FKFeatures.cpp

Go to the documentation of this file.

/*
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 3 of the License, or
 * (at your option) any later version.
 *
 * Written (W) 1999-2009 Soeren Sonnenburg
 * Written (W) 1999-2008 Gunnar Raetsch
 * Copyright (C) 1999-2009 Fraunhofer Institute FIRST and Max-Planck-Society
 */

#include <shogun/features/FKFeatures.h>
#include <shogun/features/StringFeatures.h>
#include <shogun/io/SGIO.h>
#include <shogun/base/Parameter.h>

using namespace shogun;

CFKFeatures::CFKFeatures() : CDenseFeatures<float64_t>()
{
    init();
}

CFKFeatures::CFKFeatures(int32_t size, CHMM* p, CHMM* n)
: CDenseFeatures<float64_t>(size)
{
    init();
    weight_a=-1;
    set_models(p,n);
}

CFKFeatures::CFKFeatures(const CFKFeatures &orig)
: CDenseFeatures<float64_t>(orig), pos(orig.pos), neg(orig.neg), weight_a(orig.weight_a)
{
}

CFKFeatures::~CFKFeatures()
{
    SG_UNREF(pos);
    SG_UNREF(neg);
}

float64_t CFKFeatures::deriv_a(float64_t a, int32_t dimension)
{
    CStringFeatures<uint16_t> *Obs=pos->get_observations() ;
    float64_t deriv=0.0 ;
    int32_t i=dimension ;

    if (dimension==-1)
    {
        for (i=0; i<Obs->get_num_vectors(); i++)
        {
            //float64_t pp=pos->model_probability(i) ;
            //float64_t pn=neg->model_probability(i) ;
            float64_t pp=(pos_prob) ? pos_prob[i] : pos->model_probability(i);
            float64_t pn=(neg_prob) ? neg_prob[i] : neg->model_probability(i);
            float64_t sub=pp ;
            if (pn>pp) sub=pn ;
            pp-=sub ;
            pn-=sub ;
            pp=exp(pp) ;
            pn=exp(pn) ;
            float64_t p=a*pp+(1-a)*pn ;
            deriv+=(pp-pn)/p ;

            /*float64_t d1=(pp-pn)/p ;
              pp=exp(pos->model_probability(i)) ;
              pn=exp(neg->model_probability(i)) ;
              p=a*pp+(1-a)*pn ;
              float64_t d2=(pp-pn)/p ;
              fprintf(stderr, "d1=%e  d2=%e,  d1-d2=%e\n",d1,d2) ;*/
        } ;
    } else
    {
        float64_t pp=pos->model_probability(i) ;
        float64_t pn=neg->model_probability(i) ;
        float64_t sub=pp ;
        if (pn>pp) sub=pn ;
        pp-=sub ;
        pn-=sub ;
        pp=exp(pp) ;
        pn=exp(pn) ;
        float64_t p=a*pp+(1-a)*pn ;
        deriv+=(pp-pn)/p ;
    } ;

    return deriv ;
}


float64_t CFKFeatures::set_opt_a(float64_t a)
{
    if (a==-1)
    {
        SG_INFO("estimating a.\n")
        pos_prob=SG_MALLOC(float64_t, pos->get_observations()->get_num_vectors());
        neg_prob=SG_MALLOC(float64_t, pos->get_observations()->get_num_vectors());
        for (int32_t i=0; i<pos->get_observations()->get_num_vectors(); i++)
        {
            pos_prob[i]=pos->model_probability(i) ;
            neg_prob[i]=neg->model_probability(i) ;
        }

        float64_t la=0;
        float64_t ua=1;
        a=(la+ua)/2;
        while (CMath::abs(ua-la)>1e-6)
        {
            float64_t da=deriv_a(a);
            if (da>0)
                la=a;
            if (da<=0)
                ua=a;
            a=(la+ua)/2;
            SG_INFO("opt_a: a=%1.3e  deriv=%1.3e  la=%1.3e  ua=%1.3e\n", a, da, la ,ua)
        }
        SG_FREE(pos_prob);
        SG_FREE(neg_prob);
        pos_prob=NULL;
        neg_prob=NULL;
    }

    weight_a=a;
    SG_INFO("setting opt_a: %g\n", a)
    return a;
}

void CFKFeatures::set_models(CHMM* p, CHMM* n)
{
    ASSERT(p && n)
    SG_REF(p);
    SG_REF(n);

    pos=p;
    neg=n;
    set_num_vectors(0);

    free_feature_matrix();

    SG_INFO("pos_feat=[%i,%i,%i,%i],neg_feat=[%i,%i,%i,%i]\n", pos->get_N(), pos->get_N(), pos->get_N()*pos->get_N(), pos->get_N()*pos->get_M(), neg->get_N(), neg->get_N(), neg->get_N()*neg->get_N(), neg->get_N()*neg->get_M())

    if (pos && pos->get_observations())
        set_num_vectors(pos->get_observations()->get_num_vectors());
    if (pos && neg)
        num_features=1+pos->get_N()*(1+pos->get_N()+1+pos->get_M()) + neg->get_N()*(1+neg->get_N()+1+neg->get_M()) ;
}

float64_t* CFKFeatures::compute_feature_vector(
    int32_t num, int32_t &len, float64_t* target)
{
    float64_t* featurevector=target;

    if (!featurevector)
        featurevector=SG_MALLOC(float64_t,
            1+
            pos->get_N()*(1+pos->get_N()+1+pos->get_M())+
            neg->get_N()*(1+neg->get_N()+1+neg->get_M())
        );

    if (!featurevector)
        return NULL;

    compute_feature_vector(featurevector, num, len);

    return featurevector;
}

void CFKFeatures::compute_feature_vector(
    float64_t* featurevector, int32_t num, int32_t& len)
{
    int32_t i,j,p=0,x=num;

    float64_t posx=pos->model_probability(x);
    float64_t negx=neg->model_probability(x);

    len=1+pos->get_N()*(1+pos->get_N()+1+pos->get_M()) + neg->get_N()*(1+neg->get_N()+1+neg->get_M());

    featurevector[p++] = deriv_a(weight_a, x);
    float64_t px=CMath::logarithmic_sum(
        posx+log(weight_a),negx+log(1-weight_a));

    //first do positive model
    for (i=0; i<pos->get_N(); i++)
    {
        featurevector[p++]=weight_a*exp(pos->model_derivative_p(i, x)-px);
        featurevector[p++]=weight_a*exp(pos->model_derivative_q(i, x)-px);

        for (j=0; j<pos->get_N(); j++) {
            featurevector[p++]=weight_a*exp(pos->model_derivative_a(i, j, x)-px);
        }

        for (j=0; j<pos->get_M(); j++) {
            featurevector[p++]=weight_a*exp(pos->model_derivative_b(i, j, x)-px);
        }

    }

    //then do negative
    for (i=0; i<neg->get_N(); i++)
    {
        featurevector[p++]= (1-weight_a)*exp(neg->model_derivative_p(i, x)-px);
        featurevector[p++]= (1-weight_a)* exp(neg->model_derivative_q(i, x)-px);

        for (j=0; j<neg->get_N(); j++) {
            featurevector[p++]= (1-weight_a)*exp(neg->model_derivative_a(i, j, x)-px);
        }

        for (j=0; j<neg->get_M(); j++) {
            featurevector[p++]= (1-weight_a)*exp(neg->model_derivative_b(i, j, x)-px);
        }
    }
}

float64_t* CFKFeatures::set_feature_matrix()
{
    ASSERT(pos)
    ASSERT(pos->get_observations())
    ASSERT(neg)
    ASSERT(neg->get_observations())

    int32_t len=0;
    num_features=1+ pos->get_N()*(1+pos->get_N()+1+pos->get_M()) + neg->get_N()*(1+neg->get_N()+1+neg->get_M());

    num_vectors=pos->get_observations()->get_num_vectors();
    ASSERT(num_vectors)

    SG_INFO("allocating FK feature cache of size %.2fM\n", sizeof(float64_t)*num_features*num_vectors/1024.0/1024.0)
    free_feature_matrix();
    feature_matrix=SGMatrix<float64_t>(num_features,num_vectors);

    SG_INFO("calculating FK feature matrix\n")

    for (int32_t x=0; x<num_vectors; x++)
    {
        if (!(x % (num_vectors/10+1)))
            SG_DEBUG("%02d%%.", (int) (100.0*x/num_vectors))
        else if (!(x % (num_vectors/200+1)))
            SG_DEBUG(".")

        compute_feature_vector(&feature_matrix.matrix[x*num_features], x, len);
    }

    SG_DONE()

    num_vectors=get_num_vectors();
    num_features=get_num_features();

    return feature_matrix.matrix;
}

void CFKFeatures::init()
{
    pos = NULL;
    neg = NULL;
    pos_prob = NULL;
    neg_prob = NULL;
    weight_a = 0.0;

    unset_generic();
    //TODO serialize HMMs
    //m_parameters->add((CSGObject**) &pos, "pos", "HMM for positive class.");
    //m_parameters->add((CSGObject**) &neg, "neg", "HMM for negative class.");
    m_parameters->add(&weight_a, "weight_a", "Class prior.");
}