will add this track with w=0.";
      // edm::LogInfo("AdaptiveVertexFitter" ) << "pt=" << (**i).track().pt();
    }else {
      weight = getWeight ( chi2Res.second );
    }
    
    RefCountedVertexTrack vTrData
       = vTrackFactory.vertexTrack(*i, seed, weight );

    #ifdef STORE_WEIGHTS
    map < string, dataharvester::MultiType > m;
    m["chi2"]=chi2;
    m["w"]=theAssProbComputer->weight(chi2);
    m["T"]=theAssProbComputer->currentTemp();
    m["n"]=iter;
    m["pos"]="reweight";
    m["id"]=getId ( *i );
    dataharvester::Writer::file("w.txt").save ( m );
    #endif

    finalTracks.push_back(vTrData);
  }
  sortByDistanceToRefPoint( finalTracks, vertex.position() );
  // cout << "[AdaptiveVertexFitter] /now reweight" << endl;
  return finalTracks;
}

vector<AdaptiveVertexFitter::RefCountedVertexTrack>
AdaptiveVertexFitter::weightTracks(
                    const vector<RefCountedLinearizedTrackState> & lTracks,
                    const VertexState & seed ) const
{
  theNr++;
  CachingVertex<5> seedvtx ( seed, vector<RefCountedVertexTrack> (), 0. );
  theAssProbComputer->resetAnnealing();

  vector<RefCountedVertexTrack> finalTracks;
  VertexTrackFactory<5> vTrackFactory;
  #ifdef STORE_WEIGHTS
  iter++;
  #endif
  for(vector<RefCountedLinearizedTrackState>::const_iterator i
        = lTracks.begin(); i != lTracks.end(); i++)
  {

    double weight = 0.;
    pair<bool, double> chi2Res = theComp->estimate ( seedvtx, *i, std::distance(lTracks.begin(),i) );
    if (!chi2Res.first) {
      // cout << "[AdaptiveVertexFitter] Aiee! " << endl;
      LogInfo ("AdaptiveVertexFitter" ) << "When weighting a track, chi2 calculation failed;"
                                           << " will add with w=0.";
    } else {
      weight = getWeight ( chi2Res.second );
    }
    RefCountedVertexTrack vTrData
       = vTrackFactory.vertexTrack(*i, seed, weight );
    #ifdef STORE_WEIGHTS
    map < string, dataharvester::MultiType > m;
    m["chi2"]=chi2;
    m["w"]=theAssProbComputer->weight(chi2);
    m["T"]=theAssProbComputer->currentTemp();
    m["n"]=iter;
    m["id"]=getId ( *i );
    m["pos"]="weight";
    dataharvester::Writer::file("w.txt").save ( m );
    #endif
    finalTracks.push_back(vTrData);
  }
  return finalTracks;
}

vector<AdaptiveVertexFitter::RefCountedVertexTrack>
AdaptiveVertexFitter::reWeightTracks(
                            const vector<RefCountedVertexTrack> & tracks,
                            const CachingVertex<5> & seed) const
{
  vector<RefCountedLinearizedTrackState> lTracks;
  for(vector<RefCountedVertexTrack>::const_iterator i = tracks.begin();
    i != tracks.end(); i++)
  {
    lTracks.push_back((**i).linearizedTrack());
  }

  return reWeightTracks(lTracks, seed);
}


/*
 * The method where the vertex fit is actually done!
 */

CachingVertex<5>
AdaptiveVertexFitter::fit( const vector<RefCountedVertexTrack> & tracks,
                          const VertexState & priorSeed,
                          bool withPrior) const
{
  // cout << "[AdaptiveVertexFit] fit with " << tracks.size() << endl;
  theAssProbComputer->resetAnnealing();

  vector<RefCountedVertexTrack> initialTracks;
  GlobalPoint priorVertexPosition = priorSeed.position();
  GlobalError priorVertexError = priorSeed.error();

  CachingVertex<5> returnVertex( priorVertexPosition,priorVertexError,
                              initialTracks,0);
  if (withPrior)
  {
    returnVertex = CachingVertex<5>(priorVertexPosition,priorVertexError,
                    priorVertexPosition,priorVertexError,initialTracks,0);
  }

  std::vector<RefCountedVertexTrack> globalVTracks = tracks;
  // sort the tracks, according to distance to seed!
  sortByDistanceToRefPoint (globalVTracks, priorSeed.position() );

  // main loop through all the VTracks
  int lpStep = 0; int step = 0;

  CachingVertex<5> initialVertex = returnVertex;

  GlobalPoint newPosition = priorVertexPosition;
  GlobalPoint previousPosition = newPosition;

  int ns_trks=0; // number of significant tracks.
  // If we have only two significant tracks, we return an invalid vertex

  // cout << "[AdaptiveVertexFit] start " << tracks.size() << endl;
  /*
  for ( vector< RefCountedVertexTrack >::const_iterator 
        i=globalVTracks.begin(); i!=globalVTracks.end() ; ++i )
  {
    cout << "  " << (**i).linearizedTrack()->track().initialFreeState().momentum() << endl;
  }*/
  do {
    ns_trks=0;
    CachingVertex<5> fVertex = initialVertex;
    // cout << "[AdaptiveVertexFit] step " << step << " at " << fVertex.position() << endl;
    if ((previousPosition - newPosition).transverse() > theMaxLPShift)
    {
      // relinearize and reweight.
      // (reLinearizeTracks also reweights tracks)
      // cout << "[AdaptiveVertexFit] relinearize at " << returnVertex.position() << endl;
      if (gsfIntermediarySmoothing_) returnVertex = theSmoother->smooth(returnVertex);
      globalVTracks = reLinearizeTracks( globalVTracks, returnVertex );
      lpStep++;
    } else if (step) {
      // reweight, if it is not the first step
      // cout << "[AdaptiveVertexFit] reweight at " << returnVertex.position() << endl;
      if (gsfIntermediarySmoothing_) returnVertex = theSmoother->smooth(returnVertex);
      globalVTracks = reWeightTracks( globalVTracks, returnVertex );
    }
    // cout << "[AdaptiveVertexFit] relinarized, reweighted" << endl;
    // update sequentially the vertex estimate
    CachingVertex<5> nVertex;
    for(vector<RefCountedVertexTrack>::const_iterator i
          = globalVTracks.begin(); i != globalVTracks.end(); i++)
    {
      if ((**i).weight() > 0.) nVertex = theUpdator->add( fVertex, *i );
    else  nVertex = fVertex;
      if (nVertex.isValid()) {
        if ( (**i).weight() >= theWeightThreshold ) ns_trks++;
        
        if ( fabs ( nVertex.position().z() ) > 10000. ||
             nVertex.position().perp()>120.)
        {
          // were more than 100 m off!!
          LogInfo ("AdaptiveVertexFitter" ) << "Vertex candidate just took off to " << nVertex.position()
                        << "! Will discard this update!";
//      //<< "track pt was " << (**i).linearizedTrack()->track().pt()
//                       << "track momentum was " << (**i).linearizedTrack()->track().initialFreeState().momentum()
//                       << "track position was " << (**i).linearizedTrack()->track().initialFreeState().position()
//                       << "track chi2 was " << (**i).linearizedTrack()->track().chi2()
//                       << "track ndof was " << (**i).linearizedTrack()->track().ndof()
//                       << "track w was " << (**i).weight()
//                       << "track schi2 was " << (**i).smoothedChi2();
        } else {
            fVertex = nVertex;
        }
      } else {
        LogInfo("RecoVertex/AdaptiveVertexFitter") 
          << "The updator returned an invalid vertex when adding track "
          << i-globalVTracks.begin() 
            << ".\n Your vertex might just have lost one good track.";
      }
    }
    previousPosition = newPosition;
    newPosition = fVertex.position();
    returnVertex = fVertex;
    theAssProbComputer->anneal();
    step++;
    if ( step >= theMaxStep ) break;

  } while (
      // repeat as long as
      // - vertex moved too much or
      // - we're not yet annealed
         ( ((previousPosition - newPosition).mag() > theMaxShift) ||
           (!(theAssProbComputer->isAnnealed()) ) ) ) ;

  if ( theWeightThreshold > 0. &&  ns_trks < 2 && !withPrior ) 
  {
    LogDebug("AdaptiveVertexFitter") 
      << "fewer than two significant tracks (w>" << theWeightThreshold << ")."
      << " Fitted vertex is invalid.";
    return CachingVertex<5>(); // return invalid vertex
  }

  #ifdef STORE_WEIGHTS
  map < string, dataharvester::MultiType > m;
  m["chi2"]=chi2;
  m["w"]=theAssProbComputer->weight(chi2);
  m["T"]=theAssProbComputer->currentTemp();
  m["n"]=iter;
  m["id"]=getId ( *i );
  m["pos"]="final";
  dataharvester::Writer::file("w.txt").save ( m );
  #endif
  // cout << "[AdaptiveVertexFit] /fit" << endl;
  return theSmoother->smooth( returnVertex );
}