NICE_VisLearning/regression/regcombination/RegPreRandomForests.cpp

/**
* @file RegPreRandomForests.cpp
* @brief Combination of a regression method with a pre-clustering using a random forest
* @author Sven Sickert
* @date 07/12/2013

*/

#include "vislearning/regression/regcombination/RegPreRandomForests.h"

#include <iostream>
#include <assert.h>

using namespace OBJREC;
using namespace std;
using namespace NICE;

RegPreRandomForests::RegPreRandomForests(const Config * conf,
          const string & section,
          RegressionAlgorithm *_leafRegressionPrototype )
        : leafRegressionPrototype(_leafRegressionPrototype)
{
  string cluster_section = conf->gS ( section, "cluster_section", "RandomForest" );
  mEx = conf->gI ( "RTBRandom", "min_examples", 500 );
  randomforest = new RegRandomForests( conf, cluster_section );
}

RegPreRandomForests::~RegPreRandomForests()
{
  // delete the random forest
  if ( randomforest != NULL )
    delete randomforest;
  
  // delete all regression methods in the leafs
  for ( map<RegressionNode *, RegressionAlgorithm * >::const_iterator it = leafRegressions.begin();
        it != leafRegressions.end(); it++ )
  {
    RegressionAlgorithm * lr = it->second;
    if ( lr != NULL )
      delete lr;
  }
  
  // delete regression prototype
  if ( leafRegressionPrototype != NULL )
    delete leafRegressionPrototype;
}

void RegPreRandomForests::teach ( const VVector & X, const Vector & y )
{
  randomforest->teach ( X, y );

  if ( leafRegressionPrototype != NULL )
  {
    vector<RegressionNode *> leafNodes;
    randomforest->getAllLeafNodes ( leafNodes );
        
    int lsize = leafNodes.size();
    int leafNo = 0;
    cerr << "leafnodes: " << lsize << endl;
        
    #pragma omp parallel for
    for ( int l = 0; l < lsize; l++ )
    {
      leafNo++;

      RegressionNode *node = leafNodes[l];

      if ( !node->isLeaf() ){
        fprintf( stderr, "RegPreRandomForests::predict: ID #%d not a leaf node!", leafNo );
        continue;
      }

      vector<int> leafTrainInds = node->trainExamplesIndices;
      cerr << "Teaching regression method for leaf " << leafNo-1 << "..." << endl;
      cerr << "examples in leave: " << leafTrainInds.size() << endl;
      assert ( leafTrainInds.size() > 0 );
          
      sort ( leafTrainInds.begin(), leafTrainInds.end() );

      NICE::VVector leafTrainData;
      vector<double> tmpVals;

      for ( int i = 0; i < (int)leafTrainInds.size(); i++ )
      {
        if ( leafTrainInds[i] >= 0 && leafTrainInds[i] < (int)y.size() )
        {
          leafTrainData.push_back( X[ leafTrainInds[i] ] );
          tmpVals.push_back( y[ leafTrainInds[i] ] );
        }
      }

      if (leafTrainData.size() <= 0 ) continue;

      NICE::Vector leafTrainVals( tmpVals );

      RegressionAlgorithm *lr = leafRegressionPrototype->clone();

      lr->teach( leafTrainData, leafTrainVals );

      leafRegressions.insert ( pair< RegressionNode *, RegressionAlgorithm *> ( node, lr ) );
    }
  }
}


double RegPreRandomForests::predict ( const Vector & x )
{
  double pred = 0.0;
  
  vector<RegressionNode *> leafNodes;
  
  // traverse the forest and obtain all innvolved leaf nodes
  randomforest->getLeafNodes ( x, leafNodes );
  
  for ( vector<RegressionNode *>::const_iterator it = leafNodes.begin();
        it != leafNodes.end(); it++ )
  {
    RegressionNode *node = *it;
    map<RegressionNode *, RegressionAlgorithm *>::const_iterator leafRegressionIt =
      leafRegressions.find( node );
    
    if ( leafRegressionIt == leafRegressions.end() )
    {
      // this leaf has no associated regression method
      // -> we will use the random forest result
      pred += node->predVal;
      continue;
    }
    
    RegressionAlgorithm *leafRegression = leafRegressionIt->second;
    pred += leafRegression->predict( x );
  }
  
  pred /= leafNodes.size();
  
  return pred;
}

void RegPreRandomForests::clear ()
{
  map<RegressionNode *, RegressionAlgorithm *>::iterator iter;
  for ( iter = leafRegressions.begin(); iter != leafRegressions.end(); iter++ )
  {
    iter->second->clear();
  }
  randomforest->clear();
}

void RegPreRandomForests::store ( ostream & os, int format ) const
{
  cerr << "RegPreRandomForest::store: not yet implemented" << endl;
}

void RegPreRandomForests::restore ( istream& is, int format )
{
  cerr << "RegPreRandomForest::restore: not yet implemented" << endl;
}