data_accessor_splitting.hpp

Go to the documentation of this file.

#ifndef FSTDATAACCESSORSPLITTING_H
#define FSTDATAACCESSORSPLITTING_H
 /* 
=========================================================================
Copyright and license:
  * FST4 software (with the exception of any bundled or externally linked
    third-party components) is Copyright (C) 2000-present Institute of
    Information Theory and Automation of the Czech Academy of Sciences (UTIA).
  * FST4 is free software: you can redistribute it and/or modify it under
    the terms of the GNU Affero General Public License as published by the
    Free Software Foundation, version 3. The full text of the license is in
    the file LICENSE in the distribution root and at
    https://www.gnu.org/licenses/agpl-3.0.html.
  * FST4 is distributed in the hope that it will be useful, but WITHOUT ANY
    WARRANTY; without even the implied warranty of MERCHANTABILITY or
    FITNESS FOR A PARTICULAR PURPOSE. See the GNU Affero General Public
    License for more details.
  * Closed-source use without the AGPL's copyleft obligations requires a
    commercial license from UTIA; in case of interest write to
    utia@utia.cas.cz. LICENSING.md in the distribution root explains both
    options, the contributor license agreement that every contribution
    requires, and the contacts (technical questions and pull requests:
    somol@utia.cas.cz, with "FST" in the subject line).
========================================================================== */
 
#include <memory>
#include <cassert>
#include <iostream>
#include <vector>
#include "error.hpp"
#include "global.hpp"
#include "data_accessor.hpp"
#include "data_splitter.hpp"
 
/*============== Template parameter type naming conventions ==============
--------- Numeric types: -------------------------------------------------
DATATYPE - data sample values - usually real numbers (but may be integers
          in text processing etc.)
REALTYPE - must be real numbers - for representing intermediate results of 
          calculations like mean, covariance etc.
IDXTYPE - index values for enumeration of data samples - (nonnegative) integers, 
          extent depends on numbers of samples in data
DIMTYPE - index values for enumeration of features (dimensions), or classes (not 
          class sizes) - (nonnegative) integers, usually lower extent than IDXTYPE, 
          but be aware of expressions like _classes*_features*_features ! 
          in linearized representations of feature matrices for all classes
BINTYPE - feature selection marker type - represents ca. <10 different feature 
          states (selected, deselected, sel./desel. temporarily 1st nested loop, 2nd...)
RETURNTYPE - criterion value: real value, but may be extended in future to support 
          multiple values 
--------- Class types: ---------------------------------------------------
SUBSET       - class of class type Subset 
CLASSIFIER   - class implementing interface defined in abstract class Classifier 
EVALUATOR    - class implementing interface defined in abstract class Sequential_Step 
DISTANCE     - class implementing interface defined in abstract class Distance 
DATAACCESSOR - class implementing interface defined in abstract class Data_Accessor 
INTERVALCONTAINER - class of class type Data_Intervaller 
CONTAINER    - STL container of class type Data_Interval  
========================================================================== */
 
namespace FST {

template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
class Data_Accessor_Splitting : public Data_Accessor<DATATYPE,IDXTYPE>
{
        // NOTE: the calling convention is the same for every nested data access loop (as in cross-validation):
        //   all loops call the same methods (just get*Split, get*Block, get*Pattern), and the loop level
        //   they refer to is selected beforehand with setSplittingDepth()
        // Access to data is defined using indexes, assuming data is accessible by random access
        // Support for arbitrary data splitting (train/test, cross-validation, hold-out, etc., by means
        // of arbitrary Data_Splitter objects) is implemented here
        public:
                typedef std::shared_ptr<Data_Splitter<INTERVALCONTAINER,IDXTYPE> > PSplitter;  
                typedef std::shared_ptr<std::vector<PSplitter> > PSplitters;  
                typedef const DATATYPE* PPattern;  
                Data_Accessor_Splitting(const PSplitters _dsp);
                virtual ~Data_Accessor_Splitting(){notify("Data_Accessor_Splitting destructor.");}
                
                virtual unsigned int getNoOfClasses() const override;          
                virtual unsigned int getNoOfFeatures() const override;         
                
                virtual IDXTYPE getClassSize(const unsigned int c) const override;    
                virtual IDXTYPE getClassSizeSum() const override;    
                
                virtual void setClass(const int c) override {assert(c>=0); active_class = c;} 
                virtual int  getClass() const override {return active_class;}

                void setSplittingDepth(const unsigned int depth);
                unsigned int getSplittingDepth() const {return splitting_depth;} 
                        
                virtual unsigned int getNoOfSplits() const override;  
                virtual bool getFirstSplit() override;                
                virtual bool getNextSplit() override;                 
                virtual unsigned int getSplitIndex() const override;  
 
                // NOTE: in the block accessing methods below, use a different loopdepth whenever two or more loops
                //       (of any DataPart type) overlap in time, otherwise their cursor state collides

                virtual IDXTYPE getNoOfBlocks(const DataPart ofwhat) const override;
                virtual bool getFirstBlock(const DataPart ofwhat, PPattern &firstpattern, IDXTYPE &patterns, const unsigned int loopdepth=0) override=0;
                virtual bool getNextBlock(const DataPart ofwhat, PPattern &firstpattern, IDXTYPE &patterns, const unsigned int loopdepth=0) override=0;
                virtual IDXTYPE getBlockIndex(const unsigned int loopdepth=0) const override;
                virtual IDXTYPE getNoOfPatterns(const DataPart ofwhat) const override;
                virtual void substitute(const DataPart source, const DataPart target) override;
                virtual void resubstitute() override;
 
                virtual std::ostream& print(std::ostream& os) const override; 
        protected:
                Data_Accessor_Splitting(const Data_Accessor_Splitting &da);      // to be used in clone() in derived classes
        protected:
                typedef std::vector<unsigned int> CLASSSIZES;  
                void initialize(const unsigned int _features, const CLASSSIZES &_classes);
                CLASSSIZES classes;  
                unsigned int features;  
 
                // DataPart mapping
                DataPart mappedTRAIN;      
                DataPart mappedTEST;       
                DataPart mappedTRAINTEST;  
                DataPart mappedALL;        
                DataPart mappedDataPart(const DataPart ofwhat) const;  
 
                // splitting support
                PSplitters dsp;  
                typedef const Data_Interval<IDXTYPE>* DATAINTERVAL;  
                virtual bool getFirstBlock(const DataPart ofwhat, DATAINTERVAL &tmp, const unsigned int loopdepth=0);
                virtual bool getNextBlock(const DataPart ofwhat, DATAINTERVAL &tmp, const unsigned int loopdepth=0);
 
                typedef std::shared_ptr<INTERVALCONTAINER> PIntervaller;  
                class DataSplit {
                public:
                        DataSplit() {_train.reset(); _test.reset(); _reduced_train.reset(); _reduced_test.reset(); notify("DataSplit empty constructor.");} 
                        DataSplit(const DataSplit& cs);      
                        DataSplit(const DataSplit& cs, int); 
                        ~DataSplit() {notify("DataSplit destructor.");}
                        PIntervaller _train;          
                        PIntervaller _test;           
                        PIntervaller _reduced_train;  
                        PIntervaller _reduced_test;   
                        Data_Interval<IDXTYPE> _all;   
                };
                std::vector<std::vector<DataSplit> > splits;    
 
                std::vector<IDXTYPE> enum_split;           
                std::vector<std::vector<IDXTYPE> > enum_block;  
                std::vector<std::vector<DataPart> > tt_phase;   
 
                unsigned int splitting_depth;           
 
                int active_class;                     
                
                // for debugging
                bool _initialize_called;  
                bool is_initialized() const {return _initialize_called;}  
                void assert_splits(const int splitting_check=-1) const;
};
 
 
// ============================================ PUBLIC ===============================================
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::Data_Accessor_Splitting(const PSplitters _dsp) :
        mappedTRAIN(TRAIN), mappedTEST(TEST), mappedTRAINTEST(TRAINTEST), mappedALL(ALL)
{
        notify("Data_Accessor_Splitting constructor.");
        _initialize_called=false;
        features=0;
        assert(_dsp->size()>0);
        dsp=_dsp;
        active_class=-1; 
        splitting_depth=0;
        splits.assign(_dsp->size(),std::vector<DataSplit>()); 
        enum_split.assign(_dsp->size(),0);
        enum_block.assign(_dsp->size(),std::vector<IDXTYPE>()); 
        tt_phase.assign(_dsp->size(),std::vector<DataPart>()); 
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
unsigned int Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getNoOfClasses() const
{
        return classes.size();
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
unsigned int Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getNoOfFeatures() const
{
        return features;
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
IDXTYPE Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getClassSize(const unsigned int c) const
{
        assert(c<classes.size());
        return classes[c];
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
IDXTYPE Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getClassSizeSum() const
{
        IDXTYPE patterns=0;
        for(unsigned int c=0;c<classes.size();c++) patterns+=classes[c];
        return patterns;
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
void Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::setSplittingDepth(const unsigned int depth)
{
#ifdef DEBUG
        assert(depth<dsp->size());
        assert(enum_split.size()==dsp->size());
        for(unsigned int n=0;n<depth;n++) assert(enum_split[n]>0);
#endif
        splitting_depth=depth;
} 
 
// NOTE: precondition: initialize() must have completed before this is meaningful — it assigns
//       the per-class interval containers and calls set_n() on the depth-0 splitters (deeper
//       splitters get set_n() in the course of split enumeration). The asserts below make a
//       violation fail fast in debug builds.
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
unsigned int Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getNoOfSplits() const
{
        assert(dsp); // consider also checking the vector contents
        assert(splitting_depth<dsp->size());
        assert(splitting_depth<splits.size() && !splits[splitting_depth].empty()); // initialize() has run
        return (*dsp)[splitting_depth]->getNoOfSplits();
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
bool Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getFirstSplit()
{
#ifdef DEBUG
        assert(dsp); // consider also checking the vector contents
        assert(is_initialized());
        assert(splitting_depth<(unsigned int)dsp->size());
        for(unsigned int n=0;n<splitting_depth;n++) assert(enum_split[n]>0);
        assert_splits(/*n not set in splitters yet*/);
#endif
        bool result=true;
        if(splitting_depth>0)
                for(unsigned int c=0; c<classes.size(); c++)
                        {splits[splitting_depth][c]._all.startidx=0; (*dsp)[splitting_depth]->set_n(c, splits[splitting_depth][c]._all.count=splits[splitting_depth-1][c]._train->sum() );}
        else
                for(unsigned int c=0; c<classes.size(); c++)
                        {splits[splitting_depth][c]._all.startidx=0; (*dsp)[splitting_depth]->set_n(c, splits[splitting_depth][c]._all.count=classes[c] );}
 
        if(!(*dsp)[splitting_depth]->makeFirstSplit()) result=false;
        else 
        if(splitting_depth>0) for(unsigned int c=0; c<classes.size(); c++) {
                // translate the new level's intervals (relative to the enclosing level's TRAIN part) into
                // absolute indexes through the enclosing level's ABSOLUTE train list (_reduced_train): at
                // depth 1 that list is the depth-0 _train itself, at depth 2 and deeper _train would be
                // relative to depth-2's train part and the translation would land on the wrong samples
                splits[splitting_depth-1][c]._reduced_train->reduce(splits[splitting_depth][c]._train,splits[splitting_depth][c]._reduced_train);
                splits[splitting_depth-1][c]._reduced_train->reduce(splits[splitting_depth][c]._test,splits[splitting_depth][c]._reduced_test);
        }
        if(result) enum_split[splitting_depth]=1; else enum_split[splitting_depth]=0;
        for(unsigned int n=splitting_depth+1;n<dsp->size();n++) enum_split[n]=0; // reset/cancel all deeper splits
        
        return result;
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
bool Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getNextSplit()
{
#ifdef DEBUG
        assert(dsp); // consider also checking the vector contents
        assert(is_initialized());
        assert(splitting_depth<dsp->size());
        for(unsigned int n=0;n<=splitting_depth;n++) assert(enum_split[n]>0);
        assert_splits(splitting_depth);
#endif
        bool result=true;
        if(!(*dsp)[splitting_depth]->makeNextSplit()) result=false;
        else 
        if(splitting_depth>0) for(unsigned int c=0; c<classes.size(); c++) {
                splits[splitting_depth-1][c]._reduced_train->reduce(splits[splitting_depth][c]._train,splits[splitting_depth][c]._reduced_train);
                splits[splitting_depth-1][c]._reduced_train->reduce(splits[splitting_depth][c]._test,splits[splitting_depth][c]._reduced_test);
        }
        if(result) enum_split[splitting_depth]++; else enum_split[splitting_depth]=0;
        for(unsigned int n=splitting_depth+1;n<dsp->size();n++) enum_split[n]=0; // reset/cancel all deeper splits
                
        return result;
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
unsigned int Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getSplitIndex() const
{
        assert(splitting_depth<dsp->size());
        return enum_split[splitting_depth];
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
IDXTYPE Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getNoOfBlocks(const DataPart ofwhat) const
{
        assert(dsp); // consider also checking the vector contents
        assert(is_initialized());
        assert(active_class>=0 && active_class<(int)classes.size());
        assert(splitting_depth<dsp->size());
        assert_splits(splitting_depth);
        switch(mappedDataPart(ofwhat)) {
                case TRAIN: {return splits[splitting_depth][active_class]._reduced_train->size();}
                case TEST: {return splits[splitting_depth][active_class]._reduced_test->size();}
                case TRAINTEST: {return splits[splitting_depth][active_class]._reduced_train->size()+splits[splitting_depth][active_class]._reduced_test->size();}
                case ALL: {if(splitting_depth>0) return splits[splitting_depth-1][active_class]._reduced_train->size(); else return 1;}
                case NONE: {return 0;}
        } 
        return 0;
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
IDXTYPE Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getBlockIndex(const unsigned int loopdepth) const
{
        assert(dsp); // consider also checking the vector contents
        assert(splitting_depth<dsp->size());
        assert(loopdepth<enum_block[splitting_depth].size());
        return enum_block[splitting_depth][loopdepth];
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
IDXTYPE Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getNoOfPatterns(const DataPart ofwhat) const
{
        assert(is_initialized());
        assert(splitting_depth<dsp->size());
        assert(active_class>=0 && active_class<(int)classes.size());
        assert_splits(splitting_depth);
        switch(mappedDataPart(ofwhat)) {
                case TRAIN: {return splits[splitting_depth][active_class]._reduced_train->sum();}
                case TEST: {return splits[splitting_depth][active_class]._reduced_test->sum();}
                case TRAINTEST: {return splits[splitting_depth][active_class]._reduced_train->sum()+splits[splitting_depth][active_class]._reduced_test->sum();}
                case ALL: {if(splitting_depth>0) return splits[splitting_depth-1][active_class]._reduced_train->sum(); 
                        else return classes[active_class];}
                case NONE: {return 0;}
        } 
        return 0;
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
void Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::substitute(const DataPart source, const DataPart target)
{
        assert(source==TRAIN || source==TEST || source==TRAINTEST || source==ALL);
        assert(target==TRAIN || target==TEST || target==TRAINTEST || target==ALL);
        switch(source){
                case TRAIN:     mappedTRAIN=target; break;
                case TEST: mappedTEST=target; break;
                case TRAINTEST: mappedTRAINTEST=target; break;
                case ALL: mappedALL=target; break;
                case NONE: break;
        }
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
void Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::resubstitute()
{
        mappedTRAIN=TRAIN;
        mappedTEST=TEST;
        mappedTRAINTEST=TRAINTEST;
        mappedALL=ALL;
}
 
 
// ============================================ PROTECTED ===============================================
 
// strong copy-constructor, creates 1:1 copy of all structures
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::Data_Accessor_Splitting(const Data_Accessor_Splitting &da) :
        classes(da.classes),
        features(da.features),
        mappedTRAIN(da.mappedTRAIN), 
        mappedTEST(da.mappedTEST), 
        mappedTRAINTEST(da.mappedTRAINTEST), 
        mappedALL(da.mappedALL),
        splits(da.splits),
        enum_split(da.enum_split),
        enum_block(da.enum_block),
        tt_phase(da.tt_phase),
        splitting_depth(da.splitting_depth),
        active_class(da.active_class),
        _initialize_called(da._initialize_called)
{
        notify("Data_Accessor_Splitting() copy constructor.");
        notify("DAS::copy-constructor, before da.assert_splits()");
        da.assert_splits();
        notify("DAS::copy-constructor, after da.assert_splits()");
        assert(da.dsp);
        assert(da.dsp->size()>0);
        dsp.reset(new std::vector<PSplitter>);
        for(unsigned int n=0; n<da.dsp->size();n++)
        {
                
                PSplitter ps(dynamic_cast<Data_Splitter<INTERVALCONTAINER,IDXTYPE>* >((*da.dsp)[n]->stateless_clone()));
                for(unsigned int c=0; c<classes.size(); c++)
                        ps->assign(c,splits[n][c]._train,splits[n][c]._test);
                dsp->push_back(ps);
        }       
        notify("DAS::copy-constructor, before assert_splits()");
        assert_splits();
        notify("DAS::copy-constructor, after assert_splits()");
}
 
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
void Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::initialize(const unsigned int _features, const CLASSSIZES &_classes)
{
        notify("Data_Accessor_Splitting()::initialize.");
        if(is_initialized()) notify("Re-initializing Data_Accessor_Splitting.");
        classes.resize( _classes.size() );
        copy(_classes.begin(), _classes.end(), classes.begin() );
        features=_features;
        
        for(unsigned int n=0;n<dsp->size();n++) // number of nested splitting levels
        {
                splits[n].clear();
                notify("Data_Accessor_Splitting()::initialize to resize splits.");
                splits[n].resize(classes.size());
                notify("Data_Accessor_Splitting()::initialize splits resized.");
                for(unsigned int c=0;c<classes.size();c++) 
                {
                        DataSplit &cs=splits[n][c];
                        cs._train.reset(new INTERVALCONTAINER());
                        cs._test.reset(new INTERVALCONTAINER());
                        (*dsp)[n]->assign(c,cs._train,cs._test);
                        if(n==0) {
                                (*dsp)[0]->set_n(c,classes[c]);
                                cs._reduced_train=cs._train;
                                cs._reduced_test=cs._test;
                        } else {
                                cs._reduced_train.reset(new INTERVALCONTAINER());
                                cs._reduced_test.reset(new INTERVALCONTAINER());
                        }
                        // NOTE: set_n() is done here for depth 0 only, where the sample count per class is
                        //       simply the class size. At deeper levels the count depends on the enclosing
                        //       level's current TRAIN part, so getFirstSplit() re-issues set_n() on the
                        //       splitter of that level before every split enumeration.
                }
        }
        _initialize_called=true;
        notify("Data_Accessor_Splitting()::initialize finished.");
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
DataPart Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::mappedDataPart(const DataPart ofwhat) const
{
        switch(ofwhat){
                case TRAIN:     return mappedTRAIN;
                case TEST: return mappedTEST;
                case TRAINTEST: return mappedTRAINTEST;
                case ALL: return mappedALL;
                case NONE: return NONE;
        }
        return ofwhat;
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
bool Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getFirstBlock(const DataPart ofwhat, DATAINTERVAL &tmp, const unsigned int loopdepth)
{
#ifdef DEBUG
        assert(dsp); // consider also checking the vector contents
        assert(is_initialized());
        assert(splitting_depth<dsp->size());
        assert(active_class>=0 && active_class<(int)classes.size());
        for(unsigned int n=0;n<=splitting_depth;n++) assert(enum_split[n]>0);
        assert_splits(splitting_depth);
#endif
        if(enum_block[splitting_depth].size()<=loopdepth) enum_block[splitting_depth].resize(loopdepth+1);
        switch(mappedDataPart(ofwhat)) {
                case TRAIN: {tmp=splits[splitting_depth][active_class]._reduced_train->getFirstBlock(loopdepth); break;}
                case TEST: {tmp=splits[splitting_depth][active_class]._reduced_test->getFirstBlock(loopdepth); break;}
                case TRAINTEST: {
                        if(tt_phase[splitting_depth].size()<=loopdepth) tt_phase[splitting_depth].resize(loopdepth+1);
                        tt_phase[splitting_depth][loopdepth]=TRAIN; 
                        tmp=splits[splitting_depth][active_class]._reduced_train->getFirstBlock(loopdepth); 
                        if(tmp==nullptr) {
                                tt_phase[splitting_depth][loopdepth]=TEST; 
                                tmp=splits[splitting_depth][active_class]._reduced_test->getFirstBlock(loopdepth); 
                        }
                        if(tmp==nullptr) tt_phase[splitting_depth][loopdepth]=NONE;
                        break;}
                case ALL: {
                        if(splitting_depth>0) tmp=splits[splitting_depth-1][active_class]._reduced_train->getFirstBlock(loopdepth);
                        else tmp=&(splits[0][active_class]._all);
                        break;}
                case NONE: {tmp=nullptr;}
        }
        if(tmp!=nullptr) enum_block[splitting_depth][loopdepth]=1; else {enum_block[splitting_depth][loopdepth]=0; return false;}
        return true;
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
bool Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::getNextBlock(const DataPart ofwhat, DATAINTERVAL &tmp, const unsigned int loopdepth)
{
#ifdef DEBUG
        assert(dsp); // consider also checking the vector contents
        assert(is_initialized());
        assert(splitting_depth<dsp->size());
        assert(active_class>=0 && active_class<(int)classes.size());
        assert(loopdepth<enum_block[splitting_depth].size());
        for(unsigned int n=0;n<=splitting_depth;n++) assert(enum_split[n]>0);
        assert(enum_block[splitting_depth][loopdepth]>0);
        assert_splits(splitting_depth);
#endif
        switch(mappedDataPart(ofwhat)) {
                case TRAIN: {tmp=splits[splitting_depth][active_class]._reduced_train->getNextBlock(loopdepth); break;}
                case TEST: {tmp=splits[splitting_depth][active_class]._reduced_test->getNextBlock(loopdepth); break;}
                case TRAINTEST: {
                        assert(loopdepth<tt_phase[splitting_depth].size());
                        assert(tt_phase[splitting_depth][loopdepth]==TRAIN || tt_phase[splitting_depth][loopdepth]==TEST);
                        if(tt_phase[splitting_depth][loopdepth]==TRAIN) {
                                tmp=splits[splitting_depth][active_class]._reduced_train->getNextBlock(loopdepth); 
                                if(tmp==nullptr) {
                                        tt_phase[splitting_depth][loopdepth]=TEST;
                                        tmp=splits[splitting_depth][active_class]._reduced_test->getFirstBlock(loopdepth);
                                }
                        } else {
                                tmp=splits[splitting_depth][active_class]._reduced_test->getNextBlock(loopdepth);
                        }
                        if(tmp==nullptr) tt_phase[splitting_depth][loopdepth]=NONE;
                        break;}
                case ALL: {if(splitting_depth>0) tmp=splits[splitting_depth-1][active_class]._reduced_train->getNextBlock(loopdepth); else tmp=nullptr; break;}
                case NONE: {tmp=nullptr;}
        }
        if(tmp!=nullptr) enum_block[splitting_depth][loopdepth]++; else {enum_block[splitting_depth][loopdepth]=0; return false;}
        return true;
}

template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::DataSplit::DataSplit(const DataSplit& cs)
{
        notify("DataSplit strong copy constructor.");
        if(cs._train) _train.reset(new INTERVALCONTAINER(*cs._train));
        if(cs._test) _test.reset(new INTERVALCONTAINER(*cs._test));
        if(cs._reduced_train == cs._train) // && cs._reduced_test == cs._test
        {
                notify("DataSplit _reduced_* to be linked");
                _reduced_train=_train;
                _reduced_test=_test;
        } else {
                notify("DataSplit _reduced_* to be re-allocated");
                if(cs._reduced_train) _reduced_train.reset(new INTERVALCONTAINER(*cs._reduced_train));
                if(cs._reduced_test) _reduced_test.reset(new INTERVALCONTAINER(*cs._reduced_test));
        }
        _all.startidx=cs._all.startidx; _all.count=cs._all.count;
}

template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::DataSplit::DataSplit(const DataSplit& cs, int)
{
        notify("DataSplit weak copy constructor.");
        _train=cs._train;
        _test=cs._test;
        _reduced_train=cs._reduced_train;
        _reduced_test=cs._reduced_test;
        _all.startidx=cs._all.startidx; _all.count=cs._all.count;
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
std::ostream& Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::print(std::ostream& os) const 
{
        if(dsp) {
                if(dsp->size()>0) {
                        os << "Data_Accessor_Splitting() splitters: ";
                        typename std::vector<PSplitter>::const_iterator iter;
                        for(iter=dsp->begin();iter!=dsp->end();iter++) {
                                if(*iter) os << **iter << " "; else os << "nullptr ";
                        }
                } else os << "Data_Accessor_Splitting()"; 
        }
        return os;
}
 
template<typename DATATYPE, typename IDXTYPE, class INTERVALCONTAINER>
void Data_Accessor_Splitting<DATATYPE,IDXTYPE,INTERVALCONTAINER>::assert_splits(const int splitting_check) const
{
        // assumes Data_Accessor_Splitting() has been initialized; all data structures are checked to be valid and non-empty
#ifdef DEBUG
        assert(features>0); // No. of features
        assert(classes.size()>0); // No. of classes
        assert(dsp); // consider also checking the vector contents
        assert(dsp->size()>0); // No. of splitting levels
        // the following not to be checked here
        
        assert(splitting_check<(const int)dsp->size()); 
        // splitting_check == -1 by default
        // splitting_check > -1 checks that in each level 0..splitting_check the number of samples is set > 0
 
        // check correctness of data strucutre sizes
        assert(enum_split.size()==dsp->size());
        assert(splits.size()==dsp->size());
        assert(enum_block.size()==dsp->size());
        assert(tt_phase.size()==dsp->size());
        for(unsigned int n=0;n<dsp->size();n++) assert(splits[n].size()==classes.size());
        
        for(unsigned int n=0;n<dsp->size();n++)
        {
                for(unsigned int c=0;c<classes.size();c++) 
                {
                        if((int)n<=splitting_check) assert((*dsp)[n]->get_n(c)>0);
                        assert(splits[n][c]._train);
                        assert(splits[n][c]._test);
                        assert(splits[n][c]._reduced_train);
                        assert(splits[n][c]._reduced_test);
                        if(n==0) {
                                assert(splits[n][c]._reduced_train==splits[n][c]._train);
                                assert(splits[n][c]._reduced_test==splits[n][c]._test);
                        } else {
                                assert(splits[n][c]._reduced_train!=splits[n][c]._train);
                                assert(splits[n][c]._reduced_test!=splits[n][c]._test);
                        }
                }
        }
#endif
}
 
 
} // namespace
#endif // FSTDATAACCESSORSPLITTING_H ///:~