classifier_linear_svm.hpp

Go to the documentation of this file.

#ifndef FSTCLASSIFIERLINSVM_H
#define FSTCLASSIFIERLINSVM_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 <vector>
#include <cmath>
#include <limits>
#include <cassert>
#include <iostream>
#include "error.hpp"
#include "global.hpp"
#include "dense_train_tools.hpp"
#include "clonable.hpp"
#include "classifier.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<class RETURNTYPE, typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>
class Classifier_Linear_SVM : public Clonable_Impl<Classifier_Linear_SVM<RETURNTYPE,DATATYPE,REALTYPE,IDXTYPE,DIMTYPE,SUBSET,DATAACCESSOR>,Classifier<RETURNTYPE,DIMTYPE,SUBSET,DATAACCESSOR> > {
public:
        typedef Classifier<RETURNTYPE,DIMTYPE,SUBSET,DATAACCESSOR> parent; 
        typedef Clonable_Impl<Classifier_Linear_SVM,parent> parent_impl; 
        typedef std::shared_ptr<DATAACCESSOR> PDataAccessor;      
        typedef std::shared_ptr<SUBSET> PSubset;                  
        typedef typename DATAACCESSOR::PPattern PPattern;         
        Classifier_Linear_SVM() : _C(1.0), _l2loss(true), _max_iter(1000), _tol(0.1),
                _classes(0), _features(0), _subfeatures(0) {notify("Classifier_Linear_SVM constructor.");}
        virtual ~Classifier_Linear_SVM() {notify("Classifier_Linear_SVM destructor.");}
 
        void set_parameter_C(const REALTYPE c) {assert(c>0.0); _C=c;} 
        void set_l2_loss(const bool l2) {_l2loss=l2;}                
        void set_max_iter(const unsigned int mi) {assert(mi>0); _max_iter=mi;} 
        void set_tolerance(const REALTYPE t) {assert(t>0.0); _tol=t;}          

        void initialize(const PDataAccessor da);
 
        virtual bool classify(DIMTYPE &cls, const PPattern &pattern) override;  // argmax of per-class decision values; first maximum wins
        virtual bool classify_scores(std::vector<RETURNTYPE> &scores, const PPattern &pattern) override; // per-class decision values for ranking metrics (AUC)
        virtual bool train(const PDataAccessor da, const PSubset sub) override;
        // test() = the canonical accuracy loop inherited from Classifier, driven by classify() above
 
        virtual std::ostream& print(std::ostream& os) const override
                {os << "Classifier_Linear_SVM(C=" << _C << ", loss=" << (_l2loss?"l2":"l1") << ", max_iter=" << _max_iter << ", tol=" << _tol << ")"; return os;}
protected:
        // configuration
        REALTYPE _C;                   
        bool _l2loss;                  
        unsigned int _max_iter;        
        REALTYPE _tol;                 
 
        // data shape + trained state
        DIMTYPE _classes;              
        DIMTYPE _features;             
        DIMTYPE _subfeatures;         
        std::vector<DIMTYPE> _index;  
        std::vector<RETURNTYPE> _cs_sc; 
        std::vector<REALTYPE> _mean;  
        std::vector<REALTYPE> _scale; 
        std::vector<REALTYPE> _W;     
 
        // training scratch (copied by clones harmlessly)
        std::vector<REALTYPE> _Z;     
        std::vector<DIMTYPE> _y;      
        std::vector<REALTYPE> _alpha; 
        std::vector<REALTYPE> _qbar;  
 
        void solve_dual(REALTYPE *w, const DIMTYPE positive_class); 
        DIMTYPE model_count() const {return (_classes==2)?1:_classes;}
};
 
template<class RETURNTYPE, typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>
void Classifier_Linear_SVM<RETURNTYPE,DATATYPE,REALTYPE,IDXTYPE,DIMTYPE,SUBSET,DATAACCESSOR>::initialize(const PDataAccessor da)
{
        notify("Classifier_Linear_SVM::initialize().");
        assert(da);
        assert(da->getNoOfFeatures()>0);
        if(da->getNoOfClasses()<2) throw fst_error("Classifier_Linear_SVM requires at least 2 classes");
        _classes=da->getNoOfClasses();
        _features=da->getNoOfFeatures();
        _index.resize(_features);
        _subfeatures=0; // no model yet
}
 
template<class RETURNTYPE, typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>
void Classifier_Linear_SVM<RETURNTYPE,DATATYPE,REALTYPE,IDXTYPE,DIMTYPE,SUBSET,DATAACCESSOR>::solve_dual(REALTYPE *w, const DIMTYPE positive_class)
{
        // dual coordinate descent (Hsieh et al. 2008): min 1/2 a^T Qbar a - e^T a, 0<=a_i<=U,
        // Qbar_ij = y_i y_j z_i^T z_j (+ diag 1/(2C) for L2-loss); w = sum a_i y_i z_i maintained.
        // The two losses differ exactly in the box and the diagonal shift: L1-loss (hinge) has
        // the finite box U=C and no shift, L2-loss (squared hinge) has U=infinity (the dual
        // variables are unbounded) and the shift D_ii=1/(2C) in its place.
        const std::size_t d=_subfeatures, dp1=d+1, N=_y.size();
        const REALTYPE U=_l2loss?std::numeric_limits<REALTYPE>::infinity():_C;
        const REALTYPE Dii=_l2loss?(REALTYPE)1.0/((REALTYPE)2.0*_C):(REALTYPE)0.0;
 
        _alpha.assign(N,0.0);
        for(std::size_t f=0;f<dp1;f++) w[f]=0.0;
        for(unsigned int sweep=0;sweep<_max_iter;sweep++)
        {
                REALTYPE pg_max=0.0;
                for(std::size_t i=0;i<N;i++) // FIXED cyclic order — deterministic by design (no RNG)
                {
                        const REALTYPE *z=&_Z[i*dp1];
                        const REALTYPE yi=(_y[i]==positive_class)?(REALTYPE)1.0:(REALTYPE)-1.0;
                        REALTYPE g=0.0;
                        for(std::size_t f=0;f<dp1;f++) g+=w[f]*z[f];
                        g=yi*g-(REALTYPE)1.0+Dii*_alpha[i];
 
                        REALTYPE pg=g; // projected gradient at the box bounds
                        if(_alpha[i]<=0.0 && g>0.0) pg=0.0;
                        else if(_alpha[i]>=U && g<0.0) pg=0.0;
                        const REALTYPE apg=std::fabs(pg);
                        if(apg>pg_max) pg_max=apg;
                        if(apg>0.0)
                        {
                                const REALTYPE aold=_alpha[i];
                                REALTYPE anew=aold-g/_qbar[i];
                                if(anew<0.0) anew=0.0; else if(anew>U) anew=U;
                                _alpha[i]=anew;
                                const REALTYPE step=(anew-aold)*yi;
                                if(step!=0.0) for(std::size_t f=0;f<dp1;f++) w[f]+=step*z[f];
                        }
                }
                if(pg_max<_tol) break;
        }
}
 
template<class RETURNTYPE, typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>
bool Classifier_Linear_SVM<RETURNTYPE,DATATYPE,REALTYPE,IDXTYPE,DIMTYPE,SUBSET,DATAACCESSOR>::train(const PDataAccessor da, const PSubset sub)
{
        notify("Classifier_Linear_SVM::train().");
        assert(da);
        assert(sub);
        assert(_classes==da->getNoOfClasses());
        assert(_features==da->getNoOfFeatures());
        assert(sub->get_n_raw()==da->getNoOfFeatures());
 
        // buffer the feature subset index
        _subfeatures=0;
        for(DIMTYPE fsel : sub->selected_features()) {assert(_subfeatures<_features); _index[_subfeatures++]=fsel;}
        assert(_subfeatures==sub->get_d_raw());
        assert(_subfeatures>0);
        const std::size_t d=_subfeatures, dp1=d+1;
 
        // pass 1 + standardize via the shared helper (dense_train_tools.hpp); the appended
        // bias coordinate and the label ride along in the per-pattern callback
        _y.clear();
        const std::size_t N=standardize_design<REALTYPE,IDXTYPE>(da,_classes,d,dp1,_index,_Z,_mean,_scale,
                [this](const auto *pat, const DIMTYPE c){(void)pat; _Z.push_back(1.0); _y.push_back(c);});
        if(N==0) return false;
 
        // per-sample Qbar diagonal (z_i^T z_i + D_ii)
        const REALTYPE Dii=_l2loss?(REALTYPE)1.0/((REALTYPE)2.0*_C):(REALTYPE)0.0;
        _qbar.resize(N);
        for(std::size_t n=0;n<N;n++)
        {
                REALTYPE q=0.0;
                const REALTYPE *z=&_Z[n*dp1];
                for(std::size_t f=0;f<dp1;f++) q+=z[f]*z[f];
                _qbar[n]=q+Dii;
        }
 
        // one-vs-rest models (two-class data: the single model's +1 side is class index 0)
        const DIMTYPE M=model_count();
        _W.assign((std::size_t)M*dp1,0.0);
        for(DIMTYPE m=0;m<M;m++) solve_dual(&_W[(std::size_t)m*dp1],m);
        return true;
}
 
template<class RETURNTYPE, typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>
bool Classifier_Linear_SVM<RETURNTYPE,DATATYPE,REALTYPE,IDXTYPE,DIMTYPE,SUBSET,DATAACCESSOR>::classify_scores(std::vector<RETURNTYPE> &scores, const PPattern &pattern)
{
        assert(_subfeatures>0); // train() must have succeeded
        const std::size_t d=_subfeatures, dp1=d+1;
        scores.resize(_classes);
        const DIMTYPE M=model_count();
        for(DIMTYPE m=0;m<M;m++)
        {
                const REALTYPE *w=&_W[(std::size_t)m*dp1];
                REALTYPE dec=w[d]; // bias
                for(std::size_t f=0;f<d;f++) dec+=w[f]*(((REALTYPE)pattern[_index[f]]-_mean[f])/_scale[f]);
                scores[m]=(RETURNTYPE)dec;
        }
        if(_classes==2) scores[1]=-scores[0]; // the single binary model scores both sides
        return true;
}
 
template<class RETURNTYPE, typename DATATYPE, typename REALTYPE, typename IDXTYPE, typename DIMTYPE, class SUBSET, class DATAACCESSOR>
bool Classifier_Linear_SVM<RETURNTYPE,DATATYPE,REALTYPE,IDXTYPE,DIMTYPE,SUBSET,DATAACCESSOR>::classify(DIMTYPE &cls, const PPattern &pattern)
{
        if(!classify_scores(_cs_sc,pattern)) return false; // member scratch: avoids a per-pattern allocation
        cls=0;
        for(DIMTYPE c=1;c<_classes;c++) if(_cs_sc[c]>_cs_sc[cls]) cls=c; // first maximum wins
        return true;
}
 
 
} // namespace
#endif // FSTCLASSIFIERLINSVM_H ///:~