NeuroTrigger.cc

/**************************************************************************
 * basf2 (Belle II Analysis Software Framework)                           *
 * Author: The Belle II Collaboration                                     *
 *                                                                        *
 * See git log for contributors and copyright holders.                    *
 * This file is licensed under LGPL-3.0, see LICENSE.md.                  *
 **************************************************************************/
#include <framework/logging/Logger.h>
#include <trg/cdc/NeuroTrigger.h>
#include <trg/cdc/NeuroTriggerParameters.h>
#include <trg/cdc/dataobjects/CDCTriggerMLPData.h>
#include <cdc/geometry/CDCGeometryPar.h>
#include <framework/gearbox/Const.h>
#include <framework/gearbox/Unit.h>
#include <framework/datastore/StoreObjPtr.h>
#include <framework/datastore/StoreArray.h>
#include <trg/cdc/dbobjects/CDCTriggerNeuroConfig.h>
#include <trg/cdc/dataobjects/CDCTriggerTrack.h>
#include <string>
#include <cmath>
#include <TFile.h>
#include "boost/iostreams/filter/gzip.hpp"
#include "boost/iostreams/filtering_streambuf.hpp"
#include "boost/iostreams/filtering_stream.hpp"
#include "boost/multi_array.hpp"
#define BOOST_MULTI_ARRAY_NO_GENERATORS
 
 
 
using namespace Belle2;
using namespace CDC;
using namespace std;
 
void
NeuroTrigger::initialize(const NeuroTriggerParameters& p)
{
  // check parameters
  bool okay = true;
  // ensure that length of lists matches number of sectors
  if (p.nHidden.size() != 1 && p.nHidden.size() != p.nMLP) {
    B2ERROR("Number of nHidden lists should be 1 or " << p.nMLP);
    okay = false;
  }
  if (p.outputScale.size() != 1 && p.outputScale.size() != p.nMLP) {
    B2ERROR("Number of outputScale lists should be 1 or " << p.nMLP);
    okay = false;
  }
  bool rangeProduct = (p.phiRangeUse.size() * p.invptRangeUse.size() * p.thetaRangeUse.size() * p.SLpattern.size() == p.nMLP);
  if (!rangeProduct) {
    if (p.phiRangeUse.size() != 1 && p.phiRangeUse.size() != p.nMLP) {
      B2ERROR("Number of phiRangeUse.lists should be 1 or " << p.nMLP);
      okay = false;
    }
    if (p.invptRangeUse.size() != 1 && p.invptRangeUse.size() != p.nMLP) {
      B2ERROR("Number of invptRangeUse.lists should be 1 or " << p.nMLP);
      okay = false;
    }
    if (p.thetaRangeUse.size() != 1 && p.thetaRangeUse.size() != p.nMLP) {
      B2ERROR("Number of thetaRangeUse.lists should be 1 or " << p.nMLP);
      okay = false;
    }
    if (p.SLpattern.size() != 1 && p.SLpattern.size() != p.nMLP) {
      B2ERROR("Number of SLpattern lists should be 1 or " << p.nMLP);
      okay = false;
    }
  }
  // ensure that length of maxHitsPerSL and SLpatternMask lists matches SLpattern
  if (p.maxHitsPerSL.size() != 1 && p.maxHitsPerSL.size() != p.SLpattern.size()) {
    B2ERROR("Number of maxHitsPerSL lists should be 1 or " << p.SLpattern.size());
    okay = false;
  }
  if (p.SLpatternMask.size() != 1 && p.SLpatternMask.size() != p.SLpattern.size()) {
    B2ERROR("Number of SLpatternMask lists should be 1 or " << p.SLpattern.size());
    okay = false;
  }
  // ensure that number of target nodes is valid
  if (p.targetZ.isSet() || p.targetTheta.isSet()) {
    unsigned short nTarget = int(p.targetZ) + int(p.targetTheta);
    if (nTarget < 1) {
      B2ERROR("No outputs! Turn on either targetZ or targetTheta.");
      okay = false;
    }
  }
  // ensure that sector ranges are valid
  for (unsigned iPhi = 0; iPhi < p.phiRangeUse.size(); ++iPhi) {
    if (p.phiRangeUse[iPhi].size() != 2) {
      B2ERROR("phiRangeUse should be exactly 2 values");
      okay = false;
      continue;
    }
    if (p.phiRangeUse[iPhi][0] >= p.phiRangeUse[iPhi][1]) {
      B2ERROR("phiRangeUse should be smaller than phiRangeUse");
      okay = false;
    }
    if (p.phiRangeUse[iPhi][0] < -360. || p.phiRangeUse[iPhi][1] > 360. ||
        (p.phiRangeUse[iPhi][1] - p.phiRangeUse[iPhi][0]) > 360.) {
      B2ERROR("phiRangeUse should be in [-360, 360], with maximal width of 360");
      okay = false;
    }
  }
  for (unsigned iPt = 0; iPt < p.invptRangeUse.size(); ++iPt) {
    if (p.invptRangeUse[iPt].size() != 2) {
      B2ERROR("invptRangeUse should be exactly 2 values");
      okay = false;
    }
    if (p.invptRangeUse[iPt][0] >= p.invptRangeUse[iPt][1]) {
      B2ERROR("invptRangeUse should be smaller than invptRangeUse");
      okay = false;
    }
  }
  for (unsigned iTheta = 0; iTheta < p.thetaRangeUse.size(); ++iTheta) {
    if (p.thetaRangeUse[iTheta].size() != 2) {
      B2ERROR("thetaRangeUse should be exactly 2 values");
      okay = false;
      continue;
    }
    if (p.thetaRangeUse[iTheta][0] >= p.thetaRangeUse[iTheta][1]) {
      B2ERROR("thetaRangeUse should be smaller than thetaRangeUse");
      okay = false;
    }
    if (p.thetaRangeUse[iTheta][0] < 0. || p.thetaRangeUse[iTheta][1] > 180.) {
      B2ERROR("thetaRangeUse should be in [0, 180]");
      okay = false;
    }
  }
  int nTarget = (int) p.targetZ + (int) p.targetTheta;
  if (p.outputScale.size() == p.nMLP || p.outputScale.size() == 1) {
    for (unsigned iScale = 0; iScale < p.outputScale.size(); ++iScale) {
      if (p.outputScale[iScale].size() != 2 * nTarget) {
        B2ERROR("outputScale should be exactly " << 2 * nTarget << " values");
        okay = false;
      }
    }
  } else {
    B2ERROR("the size of outputscale should be 1 or match the number of experts");
  }
  // ensure that train sectors are valid
  if (p.phiRangeUse.size() != p.phiRangeTrain.size()) {
    B2ERROR("Number of phiRangeUse.lists and phiRangeTrain lists should be equal.");
    okay = false;
  } else {
    for (unsigned iPhi = 0; iPhi < p.phiRangeUse.size(); ++iPhi) {
      if (p.phiRangeTrain[iPhi].size() != 2) {
        B2ERROR("phiRangeTrain should be exactly 2 values.");
        okay = false;
      } else if (p.phiRangeTrain[iPhi][0] > p.phiRangeUse[iPhi][0] ||
                 p.phiRangeTrain[iPhi][1] < p.phiRangeUse[iPhi][1]) {
        B2ERROR("phiRangeTrain should be wider than phiRangeUse.or equal.");
        okay = false;
      }
    }
  }
  if (p.invptRangeUse.size() != p.invptRangeTrain.size()) {
    B2ERROR("Number of invptRangeUse.lists and invptRangeTrain lists should be equal.");
    okay = false;
  } else {
    for (unsigned iPt = 0; iPt < p.invptRangeUse.size(); ++iPt) {
      if (p.invptRangeTrain[iPt].size() != 2) {
        B2ERROR("invptRangeTrain should be exactly 2 values.");
        okay = false;
      } else if (p.invptRangeTrain[iPt][0] > p.invptRangeUse[iPt][0] ||
                 p.invptRangeTrain[iPt][1] < p.invptRangeUse[iPt][1]) {
        B2ERROR("invptRangeTrain should be wider than invptRangeUse.or equal.");
        okay = false;
      }
    }
  }
  if (p.thetaRangeUse.size() != p.thetaRangeTrain.size()) {
    B2ERROR("Number of thetaRangeUse.lists and thetaRangeTrain lists should be equal.");
    okay = false;
  } else {
    for (unsigned iTheta = 0; iTheta < p.thetaRangeUse.size(); ++iTheta) {
      if (p.thetaRangeTrain[iTheta].size() != 2) {
        B2ERROR("thetaRangeTrain should be exactly 2 values.");
        okay = false;
      } else if (p.thetaRangeTrain[iTheta][0] > p.thetaRangeUse[iTheta][0] ||
                 p.thetaRangeTrain[iTheta][1] < p.thetaRangeUse[iTheta][1]) {
        B2ERROR("thetaRangeTrain should be wider than thetaRangeUse.or equal.");
        okay = false;
      }
    }
  }
 
  if (!okay) return;
 
  // initialize MLPs
  m_MLPs.clear();
  for (unsigned iMLP = 0; iMLP < p.nMLP; ++iMLP) {
    //get indices for sector parameters
    //this is important for cases, where we have experts specialized on different geometrical sectors as well as the pattern mask. since they are all in one array, we need the specific index of the expert. E.g. p.maxhitspersl cloud look like: [<expert-trained-on-slpattern0+thetabigger90>,<expert-trained-on-slpattern1+thetabigger90>,<expert-trained-on-slpattern0+thetasmaller90>,<expert-trained-on-slpattern1+thetasmaller90>]
    vector<unsigned> indices = getRangeIndices(p, iMLP);
    //get number of nodes for each layer
    unsigned short maxHits = (p.maxHitsPerSL.size() == 1) ? p.maxHitsPerSL[0] : p.maxHitsPerSL[indices[3]];
    unsigned long SLpattern = p.SLpattern[indices[3]];
    unsigned long SLpatternMask = (p.SLpatternMask.size() == 1) ? p.SLpatternMask[0] : p.SLpatternMask[indices[3]];
    unsigned short nInput = 27 * maxHits;
    vector<NNTParam<float>> nHidden = (p.nHidden.size() == 1) ? p.nHidden[0] : p.nHidden[iMLP];
    vector<unsigned short> nNodes = {nInput};
    for (unsigned iHid = 0; iHid < nHidden.size(); ++iHid) {
      if (p.multiplyHidden) {
        nNodes.push_back(nHidden[iHid] * nNodes[0]);
      } else {
        nNodes.push_back(nHidden[iHid]);
      }
    }
    nNodes.push_back(nTarget);
    unsigned short targetVars = int(p.targetZ) + (int(p.targetTheta) << 1);
    // the parameters stored in the parameterset are not advanced enough to be vectors, they can only be single data types. the workaround was to make every variable contained in the (nested) vector an NNTParam. for the further use, those have to be converted to float vecors, which is done by the tcastvector function.
    vector<float> phiRangeUse = p.tcastvector<float>(p.phiRangeUse)[indices[0]];
    vector<float> invptRangeUse = p.tcastvector<float>(p.invptRangeUse)[indices[1]];
    vector<float> thetaRangeUse = p.tcastvector<float>(p.thetaRangeUse)[indices[2]];
    vector<float> phiRangeTrain = p.tcastvector<float>(p.phiRangeTrain)[indices[0]];
    vector<float> invptRangeTrain = p.tcastvector<float>(p.invptRangeTrain)[indices[1]];
    vector<float> thetaRangeTrain = p.tcastvector<float>(p.thetaRangeTrain)[indices[2]];
    B2DEBUG(50, "Ranges for sector " << iMLP
            << ": phiRange [" << phiRangeUse[0] << ", " << phiRangeUse[1]
            << "], invptRange [" << invptRangeUse[0] << ", " << invptRangeUse[1]
            << "], thetaRange [" << thetaRangeUse[0] << ", " << thetaRangeUse[1]
            << "], SLpattern " << SLpattern);
    //get scaling values
    vector<float> outputScale = (p.outputScale.size() == 1) ? p.tcastvector<float>(p.outputScale)[0] : p.tcastvector<float>
                                (p.outputScale)[iMLP];
    //convert phi and theta from degree to radian
    phiRangeUse[0] *= Unit::deg;
    phiRangeUse[1] *= Unit::deg;
    thetaRangeUse[0] *= Unit::deg;
    thetaRangeUse[1] *= Unit::deg;
    phiRangeTrain[0] *= Unit::deg;
    phiRangeTrain[1] *= Unit::deg;
    thetaRangeTrain[0] *= Unit::deg;
    thetaRangeTrain[1] *= Unit::deg;
    if (p.targetTheta) {
      outputScale[2 * int(p.targetZ)] *= Unit::deg;
      outputScale[2 * int(p.targetZ) + 1] *= Unit::deg;
    }
    //create new MLP
    m_MLPs.push_back(CDCTriggerMLP(nNodes, targetVars, outputScale,
                                   phiRangeUse, invptRangeUse, thetaRangeUse,
                                   phiRangeTrain, invptRangeTrain, thetaRangeTrain,
                                   maxHits, SLpattern, SLpatternMask, p.tMax,
                                   p.et_option()));
  }
 
  if (p.IDRanges.size() == p.nMLP) {
    for (auto exp : p.IDRanges) {
      // first entry is the expert number, after that follow the idranges for all the superlayers
      std::vector<float> irange = {exp.begin() + 1, exp.end()};
      m_MLPs[static_cast<int>(exp[0])].setRelID(irange);
    }
  } else if (p.IDRanges.size() == 0) {
    B2WARNING("idranges have not been initialized yet, did you forget it?");
  } else {
    B2ERROR("number of idranges should match the number of experts!");
  }
  // load some values from the geometry that will be needed for the input
  setConstants();
}
 
vector<unsigned>
NeuroTrigger::getRangeIndices(const NeuroTriggerParameters& p, unsigned isector)
{
  // the indices can be used for both rangeuse and rangetrain, because the size of those arrays should be the same (it is checked in the initialize function).
  std::vector<unsigned> indices = {0, 0, 0, 0};
  if (p.phiRangeUse.size() * p.invptRangeUse.size() * p.thetaRangeUse.size() * p.SLpattern.size() == p.nMLP) {
    indices[0] = isector % p.phiRangeUse.size();
    indices[1] = (isector / p.phiRangeUse.size()) % p.invptRangeUse.size();
    indices[2] = (isector / (p.phiRangeUse.size() * p.invptRangeUse.size())) % p.thetaRangeUse.size();
    indices[3] = isector / (p.phiRangeUse.size() * p.invptRangeUse.size() * p.thetaRangeUse.size());
  } else {
    indices[0] = (p.phiRangeUse.size() == 1) ? 0 : isector;
    indices[1] = (p.invptRangeUse.size() == 1) ? 0 : isector;
    indices[2] = (p.thetaRangeUse.size() == 1) ? 0 : isector;
    indices[3] = (p.SLpattern.size() == 1) ? 0 : isector;
  }
  return indices;
}
 
 
void
NeuroTrigger::setConstants()
{
  CDCGeometryPar& cdc = CDCGeometryPar::Instance();
  int layerId = 3;
  int nTS = 0;
  for (int iSL = 0; iSL < 9; ++iSL) {
    m_TSoffset[iSL] = nTS;
    nTS += cdc.nWiresInLayer(layerId);
    m_TSoffset[iSL + 1] = nTS;
    for (int priority = 0; priority < 2; ++ priority) {
      m_radius[iSL][priority] = cdc.senseWireR(layerId + priority);
    }
    layerId += (iSL > 0 ? 6 : 7);
  }
}
 
void
NeuroTrigger::initializeCollections(string hitCollectionName, string eventTimeName, const std::string& et_option)
{
  m_segmentHits.isRequired(hitCollectionName);
  if (!((et_option == "fastestpriority") || (et_option == "etfhwin") || (et_option == "zero") || (et_option == "fastest2d"))) {
    m_eventTime.isRequired(eventTimeName);
  }
  m_hitCollectionName = hitCollectionName;
}
 
void
NeuroTrigger::initializeCollections(string hitCollectionName)
{
  m_segmentHits.isRequired(hitCollectionName);
  m_hitCollectionName = hitCollectionName;
}
vector<int>
NeuroTrigger::selectMLPsTrain(float phi0, float invpt, float theta)
{
  vector<int> indices = {};
 
  if (m_MLPs.size() == 0) {
    B2WARNING("Trying to select MLP before initializing MLPs.");
    return indices;
  }
 
  // find all matching sectors
  for (unsigned isector = 0; isector < m_MLPs.size(); ++isector) {
    if (m_MLPs[isector].inPhiRangeTrain(phi0) && m_MLPs[isector].inInvptRangeTrain(invpt)
        && m_MLPs[isector].inThetaRangeTrain(theta)) {
      indices.push_back(isector);
    }
  }
 
  if (indices.size() == 0) {
    B2DEBUG(150, "Track does not match any sector.");
    B2DEBUG(150, "invpt=" << invpt << ", phi=" << phi0 * 180. / M_PI << ", theta=" << theta * 180. / M_PI);
  }
 
  return indices;
}
vector<int>
NeuroTrigger::selectMLPs(float phi0, float invpt, float theta)
{
  vector<int> indices = {};
 
  if (m_MLPs.size() == 0) {
    B2WARNING("Trying to select MLP before initializing MLPs.");
    return indices;
  }
 
  // find all matching sectors
  for (unsigned isector = 0; isector < m_MLPs.size(); ++isector) {
    if (m_MLPs[isector].inPhiRangeUse(phi0) && m_MLPs[isector].inInvptRangeUse(invpt)
        && m_MLPs[isector].inThetaRangeUse(theta)) {
      indices.push_back(isector);
    }
  }
 
  if (indices.size() == 0) {
    B2DEBUG(150, "Track does not match any sector.");
    B2DEBUG(150, "invpt=" << invpt << ", phi=" << phi0 * 180. / M_PI << ", theta=" << theta * 180. / M_PI);
  }
 
  return indices;
}
 
 
int
NeuroTrigger::selectMLPbyPattern(std::vector<int>& MLPs, unsigned long pattern, const bool neurotrackinputmode)
{
  if (MLPs.size() == 0) {
    return -1;
  }
 
  int bestIndex = -1;
  for (unsigned i = 0; i < MLPs.size(); ++i) {
    int isector = MLPs[i];
    unsigned long sectorPattern = m_MLPs[isector].getSLpattern();
    B2DEBUG(250, "hitPattern " << pattern << " sectorPattern " << sectorPattern);
    // no hit pattern restriction -> keep looking for exact match
    if (sectorPattern == 0) {
      B2DEBUG(250, "found match for general sector");
      bestIndex = isector;
    }
    // exact match -> keep this sector and stop searching
    if (pattern == sectorPattern) {
      B2DEBUG(250, "found match for hit pattern " << pattern);
      bestIndex = isector;
      break;
    }
  }
 
  if (bestIndex < 0) {
    if (neurotrackinputmode) {
      B2DEBUG(150, "No sector found to match pattern, using sector 0" << pattern << ".");
      bestIndex = 0;
    } else {
      B2DEBUG(150, "No sector found to match pattern " << pattern << ".");
    }
  }
  return bestIndex;
}
 
 
void
NeuroTrigger::updateTrack(const CDCTriggerTrack& track)
{
  double omega = track.getOmega(); // signed track curvature
  for (int iSL = 0; iSL < 9; ++iSL) {
    for (int priority = 0; priority < 2; ++priority) {
      m_alpha[iSL][priority] = (m_radius[iSL][priority] * abs(omega) < 2.) ?
                               asin(m_radius[iSL][priority] * omega / 2.) :
                               M_PI_2;
      m_idRef[iSL][priority] = remainder(((track.getPhi0() - m_alpha[iSL][priority]) *
                                          (m_TSoffset[iSL + 1] - m_TSoffset[iSL]) / 2. / M_PI),
                                         (m_TSoffset[iSL + 1] - m_TSoffset[iSL]));
    }
  }
}
 
void
NeuroTrigger::updateTrackFix(const CDCTriggerTrack& track)
{
  //unsigned precisionPhi = m_precision[0];
  //unsigned precisionAlpha = m_precision[0];
  unsigned precisionPhiAlpha = m_precision[0];
  unsigned precisionScale = m_precision[1];
  unsigned precisionId = m_precision[2];
 
  double omega = track.getOmega();
  long phi = round(track.getPhi0() * (1 << precisionPhiAlpha));
 
  for (int iSL = 0; iSL < 9; ++iSL) {
    for (int priority = 0; priority < 2; ++priority) {
      // LUT, calculated on the fly here
      m_alpha[iSL][priority] =
        (m_radius[iSL][priority] * abs(omega) < 2) ?
        round(asin(m_radius[iSL][priority] * omega / 2) * (1 << precisionPhiAlpha)) :
        round(M_PI_2 * (1 << precisionPhiAlpha));
      long dphi = phi - (long(m_alpha[iSL][priority]));
      m_idRef[iSL][priority] =
        long(dphi * round((m_TSoffset[iSL + 1] - m_TSoffset[iSL]) / 2. / M_PI * (1 << precisionScale)) /
             (long(1) << (precisionPhiAlpha + precisionScale - precisionId)));
      // unscale after rounding
      m_alpha[iSL][priority] /= (1 << precisionPhiAlpha);
      m_idRef[iSL][priority] /= (1 << precisionId);
    }
  }
}
 
double
NeuroTrigger::getRelId(const CDCTriggerSegmentHit& hit)
{
  int iSL = hit.getISuperLayer();
  int priority = hit.getPriorityPosition();
  // (((priority >> 1) & 1) - (priority & 1)) is 0, -1, 1, 0 for priority = 0, 1, 2, 3
  double relId = hit.getSegmentID() + 0.5 * (((priority >> 1) & 1) - (priority & 1))
                 - m_TSoffset[iSL] - m_idRef[iSL][int(priority < 3)];
  relId = remainder(relId, (m_TSoffset[iSL + 1] - m_TSoffset[iSL]));
  return relId;
}
 
 
int
NeuroTrigger::getLowestTime(unsigned isector, RelationVector<CDCTriggerSegmentHit> Hits, bool onlyAxials = false)
{
  int tlow = 9999;
  B2DEBUG(200, "looping over axials:");
  for (unsigned ihit = 0; ihit < Hits.size(); ++ihit) {
    // skip hits with negative relation weight (not selected in finder)
    if (Hits.weight(ihit) < 0) continue;
    unsigned short iSL = Hits[ihit]->getISuperLayer();
    if (iSL % 2 == 1 && onlyAxials) {continue;}
    // get shortest time of relevant hits
    B2DEBUG(200, "  check drifttime: SL" + std::to_string(iSL) + ",ID = " + std::to_string(Hits[ihit]->getSegmentID()) + ", t = " +
            std::to_string(Hits[ihit]->priorityTime()));
    double relId = getRelId(*Hits[ihit]);
    if (m_MLPs[isector].isRelevant(relId, iSL) &&
        Hits[ihit]->priorityTime() < tlow) {
      tlow = Hits[ihit]->priorityTime();
      B2DEBUG(200, "    new tlow: " << std::to_string(tlow));
    }
  }
  return tlow;
}
 
 
void
NeuroTrigger::getEventTime(unsigned isector, const CDCTriggerTrack& track, std::string et_option, const bool neuroinputmode = false)
{
 
  if (et_option != m_MLPs[isector].get_et_option()) {
    B2DEBUG(20, "Used event time option is different to the one set in the MLP"
            << LogVar("et_option", et_option) << LogVar("isector", isector)
            << LogVar("et_option_mlp", m_MLPs[isector].get_et_option()));
  }
  if (et_option == "fastestpriority") {
    B2DEBUG(200, "et_option is 'fastestpriority'");
    m_T0 = 9999;
    // find shortest time of related and relevant axial hits
    RelationVector<CDCTriggerSegmentHit> Hits =
      track.getRelationsTo<CDCTriggerSegmentHit>(m_hitCollectionName);
    m_T0 = getLowestTime(isector, Hits, false);
    if (m_T0 < 9999) {
      m_hasT0 = true;
    } else {
      m_T0 = 0;
      m_hasT0 = false;
    }
 
  } else if (et_option == "fastest2d") {
    m_T0 = 9999;
    // find shortest time of related and relevant axial hits
    RelationVector<CDCTriggerSegmentHit> Hits =
      track.getRelationsTo<CDCTriggerSegmentHit>(m_hitCollectionName);
    m_T0 = getLowestTime(isector, Hits, true);
    if (m_T0 < 9999) {
      m_hasT0 = true;
    } else {
      m_T0 = 0;
      m_hasT0 = false;
    }
  } else if (et_option == "zero") {
    m_hasT0 = true;
    m_T0 = 0;
  } else if (et_option == "etf_only") {
    bool hasT0 = (m_eventTime.isValid()) ? m_eventTime->hasBinnedEventT0(Const::CDC) : false;
    if (hasT0) {
      m_T0 = m_eventTime->getBinnedEventT0(Const::CDC);
      m_hasT0 = true;
    } else {
      B2ERROR("No ETF output, but forced to use ETF!");
      m_hasT0 = false;
      m_T0 = 0;
    }
  } else if (et_option == "etf_or_fastestpriority") {
    bool hasT0 = (m_eventTime.isValid()) ? m_eventTime->hasBinnedEventT0(Const::CDC) : false;
    if (hasT0) {
      m_T0 = m_eventTime->getBinnedEventT0(Const::CDC);
      m_hasT0 = true;
    } else {
      getEventTime(isector, track, "fastestpriority", neuroinputmode);
    }
  } else if (et_option == "min_etf_fastestpriority") {
    bool hasT0 = (m_eventTime.isValid()) ? m_eventTime->hasBinnedEventT0(Const::CDC) : false;
    int T0_etf = 9999;
    if (hasT0) {
      T0_etf = m_eventTime->getBinnedEventT0(Const::CDC);
      m_hasT0 = true;
    }
    getEventTime(isector, track, "fastestpriority", neuroinputmode);
    if (m_T0 > T0_etf) {
      m_T0 = T0_etf;
    }
  } else if (et_option == "etf_or_fastest2d") {
    bool hasT0 = (m_eventTime.isValid()) ? m_eventTime->hasBinnedEventT0(Const::CDC) : false;
    if (hasT0) {
      m_T0 = m_eventTime->getBinnedEventT0(Const::CDC);
      m_hasT0 = true;
    } else {
      getEventTime(isector, track, "fastest2d", neuroinputmode);
    }
  } else if (et_option == "etf_or_zero") {
    bool hasT0 = (m_eventTime.isValid()) ? m_eventTime->hasBinnedEventT0(Const::CDC) : false;
    if (hasT0) {
      m_T0 = m_eventTime->getBinnedEventT0(Const::CDC);
      m_hasT0 = true;
    } else {
      m_hasT0 = true;
      m_T0 = 0;
    }
  } else if (et_option == "etfcc") {
    if (track.getHasETFTime()) {
      m_T0 = track.getETF_unpacked();
      m_hasT0 = true;
    } else {
      m_T0 = 0;
      m_hasT0 = false;
    }
 
  } else if (et_option == "etfcc_or_zero") {
    if (track.getHasETFTime()) {
      m_T0 = track.getETF_unpacked();
      m_hasT0 = true;
    } else {
      m_T0 = 0;
      m_hasT0 = true;
    }
 
  } else if (et_option == "etfcc_or_fastestpriority") {
    if (track.getHasETFTime()) {
      m_T0 = track.getETF_unpacked();
      m_hasT0 = true;
    } else {
      getEventTime(isector, track, "fastestpriority", neuroinputmode);
    }
 
  } else if (et_option == "etfhwin") {
    m_T0 = track.getETF_recalced();
    m_hasT0 = true;
 
  } else {
    B2ERROR("No valid parameter for et_option (" << et_option << " )!");
  }
 
}
 
unsigned long
NeuroTrigger::getPureDriftThreshold(unsigned isector, const CDCTriggerTrack& track, const bool neurotrackinputmode)
{
  const CDCTriggerMLP& expert = m_MLPs[isector];
  unsigned long driftth = 0;
  vector<unsigned> nHits;
  nHits.assign(9, 0);
  // loop over axial hits related to input track
  RelationVector<CDCTriggerSegmentHit> axialHits =
    track.getRelationsTo<CDCTriggerSegmentHit>(m_hitCollectionName);
  for (unsigned ihit = 0; ihit < axialHits.size(); ++ ihit) {
    // skip hits with negative relation weight (not selected in finder)
    if (axialHits.weight(ihit) < 0) {
      continue;
    }
    unsigned short iSL = axialHits[ihit]->getISuperLayer();
    // // skip stereo hits (should not be related to track, but check anyway)
    if ((!neurotrackinputmode) && (iSL % 2 == 1)) continue;
    // get priority time
    int t = (m_hasT0) ? axialHits[ihit]->priorityTime() - m_T0 : 0;
    double relId = getRelId(*axialHits[ihit]);
    if (t < 0 || t > expert.getTMax()) {
      if (expert.isRelevant(relId, iSL)) {
        if (nHits[iSL] < expert.getMaxHitsPerSL()) {
          driftth |= 1 << (8 - iSL + 9 * nHits[iSL]);
          ++nHits[iSL];
        }
      }
    }
  }
  if (!neurotrackinputmode) {
    // loop over stereo hits
    for (int ihit = 0; ihit < m_segmentHits.getEntries(); ++ ihit) {
      unsigned short iSL = m_segmentHits[ihit]->getISuperLayer();
      // skip axial hits
      if (iSL % 2 == 0) continue;
      // get priority time
      int t = (m_hasT0) ? m_segmentHits[ihit]->priorityTime() - m_T0 : 0;
      double relId = getRelId(*m_segmentHits[ihit]);
      if (t < 0 || t > expert.getTMax()) {
        if (expert.isRelevant(relId, iSL)) {
          if (nHits[iSL] < expert.getMaxHitsPerSL()) {
            driftth |= 1 << (8 - iSL + 9 * nHits[iSL]);
            ++nHits[iSL];
          }
        }
      }
    }
  }
  return driftth;
}
unsigned long NeuroTrigger::getCompleteHitPattern(unsigned isector, const CDCTriggerTrack& track, const bool neurotrackinputmode)
{
  const CDCTriggerMLP& expert = m_MLPs[isector];
  unsigned long chitPattern = 0;
  vector<unsigned> nHits;
  nHits.assign(9, 0);
  // loop over axial hits related to input track
  RelationVector<CDCTriggerSegmentHit> axialHits =
    track.getRelationsTo<CDCTriggerSegmentHit>(m_hitCollectionName);
  for (unsigned ihit = 0; ihit < axialHits.size(); ++ ihit) {
    // skip hits with negative relation weight (not selected in finder)
    if (axialHits.weight(ihit) < 0) {
      continue;
    }
    unsigned short iSL = axialHits[ihit]->getISuperLayer();
    // // skip stereo hits (should not be related to track, but check anyway)
    if ((!neurotrackinputmode) && (iSL % 2 == 1)) continue;
    double relId = getRelId(*axialHits[ihit]);
    if (expert.isRelevant(relId, iSL)) {
      if (nHits[iSL] < expert.getMaxHitsPerSL()) {
        chitPattern |= 1 << (iSL + 9 * nHits[iSL]);
        ++nHits[iSL];
      }
    }
  }
  if (!neurotrackinputmode) {
    // loop over stereo hits
    for (int ihit = 0; ihit < m_segmentHits.getEntries(); ++ ihit) {
      unsigned short iSL = m_segmentHits[ihit]->getISuperLayer();
      // skip axial hits
      if (iSL % 2 == 0) continue;
      // get priority time
      double relId = getRelId(*m_segmentHits[ihit]);
      if (expert.isRelevant(relId, iSL)) {
        if (nHits[iSL] < expert.getMaxHitsPerSL()) {
          chitPattern |= 1 << (iSL + 9 * nHits[iSL]);
          ++nHits[iSL];
        }
      }
    }
  }
  return chitPattern;
}
 
unsigned long
NeuroTrigger::getInputPattern(unsigned isector, const CDCTriggerTrack& track, const bool neurotrackinputmode)
{
  const CDCTriggerMLP& expert = m_MLPs[isector];
  unsigned long hitPattern = 0;
  vector<unsigned> nHits;
  nHits.assign(9, 0);
  // loop over axial hits related to input track
  RelationVector<CDCTriggerSegmentHit> axialHits =
    track.getRelationsTo<CDCTriggerSegmentHit>(m_hitCollectionName);
  for (unsigned ihit = 0; ihit < axialHits.size(); ++ ihit) {
    // skip hits with negative relation weight (not selected in finder)
    if (axialHits.weight(ihit) < 0) {
      continue;
    }
    unsigned short iSL = axialHits[ihit]->getISuperLayer();
    // // skip stereo hits (should not be related to track, but check anyway)
    if ((!neurotrackinputmode) && (iSL % 2 == 1)) continue;
    // get priority time
    int t = (m_hasT0) ? axialHits[ihit]->priorityTime() - m_T0 : 0;
    if (t < 0) {
      t = 0;
    } else if (t > expert.getTMax()) {
      t = expert.getTMax();
    }
    double relId = getRelId(*axialHits[ihit]);
 
    if (expert.isRelevant(relId, iSL)) {
      if (nHits[iSL] < expert.getMaxHitsPerSL()) {
        hitPattern |= 1 << (iSL + 9 * nHits[iSL]);
        ++nHits[iSL];
      }
      B2DEBUG(250, "hit in SL " << iSL);
    } else {
      B2DEBUG(250, "hit in SL " << iSL << " not relevant (relId = " << relId << ")");
    }
  }
  if (!neurotrackinputmode) {
    // loop over stereo hits
    for (int ihit = 0; ihit < m_segmentHits.getEntries(); ++ ihit) {
      unsigned short iSL = m_segmentHits[ihit]->getISuperLayer();
      // skip axial hits
      if (iSL % 2 == 0) continue;
      // get priority time
      int t = (m_hasT0) ? m_segmentHits[ihit]->priorityTime() - m_T0 : 0;
      if (t < 0) {
        t = 0;
      } else if (t > expert.getTMax()) {
        t = expert.getTMax();
      }
      double relId = getRelId(*m_segmentHits[ihit]);
      if (expert.isRelevant(relId, iSL)) {
        if (nHits[iSL] < expert.getMaxHitsPerSL()) {
          hitPattern |= 1 << (iSL + 9 * nHits[iSL]);
          ++nHits[iSL];
        }
        B2DEBUG(250, "hit in SL " << iSL);
      } else {
        B2DEBUG(250, "hit in SL " << iSL << " not relevant (relId = " << relId << ")");
      }
    }
  }
  B2DEBUG(250, "hitPattern " << hitPattern);
  return hitPattern & expert.getSLpatternMask();
}
 
vector<unsigned>
NeuroTrigger::selectHitsHWSim(unsigned isector, const CDCTriggerTrack& track)
{
  const CDCTriggerMLP& expert = m_MLPs[isector];
  vector<unsigned> selectedHitIds = {};
  // prepare vectors to keep best drift times, left/right and selected hit IDs
  vector<int> tMin;
  tMin.assign(expert.nNodesLayer(0), expert.getTMax());
  vector<bool> LRknown;
  LRknown.assign(expert.nNodesLayer(0), false);
  vector<int> hitIds;
  hitIds.assign(expert.nNodesLayer(0), -1);
  vector<unsigned> nHits;
  nHits.assign(9, 0);
 
  // loop over all hits related to input track
  RelationVector<CDCTriggerSegmentHit> allHits =
    track.getRelationsTo<CDCTriggerSegmentHit>(m_hitCollectionName);
  B2DEBUG(250, "start hit loop over all related hits");
  for (unsigned ihit = 0; ihit < allHits.size(); ++ihit) {
    // skip hits with negative relation weight (not selected in finder)
    if (allHits.weight(ihit) < 0) continue;
    unsigned short iSL = allHits[ihit]->getISuperLayer();
    if (expert.getSLpatternUnmasked() != 0 &&
        !((expert.getSLpatternUnmasked() >> iSL) & 1)) {
      B2DEBUG(250, "skipping hit in SL " << iSL);
      continue;
    }
    // get priority time and apply time window cut
    int t = (m_hasT0) ? allHits[ihit]->priorityTime() - m_T0 : 0;
    if (t < 0) {
      t = 0;
    } else if (t > expert.getTMax()) {
      t = expert.getTMax();
    }
    double relId = getRelId(*allHits[ihit]);
    if (expert.isRelevant(relId, iSL)) {
      // get reference hit (worst of existing hits)
      unsigned short iRef = iSL;
      if (expert.getMaxHitsPerSL() > 1) {
        if (nHits[iSL] < expert.getMaxHitsPerSL() &&
            (expert.getSLpatternUnmasked() >> (iSL + 9 * nHits[iSL])) & 1) {
          iRef += 9 * nHits[iSL];
          ++nHits[iSL];
        } else {
          for (unsigned compare = iSL; compare < iSL + 9 * nHits[iSL]; compare += 9) {
            if ((LRknown[iRef] && !LRknown[compare]) ||
                (LRknown[iRef] == LRknown[compare] && tMin[iRef] < tMin[compare]))
              iRef = compare;
          }
        }
      }
      // choose best hit (LR known before LR unknown, then shortest drift time)
      bool useHit = false;
      if (LRknown[iRef]) {
        useHit = (allHits[ihit]->LRknown() && t <= tMin[iRef]);
      } else {
        useHit = (allHits[ihit]->LRknown() || t <= tMin[iRef]);
      }
      B2DEBUG(250, "relevant wire SL " << iSL << " LR " << allHits[ihit]->getLeftRight()
              << " t " << t << " iRef " << iRef << " useHit " << useHit);
      if (useHit) {
        // keep drift time and LR
        LRknown[iRef] = allHits[ihit]->LRknown();
        tMin[iRef] = t;
        hitIds[iRef] = allHits[ihit]->getArrayIndex();
      }
    }
  }
  // save selected hit Ids
  for (unsigned iHit = 0; iHit < hitIds.size(); ++iHit) {
    if (hitIds[iHit] >= 0) selectedHitIds.push_back(hitIds[iHit]);
  }
  return selectedHitIds;
}
vector<unsigned>
NeuroTrigger::selectHits(unsigned isector, const CDCTriggerTrack& track,
                         bool returnAllRelevant)
{
  const CDCTriggerMLP& expert = m_MLPs[isector];
  vector<unsigned> selectedHitIds = {};
  // prepare vectors to keep best drift times, left/right and selected hit IDs
  vector<int> tMin;
  tMin.assign(expert.nNodesLayer(0), expert.getTMax());
  vector<bool> LRknown;
  LRknown.assign(expert.nNodesLayer(0), false);
  vector<int> hitIds;
  hitIds.assign(expert.nNodesLayer(0), -1);
  vector<unsigned> nHits;
  nHits.assign(9, 0);
 
  // loop over axial hits related to input track
  RelationVector<CDCTriggerSegmentHit> axialHits =
    track.getRelationsTo<CDCTriggerSegmentHit>(m_hitCollectionName);
  B2DEBUG(250, "start axial hit loop");
  for (unsigned ihit = 0; ihit < axialHits.size(); ++ihit) {
    // skip hits with negative relation weight (not selected in finder)
    if (axialHits.weight(ihit) < 0) continue;
    unsigned short iSL = axialHits[ihit]->getISuperLayer();
    // skip stereo hits (should not be related to track, but check anyway)
    if (iSL % 2 == 1) continue;
    if (expert.getSLpatternUnmasked() != 0 &&
        !((expert.getSLpatternUnmasked() >> iSL) & 1)) {
      B2DEBUG(250, "skipping hit in SL " << iSL);
      continue;
    }
    // get priority time and apply time window cut
 
    int t = (m_hasT0) ? axialHits[ihit]->priorityTime() - m_T0 : 0;
    if (t < 0) {
      t = 0;
    } else if (t > expert.getTMax()) {
      t = expert.getTMax();
    }
    double relId = getRelId(*axialHits[ihit]);
    if (expert.isRelevant(relId, iSL)) {
      // get reference hit (worst of existing hits)
      unsigned short iRef = iSL;
      if (expert.getMaxHitsPerSL() > 1) {
        if (nHits[iSL] < expert.getMaxHitsPerSL() &&
            (expert.getSLpatternUnmasked() >> (iSL + 9 * nHits[iSL])) & 1) {
          iRef += 9 * nHits[iSL];
          ++nHits[iSL];
        } else {
          for (unsigned compare = iSL; compare < iSL + 9 * nHits[iSL]; compare += 9) {
            if ((LRknown[iRef] && !LRknown[compare]) ||
                (LRknown[iRef] == LRknown[compare] && tMin[iRef] < tMin[compare]))
              iRef = compare;
          }
        }
      }
      // choose best hit (LR known before LR unknown, then shortest drift time)
      bool useHit = false;
      if (LRknown[iRef]) {
        useHit = (axialHits[ihit]->LRknown() && t <= tMin[iRef]);
      } else {
        useHit = (axialHits[ihit]->LRknown() || t <= tMin[iRef]);
      }
      B2DEBUG(250, "relevant wire SL " << iSL << " LR " << axialHits[ihit]->getLeftRight()
              << " t " << t << " iRef " << iRef << " useHit " << useHit);
      if (useHit) {
        // keep drift time and LR
        LRknown[iRef] = axialHits[ihit]->LRknown();
        tMin[iRef] = t;
        hitIds[iRef] = axialHits[ihit]->getArrayIndex();
      }
      if (returnAllRelevant) selectedHitIds.push_back(axialHits[ihit]->getArrayIndex());
    }
  }
 
  // loop over stereo hits, choosing up to MaxHitsPerSL per superlayer
  B2DEBUG(250, "start stereo hit loop");
  for (int ihit = 0; ihit < m_segmentHits.getEntries(); ++ ihit) {
    unsigned short iSL = m_segmentHits[ihit]->getISuperLayer();
    // skip axial hits
    if (iSL % 2 == 0) continue;
    if (expert.getSLpatternUnmasked() != 0 &&
        !((expert.getSLpatternUnmasked() >> iSL) & 1)) {
      B2DEBUG(250, "skipping hit in SL " << iSL);
      continue;
    }
    // get priority time and apply time window cut
    int t = (m_hasT0) ? m_segmentHits[ihit]->priorityTime() - m_T0 : 0;
    if (t < 0) {
      t = 0;
    } else if (t > expert.getTMax()) {
      t = expert.getTMax();
    }
    double relId = getRelId(*m_segmentHits[ihit]);
    if (expert.isRelevant(relId, iSL)) {
      // get reference hit (worst of existing hits)
      unsigned short iRef = iSL;
      if (expert.getMaxHitsPerSL() > 1) {
        if (nHits[iSL] < expert.getMaxHitsPerSL() &&
            (expert.getSLpatternUnmasked() >> (iSL + 9 * nHits[iSL])) & 1) {
          iRef += 9 * nHits[iSL];
          ++nHits[iSL];
        } else {
          for (unsigned compare = iSL; compare < iSL + 9 * nHits[iSL]; compare += 9) {
            if ((LRknown[iRef] && !LRknown[compare]) ||
                (LRknown[iRef] == LRknown[compare] && tMin[iRef] < tMin[compare]))
              iRef = compare;
          }
        }
      }
      // choose best hit (LR known before LR unknown, then shortest drift time)
      bool useHit = false;
      if (LRknown[iRef]) {
        useHit = (m_segmentHits[ihit]->LRknown() && t <= tMin[iRef]);
      } else {
        useHit = (m_segmentHits[ihit]->LRknown() || t <= tMin[iRef]);
      }
      B2DEBUG(250, "relevant wire SL " << iSL << " LR " << m_segmentHits[ihit]->getLeftRight()
              << " t " << t << " iRef " << iRef << " useHit " << useHit);
      if (useHit) {
        // keep drift time and LR
        LRknown[iRef] = m_segmentHits[ihit]->LRknown();
        tMin[iRef] = t;
        hitIds[iRef] = ihit;
      }
      if (returnAllRelevant) selectedHitIds.push_back(ihit);
    }
  }
 
  // save selected hit Ids
  if (!returnAllRelevant) {
    for (unsigned iHit = 0; iHit < hitIds.size(); ++iHit) {
      if (hitIds[iHit] >= 0) selectedHitIds.push_back(hitIds[iHit]);
    }
  }
  return selectedHitIds;
}
 
vector<float>
NeuroTrigger::getInputVector(unsigned isector, const vector<unsigned>& hitIds)
{
  const CDCTriggerMLP& expert = m_MLPs[isector];
  // prepare empty input vector and vectors to keep best drift times
  vector<float> inputVector;
  inputVector.assign(expert.nNodesLayer(0), 0.);
  // convert hits to input values
  vector<unsigned> nHits;
  nHits.assign(9, 0);
  for (unsigned ii = 0; ii < hitIds.size(); ++ii) {
    int ihit = hitIds[ii];
    unsigned short iSL = m_segmentHits[ihit]->getISuperLayer();
    unsigned short iRef = iSL + 9 * nHits[iSL];
    ++nHits[iSL];
    int priot = m_segmentHits[ihit]->priorityTime();
    int t = (m_hasT0) ? priot - m_T0 : 0;
    if (t < 0) {
      t = 0;
    } else if (t > expert.getTMax()) {
      t = expert.getTMax();
    }
    int LR = m_segmentHits[ihit]->getLeftRight();
    double relId = getRelId(*m_segmentHits[ihit]);
    int priority = m_segmentHits[ihit]->getPriorityPosition();
    // get scaled input values (scaling such that the absolute value of all inputs is < 1)
    inputVector[3 * iRef] = expert.scaleId(relId, iSL);
    float scaleT = pow(2, floor(log2(1. / expert.getTMax())));
    inputVector[3 * iRef + 1] = (((LR >> 1) & 1) - (LR & 1)) * t * scaleT;
    inputVector[3 * iRef + 2] = m_alpha[iSL][int(priority < 3)] * 0.5;
  }
  return inputVector;
}
 
vector<float>
NeuroTrigger::runMLP(unsigned isector, const vector<float>& input)
{
  const CDCTriggerMLP& expert = m_MLPs[isector];
  vector<float> weights = expert.getWeights();
  vector<float> layerinput = input;
  vector<float> layeroutput = {};
  unsigned iw = 0;
  for (unsigned il = 1; il < expert.nLayers(); ++il) {
    //add bias input
    layerinput.push_back(1.);
    //prepare output
    layeroutput.clear();
    layeroutput.assign(expert.nNodesLayer(il), 0.);
    //loop over outputs
    for (unsigned io = 0; io < layeroutput.size(); ++io) {
      //loop over inputs
      for (unsigned ii = 0; ii < layerinput.size(); ++ii) {
        layeroutput[io] += layerinput[ii] * weights[iw++];
      }
      //apply activation function
      layeroutput[io] = tanh(layeroutput[io] / 2.);
    }
    //output is new input
    layerinput = layeroutput;
  }
  return expert.unscaleTarget(layeroutput);
}
 
vector<float>
NeuroTrigger::runMLPFix(unsigned isector, vector<float> input)
{
  unsigned precisionInput = m_precision[3];
  unsigned precisionWeights = m_precision[4];
  unsigned precisionLUT = m_precision[5];
  unsigned precisionTanh = m_precision[3];
  unsigned dp = precisionInput + precisionWeights - precisionLUT;
 
  const CDCTriggerMLP& expert = m_MLPs[isector];
  // transform inputs to fixed point (cut off to input precision)
  vector<long> inputFix(input.size(), 0);
  for (unsigned ii = 0; ii < input.size(); ++ii) {
    inputFix[ii] = long(input[ii] * (1 << precisionInput));
  }
  // transform weights to fixed point (round to weight precision)
  vector<float> weights = expert.getWeights();
  vector<long> weightsFix(weights.size(), 0);
  for (unsigned iw = 0; iw < weights.size(); ++iw) {
    weightsFix[iw] = long(round(weights[iw] * (1 << precisionWeights)));
  }
  // maximum input value for the tanh LUT
  unsigned xMax = unsigned(ceil(atanh(1. - 1. / (1 << (precisionTanh + 1))) *
                                (1 << (precisionLUT + 1))));
 
  // run MLP
  vector<long> layerinput = inputFix;
  vector<long> layeroutput = {};
  unsigned iw = 0;
  for (unsigned il = 1; il < expert.nLayers(); ++il) {
    // add bias input
    layerinput.push_back(1 << precisionInput);
    // prepare output
    layeroutput.clear();
    layeroutput.assign(expert.nNodesLayer(il), 0);
    // loop over outputs
    for (unsigned io = 0; io < layeroutput.size(); ++io) {
      // loop over inputs
      for (unsigned ii = 0; ii < layerinput.size(); ++ii) {
        layeroutput[io] += layerinput[ii] * weightsFix[iw++];
      }
      // apply activation function -> LUT, calculated on the fly here
      unsigned long bin = abs(layeroutput[io]) >> dp;
      // correction to get symmetrical rounding errors
      float x = (bin + 0.5 - 1. / (1 << (dp + 1))) / (1 << precisionLUT);
      long tanhLUT = (bin < xMax) ? long(round(tanh(x / 2.) * (1 << precisionTanh))) : (1 << precisionTanh);
      layeroutput[io] = (layeroutput[io] < 0) ? -tanhLUT : tanhLUT;
    }
    // output is new input
    layerinput = layeroutput;
  }
 
  // transform output back to float before unscaling
  vector<float> output(layeroutput.size(), 0.);
  for (unsigned io = 0; io < output.size(); ++io) {
    output[io] = layeroutput[io] / float(1 << precisionTanh);
  }
  return expert.unscaleTarget(output);
}
 
void
NeuroTrigger::save(const string& filename, const string& arrayname)
{
  B2INFO("Saving networks to file " << filename << ", array " << arrayname);
  TFile datafile(filename.c_str(), "UPDATE");
  TObjArray* MLPs = new TObjArray(m_MLPs.size());
  for (unsigned isector = 0; isector < m_MLPs.size(); ++isector) {
    MLPs->Add(&m_MLPs[isector]);
  }
  MLPs->Write(arrayname.c_str(), TObject::kSingleKey | TObject::kOverwrite);
  datafile.Close();
  MLPs->Clear();
  delete MLPs;
}
bool NeuroTrigger::loadIDHist(const std::string& filename)
{
  std::ifstream gzipfile(filename, ios_base::in | ios_base::binary);
  boost::iostreams::filtering_istream arrayStream;
  arrayStream.push(boost::iostreams::gzip_decompressor());
  arrayStream.push(gzipfile);
  CDCTriggerMLPData::HeaderSet hline;
  if (gzipfile.is_open()) {
    while (arrayStream >> hline) {
      for (unsigned i = 0; i < 18; ++i) {
        m_MLPs[hline.exPert].relevantID[i] = hline.relID[i];
      }
    }
  } else { return false;}
  return true;
}
 
 
bool
NeuroTrigger::load(const string& filename, const string& arrayname)
{
  if (filename.size() < 1) {
    m_MLPs.clear();
    m_MLPs = m_cdctriggerneuroconfig->getMLPs();
    if (m_MLPs.size() == 0) {
      B2ERROR("Could not load Neurotrigger weights from database!");
      return false;
    }
    B2INFO("Loaded Neurotrigger MLP weights from database: " +  m_cdctriggerneuroconfig->getNNName());
    B2DEBUG(100, "loaded " << m_MLPs.size() << " networks from database");
    // load some values from the geometry that will be needed for the input
    setConstants();
    return true;
  } else {
    TFile datafile(filename.c_str(), "READ");
    if (!datafile.IsOpen()) {
      B2WARNING("Could not open file " << filename);
      return false;
    }
 
    TObjArray* MLPs = (TObjArray*)datafile.Get(arrayname.c_str());
    if (!MLPs) {
      datafile.Close();
      B2WARNING("File " << filename << " does not contain key " << arrayname);
      return false;
    }
    m_MLPs.clear();
    for (int isector = 0; isector < MLPs->GetEntriesFast(); ++isector) {
      CDCTriggerMLP* expert = dynamic_cast<CDCTriggerMLP*>(MLPs->At(isector));
      if (expert) m_MLPs.push_back(*expert);
      else B2WARNING("Wrong type " << MLPs->At(isector)->ClassName() << ", ignoring this entry.");
    }
    MLPs->Clear();
    delete MLPs;
    datafile.Close();
    B2DEBUG(100, "loaded " << m_MLPs.size() << " networks");
 
    B2INFO("Loaded Neurotrigger MLP weights from file: " +  filename);
    // load some values from the geometry that will be needed for the input
    setConstants();
 
    return true;
  }
}