#include <svm289_BFS.h>

Public Member Functions
	svm_model (const svm_model &_model, FileDescPtr _fileDesc, RunLog &_log)

	svm_model (FileDescPtr _fileDesc, RunLog &_log)

	svm_model (const svm_parameter &_param, const FeatureNumList &_selFeatures, FileDescPtr _fileDesc, RunLog &_log)

	svm_model (istream &_fileName, FileDescPtr _fileDesc, RunLog &_log)

	~svm_model ()

double *	DecValues ()

void	NormalizeProbability ()
	Derining multiclass probability as done in "Recognizing Plankton Images From the SIPPER". More...

double **	PairwiseProb ()

double *	ProbEstimates ()

void	Read (istream &i, FileDescPtr fileDesc, RunLog &log)

void	Write (ostream &o)

Public Attributes
double *	dec_values

kkint32	l

kkint32 *	label

kkint32	nr_class

kkint32 *	nSV

double **	pairwise_prob

svm_parameter	param

double *	prob_estimates

double *	probA

double *	probB

double *	rho

FeatureNumList	selFeatures

FeatureVectorList	SV

double **	sv_coef

bool	weOwnSupportVectors

Detailed Description

Definition at line 146 of file svm289_BFS.h.

Constructor & Destructor Documentation

SVM289_BFS::svm_model::svm_model	(	const svm_model &	_model,
		FileDescPtr	_fileDesc,
		RunLog &	_log
	)

Definition at line 4105 of file svm289_BFS.cpp.

References dec_values, KKMLL::FeatureNumList::FeatureNumList(), l, label, nr_class, nSV, pairwise_prob, param, prob_estimates, probA, probB, rho, selFeatures, sv_coef, svm_model(), SVM289_BFS::svm_parameter::svm_parameter(), and weOwnSupportVectors.

Referenced by svm_model().

                                   :
   param               (_model.param),
   nr_class            (_model.nr_class),
   l                   (_model.l),
   SV                  (_fileDesc, false),
   sv_coef             (NULL),
   rho                 (NULL),
   probA               (NULL),
   probB               (NULL),
   label               (NULL),
   nSV                 (NULL),     // number of SVs for each class (nSV[k])
   weOwnSupportVectors (_model.weOwnSupportVectors),
   selFeatures         (_model.selFeatures),
   dec_values          (NULL),
   pairwise_prob       (NULL),
   prob_estimates      (NULL)
 
 {
   kkint32  m = nr_class - 1;
   kkint32  numBinaryCombos = nr_class * (nr_class - 1) / 2;
 
   {
     // Copy over support vectors.
     FeatureVectorList::const_iterator idx;
     SV.Owner (weOwnSupportVectors);
     for  (idx = _model.SV.begin ();  idx != _model.SV.end ();  idx++)
     {
       FeatureVectorPtr fv = *idx;
       if  (weOwnSupportVectors)
         SV.push_back (new FeatureVector (*fv));
       else
         SV.push_back (fv);
     }
   }
 
   if  (_model.sv_coef)
   {
     sv_coef = new double*[m];
     for  (kkint32 j = 0;  j < m;  j++)
     {
       sv_coef[j] = new double[l];
       for  (kkint32 i = 0;   i < l;  i++)
        sv_coef[j][i] = _model.sv_coef[j][i];
     }
   }
 
   if  (_model.rho)
   {
     // Copy over RHO
     rho = new double[numBinaryCombos];
     for  (kkint32 i = 0;  i < numBinaryCombos;  i++)
       rho[i] = _model.rho[i];
   }
 
   if  (_model.probA)
   {
     probA = new double[numBinaryCombos];
     for  (kkint32 i = 0;  i < numBinaryCombos;  i++)
       probA[i] = _model.probA[i];
   }
 
   if  (_model.probB)
   {
     probB = new double[numBinaryCombos];
     for  (kkint32 i = 0;  i < numBinaryCombos;  i++)
       probB[i] = _model.probB[i];
   }
 
   if  (_model.label)
   {
     label = new kkint32[nr_class];
     for  (kkint32 i = 0;  i < nr_class;  i++)
       label[i] = _model.label[i];
   }
 
   if  (_model.nSV)
   {
     nSV = new kkint32[nr_class];
     for (kkint32 i = 0;  i < nr_class;  i++)
       nSV[i] = _model.nSV[i];
   }
 }

SVM289_BFS::svm_model::svm_model	(	FileDescPtr	_fileDesc,
		RunLog &	_log
	)

Definition at line 4193 of file svm289_BFS.cpp.

References dec_values, KKMLL::FeatureNumList::FeatureNumList(), l, label, nr_class, nSV, pairwise_prob, param, prob_estimates, probA, probB, rho, selFeatures, sv_coef, svm_model(), SVM289_BFS::svm_parameter::svm_parameter(), and weOwnSupportVectors.

Referenced by svm_model().

                                  :
    param               (),
    nr_class            (0),
    l                   (0),
    SV                  (_fileDesc, true),
    sv_coef             (NULL),
    rho                 (NULL),
    probA               (NULL),
    probB               (NULL),
    label               (NULL),
    nSV                 (NULL),     // number of SVs for each class (nSV[k])
    weOwnSupportVectors (false),
    selFeatures         (_fileDesc),
    dec_values          (NULL),
    pairwise_prob       (NULL),
    prob_estimates      (NULL)
 {
 }

SVM289_BFS::svm_model::svm_model	(	const svm_parameter &	_param,
		const FeatureNumList &	_selFeatures,
		FileDescPtr	_fileDesc,
		RunLog &	_log
	)

Definition at line 4215 of file svm289_BFS.cpp.

References dec_values, KKMLL::FeatureNumList::FeatureNumList(), l, label, nr_class, nSV, pairwise_prob, param, prob_estimates, probA, probB, rho, selFeatures, sv_coef, svm_model(), SVM289_BFS::svm_parameter::svm_parameter(), and weOwnSupportVectors.

Referenced by svm_model(), and SVM289_BFS::svm_train().

                                   :
    param               (_param),
    nr_class            (0),
    l                   (0),
    SV                  (_fileDesc, true),
    sv_coef             (NULL),
    rho                 (NULL),
    probA               (NULL),
    probB               (NULL),
    label               (NULL),
    nSV                 (NULL),     // number of SVs for each class (nSV[k])
    weOwnSupportVectors (false),
    selFeatures         (_selFeatures),
    dec_values          (NULL),
    pairwise_prob       (NULL),
    prob_estimates      (NULL)
 
 {
 }

SVM289_BFS::svm_model::svm_model	(	istream &	_fileName,
		FileDescPtr	_fileDesc,
		RunLog &	_log
	)

Definition at line 4240 of file svm289_BFS.cpp.

References dec_values, KKMLL::FeatureNumList::FeatureNumList(), l, label, nr_class, nSV, pairwise_prob, param, prob_estimates, probA, probB, Read(), rho, selFeatures, sv_coef, svm_model(), SVM289_BFS::svm_parameter::svm_parameter(), and weOwnSupportVectors.

Referenced by svm_model().

                                   :
    param               (),
    nr_class            (0),
    l                   (0),
    SV                  (_fileDesc, true),
    sv_coef             (NULL),
    rho                 (NULL),
    probA               (NULL),
    probB               (NULL),
    label               (NULL),
    nSV                 (NULL),     // number of SVs for each class (nSV[k])
    weOwnSupportVectors (false),
    selFeatures         (_fileDesc),
    dec_values          (NULL),
    pairwise_prob       (NULL),
    prob_estimates      (NULL)
 {
   Read (_in, _fileDesc, _log);
 }

SVM289_BFS::svm_model::~svm_model ( )

Definition at line 4266 of file svm289_BFS.cpp.

References dec_values, label, nr_class, nSV, pairwise_prob, prob_estimates, probA, probB, rho, and sv_coef.

 {
   if  (weOwnSupportVectors)
     SV.Owner (true);
   else
     SV.Owner (false);
 
   kkint32  i;
 
   if  (sv_coef)
   {
     for  (i = 0;  i < (nr_class - 1);  i++)
       delete sv_coef[i];
     delete  sv_coef;
     sv_coef = NULL;
   }
 
   delete  rho;     rho   = NULL;
   delete  probA;   probA = NULL;
   delete  probB;   probB = NULL;
   delete  label;   label = NULL;
   delete  nSV;     nSV   = NULL;
 
   if  (pairwise_prob)
   {
     for  (i = 0;  i < (nr_class - 1);  i++)
     {
       delete  pairwise_prob[i];
       pairwise_prob[i] = NULL;
     }
     delete  pairwise_prob;
     pairwise_prob = NULL;
   }
 
   delete  dec_values;
   dec_values = NULL;
   delete  prob_estimates;
   prob_estimates = NULL;
 }

Member Function Documentation

double * SVM289_BFS::svm_model::DecValues ( )

Definition at line 4307 of file svm289_BFS.cpp.

References dec_values, and nr_class.

Referenced by SVM289_BFS::svm_predict_probability().

 {
   if  (!dec_values)
     dec_values = new double[nr_class * (nr_class - 1) / 2];
   return  dec_values;
 }

void SVM289_BFS::svm_model::NormalizeProbability ( )

Derining multiclass probability as done in "Recognizing Plankton Images From the SIPPER".

Definition at line 4619 of file svm289_BFS.cpp.

References nr_class, pairwise_prob, prob_estimates, and ProbEstimates().

Referenced by SVM289_BFS::svm_predict_probability().

 {
   // Make sure that the ProbEstimates array exists.
   ProbEstimates ();
 
   if  (pairwise_prob == NULL)
     return;
 
   kkint32  x = 0;
   kkint32  y = 0;
   double  totalProb = 0.0;
 
   for  (x = 0;  x < nr_class;  x++)
   {
     prob_estimates[x] = 1.0;
     for  (y = 0;  y < nr_class;  y++)
     {
       if  (x != y)
         prob_estimates[x] *= pairwise_prob[x][y];
     }
 
     totalProb += prob_estimates[x];
   }
 
   for  (x = 0;  x < nr_class;  x++)
     prob_estimates[x] /= totalProb;
 }  /* NormalizeProbability */

double ** SVM289_BFS::svm_model::PairwiseProb ( )

Definition at line 4323 of file svm289_BFS.cpp.

References nr_class, and pairwise_prob.

Referenced by SVM289_BFS::svm_predict_probability().

 {
   if  (!pairwise_prob)
   {
     pairwise_prob = new double*[nr_class];
     for  (kkint32 x = 0;  x < nr_class;  x++)
       pairwise_prob[x] = new double[nr_class];
   }
   return  pairwise_prob;
 }

double * SVM289_BFS::svm_model::ProbEstimates ( )

Definition at line 4315 of file svm289_BFS.cpp.

References nr_class, and prob_estimates.

Referenced by NormalizeProbability(), and SVM289_BFS::svm_predict_probability().

 {
   if  (!prob_estimates)
     prob_estimates = new double[nr_class];
   return  prob_estimates;
 }

void SVM289_BFS::svm_model::Read	(	istream &	i,
		FileDescPtr	fileDesc,
		RunLog &	log
	)

Definition at line 4430 of file svm289_BFS.cpp.

References label, nSV, probA, probB, rho, and weOwnSupportVectors.

Referenced by svm_model().

 {
   // read parameters
   delete  rho;    rho   = NULL;
   delete  probA;  probA = NULL;
   delete  probB;  probB = NULL;
   delete  label;  label = NULL;
   delete  nSV;    nSV   = NULL;
 
   SV.DeleteContents ();
 
   kkint32  buffLen = 80 * 1024;
   char*  buff = new char[buffLen];
 
   bool  eof = false;
   bool  eol = false;
 
   kkint32  numBinaryCombos = 0;
 
   while  (!in.eof ())
   {
     in.getline (buff, sizeof (buffLen));
 
     KKStr line = buff;
 
     if  (line.SubStrPart (0, 1) == "//")
       continue;
 
     KKStr fieldName = line.ExtractToken2 ("\t\n\r");
     if  (fieldName.EqualIgnoreCase ("</Svm_Model>"))
       break;
 
     if  (fieldName.EqualIgnoreCase ("svm_type"))
     {
       param.svm_type = SVM_Type_FromStr (line);
       if  (param.svm_type == SVM_Type::SVM_NULL)
       {
         KKStr errorMsg = "SVM289_BFS::svm_model::Read   ***ERROR***  Invalid SVM_Type[" + line + "].";
         log.Level (-1) << endl << errorMsg << endl << endl;
         delete  buff;
         buff = NULL;
         throw  errorMsg;
       }
     }
 
     else if  (fieldName.EqualIgnoreCase ("kernel_type") == 0)
     {    
       param.kernel_type = Kernel_Type_FromStr (line);
       if  (param.kernel_type == Kernel_NULL)
       {
         KKStr errorMsg = "SVM289_BFS::svm_model::Read   ***ERROR***  Invalid kernel_type[" + line + "].";
         log.Level (-1) << endl << errorMsg << endl << endl;
         delete[]  buff;  buff = NULL;
         throw  errorMsg;
       }
     }
 
     else if  (fieldName.EqualIgnoreCase ("degree"))
       param.degree = line.ExtractTokenInt ("\t\n\r");
 
     else if  (fieldName.EqualIgnoreCase ("gamma"))
       param.gamma = line.ExtractTokenDouble ("\t\n\r");
 
     else if  (fieldName.EqualIgnoreCase ("coef0"))
       param.coef0 = line.ExtractTokenDouble ("\t\n\r");
 
     else if  (fieldName.EqualIgnoreCase ("nr_class"))
     {
       nr_class = line.ExtractTokenInt ("\t\n\r");
       numBinaryCombos = nr_class * (nr_class - 1) / 2;
     }
 
     else if  (fieldName.EqualIgnoreCase ("total_sv"))
       l = line.ExtractTokenInt ("\t\n\r");
 
     else if  (fieldName.EqualIgnoreCase ("rho"))
     {
       rho = new double[numBinaryCombos];
       for (kkint32 i = 0;  i < numBinaryCombos;  i++)
         rho[i] = line.ExtractTokenDouble ("\t\n\r");
     }
 
     else if  (fieldName.EqualIgnoreCase ("label"))
     {
       label = new kkint32[nr_class];
       for (kkint32 i=0;  i < nr_class;  i++)
         label[i] = line.ExtractTokenInt ("\t\n\r");
     }
 
     else if  (fieldName.EqualIgnoreCase ("probA"))
     {
       probA = new double[numBinaryCombos];
       for (kkint32 i = 0;  i < numBinaryCombos;  i++)
         probA[i] = line.ExtractTokenDouble ("\t\n\r");
     }
 
     else if  (fieldName.EqualIgnoreCase ("probB"))
     {
       probB = new double[numBinaryCombos];
       for  (kkint32 i = 0;  i < numBinaryCombos;  i++)
         probB[i] = line.ExtractTokenDouble ("\t\n\r");
     }
 
     else if  (fieldName.EqualIgnoreCase ("nr_sv"))
     {
       nSV = new kkint32[nr_class];
       for (kkint32 i = 0;  i < nr_class;  i++)
         nSV[i] = line.ExtractTokenInt ("\t\n\r");
     }
 
     else if  (fieldName.EqualIgnoreCase ("SelFeatures"))
     {
       bool  valid = false;
       selFeatures = FeatureNumList (line, valid);
     }
 
 
     else if  (fieldName.EqualIgnoreCase ("SupportVector"))
     {
       // read sv_coef and SV
 
       kkint32 m = nr_class - 1;
       kkint32 i, j;
 
       if  (!sv_coef)
       {
         sv_coef = new double*[m];
         for  (i = 0;  i < m;  i++)
         {
           sv_coef[i] = new double[l];
           for  (j = 0;  j < l;  j++)
             sv_coef[i][j] = 0.0;
         }
       }
 
       if  (SV.QueueSize () >= l)
       {
         KKStr errorMsg = "SVM289_BFS::svm_model::Read   ***ERROR***  To many Support Vector's Defined.";
         log.Level (-1) << endl << errorMsg << endl << endl;
         delete[]  buff;
         throw  errorMsg;
       }
 
       // We are now going to process one line per SV.
       j = 0;
       eof = false;
 
       KKStr  imageFileName = line.ExtractToken2 ("\t");
       //model->SV[i] = &x_space[j];
       FeatureVectorPtr  fv = new FeatureVector (fileDesc->NumOfFields ());
       fv->ExampleFileName (imageFileName);
 
       for  (j = 0;  (j < (nr_class - 1))  &&  (!eol);  j++)
         sv_coef[j][i] = line.ExtractTokenDouble ("\t");
 
       if  (param.kernel_type == PRECOMPUTED)
       {
         log.Level (-1) << endl << endl
                        << "SVM289_BFS::svm_model::Read  ***ERROR***    PRECOMPUTED   Can not Handle." << endl
                        << endl;
       }
       else
       {
         for  (kkuint32 zed = 0;  (zed < fileDesc->NumOfFields ())  &&  (!eol);  zed++)
         {
           KKStr  featureField = line.ExtractToken2 ("\t");
           kkint32  featureNum   = featureField.ExtractTokenInt (":");
           float  featureValue = (float)featureField.ExtractTokenDouble ("\t\n\r");
           fv->FeatureData (featureNum, featureValue);
         }
       }
       SV.PushOnBack (fv);
     }
   }
 
   delete[]  buff;
   buff = NULL;
   weOwnSupportVectors = true;  // XXX
   SV.Owner (true);
 }  /* Read */

void SVM289_BFS::svm_model::Write ( ostream & o )

Definition at line 4335 of file svm289_BFS.cpp.

References KKMLL::FeatureVector::ExampleFileName(), SVM289_BFS::svm_parameter::kernel_type, l, label, nr_class, nSV, KKB::operator<<(), param, SVM289_BFS::PRECOMPUTED, probA, and probB.

 {
   o << "<Svm_Model>"  << endl;
   o << "svm_type"    << "\t" << SVM_Type_ToStr    (param.svm_type)    << endl;
   o << "kernel_type" << "\t" << Kernel_Type_ToStr (param.kernel_type) << endl;
   
   if  (param.kernel_type == POLY)
     o << "degree" << "\t" << param.degree << endl;
 
   if  (param.kernel_type == POLY || param.kernel_type == RBF || param.kernel_type == SIGMOID)
     o << "gamma" << "\t" << param.gamma << endl;
 
   if  (param.kernel_type == POLY || param.kernel_type == SIGMOID)
     o << "coef0" << "\t" << param.coef0 << endl;
 
   o << "SelFeatures" << "\t" << selFeatures.ToCommaDelStr () << endl;
 
   o << "nr_class" << "\t" << nr_class << endl;
   kkint32  numBinaryCombos = nr_class * (nr_class - 1) / 2;
 
   o << "total_sv" << "\t" << l << endl;
   
   {
     o << "rho";
     for  (kkint32 i = 0;  numBinaryCombos;  i++)
       o << "\t" << rho[i];
     o << endl;
   }
   
   if  (label)
   {
     o << "label";
     for  (kkint32 i = 0;  i < nr_class;  i++)
       o << "\t" << label[i];
     o << endl;
   }
 
   if  (probA) // regression has probA only
   {
     o << "probA";
     for  (kkint32 i = 0;  i < numBinaryCombos;  i++)
       o << "\t" << probA[i];
     o << endl;
   }
 
   if  (probB)
   {
     o << "probB";
     for  (kkint32 i = 0;  i < numBinaryCombos;  i++)
       o << "\t" << probB[i];
     o << endl;
   }
 
   if  (nSV)
   {
     o << "nr_sv";
     for  (kkint32 i = 0;  i < nr_class;  i++)
       o << "\t" << nSV[i];
     o << endl;
   }
 
   for  (kkint32 i = 0;  i < l;  i++)
   {
     const  FeatureVector&  p = SV[i];
     o << "SupportVector" << "\t" << p.ExampleFileName ();
 
     kkint32  origPrec = (kkint32)o.precision ();
     o.precision (16);
     for  (kkint32 j = 0;  j < nr_class - 1;  j++)
     {
       //fprintf (fp, "%.16g ", sv_coef[j][i]);
       o << "\t" << sv_coef[j][i];
     }
 
     //const svm_node *p = SV[i];
     o.precision (8);
 
     if  (param.kernel_type == PRECOMPUTED)
     {
       //fprintf(fp,"0:%d ",(kkint32)(p->value));
       o << "\t" << p.FeatureData (0);
     }
     else
     {
       for  (kkint32 zed = 0;  zed < p.NumOfFeatures ();  zed++)
         o << "\t" << zed << ":" << p.FeatureData (zed);
     }
     o << endl;
   }
 
   o << "</Svm_Model>"  << endl;
 }  /* Write */