emtf::Forest Class Reference

#include <Forest.h>

Public Member Functions
void	appendCorrection (std::vector< Event *> &eventsp, int treenum)
void	appendCorrection (Event *e, int treenum)
void	doRegression (int nodeLimit, int treeLimit, double learningRate, LossFunction *l, const char *savetreesdirectory, bool saveTrees)
void	doStochasticRegression (int nodeLimit, int treeLimit, double learningRate, double fraction, LossFunction *l)
	Forest ()
	Forest (std::vector< Event *> &trainingEvents)
	Forest (const Forest &forest)
	Forest (Forest &&forest)=default
void	generate (int numTrainEvents, int numTestEvents, double sigma)
std::vector< Event * >	getTrainingEvents ()
Tree *	getTree (unsigned int i)
void	listEvents (std::vector< std::vector< Event *> > &e)
	!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! More...
void	loadForestFromXML (const char *directory, unsigned int numTrees)
void	loadFromCondPayload (const L1TMuonEndCapForest::DForest &payload)
Forest &	operator= (const Forest &forest)
void	predictEvent (Event *e, unsigned int trees)
void	predictEvents (std::vector< Event *> &eventsp, unsigned int trees)
void	prepareRandomSubsample (double fraction)
void	rankVariables (std::vector< int > &rank)
void	saveSplitValues (const char *savefilename)
void	setTrainingEvents (std::vector< Event *> &trainingEvents)
unsigned int	size ()
void	sortEventVectors (std::vector< std::vector< Event *> > &e)
void	updateEvents (Tree *tree)
void	updateRegTargets (Tree *tree, double learningRate, LossFunction *l)
	~Forest ()

Private Attributes
std::vector< std::vector< Event * > >	events
std::vector< std::vector< Event * > >	subSample
std::vector< Tree * >	trees

Detailed Description

Definition at line 12 of file Forest.h.

Constructor & Destructor Documentation

Forest() [1/4]

Forest::Forest

(

)

Definition at line 40 of file Forest.cc.

References events.

{ events = std::vector<std::vector<Event*>>(1); }

Forest() [2/4]

Forest::Forest

(

std::vector< Event *> &

trainingEvents

)

Definition at line 46 of file Forest.cc.

References setTrainingEvents().

{ setTrainingEvents(trainingEvents); }

~Forest()

Forest::~Forest

(

)

Definition at line 52 of file Forest.cc.

References mps_fire::i, and trees.

                {
  // When the forest is destroyed it will delete the trees as well as the
  // events from the training and testing sets.
  // The user may want the events to remain after they destroy the forest
  // this should be changed in future upgrades.

  for (unsigned int i = 0; i < trees.size(); i++) {
    if (trees[i])
      delete trees[i];
  }
}

Forest() [3/4]

Forest::Forest

(

const Forest &

forest

)

Definition at line 64 of file Forest.cc.

References HcalDetIdTransform::transform(), compare::tree, and trees.

                                   {
  transform(forest.trees.cbegin(), forest.trees.cend(), back_inserter(trees), [](const Tree* tree) {
    return new Tree(*tree);
  });
}

Forest() [4/4]

emtf::Forest::Forest ( Forest &&  forest )

default

Member Function Documentation

appendCorrection() [1/2]

void Forest::appendCorrection	(	std::vector< Event *> &	eventsp,
		int	treenum
	)

Definition at line 426 of file Forest.cc.

References trees, and updateEvents().

Referenced by predictEvent(), and predictEvents().

                                                                     {
  // Update the prediction by appending the next correction.

  Tree* tree = trees[treenum];
  tree->filterEvents(eventsp);

  // Update the events with their new prediction.
  updateEvents(tree);
}

appendCorrection() [2/2]

void Forest::appendCorrection	(	Event *	e,
		int	treenum
	)

Definition at line 463 of file Forest.cc.

References MillePedeFileConverter_cfg::e, l1tPhase2CaloJetEmulator_cfi::fit, emtf::Node::getFitValue(), and trees.

                                                   {
  // Update the prediction by appending the next correction.

  Tree* tree = trees[treenum];
  Node* terminalNode = tree->filterEvent(e);

  // Update the event with its new prediction.
  double fit = terminalNode->getFitValue();
  e->predictedValue += fit;
}

doRegression()

void Forest::doRegression	(	int	nodeLimit,
		int	treeLimit,
		double	learningRate,
		LossFunction *	l,
		const char *	savetreesdirectory,
		bool	saveTrees
	)

Definition at line 355 of file Forest.cc.

References DummyCfis::c, events, mps_fire::i, MainPageGenerator::l, alignCSCRings::s, sortEventVectors(), contentValuesCheck::ss, AlCaHLTBitMon_QueryRunRegistry::string, trees, and updateRegTargets().

                                          {
  // Build the forest using the training sample.

  //std::cout << std::endl << "--Building Forest..." << std::endl << std::endl;

  // The trees work with a matrix of events where the rows have the same set of events. Each row however
  // is sorted according to the feature variable given by event->data[row].
  // If we only had one set of events we would have to sort it according to the
  // feature variable every time we want to calculate the best split point for that feature.
  // By keeping sorted copies we avoid the sorting operation during splint point calculation
  // and save computation time. If we do not sort each of the rows the regression will fail.
  //std::cout << "Sorting event vectors..." << std::endl;
  sortEventVectors(events);

  // See how long the regression takes.
  TStopwatch timer;
  timer.Start(kTRUE);

  for (unsigned int i = 0; i < (unsigned)treeLimit; i++) {
    // std::cout << "++Building Tree " << i << "... " << std::endl;
    Tree* tree = new Tree(events);
    trees.push_back(tree);
    tree->buildTree(nodeLimit);

    // Update the targets for the next tree to fit.
    updateRegTargets(tree, learningRate, l);

    // Save trees to xml in some directory.
    std::ostringstream ss;
    ss << savetreesdirectory << "/" << i << ".xml";
    std::string s = ss.str();
    const char* c = s.c_str();

    if (saveTrees)
      tree->saveToXML(c);
  }
  //std::cout << std::endl;
  //std::cout << std::endl << "Done." << std::endl << std::endl;

  //    std::cout << std::endl << "Total calculation time: " << timer.RealTime() << std::endl;
}

doStochasticRegression()

void Forest::doStochasticRegression	(	int	nodeLimit,
		int	treeLimit,
		double	learningRate,
		double	fraction,
		LossFunction *	l
	)

Definition at line 548 of file Forest.cc.

References DummyCfis::c, events, HLT_2024v13_cff::fraction, mps_fire::i, MainPageGenerator::l, prepareRandomSubsample(), alignCSCRings::s, sortEventVectors(), contentValuesCheck::ss, AlCaHLTBitMon_QueryRunRegistry::string, subSample, trees, and updateRegTargets().

                                                                                         {
  // If the fraction of events to use is one then this algorithm is slower than doRegression due to the fact
  // that we have to sort the events every time we extract a subsample. Without random sampling we simply
  // use all of the events and keep them sorted.

  // Anyways, this algorithm uses a portion of the events to train each tree. All of the events are updated
  // afterwards with the results from the subsample built tree.

  // Prepare some things.
  sortEventVectors(events);
  trees = std::vector<Tree*>(treeLimit);

  // See how long the regression takes.
  TStopwatch timer;
  timer.Start(kTRUE);

  // Output the current settings.
  // std::cout << std::endl << "Running stochastic regression ... " << std::endl;
  //std::cout << "# Nodes: " << nodeLimit << std::endl;
  //std::cout << "Learning Rate: " << learningRate << std::endl;
  //std::cout << "Bagging Fraction: " << fraction << std::endl;
  //std::cout << std::endl;

  for (unsigned int i = 0; i < (unsigned)treeLimit; i++) {
    // Build the tree using a random subsample.
    prepareRandomSubsample(fraction);
    trees[i] = new Tree(subSample);
    trees[i]->buildTree(nodeLimit);

    // Fit all of the events based upon the tree we built using
    // the subsample of events.
    trees[i]->filterEvents(events[0]);

    // Update the targets for the next tree to fit.
    updateRegTargets(trees[i], learningRate, l);

    // Save trees to xml in some directory.
    std::ostringstream ss;
    ss << "trees/" << i << ".xml";
    std::string s = ss.str();
    const char* c = s.c_str();

    trees[i]->saveToXML(c);
  }

  //std::cout << std::endl << "Done." << std::endl << std::endl;

  //std::cout << std::endl << "Total calculation time: " << timer.RealTime() << std::endl;
}

generate()

void emtf::Forest::generate	(	int	numTrainEvents,
		int	numTestEvents,
		double	sigma
	)

getTrainingEvents()

std::vector< Event * > Forest::getTrainingEvents

(

)

Definition at line 107 of file Forest.cc.

References events.

{ return events[0]; }

getTree()

Tree * Forest::getTree

(

unsigned int

i

)

Definition at line 114 of file Forest.cc.

References mps_fire::i, and trees.

Referenced by L1TMuonEndCapForestESProducer::produce().

                                    {
  if (/*i>=0 && */ i < trees.size())
    return trees[i];
  else {
    //std::cout << i << "is an invalid input for getTree. Out of range." << std::endl;
    return nullptr;
  }
}

listEvents()

void Forest::listEvents

(

std::vector< std::vector< Event *> > &

e

)

!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

Definition at line 142 of file Forest.cc.

References gather_cfg::cout, MillePedeFileConverter_cfg::e, mps_fire::i, and dqmiolumiharvest::j.

                                                       {
  // Simply list the events in each event vector. We have multiple copies
  // of the events vector. Each copy is sorted according to a different
  // determining variable.
  std::cout << std::endl << "Listing Events... " << std::endl;

  for (unsigned int i = 0; i < e.size(); i++) {
    std::cout << std::endl << "Variable " << i << " vector contents: " << std::endl;
    for (unsigned int j = 0; j < e[i].size(); j++) {
      e[i][j]->outputEvent();
    }
    std::cout << std::endl;
  }
}

loadForestFromXML()

void Forest::loadForestFromXML	(	const char *	directory,
		unsigned int	numTrees
	)

Definition at line 477 of file Forest.cc.

References createBeamHaloJobs::directory, contentValuesFiles::fullPath, mps_fire::i, contentValuesCheck::ss, and trees.

                                                                           {
  // Load a forest that has already been created and stored into XML somewhere.

  // Initialize the vector of trees.
  trees = std::vector<Tree*>(numTrees);

  // Load the Forest.
  // std::cout << std::endl << "Loading Forest from XML ... " << std::endl;
  for (unsigned int i = 0; i < numTrees; i++) {
    trees[i] = new Tree();

    std::stringstream ss;
    ss << directory << "/" << i << ".xml";

    trees[i]->loadFromXML(edm::FileInPath(ss.str().c_str()).fullPath().c_str());
  }

  //std::cout << "Done." << std::endl << std::endl;
}

loadFromCondPayload()

void Forest::loadFromCondPayload

(

const L1TMuonEndCapForest::DForest &

payload

)

Definition at line 497 of file Forest.cc.

References mps_fire::i, and trees.

                                                                         {
  // Load a forest that has already been created and stored in CondDB.
  // Initialize the vector of trees.
  unsigned int numTrees = forest.size();

  // clean-up leftovers from previous initialization (if any)
  for (unsigned int i = 0; i < trees.size(); i++) {
    if (trees[i])
      delete trees[i];
  }

  trees = std::vector<Tree*>(numTrees);

  // Load the Forest.
  for (unsigned int i = 0; i < numTrees; i++) {
    trees[i] = new Tree();
    trees[i]->loadFromCondPayload(forest[i]);
  }
}

operator=()

Forest & Forest::operator=

(

const Forest &

forest

)

Definition at line 70 of file Forest.cc.

References mps_fire::i, HcalDetIdTransform::transform(), compare::tree, and trees.

                                              {
  for (unsigned int i = 0; i < trees.size(); i++) {
    if (trees[i])
      delete trees[i];
  }
  trees.resize(0);

  transform(forest.trees.cbegin(), forest.trees.cend(), back_inserter(trees), [](const Tree* tree) {
    return new Tree(*tree);
  });
  return *this;
}

predictEvent()

void Forest::predictEvent	(	Event *	e,
		unsigned int	trees
	)

Definition at line 440 of file Forest.cc.

References appendCorrection(), MillePedeFileConverter_cfg::e, mps_fire::i, and trees.

Referenced by PtAssignmentEngine2016::calculate_pt_xml(), and PtAssignmentEngine2017::calculate_pt_xml().

                                                         {
  // Predict values for eventsp by running them through the forest up to numtrees.

  //std::cout << "Using " << numtrees << " trees from the forest to predict events ... " << std::endl;
  if (numtrees > trees.size()) {
    //std::cout << std::endl << "!! Input greater than the forest size. Using forest.size() = " << trees.size() << " to predict instead." << std::endl;
    numtrees = trees.size();
  }

  // just like in line #2470 of https://root.cern.ch/doc/master/MethodBDT_8cxx_source.html for gradient boosting
  e->predictedValue = trees[0]->getBoostWeight();

  // i iterates through the trees in the forest. Each tree corrects the last prediction.
  for (unsigned int i = 0; i < numtrees; i++) {
    //std::cout << "++Tree " << i << "..." << std::endl;
    appendCorrection(e, i);
  }
}

predictEvents()

void Forest::predictEvents	(	std::vector< Event *> &	eventsp,
		unsigned int	trees
	)

Definition at line 406 of file Forest.cc.

References appendCorrection(), mps_fire::i, and trees.

                                                                            {
  // Predict values for eventsp by running them through the forest up to numtrees.

  //std::cout << "Using " << numtrees << " trees from the forest to predict events ... " << std::endl;
  if (numtrees > trees.size()) {
    //std::cout << std::endl << "!! Input greater than the forest size. Using forest.size() = " << trees.size() << " to predict instead." << std::endl;
    numtrees = trees.size();
  }

  // i iterates through the trees in the forest. Each tree corrects the last prediction.
  for (unsigned int i = 0; i < numtrees; i++) {
    //std::cout << "++Tree " << i << "..." << std::endl;
    appendCorrection(eventsp, i);
  }
}

prepareRandomSubsample()

void Forest::prepareRandomSubsample

(

double

fraction

)

Definition at line 521 of file Forest.cc.

References events, HLT_2024v13_cff::fraction, mps_fire::i, emtf::shuffle(), sortEventVectors(), subSample, and findQualityFiles::v.

Referenced by doStochasticRegression().

                                                   {
  // We use this for Stochastic Gradient Boosting. Basically you
  // take a subsample of the training events and build a tree using
  // those. Then use the tree built from the subsample to update
  // the predictions for all the events.

  subSample = std::vector<std::vector<Event*>>(events.size());
  size_t subSampleSize = fraction * events[0].size();

  // Randomize the first subSampleSize events in events[0].
  shuffle(events[0].begin(), events[0].end(), subSampleSize);

  // Get a copy of the random subset we just made.
  std::vector<Event*> v(events[0].begin(), events[0].begin() + subSampleSize);

  // Initialize and sort the subSample collection.
  for (unsigned int i = 0; i < subSample.size(); i++) {
    subSample[i] = v;
  }

  sortEventVectors(subSample);
}

rankVariables()

void Forest::rankVariables

(

std::vector< int > &

rank

)

Definition at line 195 of file Forest.cc.

References events, mps_fire::i, dqmiolumiharvest::j, SiStripPI::max, edm::second(), jetUpdater_cfi::sort, trees, findQualityFiles::v, and w().

                                               {
  // This function ranks the determining variables according to their importance
  // in determining the fit. Use a low learning rate for better results.
  // Separates completely useless variables from useful ones well,
  // but isn't the best at separating variables of similar importance.
  // This is calculated using the error reduction on the training set. The function
  // should be changed to use the testing set, but this works fine for now.
  // I will try to change this in the future.

  // Initialize the vector v, which will store the total error reduction
  // for each variable i in v[i].
  std::vector<double> v(events.size(), 0);

  //std::cout << std::endl << "Ranking Variables by Net Error Reduction... " << std::endl;

  for (unsigned int j = 0; j < trees.size(); j++) {
    trees[j]->rankVariables(v);
  }

  double max = *std::max_element(v.begin(), v.end());

  // Scale the importance. Maximum importance = 100.
  for (unsigned int i = 0; i < v.size(); i++) {
    v[i] = 100 * v[i] / max;
  }

  // Change the storage format so that we can keep the index
  // and the value associated after sorting.
  std::vector<std::pair<double, int>> w(events.size());

  for (unsigned int i = 0; i < v.size(); i++) {
    w[i] = std::pair<double, int>(v[i], i);
  }

  // Sort so that we can output in order of importance.
  std::sort(w.begin(), w.end());

  // Output the results.
  for (int i = (v.size() - 1); i >= 0; i--) {
    rank.push_back(w[i].second);
    // std::cout << "x" << w[i].second  << ": " << w[i].first  << std::endl;
  }

  // std::cout << std::endl << "Done." << std::endl << std::endl;
}

saveSplitValues()

void Forest::saveSplitValues

(

const char *

savefilename

)

Definition at line 245 of file Forest.cc.

References events, mps_fire::i, dqmiolumiharvest::j, jetUpdater_cfi::sort, contentValuesCheck::ss, trees, tier0::unique(), and findQualityFiles::v.

                                                     {
  // This function gathers all of the split values from the forest and puts them into lists.

  std::ofstream splitvaluefile;
  splitvaluefile.open(savefilename);

  // Initialize the matrix v, which will store the list of split values
  // for each variable i in v[i].
  std::vector<std::vector<double>> v(events.size(), std::vector<double>());

  //std::cout << std::endl << "Gathering split values... " << std::endl;

  // Gather the split values from each tree in the forest.
  for (unsigned int j = 0; j < trees.size(); j++) {
    trees[j]->getSplitValues(v);
  }

  // Sort the lists of split values and remove the duplicates.
  for (unsigned int i = 0; i < v.size(); i++) {
    std::sort(v[i].begin(), v[i].end());
    v[i].erase(unique(v[i].begin(), v[i].end()), v[i].end());
  }

  // Output the results after removing duplicates.
  // The 0th variable is special and is not used for splitting, so we start at 1.
  for (unsigned int i = 1; i < v.size(); i++) {
    TString splitValues;
    for (unsigned int j = 0; j < v[i].size(); j++) {
      std::stringstream ss;
      ss.precision(14);
      ss << std::scientific << v[i][j];
      splitValues += ",";
      splitValues += ss.str().c_str();
    }

    splitValues = splitValues(1, splitValues.Length());
    splitvaluefile << splitValues << std::endl << std::endl;
    ;
  }
}

setTrainingEvents()

void Forest::setTrainingEvents

(

std::vector< Event *> &

trainingEvents

)

Definition at line 87 of file Forest.cc.

References MillePedeFileConverter_cfg::e, events, and mps_fire::i.

Referenced by Forest().

                                                                {
  // tell the forest which events to use for training

  Event* e = trainingEvents[0];
  // Unused variable
  // unsigned int numrows = e->data.size();

  // Reset the events matrix.
  events = std::vector<std::vector<Event*>>();

  for (unsigned int i = 0; i < e->data.size(); i++) {
    events.push_back(trainingEvents);
  }
}

size()

unsigned int Forest::size

(

void

)

Definition at line 127 of file Forest.cc.

References trees.

Referenced by ntupleDataFormat._Collection::__iter__(), ntupleDataFormat._Collection::__len__(), and L1TMuonEndCapForestESProducer::produce().

                          {
  // Return the number of trees in the forest.
  return trees.size();
}

sortEventVectors()

void Forest::sortEventVectors

(

std::vector< std::vector< Event *> > &

e

)

Definition at line 181 of file Forest.cc.

References compareEvents(), MillePedeFileConverter_cfg::e, mps_fire::i, jetUpdater_cfi::sort, and emtf::Event::sortingIndex.

Referenced by doRegression(), doStochasticRegression(), and prepareRandomSubsample().

                                                             {
  // When a node chooses the optimum split point and split variable it needs
  // the events to be sorted according to the variable it is considering.

  for (unsigned int i = 0; i < e.size(); i++) {
    Event::sortingIndex = i;
    std::sort(e[i].begin(), e[i].end(), compareEvents);
  }
}

updateEvents()

void Forest::updateEvents

(

Tree *

tree

)

Definition at line 328 of file Forest.cc.

References MillePedeFileConverter_cfg::e, l1tPhase2CaloJetEmulator_cfi::fit, ALPAKA_ACCELERATOR_NAMESPACE::vertexFinder::it, dqmiolumiharvest::j, and findQualityFiles::v.

Referenced by appendCorrection().

                                    {
  // Prepare the test events for the next tree.

  // Get the list of terminal nodes for this tree.
  std::list<Node*>& tn = tree->getTerminalNodes();

  // Loop through the terminal nodes.
  for (std::list<Node*>::iterator it = tn.begin(); it != tn.end(); it++) {
    std::vector<Event*>& v = (*it)->getEvents()[0];
    double fit = (*it)->getFitValue();

    // Loop through each event in the terminal region and update the
    // the global event it maps to.
    for (unsigned int j = 0; j < v.size(); j++) {
      Event* e = v[j];
      e->predictedValue += fit;
    }

    // Release memory.
    (*it)->getEvents() = std::vector<std::vector<Event*>>();
  }
}

updateRegTargets()

void Forest::updateRegTargets	(	Tree *	tree,
		double	learningRate,
		LossFunction *	l
	)

Definition at line 289 of file Forest.cc.

References MillePedeFileConverter_cfg::e, l1tPhase2CaloJetEmulator_cfi::fit, ALPAKA_ACCELERATOR_NAMESPACE::vertexFinder::it, dqmiolumiharvest::j, MainPageGenerator::l, and findQualityFiles::v.

Referenced by doRegression(), and doStochasticRegression().

                                                                              {
  // Prepare the global vector of events for the next tree.
  // Update the fit for each event and set the new target value
  // for the next tree.

  // Get the list of terminal nodes for this tree.
  std::list<Node*>& tn = tree->getTerminalNodes();

  // Loop through the terminal nodes.
  for (std::list<Node*>::iterator it = tn.begin(); it != tn.end(); it++) {
    // Get the events in the current terminal region.
    std::vector<Event*>& v = (*it)->getEvents()[0];

    // Fit the events depending on the loss function criteria.
    double fit = l->fit(v);

    // Scale the rate at which the algorithm converges.
    fit = learningRate * fit;

    // Store the official fit value in the terminal node.
    (*it)->setFitValue(fit);

    // Loop through each event in the terminal region and update the
    // the target for the next tree.
    for (unsigned int j = 0; j < v.size(); j++) {
      Event* e = v[j];
      e->predictedValue += fit;
      e->data[0] = l->target(e);
    }

    // Release memory.
    (*it)->getEvents() = std::vector<std::vector<Event*>>();
  }
}

Member Data Documentation

events

std::vector<std::vector<Event*> > emtf::Forest::events

private

Definition at line 66 of file Forest.h.

Referenced by eventsfwlite.Events::__getattr__(), eventsfwlite.Events::__getitem__(), eventsfwlite.Events::__len__(), doRegression(), doStochasticRegression(), Forest(), getTrainingEvents(), looper.Looper::loop(), prepareRandomSubsample(), looper.Looper::process(), rankVariables(), saveSplitValues(), and setTrainingEvents().

subSample

std::vector<std::vector<Event*> > emtf::Forest::subSample

private

Definition at line 67 of file Forest.h.

Referenced by doStochasticRegression(), and prepareRandomSubsample().

trees

std::vector<Tree*> emtf::Forest::trees

private

Definition at line 68 of file Forest.h.

Referenced by appendCorrection(), doRegression(), doStochasticRegression(), Forest(), getTree(), loadForestFromXML(), loadFromCondPayload(), operator=(), predictEvent(), predictEvents(), rankVariables(), saveSplitValues(), size(), and ~Forest().