/**
* @file IL_AL.cpp
* @brief Incrementally train the GP HIK classifier using the predictive variance and its approximations to select new samples
* @author Alexander Freytag
* @date 09-05-2012
*/
#include <vector>
#include <stdlib.h>
#include <time.h>
#include <set>


#include <core/basics/Config.h>
#include <core/basics/StringTools.h>

#include <core/vector/SparseVectorT.h>
#include <core/vector/VectorT.h>

//----------

#include "vislearning/baselib/ProgressBar.h"

#include <vislearning/classifier/kernelclassifier/KCGPRegOneVsAll.h>
#include <vislearning/classifier/fpclassifier/gphik/FPCGPHIK.h>
#include "vislearning/cbaselib/MultiDataset.h"
#include <vislearning/cbaselib/LabeledSet.h>
#include "vislearning/cbaselib/ClassificationResults.h"
#include <vislearning/baselib/Globals.h>

#include <vislearning/math/kernels/KernelData.h>

//----------

#include "gp-hik-exp/progs/datatools.h"


//----------

// #include <incrementallearning/IL_Framework_Generic.h>


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

enum verbose_level {NONE = 0, LOW = 1,  MEDIUM = 2, EVERYTHING = 3};
enum QueryStrategy{
      RANDOM = 0,
      GPMEAN,
      GPPREDVAR,
      GPHEURISTIC
    }; 
    
std::string convertInt(int number)
{
   stringstream ss;//create a stringstream
   ss << number;//add number to the stream
   return ss.str();//return a string with the contents of the stream
}

/**
    Computes from randomly or deterministically choosen trainimages kernelmatrizes and evaluates their performance, using ROI-optimization
*/
int main ( int argc, char **argv )
{
  std::cout.precision ( 10 );
  std::cerr.precision ( 10 );

  NICE::Config conf ( argc, argv );
  int trainExPerClass = conf.gI ( "GP_IL", "trainExPerClass", 10 );
  int incrementalAddSize = conf.gI("GP_IL", "incrementalAddSize", 1);
  int nrOfIncrements = conf.gI("GP_IL", "nrOfIncrements", 9);
  int num_runs = conf.gI ( "GP_IL", "num_runs", 10 );  
  bool do_classification = conf.gB ( "GP_IL", "do_classification", true );
  
  double squaredNoise = pow(  conf.gD("FPCGPHIK", "noise", 0.01) , 2);

  string queryStrategyString = conf.gS( "main", "queryStrategy", "random");
  QueryStrategy queryStrategy;
  if (queryStrategyString.compare("gpMean") == 0)
  {
    queryStrategy = GPMEAN;
  }
  else if (queryStrategyString.compare("gpPredVar") == 0)
  {
    queryStrategy = GPPREDVAR;
  }
  else if (queryStrategyString.compare("gpHeuristic") == 0)
  {
    queryStrategy = GPHEURISTIC;
  }  
  else
  {
    queryStrategy = RANDOM;
  }
 
  
  int verbose_int = conf.gI ( "GP_IL", "verbose", 0 );
  verbose_level verbose ( NONE );
  switch ( verbose_int )
  {
    case 0:
      verbose = NONE;
      break;
    case 1:
      verbose = LOW;
      break;
    case 2:
      verbose = MEDIUM;
      break;
    case 3:
      verbose = EVERYTHING;
      break;
  }



  /* initialize random seed: */
  srand ( time ( NULL ) ); //with 0 for reproductive results
//    srand ( 0 ); //with 0 for reproductive results

  // ===========================  INIT =========================== 
  
  std::vector<std::vector<double> > recognitions_rates(nrOfIncrements+1);
  std::vector<std::vector<float> > classification_times(nrOfIncrements+1);
  std::vector<std::vector<float> > IL_training_times(nrOfIncrements);
  
 int nrOfClassesUsed;
  
  for ( int run = 0; run < num_runs; run++ )
  {
    std::cerr << "run: " << run << std::endl;    
    
    //15-scenes settings
    std::string ext = conf.gS("main", "ext", ".txt");
    std::cerr << "Using cache extension: " << ext << std::endl;

    OBJREC::MultiDataset md ( &conf );
    
    std::cerr << "now read the dataset" << std::endl;
   
    // read training set
    vector< NICE::Vector > trainDataOrig;
    Vector y;
    string trainRun ( "train" + convertInt( run ) );
    std::cerr << "look for " << trainRun << std::endl;
    const LabeledSet *train = md[ trainRun ]; //previously, we only selected "train", no we select the permutation for this run
    
    LabeledSet::Permutation orderTrain;
    train->getPermutation(orderTrain); 
    std::vector<string> filenamesTraining;
    for ( LabeledSet::Permutation::const_iterator i = orderTrain.begin(); i != orderTrain.end(); i++)
    {
      string filename((i->second)->img());
      filenamesTraining.push_back(filename);
    }    

    readData< std::vector< NICE::Vector >, NICE::Vector >  ( conf, *train, trainDataOrig, y, ext );
    
    std::set<int> classesAvailable;
    for ( uint i = 0; i < y.size(); i++)
    {
      //automatically check for duplicates
      classesAvailable.insert( y[i] );
    }
    
    int numberOfClasses =  classesAvailable.size();
    
    std::map<int,int> nrExamplesPerClassInDataset;
    std::map<int,std::vector<int> > examplesPerClassInDataset;
    
    for (std::set<int>::const_iterator it = classesAvailable.begin(); it != classesAvailable.end(); it++)
    {
      nrExamplesPerClassInDataset.insert(std::pair<int,int>(*it,0));
      examplesPerClassInDataset.insert(std::pair<int,std::vector<int> >(*it,std::vector<int>(0)));
    }

    for ( uint i = 0; i < y.size(); i++ )
    {
      (examplesPerClassInDataset.find( y[i] )->second).push_back(i);
    }
    
    for (std::map<int,std::vector<int> >::const_iterator it = examplesPerClassInDataset.begin(); it != examplesPerClassInDataset.end(); it++)
    {
      nrExamplesPerClassInDataset.find(it->first)->second = it->second.size();
    }
    
    for ( std::map<int,int>::const_iterator it =  nrExamplesPerClassInDataset.begin(); it != nrExamplesPerClassInDataset.end(); it++)
    {
      cerr << it->first << ": " << it->second << endl;
    }    
       
    Examples examples;   
    
    //count how many examples of every class we have while actively selecting new examples
    //NOTE works only if we have subsequent class numbers
    NICE::Vector pickedExamplesPerClass( classesAvailable.size(), trainExPerClass);
    
    std::map<int,std::vector<int> > examplesPerClassInDatasetTmp (examplesPerClassInDataset);
    
    //chose examples for every class used for training
    //we will always use the first examples from each class, since the dataset comes already randomly ordered
    for (std::set<int>::const_iterator clIt = classesAvailable.begin(); clIt != classesAvailable.end(); clIt++)
    {
      std::map<int,std::vector<int> >::iterator exIt = examplesPerClassInDatasetTmp.find(*clIt);
      std::cerr << "pick training examples for class " << *clIt << std::endl;
      
      for (int i = 0; i < trainExPerClass; i++)
      {
        std::cerr << "i: " << i << std::endl;
         int exampleIndex ( 0 ); //old: rand() % ( exIt->second.size() ) );
         std::cerr << "pick example " << exIt->second[exampleIndex] << " - " << y[exIt->second[exampleIndex] ] << " -- " << filenamesTraining[exIt->second[exampleIndex]] <<  std::endl;
         
        Example example;
        NICE::Vector & xTrain = trainDataOrig[exIt->second[exampleIndex]];
        example.svec = new SparseVector(xTrain);
        //let's take this example and its corresponding label (which should be *clIt)
        examples.push_back ( pair<int, Example> ( y[exIt->second[exampleIndex] ], example ) ); 
        //
        exIt->second.erase(exIt->second.begin()+exampleIndex);
      }
    }    
     
    
    std::vector<string> filenamesUnlabeled;
    filenamesUnlabeled.clear();
    //which examples are left to be actively chosen lateron?
    std::vector<int> unlabeledExamples( y.size() - trainExPerClass*classesAvailable.size() );
    int exCnt( 0 );
    for (std::set<int>::const_iterator clIt = classesAvailable.begin(); clIt != classesAvailable.end(); clIt++ )
    {
      std::map<int,std::vector<int> >::iterator exIt = examplesPerClassInDatasetTmp.find(*clIt);
      //list all examples of this specific class
      for (std::vector<int>::const_iterator it = exIt->second.begin(); it != exIt->second.end(); it++)
      {
        unlabeledExamples[exCnt] = *it;
        exCnt++;
        filenamesUnlabeled.push_back( filenamesTraining[*it] );
      }
    }
    
    time_t  prep_start_time = clock();

    FPCGPHIK * classifier  = new FPCGPHIK( &conf );
    
    FeaturePool fp; // will be ignored
    classifier->train ( fp, examples );

    float time_preparation = ( float ) ( clock() - prep_start_time ) ;
    std::cerr << "Time for initial training: " << time_preparation / CLOCKS_PER_SEC << std::endl;
    
    nrOfClassesUsed = classesAvailable.size();
    
      // ------------------ TESTING
    string testRun ( "test" + convertInt( run ) );
    const LabeledSet *test = md[ testRun ]; //previously, we only selected "test", no we select the permutation for this run
    VVector testData;
    Vector yTest;
    readData< VVector, Vector > ( conf, *test, testData, yTest, ext );
    

    NICE::Matrix confusionMatrix ( numberOfClasses, numberOfClasses );
    confusionMatrix.set ( 0.0 );

    time_t  start_time = clock();

    std::vector<int> chosen_examples_per_class ( numberOfClasses );
    
    std::cerr << "Current statistic about picked examples per class: " << pickedExamplesPerClass << std::endl;

    if ( do_classification )
    {
      for ( uint i = 0 ; i < testData.size(); i++ )
      {
        Example example;
        const Vector & xstar = testData[i];
        SparseVector xstar_sparse ( xstar );
        OBJREC::ClassificationResult result;
        example.svec = &xstar_sparse;
        
        result = classifier->classify( example );
//         cerr << "[" << i << " / " << testData.size() << "] " << result.classno << " " << yTest[i] << std::endl;
        
        result.classno_groundtruth = yTest[i];
        confusionMatrix ( result.classno_groundtruth , result.classno ) ++;
      }

      float time_classification = ( float ) ( clock() - start_time ) ;
      if ( verbose >= LOW )
        cerr << "Time for Classification with " << nrOfClassesUsed*trainExPerClass << " training-examples: " << time_classification / CLOCKS_PER_SEC << " [s]" << endl;
      ( classification_times[0] ).push_back ( time_classification / CLOCKS_PER_SEC );

      confusionMatrix.normalizeRowsL1();
      double avg_recognition_rate = 0.0;
      for ( int i = 0 ; i < ( int ) confusionMatrix.rows(); i++ )
      {
        if ( verbose >= MEDIUM )
        {
          std::cerr << "Class no: " <<  i  << " : " << confusionMatrix ( i, i ) << std::endl;
        }
        avg_recognition_rate += confusionMatrix ( i, i );
      }

      avg_recognition_rate /= confusionMatrix.rows();

      std::cerr << confusionMatrix;
      std::cerr << "avg recognition rate " << avg_recognition_rate*100 << " % -- " << examples.size() << " training examples used" << std::endl << std::endl;

      recognitions_rates[0].push_back ( avg_recognition_rate*100 );
    }

    //Now start the Incremental-Learning-Part
    
    for (int incrementationStep = 0; incrementationStep < nrOfIncrements; incrementationStep++)
    {
      //chose examples for every class used for training
      Examples newExamples;
      
      
      //simply count how many possible example we have 
      int nrOfPossibleExamples(  unlabeledExamples.size() );
      
      if (queryStrategy == RANDOM)
      {
        std::cerr << "print chosen examples: " << std::endl;        
        for (int i = 0; i < incrementalAddSize; i++)
        {        
          int exampleIndex ( rand() % ( unlabeledExamples.size() ) );
          
          Example newExample;
          NICE::Vector & xTrain = trainDataOrig[ unlabeledExamples[exampleIndex] ];
          newExample.svec = new SparseVector( xTrain ); 
          int label( y[ unlabeledExamples[exampleIndex] ] );
          newExamples.push_back ( pair<int, Example> ( label, newExample ) );
          unlabeledExamples.erase( unlabeledExamples.begin()+exampleIndex );
          std::cerr << exampleIndex+1 << " / " << incrementalAddSize << " : " <<  filenamesUnlabeled[ exampleIndex ] << std::endl;          
          filenamesUnlabeled.erase( filenamesUnlabeled.begin()+exampleIndex );
          pickedExamplesPerClass[label]++;
        }
      }// end computation for RANDOM
      else if ( (queryStrategy == GPMEAN) || (queryStrategy == GPPREDVAR) || (queryStrategy == GPHEURISTIC) )
      {
        //compute uncertainty values for all examples according to the query strategy
        std::vector<std::pair<int,double> > scores;
        scores.clear();
        time_t  unc_pred_start_time = clock();
//         std::cerr << "possible examples to query: " << unlabeledExamples.size() << std::endl;
        for (uint exIndex = 0; exIndex < unlabeledExamples.size(); exIndex++)
        {
            Example example;    
            NICE::Vector & xTrain = trainDataOrig[ unlabeledExamples[exIndex] ];
            SparseVector xTrainSparse ( xTrain );
            example.svec = &xTrainSparse;
            
            if (queryStrategy == GPMEAN)
            {
              ClassificationResult r = classifier->classify( example );
              double bestScore( numeric_limits<double>::max() );
              for( int clCnt = 0; clCnt < nrOfClassesUsed; clCnt++)
              {
                if ( fabs(r.scores[clCnt]) < bestScore )
                  bestScore = fabs(r.scores[clCnt]);
              }
              scores.push_back( std::pair<int,double> ( exIndex, bestScore ) );
            }
            else if (queryStrategy == GPPREDVAR)
            {
              double singleUncertainty;
              //use the pred variance computation specified in the config file
              classifier->predictUncertainty( example, singleUncertainty );
              //take the maximum of the scores for the predictive variance
              scores.push_back( std::pair<int,double> ( exIndex, singleUncertainty ) );
            }
            else if (queryStrategy == GPHEURISTIC)
            {
              double singleUncertainty;
              //use the pred variance computation specified in the config file
              classifier->predictUncertainty( example, singleUncertainty );
              //compute the mean values for every class
              ClassificationResult r = classifier->classify( example );
              NICE::Vector heuristicValues ( r.scores.size(), 0);
              for ( int tmp = 0; tmp < heuristicValues.size(); tmp++ )
              {
                heuristicValues[tmp] = fabs(r.scores[tmp]) / sqrt( squaredNoise + singleUncertainty );
              }              
              //take the minimum of the scores for the heuristic measure
              scores.push_back( std::pair<int,double> ( exIndex, heuristicValues.Min()) );
            }
        }
        float time_score_computation = ( float ) ( clock() - unc_pred_start_time ) ;
          
        //pick the ones with best score
        //we could speed this up using a more sophisticated search method
        
        if (queryStrategy == GPPREDVAR) //take the maximum of the scores for the predictive variance
        {
          std::set<int> chosenExamplesForThisRun;
          chosenExamplesForThisRun.clear();          
          for (int i = 0; i < incrementalAddSize; i++)
          {
            std::vector<std::pair<int,double> >::iterator bestExample = scores.begin();
            
            for (std::vector<std::pair<int,double> >::iterator jIt = scores.begin(); jIt !=scores.end(); jIt++)
            {
              if (jIt->second > bestExample->second)
                bestExample = jIt;
            }
            Example newExample;    
            NICE::Vector & xTrain = trainDataOrig[ unlabeledExamples[bestExample->first] ]; 
            newExample.svec = new SparseVector( xTrain ); 
            //actually this is the ACTIVE LEARNING step (query a label)
            int label( y[ unlabeledExamples[bestExample->first] ] );
            newExamples.push_back ( pair<int, Example> ( label, newExample ) );    
            //remember the index, to safely remove this example afterwards from unlabeledExamples
            chosenExamplesForThisRun.insert(bestExample->first);
            scores.erase(bestExample);
            pickedExamplesPerClass[label]++;
          }
          
          std::cerr << "print chosen examples: " << std::endl;
          int tmpCnt(0);
          for (std::set<int>::const_iterator it = chosenExamplesForThisRun.begin(); it != chosenExamplesForThisRun.end(); it++, tmpCnt++)
          {
            std::cerr << tmpCnt+1 << " / " << incrementalAddSize << " : " <<  filenamesUnlabeled[ *it ] << std::endl;
          }   
          
          //delete the queried examples from the set of unlabeled ones
          //do this in an decreasing order in terms of indices to ensure valid access
          for (std::set<int>::const_reverse_iterator it = chosenExamplesForThisRun.rbegin(); it != chosenExamplesForThisRun.rend(); it++)
          {
            unlabeledExamples.erase( unlabeledExamples.begin()+(*it) );             
          }          
        }
        else //take the minimum of the scores for the heuristic and the gp mean (minimum margin)
        {
          std::set<int> chosenExamplesForThisRun;
          chosenExamplesForThisRun.clear();
          for (int i = 0; i < incrementalAddSize; i++)
          {
            std::vector<std::pair<int,double> >::iterator bestExample = scores.begin();
            
            for (std::vector<std::pair<int,double> >::iterator jIt = scores.begin(); jIt !=scores.end(); jIt++)
            {
              if (jIt->second < bestExample->second)
                bestExample = jIt;
            }
            Example newExample;    
            NICE::Vector & xTrain = trainDataOrig[ unlabeledExamples[bestExample->first] ];
            newExample.svec = new SparseVector( xTrain ); 
            //actually this is the ACTIVE LEARNING step (query a label)
            int label( y[ unlabeledExamples[bestExample->first] ] );
            newExamples.push_back ( pair<int, Example> ( label, newExample ) );           
            //remember the index, to safely remove this example afterwards from unlabeledExamples
            chosenExamplesForThisRun.insert(bestExample->first);
            scores.erase(bestExample);
            pickedExamplesPerClass[label]++;
          }    
          
          std::cerr << "print chosen examples: " << std::endl;
          int tmpCnt(0);
          for (std::set<int>::const_iterator it = chosenExamplesForThisRun.begin(); it != chosenExamplesForThisRun.end(); it++, tmpCnt++)
          {
            std::cerr << tmpCnt+1 << " / " << incrementalAddSize << " : " <<  filenamesUnlabeled[ *it ] << std::endl;
          }           
          
          //delete the queried example from the set of unlabeled ones
          //do this in an decreasing order in terms of indices to ensure valid access
          for (std::set<int>::const_reverse_iterator it = chosenExamplesForThisRun.rbegin(); it != chosenExamplesForThisRun.rend(); it++)
          {
            unlabeledExamples.erase( unlabeledExamples.begin()+(*it) );  
            filenamesUnlabeled.erase( filenamesUnlabeled.begin()+(*it) );
          }

        }
      
        std::cerr << "Time used to compute query-scores for " <<  nrOfPossibleExamples << " examples: " << time_score_computation / CLOCKS_PER_SEC << " [s]" << std::endl;
      } // end computation for GPMEAN, GPPREDVAR, or GPHEURISTIC
           
      
      std::cerr << "Current statistic about picked examples per class: " << pickedExamplesPerClass << std::endl;
          
      time_t  IL_add_start_time = clock();

      classifier->addMultipleExamples( newExamples );
      
      //remove the memory used in newExamples
      for ( uint tmp = 0; tmp < newExamples.size(); tmp++ )
      {
        delete newExamples[tmp].second.svec;
        newExamples[tmp].second.svec = NULL;
      }
      
      float time_IL_add = ( float ) ( clock() - IL_add_start_time ) ;
      std::cerr << "Time for IL-adding of " << incrementalAddSize << " examples to already " <<  nrOfClassesUsed*trainExPerClass+incrementalAddSize*incrementationStep << "  training-examples: " << time_IL_add / CLOCKS_PER_SEC << " [s]" << std::endl;
      IL_training_times[incrementationStep].push_back(time_IL_add / CLOCKS_PER_SEC);

           
      //do the classification for evaluating the benefit of new examples
      if ( do_classification )
      {
        confusionMatrix.set( 0.0 );
        for ( uint i = 0 ; i < testData.size(); i++ )
        {
          Example example;
          const Vector & xstar = testData[i];
          SparseVector xstar_sparse ( xstar );
          example.svec = &xstar_sparse;
          OBJREC::ClassificationResult result;
          
          result = classifier->classify( example );
          
          result.classno_groundtruth = yTest[i];
          confusionMatrix ( result.classno_groundtruth , result.classno ) ++;
        }     


        float time_classification = ( float ) ( clock() - start_time ) ;
        if ( verbose >= LOW )
          std::cerr << "Time for Classification with " << nrOfClassesUsed*trainExPerClass+incrementalAddSize*(incrementationStep+1) << " training-examples: " << time_classification / CLOCKS_PER_SEC << " [s]" << std::endl;
        ( classification_times[incrementationStep+1] ).push_back ( time_classification / CLOCKS_PER_SEC );

        confusionMatrix.normalizeRowsL1();
        double avg_recognition_rate = 0.0;
        for ( int i = 0 ; i < ( int ) confusionMatrix.rows(); i++ )
        {
          if ( verbose >= MEDIUM )
          {
            std::cerr << "Class no: " <<  i  << " : " << confusionMatrix ( i, i ) << std::endl;
          }
          avg_recognition_rate += confusionMatrix ( i, i );
        }

        avg_recognition_rate /= confusionMatrix.rows();

        std::cerr << confusionMatrix;
        std::cerr << "avg recognition rate " << avg_recognition_rate*100 << " % -- " << nrOfClassesUsed*trainExPerClass+incrementalAddSize*(incrementationStep+1) << " training examples used" << std::endl  << std::endl;

        recognitions_rates[incrementationStep+1].push_back ( avg_recognition_rate*100 );
      } //classification after IL adding
    } //IL adding of different classes
    std::cerr << "Final statistic about picked examples per class: " << pickedExamplesPerClass << std::endl;
    
    //don't waste memory!
    delete classifier;
    for ( int tmp = 0; tmp < examples.size(); tmp++ )
    {
      delete examples[tmp].second.svec;
      examples[tmp].second.svec = NULL;
    }
  }//runs 


  std::cerr << "no of classes used: " << nrOfClassesUsed << " incrementalAddSize: " << incrementalAddSize << std::endl;
  
  // ================= EVALUATION ========================0

  if ( do_classification )
  {
    std::cerr << "========================" << std::endl;
    std::cerr << "content of classification_times: " << std::endl;
    for ( std::vector<std::vector<float> >::const_iterator it = classification_times.begin(); it != classification_times.end(); it++ )
    {
      for ( std::vector<float> ::const_iterator jt = ( *it ).begin(); jt != ( *it ).end(); jt++ )
      {
        std::cerr << *jt << " ";
      }
      std::cerr << std::endl;
    }

    std::vector<float> mean_classification_times;
    std::vector<float> std_dev_classification_times;
    for ( std::vector<std::vector<float> >::const_iterator it = classification_times.begin(); it != classification_times.end(); it++ )
    {
      float mean_classification_time ( 0.0 );
      for ( std::vector<float>::const_iterator itRun = it->begin(); itRun != it->end(); itRun++ )
      {
        mean_classification_time += *itRun;
      }
      mean_classification_time /= it->size();
      mean_classification_times.push_back ( mean_classification_time );

      double std_dev_classification_time ( 0.0 );
      for ( std::vector<float>::const_iterator itRun = it->begin(); itRun != it->end(); itRun++ )
      {
        std_dev_classification_time += pow ( *itRun - mean_classification_time, 2 );
      }
      std_dev_classification_time /= it->size();
      std_dev_classification_time = sqrt ( std_dev_classification_time );
      std_dev_classification_times.push_back ( std_dev_classification_time );
    }
    
    int datasize ( nrOfClassesUsed*trainExPerClass );
    for ( uint i = 0; i < mean_classification_times.size(); i++)
    {
      std::cerr << "size: " << datasize << " mean classification time: " << mean_classification_times[i] << " std_dev classification time: " << std_dev_classification_times[i] << std::endl;
      datasize += incrementalAddSize ;
    }
  }
  else
  {
    std::cerr << "========================" << std::endl;
    std::cerr << "No classification done therefor no classification times available." << std::endl;
  }

  std::cerr << "========================" << std::endl;
  std::cerr << "content of IL_training_times: " << std::endl;
  for ( std::vector<std::vector<float> >::const_iterator it = IL_training_times.begin(); it != IL_training_times.end(); it++ )
  {
    for ( std::vector<float> ::const_iterator jt = ( *it ).begin(); jt != ( *it ).end(); jt++ )
    {
      std::cerr << *jt << " ";
    }
    std::cerr << std::endl;
  }

  std::vector<float> mean_IL_training_times;
  std::vector<float> std_dev_IL_training_times;
  for ( std::vector<std::vector<float> >::const_iterator it = IL_training_times.begin(); it != IL_training_times.end(); it++ )
  {  
    float mean_IL_training_time ( 0.0 );
    for ( std::vector<float>::const_iterator itRun = it->begin(); itRun != it->end(); itRun++ )
    {
      mean_IL_training_time += *itRun;
    }
    mean_IL_training_time /= it->size();
    mean_IL_training_times.push_back ( mean_IL_training_time );

    double std_dev_IL_training_time ( 0.0 );
    for ( std::vector<float>::const_iterator itRun = it->begin(); itRun != it->end(); itRun++ )
    {
      std_dev_IL_training_time += pow ( *itRun - mean_IL_training_time, 2 );
    }
    std_dev_IL_training_time /= it->size();
    std_dev_IL_training_time = sqrt ( std_dev_IL_training_time );
    std_dev_IL_training_times.push_back ( std_dev_IL_training_time );
  }

  int datasize ( nrOfClassesUsed*trainExPerClass );
  for ( uint i = 0; i < mean_IL_training_times.size(); i++)
  {
    cerr << "size: " << datasize << " and adding " << incrementalAddSize << " mean IL_training time: " << mean_IL_training_times[i] << " std_dev IL_training time: " << std_dev_IL_training_times[i] << endl;
    datasize += incrementalAddSize ;
  }

  if ( do_classification )
  {
    std::cerr << "========================" << std::endl;
    std::cerr << "content of recognition_rates: " << std::endl;
    for ( std::vector<std::vector<double> >::const_iterator it = recognitions_rates.begin(); it != recognitions_rates.end(); it++ )
    {
      for ( std::vector<double> ::const_iterator jt = ( *it ).begin(); jt != ( *it ).end(); jt++ )
      {
        std::cerr << *jt << " ";
      }
      std::cerr << std::endl;
    }

    std::cerr << "calculating final results " << std::endl;
    std::vector<double> mean_recs;
    std::vector<double> std_dev_recs;
    for (std::vector<std::vector<double> >::const_iterator it = recognitions_rates.begin(); it != recognitions_rates.end(); it++ )
    {
      double mean_rec ( 0.0 );
      for ( std::vector<double>::const_iterator itRun = it->begin(); itRun != it->end(); itRun++ )
      {
        mean_rec += *itRun;
      }
      mean_rec /= it->size();
      mean_recs.push_back ( mean_rec );

      double std_dev_rec ( 0.0 );
      for ( std::vector<double>::const_iterator itRun = it->begin(); itRun != it->end(); itRun++ )
      {
        std_dev_rec += pow ( *itRun - mean_rec, 2 );
      }
      std_dev_rec /= it->size();
      std_dev_rec = sqrt ( std_dev_rec );
      std_dev_recs.push_back ( std_dev_rec );
    }

    int datasize ( nrOfClassesUsed*trainExPerClass );
    for ( uint i = 0; i < recognitions_rates.size(); i++)
    {
      std::cerr << "size: " << datasize << " mean_IL: " << mean_recs[i] << " std_dev_IL: " << std_dev_recs[i] << std::endl;
      datasize += incrementalAddSize ;
    }
  }
  else
  {
    std::cerr << "========================" << std::endl;
    std::cerr << "No classification done therefor no classification times available." << std::endl;
  }

  return 0;
}