#ifdef NICE_USELIB_CPPUNIT

#include <string>
#include <exception>
#include <iostream>
#include <fstream>

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

#include <core/basics/Timer.h>

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

#include <vislearning/cbaselib/ClassificationResults.h>
#include <vislearning/classifier/kernelclassifier/KCGPRegOneVsAll.h>

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

#include "vislearning/classifier/fpclassifier/gphik/FPCGPHIK.h"

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

#include "TestFPCGPHIK.h"


const bool verbose = false;
const bool verboseStartEnd = true;

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

CPPUNIT_TEST_SUITE_REGISTRATION( TestFPCGPHIK );

void TestFPCGPHIK::setUp() {
}

void TestFPCGPHIK::tearDown() {
}

void myClassifierTest( FPCGPHIK & classifier, const Matrix & mX, const Vector & vY )
{
  
  if (verboseStartEnd)
    std::cerr << "================== TestFPCGPHIK::myClassifierTest ===================== " << std::endl;
  
  Examples examples;

  for ( uint i = 0 ; i < vY.size() ; i++ )
    if ( i % 2 == 1 )
    {
      Example example;
      example.svec = new SparseVector;
      example.svec->setDim(3);
      example.svec->set ( 0, mX(i,0) );
      example.svec->set ( 1, mX(i,1) );
      example.svec->set ( 2, 1.0-mX(i,0)-mX(i,1) );
      examples.push_back ( pair<int, Example> ( vY[i], example ) );
    }

  FeaturePool fp; // will be ignored

  if ( verbose )
    std::cerr << "preparation done." << std::endl;

  if ( verbose ) 
    std::cerr << "learning ..." << std::endl;
  classifier.train ( fp, examples ); 

  if ( verbose )
    std::cerr << "testing ..." << std::endl;
  for ( uint i = 0 ; i < vY.size() ; i++ )
    if ( i % 2 == 0 )
    {
      Example example;
      example.svec = new SparseVector;
      example.svec->setDim(3);
      example.svec->set ( 0, mX(i,0) );
      example.svec->set ( 1, mX(i,1) );
      example.svec->set ( 2, 1.0-mX(i,0)-mX(i,1) );
      ClassificationResult r = classifier.classify ( example );
      if (verbose)
      {
        r.scores >> std::cerr;
        std::cerr << "predicted uncertainty: " << r.uncertainty << std::endl;
      }
    } 

  examples.clean();
  
  if (verboseStartEnd)
    std::cerr << "================== TestFPCGPHIK::myClassifierTest done ===================== " << std::endl;

}

void myClassifierStoreRestoreTest( FPCGPHIK & classifier, const Matrix & mX, const Vector & vY )
{
  
  if (verboseStartEnd)
    std::cerr << "================== TestFPCGPHIK::myClassifierStoreRestoreTest ===================== " << std::endl;
  
  Examples examples;

  for ( uint i = 0 ; i < vY.size() ; i++ )
    if ( i % 2 == 1 )
    {
      Example example;
      example.svec = new SparseVector;
      example.svec->setDim(3);
      example.svec->set ( 0, mX(i,0) );
      example.svec->set ( 1, mX(i,1) );
      example.svec->set ( 2, 1.0-mX(i,0)-mX(i,1) );
      examples.push_back ( pair<int, Example> ( vY[i], example ) );
    }

  FeaturePool fp; // will be ignored

  if ( verbose )
    std::cerr << "preparation done." << std::endl;

  if ( verbose ) 
    std::cerr << "learning ..." << std::endl;
  classifier.train ( fp, examples ); 
  
  if ( verbose ) 
    std::cerr << "storing ..." << std::endl;  
  //test the store-functionality  
  string destination("/tmp/GPHIK_store.txt");
  
  std::filebuf fb;
  fb.open (destination.c_str(),ios::out);
  std::ostream os(&fb);
//   
  classifier.store(os);  
//   
  fb.close();
  
  if ( verbose ) 
    std::cerr << "loading ..." << std::endl;  
  
  Config confTmp;
  FPCGPHIK classifierRestored(&confTmp);
  
  std::filebuf fbIn;
  fbIn.open (destination.c_str(),ios::in);
  std::istream is(&fbIn);
//   
  classifierRestored.restore(is);
//   
  fbIn.close();    

  if ( verbose )
    std::cerr << "testing ..." << std::endl;
  for ( uint i = 0 ; i < vY.size() ; i++ )
    if ( i % 2 == 0 )
    {
      Example example;
      example.svec = new SparseVector;
      example.svec->setDim(3);
      example.svec->set ( 0, mX(i,0) );
      example.svec->set ( 1, mX(i,1) );
      example.svec->set ( 2, 1.0-mX(i,0)-mX(i,1) );
      ClassificationResult rOrig = classifier.classify ( example );
      ClassificationResult rRestored = classifierRestored.classify ( example );
      
      //scores are of type FullVector
      //we use the [] operator, since there are no iterators given in FullVector.h
      bool equal(true);
      for (int i = 0; i< rOrig.scores.size(); i++)
      {
        if ( fabs(rOrig.scores[i] - rRestored.scores[i]) > 10-6)
        {
          equal = false;
          break;
        }        
      }
      
      CPPUNIT_ASSERT_EQUAL ( equal, true ); 
    } 

  examples.clean();
  
  if (verboseStartEnd)
    std::cerr << "================== TestFPCGPHIK::myClassifierStoreRestoreTest done ===================== " << std::endl;

}

void myClassifierILTest( FPCGPHIK & classifierRetrain, FPCGPHIK & classifierIL, const Matrix & mX, const Vector & vY )
{
 
  if (verboseStartEnd)
    std::cerr << "================== TestFPCGPHIK::myClassifierILTest ===================== " << std::endl;
  
  Examples examples;
  
  if (verbose)
    std::cerr << "vY: " << vY << std::endl;

  for ( uint i = 0 ; i < vY.size() ; i++ )
  {
    if ( i % 4 == 1 )
    {
      Example example;
      example.svec = new SparseVector;
      example.svec->setDim(3);
      example.svec->set ( 0, mX(i,0) );
      example.svec->set ( 1, mX(i,1) );
      example.svec->set ( 2, 1.0-mX(i,0)-mX(i,1) );
      examples.push_back ( pair<int, Example> ( vY[i], example ) );
    }
  }

  if (verbose)
    std::cerr << "examples.size(): " << examples.size()  << std::endl;

  FeaturePool fp; // will be ignored

  if ( verbose )
    std::cerr << "preparation done." << std::endl;

  if ( verbose ) 
    std::cerr << "learning ..." << std::endl;
  classifierIL.train ( fp, examples ); 
  
  //choose next example(s)
  
  Examples newExamples;
  for ( uint i = 0 ; i < vY.size() ; i++ )
  {
    if ( i % 4 == 3 )
    {
      Example example;
      example.svec = new SparseVector;
      example.svec->setDim(3);
      example.svec->set ( 0, mX(i,0) );
      example.svec->set ( 1, mX(i,1) );
      example.svec->set ( 2, 1.0-mX(i,0)-mX(i,1) );
      newExamples.push_back ( pair<int, Example> ( vY[i], example ) );
    }  
  }

//   if ( verbose ) 
    std::cerr << std::endl << " =============== " << std::endl << "incremental learning ..." << std::endl;
  
  // add them to classifierIL
//   std::cerr << "We add several new examples" << std::endl;
  Timer t;
  t.start();  
//   for (uint i = 0; i < newExamples.size(); i++)
  for (uint i = 0; i < 1; i++)
  {
    classifierIL.addExample( newExamples[i].second, newExamples[i].first);      
  }  
  
  t.stop();  
  std::cerr << "Time used for incremental training: " << t.getLast() << std::endl;

  //add the new features to feature pool needed for batch training
//   for (uint i = 0; i < newExamples.size(); i++)
  for (uint i = 0; i < 2; i++)
  {  
    examples.push_back( newExamples[i] );
  }
  
  std::cerr << std::endl << " =============== " << std::endl << "We train the second classifier from the scratch with the additional new example" << std::endl;
  t.start(); 
  
  classifierRetrain.train ( fp, examples );  
  
  t.stop();  
  std::cerr << "Time used for batch training: " << t.getLast() << std::endl;  
  
  //evaluate both and compare the resulting scores
//  if ( verbose )
    std::cerr << "testing ..." << std::endl;
  for ( uint i = 0 ; i < vY.size() ; i++ )
    if ( i % 2 == 0 )
    {
      Example example;
      example.svec = new SparseVector;
      example.svec->setDim(3);
      example.svec->set ( 0, mX(i,0) );
      example.svec->set ( 1, mX(i,1) );
      example.svec->set ( 2, 1.0-mX(i,0)-mX(i,1) );
      ClassificationResult resultIL = classifierIL.classify ( example );
      ClassificationResult resultBatch = classifierRetrain.classify ( example );
      
      if (verbose)
      {
        std::cerr << "result of IL classifier: " << std::endl;
        resultIL.scores >> std::cerr;
        
        std::cerr << "result of batch classifier: " << std::endl;
        resultBatch.scores >> std::cerr;
      }
      
      //scores are of type FullVector
      //we use the [] operator, since there are no iterators given in FullVector.h
      bool equal(true);
      for (int i = 0; i< resultIL.scores.size(); i++)
      {
        if ( fabs(resultIL.scores[i] - resultBatch.scores[i]) > 10e-3)
        {
          equal = false;
          break;
        }        
      }
      
      CPPUNIT_ASSERT_EQUAL ( equal, true ); 
    } 

  examples.clean();
  
  if (verboseStartEnd)
    std::cerr << "================== TestFPCGPHIK::myClassifierILTest done ===================== " << std::endl;
}

void TestFPCGPHIK::testFPCGPHIK() 
{
  if (verboseStartEnd)
    std::cerr << "================== TestFPCGPHIK::testFPCGPHIK ===================== " << std::endl;

  Config conf;
  conf.sD( "FPCGPHIK", "noise", 0.01 );
  conf.sD( "FPCGPHIK", "parameter_lower_bound", 0.5 );
  conf.sD( "FPCGPHIK", "parameter_upper_bound", 3.5 );
  conf.sI( "FPCGPHIK", "uncertaintyPrediction", 1);
//   conf.sS( "FPCGPHIK", "optimization_method", "none");
  conf.sS( "FPCGPHIK", "optimization_method", "downhillsimplex");
  conf.sB( "FPCGPHIK", "uncertaintyPredictionForClassification", true);

  FPCGPHIK * classifier  = new FPCGPHIK ( &conf );
  
  Matrix mX;
  Vector vY;
  Vector vY_multi;

//   ifstream ifs ("toyExample1.data", ios::in);
//   ifstream ifs ("toyExampleLargeScale.data", ios::in);
  ifstream ifs ("toyExampleLargeLargeScale.data", ios::in);
  CPPUNIT_ASSERT ( ifs.good() );
  ifs >> mX;
  ifs >> vY;
  ifs >> vY_multi;
  ifs.close();
  
  if (verbose)
  {
    std::cerr << "data loaded: mX" << std::endl;
    std::cerr << mX << std::endl;
    std::cerr << "vY: " << std::endl;
    std::cerr << vY << std::endl;
    std::cerr << "vY_multi: " << std::endl;
    std::cerr << vY_multi << std::endl;
  }

  if ( verbose )
    std::cerr << "Binary classification test " << std::endl; 

  myClassifierTest ( *classifier, mX, vY );
  
  // ... we remove nothing here since we are only interested in store and restore :)
  myClassifierStoreRestoreTest ( *classifier, mX, vY );
  
  // ... remove previously computed things and start again, this time with incremental settings
  if (classifier != NULL)
    delete classifier;
  
  classifier  = new FPCGPHIK ( &conf );
    FPCGPHIK * classifierBatch = new FPCGPHIK ( &conf ); 
  
  myClassifierILTest( *classifierBatch, *classifier, mX, vY );
  
  if (classifier != NULL)
    delete classifier;
  if (classifierBatch != NULL)
    delete classifierBatch;
  
  classifier  = new FPCGPHIK ( &conf );
  classifierBatch = new FPCGPHIK ( &conf );  

  if ( verbose )
    std::cerr << "Multi-class classification test " << std::endl; 
  myClassifierTest ( *classifier, mX, vY_multi );
  
  // ... we remove nothing here since we are only interested and store and restore :)
//   
//   myClassifierStoreRestoreTest ( classifier, mX, vY_multi );
  
  // ... remove previously computed things and start again, this time with incremental settings
  if (classifier != NULL)
    delete classifier;
  if (classifierBatch != NULL)
    delete classifierBatch;
  
  classifier  = new FPCGPHIK ( &conf, "GPHIKClassifier" /* section */);
  classifierBatch = new FPCGPHIK ( &conf, "GPHIKClassifier" /* section */ ); 
  
  myClassifierILTest( *classifierBatch, *classifier, mX, vY_multi );
  
  if (classifier != NULL)
    delete classifier;
  if (classifierBatch != NULL)
    delete classifierBatch;  
  
  if (verboseStartEnd)
    std::cerr << "================== TestFPCGPHIK::testFPCGPHIK done ===================== " << std::endl;
 
}

void TestFPCGPHIK::testGPHIKVariance()
{
  if (verboseStartEnd)
    std::cerr << "================== TestFPCGPHIK::testGPHIKVariance ===================== " << std::endl;

  double noise (0.01);
  
  Config conf;
  conf.sD( "GPHIKClassifier", "noise", noise );
  conf.sD( "GPHIKClassifier", "parameter_lower_bound", 1.0 );
  conf.sD( "GPHIKClassifier", "parameter_upper_bound", 1.0 );
  conf.sS( "GPHIKClassifier", "varianceApproximation", "approximate_rough");
  conf.sB( "GPHIKClassifier", "learn_balanced", true);
  conf.sB( "GPHIKClassifier", "uncertaintyPredictionForClassification", true);
  
  FPCGPHIK classifier ( &conf );
  
  Config confVarApproxQuant(conf);
  confVarApproxQuant.sB( "GPHIKClassifier", "use_quantization", true );
  FPCGPHIK classifierQuant ( &confVarApproxQuant );  

  Config confVarApproxFine1(conf);
  confVarApproxFine1.sS( "GPHIKClassifier", "varianceApproximation", "approximate_fine");  
  confVarApproxFine1.sI( "GPHIKClassifier", "nrOfEigenvaluesToConsiderForVarApprox", 1);
  
  FPCGPHIK classifierVarApproxFine1 ( &confVarApproxFine1 );  

  Config confVarApproxFine2(conf);
  confVarApproxFine2.sS( "GPHIKClassifier", "varianceApproximation", "approximate_fine");    
  confVarApproxFine2.sI( "GPHIKClassifier", "nrOfEigenvaluesToConsiderForVarApprox", 2);
  
  FPCGPHIK classifierVarApproxFine2 ( &confVarApproxFine2 );    
  
  Config confExact(conf);
  confExact.sS( "GPHIKClassifier", "varianceApproximation", "exact");   
  
  FPCGPHIK classifierVarExact ( &confExact );
  
  NICE::Matrix mX;
  NICE::Vector vY;
  NICE::Vector vY_multi;

  ifstream ifs ("toyExample2.data", ios::in);
  CPPUNIT_ASSERT ( ifs.good() );
  ifs >> mX;
  ifs >> vY;
  ifs >> vY_multi;
  ifs.close();

  if (verbose)
  {
    std::cerr << "data loaded: mX" << std::endl;
    std::cerr << mX << std::endl;
    std::cerr << "vY: " << std::endl;
    std::cerr << vY << std::endl;
    std::cerr << "vY_multi: " << std::endl;
    std::cerr << vY_multi << std::endl;
  }
  
  Examples examples;

  for ( uint i = 0 ; i < vY.size() ; i++ )
    if ( i % 2 == 0 )
    {
      Example example;
      example.svec = new SparseVector;
      example.svec->setDim(3);
      example.svec->set ( 0, mX(i,0) );
      example.svec->set ( 1, mX(i,1) );
      example.svec->set ( 2, 1.0-mX(i,0)-mX(i,1) );
      examples.push_back ( pair<int, Example> ( vY_multi[i], example ) );
    }

  FeaturePool fp; // will be ignored

  if ( verbose )
    std::cerr << "preparation for variance testing done." << std::endl;

  if ( verbose ) 
    std::cerr << "learning for variance testing ..." << std::endl;
  classifier.train ( fp, examples ); 
  classifierQuant.train ( fp, examples );
  classifierVarApproxFine1.train ( fp, examples ); 
  classifierVarApproxFine2.train ( fp, examples ); 
  classifierVarExact.train ( fp, examples ); 

  if ( verbose )
    std::cerr << "testing for variance testing ..." << std::endl;
  
  for ( uint i = 0 ; i < vY_multi.size() ; i++ )
    if ( i % 2 == 1 )
    {
      Example example;
      example.svec = new SparseVector;
      example.svec->setDim(3);
      example.svec->set ( 0, mX(i,0) );
      example.svec->set ( 1, mX(i,1) );
      example.svec->set ( 2, 1.0-mX(i,0)-mX(i,1) );
      ClassificationResult r = classifier.classify ( example );
      ClassificationResult rQuant = classifierQuant.classify ( example );
      ClassificationResult rVarApproxFine1 = classifierVarApproxFine1.classify ( example );
      ClassificationResult rVarApproxFine2 = classifierVarApproxFine2.classify ( example );
      ClassificationResult rExact = classifierVarExact.classify ( example );
      
      if (verbose)
      {
        std::cerr << "approxUnc: " << r.uncertainty << " approxUncQuant: " << rQuant.uncertainty<< " approxUncFine1: " << rVarApproxFine1.uncertainty << " approxUncFine2: " << rVarApproxFine2.uncertainty << " exactUnc: " << rExact.uncertainty << std::endl;
      }

      CPPUNIT_ASSERT ( r.uncertainty <=  (1.0 + noise) ); //using the "standard" HIK, this is the upper bound
      CPPUNIT_ASSERT ( r.uncertainty >  rVarApproxFine1.uncertainty);
      CPPUNIT_ASSERT ( rQuant.uncertainty >  rVarApproxFine1.uncertainty);
      CPPUNIT_ASSERT ( rVarApproxFine1.uncertainty >  rVarApproxFine2.uncertainty);
      CPPUNIT_ASSERT ( rVarApproxFine2.uncertainty >  rExact.uncertainty);
      
    } 

  examples.clean();  
  
  
  if (verboseStartEnd)
    std::cerr << "================== TestFPCGPHIK::testGPHIKVariance done ===================== " << std::endl;
  
}

#endif