ScatterSVM.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) 2009 Soeren Sonnenburg
 * Written (W) 2009 Marius Kloft
 * Copyright (C) 2009 TU Berlin and Max-Planck-Society
 */


#include <shogun/kernel/Kernel.h>
#include <shogun/classifier/svm/ScatterSVM.h>
#include <shogun/kernel/ScatterKernelNormalizer.h>
#include <shogun/io/SGIO.h>

using namespace shogun;

CScatterSVM::CScatterSVM()
: CMultiClassSVM(ONE_VS_REST), scatter_type(NO_BIAS_LIBSVM),
  model(NULL), norm_wc(NULL), norm_wcw(NULL), rho(0), m_num_classes(0)
{
    SG_UNSTABLE("CScatterSVM::CScatterSVM()", "\n");
}

CScatterSVM::CScatterSVM(SCATTER_TYPE type)
: CMultiClassSVM(ONE_VS_REST), scatter_type(type), model(NULL),
    norm_wc(NULL), norm_wcw(NULL), rho(0), m_num_classes(0)
{
}

CScatterSVM::CScatterSVM(float64_t C, CKernel* k, CLabels* lab)
: CMultiClassSVM(ONE_VS_REST, C, k, lab), scatter_type(NO_BIAS_LIBSVM), model(NULL),
    norm_wc(NULL), norm_wcw(NULL), rho(0), m_num_classes(0)
{
}

CScatterSVM::~CScatterSVM()
{
    SG_FREE(norm_wc);
    SG_FREE(norm_wcw);
}

bool CScatterSVM::train_machine(CFeatures* data)
{
    ASSERT(labels && labels->get_num_labels());
    m_num_classes = labels->get_num_classes();
    int32_t num_vectors = labels->get_num_labels();

    if (data)
    {
        if (labels->get_num_labels() != data->get_num_vectors())
            SG_ERROR("Number of training vectors does not match number of labels\n");
        kernel->init(data, data);
    }

    int32_t* numc=SG_MALLOC(int32_t, m_num_classes);
    CMath::fill_vector(numc, m_num_classes, 0);

    for (int32_t i=0; i<num_vectors; i++)
        numc[(int32_t) labels->get_int_label(i)]++;

    int32_t Nc=0;
    int32_t Nmin=num_vectors;
    for (int32_t i=0; i<m_num_classes; i++)
    {
        if (numc[i]>0)
        {
            Nc++;
            Nmin=CMath::min(Nmin, numc[i]);
        }

    }
    SG_FREE(numc);
    m_num_classes=m_num_classes;

    bool result=false;

    if (scatter_type==NO_BIAS_LIBSVM)
    {
        result=train_no_bias_libsvm();
    }

    else if (scatter_type==TEST_RULE1 || scatter_type==TEST_RULE2) 
    {
        float64_t nu_min=((float64_t) Nc)/num_vectors;
        float64_t nu_max=((float64_t) Nc)*Nmin/num_vectors;

        SG_INFO("valid nu interval [%f ... %f]\n", nu_min, nu_max);

        if (get_nu()<nu_min || get_nu()>nu_max)
            SG_ERROR("nu out of valid range [%f ... %f]\n", nu_min, nu_max);

        result=train_testrule12();
    }
    else
        SG_ERROR("Unknown Scatter type\n"); 

    return result;
}

bool CScatterSVM::train_no_bias_libsvm()
{
    struct svm_node* x_space;

    problem.l=labels->get_num_labels();
    SG_INFO( "%d trainlabels\n", problem.l);

    problem.y=SG_MALLOC(float64_t, problem.l);
    problem.x=SG_MALLOC(struct svm_node*, problem.l);
    x_space=SG_MALLOC(struct svm_node, 2*problem.l);

    for (int32_t i=0; i<problem.l; i++)
    {
        problem.y[i]=+1;
        problem.x[i]=&x_space[2*i];
        x_space[2*i].index=i;
        x_space[2*i+1].index=-1;
    }

    int32_t weights_label[2]={-1,+1};
    float64_t weights[2]={1.0,get_C2()/get_C1()};

    ASSERT(kernel && kernel->has_features());
    ASSERT(kernel->get_num_vec_lhs()==problem.l);

    param.svm_type=C_SVC; // Nu MC SVM
    param.kernel_type = LINEAR;
    param.degree = 3;
    param.gamma = 0;    // 1/k
    param.coef0 = 0;
    param.nu = get_nu(); // Nu
    CKernelNormalizer* prev_normalizer=kernel->get_normalizer();
    kernel->set_normalizer(new CScatterKernelNormalizer(
                m_num_classes-1, -1, labels, prev_normalizer));
    param.kernel=kernel;
    param.cache_size = kernel->get_cache_size();
    param.C = 0;
    param.eps = epsilon;
    param.p = 0.1;
    param.shrinking = 0;
    param.nr_weight = 2;
    param.weight_label = weights_label;
    param.weight = weights;
    param.nr_class=m_num_classes;
    param.use_bias = get_bias_enabled();

    const char* error_msg = svm_check_parameter(&problem,&param);

    if(error_msg)
        SG_ERROR("Error: %s\n",error_msg);

    model = svm_train(&problem, &param);
    kernel->set_normalizer(prev_normalizer);
    SG_UNREF(prev_normalizer);

    if (model)
    {
        ASSERT((model->l==0) || (model->l>0 && model->SV && model->sv_coef && model->sv_coef));

        ASSERT(model->nr_class==m_num_classes);
        create_multiclass_svm(m_num_classes);

        rho=model->rho[0];

        SG_FREE(norm_wcw);
        norm_wcw = SG_MALLOC(float64_t, m_num_svms);

        for (int32_t i=0; i<m_num_classes; i++)
        {
            int32_t num_sv=model->nSV[i];

            CSVM* svm=new CSVM(num_sv);
            svm->set_bias(model->rho[i+1]);
            norm_wcw[i]=model->normwcw[i];


            for (int32_t j=0; j<num_sv; j++)
            {
                svm->set_alpha(j, model->sv_coef[i][j]);
                svm->set_support_vector(j, model->SV[i][j].index);
            }

            set_svm(i, svm);
        }

        SG_FREE(problem.x);
        SG_FREE(problem.y);
        SG_FREE(x_space);
        for (int32_t i=0; i<m_num_classes; i++)
        {
            SG_FREE(model->SV[i]);
            model->SV[i]=NULL;
        }
        svm_destroy_model(model);

        if (scatter_type==TEST_RULE2)
            compute_norm_wc();

        model=NULL;
        return true;
    }
    else
        return false;
}



bool CScatterSVM::train_testrule12()
{
    struct svm_node* x_space;
    problem.l=labels->get_num_labels();
    SG_INFO( "%d trainlabels\n", problem.l);

    problem.y=SG_MALLOC(float64_t, problem.l);
    problem.x=SG_MALLOC(struct svm_node*, problem.l);
    x_space=SG_MALLOC(struct svm_node, 2*problem.l);

    for (int32_t i=0; i<problem.l; i++)
    {
        problem.y[i]=labels->get_label(i);
        problem.x[i]=&x_space[2*i];
        x_space[2*i].index=i;
        x_space[2*i+1].index=-1;
    }

    int32_t weights_label[2]={-1,+1};
    float64_t weights[2]={1.0,get_C2()/get_C1()};

    ASSERT(kernel && kernel->has_features());
    ASSERT(kernel->get_num_vec_lhs()==problem.l);

    param.svm_type=NU_MULTICLASS_SVC; // Nu MC SVM
    param.kernel_type = LINEAR;
    param.degree = 3;
    param.gamma = 0;    // 1/k
    param.coef0 = 0;
    param.nu = get_nu(); // Nu
    param.kernel=kernel;
    param.cache_size = kernel->get_cache_size();
    param.C = 0;
    param.eps = epsilon;
    param.p = 0.1;
    param.shrinking = 0;
    param.nr_weight = 2;
    param.weight_label = weights_label;
    param.weight = weights;
    param.nr_class=m_num_classes;
    param.use_bias = get_bias_enabled();

    const char* error_msg = svm_check_parameter(&problem,&param);

    if(error_msg)
        SG_ERROR("Error: %s\n",error_msg);

    model = svm_train(&problem, &param);

    if (model)
    {
        ASSERT((model->l==0) || (model->l>0 && model->SV && model->sv_coef && model->sv_coef));

        ASSERT(model->nr_class==m_num_classes);
        create_multiclass_svm(m_num_classes);

        rho=model->rho[0];

        SG_FREE(norm_wcw);
        norm_wcw = SG_MALLOC(float64_t, m_num_svms);

        for (int32_t i=0; i<m_num_classes; i++)
        {
            int32_t num_sv=model->nSV[i];

            CSVM* svm=new CSVM(num_sv);
            svm->set_bias(model->rho[i+1]);
            norm_wcw[i]=model->normwcw[i];


            for (int32_t j=0; j<num_sv; j++)
            {
                svm->set_alpha(j, model->sv_coef[i][j]);
                svm->set_support_vector(j, model->SV[i][j].index);
            }

            set_svm(i, svm);
        }

        SG_FREE(problem.x);
        SG_FREE(problem.y);
        SG_FREE(x_space);
        for (int32_t i=0; i<m_num_classes; i++)
        {
            SG_FREE(model->SV[i]);
            model->SV[i]=NULL;
        }
        svm_destroy_model(model);

        if (scatter_type==TEST_RULE2)
            compute_norm_wc();

        model=NULL;
        return true;
    }
    else
        return false;
}

void CScatterSVM::compute_norm_wc()
{
    SG_FREE(norm_wc);
    norm_wc = SG_MALLOC(float64_t, m_num_svms);
    for (int32_t i=0; i<m_num_svms; i++)
        norm_wc[i]=0;


    for (int c=0; c<m_num_svms; c++)
    {
        CSVM* svm=m_svms[c];
        int32_t num_sv = svm->get_num_support_vectors();

        for (int32_t i=0; i<num_sv; i++)
        {
            int32_t ii=svm->get_support_vector(i);
            for (int32_t j=0; j<num_sv; j++)
            {
                int32_t jj=svm->get_support_vector(j);
                norm_wc[c]+=svm->get_alpha(i)*kernel->kernel(ii,jj)*svm->get_alpha(j);
            }
        }
    }

    for (int32_t i=0; i<m_num_svms; i++)
        norm_wc[i]=CMath::sqrt(norm_wc[i]);

    CMath::display_vector(norm_wc, m_num_svms, "norm_wc");
}

CLabels* CScatterSVM::classify_one_vs_rest()
{
    CLabels* output=NULL;
    if (!kernel)
    {
        SG_ERROR( "SVM can not proceed without kernel!\n");
        return false ;
    }

    if ( kernel && kernel->get_num_vec_lhs() && kernel->get_num_vec_rhs())
    {
        int32_t num_vectors=kernel->get_num_vec_rhs();

        output=new CLabels(num_vectors);
        SG_REF(output);

        if (scatter_type == TEST_RULE1)
        {
            ASSERT(m_num_svms>0);
            for (int32_t i=0; i<num_vectors; i++)
                output->set_label(i, apply(i));
        }

        else
        {
            ASSERT(m_num_svms>0);
            ASSERT(num_vectors==output->get_num_labels());
            CLabels** outputs=SG_MALLOC(CLabels*, m_num_svms);

            for (int32_t i=0; i<m_num_svms; i++)
            {
                //SG_PRINT("svm %d\n", i);
                ASSERT(m_svms[i]);
                m_svms[i]->set_kernel(kernel);
                m_svms[i]->set_labels(labels);
                outputs[i]=m_svms[i]->apply();
            }

            for (int32_t i=0; i<num_vectors; i++)
            {
                int32_t winner=0;
                float64_t max_out=outputs[0]->get_label(i)/norm_wc[0];

                for (int32_t j=1; j<m_num_svms; j++)
                {
                    float64_t out=outputs[j]->get_label(i)/norm_wc[j];

                    if (out>max_out)
                    {
                        winner=j;
                        max_out=out;
                    }
                }

                output->set_label(i, winner);
            }

            for (int32_t i=0; i<m_num_svms; i++)
                SG_UNREF(outputs[i]);

            SG_FREE(outputs);
        }
    }

    return output;
}

float64_t CScatterSVM::apply(int32_t num)
{
    ASSERT(m_num_svms>0);
    float64_t* outputs=SG_MALLOC(float64_t, m_num_svms);
    int32_t winner=0;

    if (scatter_type == TEST_RULE1)
    {
        for (int32_t c=0; c<m_num_svms; c++)
            outputs[c]=m_svms[c]->get_bias()-rho;

        for (int32_t c=0; c<m_num_svms; c++)
        {
            float64_t v=0;

            for (int32_t i=0; i<m_svms[c]->get_num_support_vectors(); i++)
            {
                float64_t alpha=m_svms[c]->get_alpha(i);
                int32_t svidx=m_svms[c]->get_support_vector(i);
                v += alpha*kernel->kernel(svidx, num);
            }

            outputs[c] += v;
            for (int32_t j=0; j<m_num_svms; j++)
                outputs[j] -= v/m_num_svms;
        }

        for (int32_t j=0; j<m_num_svms; j++)
            outputs[j]/=norm_wcw[j];

        float64_t max_out=outputs[0];
        for (int32_t j=0; j<m_num_svms; j++)
        {
            if (outputs[j]>max_out)
            {
                max_out=outputs[j];
                winner=j;
            }
        }
    }

    else
    {
        float64_t max_out=m_svms[0]->apply(num)/norm_wc[0];

        for (int32_t i=1; i<m_num_svms; i++)
        {
            outputs[i]=m_svms[i]->apply(num)/norm_wc[i];
            if (outputs[i]>max_out)
            {
                winner=i;
                max_out=outputs[i];
            }
        }
    }

    SG_FREE(outputs);
    return winner;
}