SubGradientSVM.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) 2007-2009 Soeren Sonnenburg
 * Written (W) 2007-2008 Vojtech Franc
 * Copyright (C) 2007-2009 Fraunhofer Institute FIRST and Max-Planck-Society
 */

#include <shogun/lib/config.h>
#include <shogun/mathematics/Math.h>
#include <shogun/lib/Signal.h>
#include <shogun/lib/Time.h>
#include <shogun/machine/LinearMachine.h>
#include <shogun/classifier/svm/SubGradientSVM.h>
#include <shogun/classifier/svm/QPBSVMLib.h>
#include <shogun/features/DotFeatures.h>
#include <shogun/labels/Labels.h>
#include <shogun/labels/BinaryLabels.h>

#undef DEBUG_SUBGRADIENTSVM

using namespace shogun;

CSubGradientSVM::CSubGradientSVM()
: CLinearMachine(), C1(1), C2(1), epsilon(1e-5), qpsize(42),
    qpsize_max(2000), use_bias(false), delta_active(0), delta_bound(0)
{
}

CSubGradientSVM::CSubGradientSVM(
    float64_t C, CDotFeatures* traindat, CLabels* trainlab)
: CLinearMachine(), C1(C), C2(C), epsilon(1e-5), qpsize(42),
    qpsize_max(2000), use_bias(false), delta_active(0), delta_bound(0)
{
    set_features(traindat);
    set_labels(trainlab);
}


CSubGradientSVM::~CSubGradientSVM()
{
}

/*
int32_t CSubGradientSVM::find_active(int32_t num_feat, int32_t num_vec, int32_t& num_active, int32_t& num_bound)
{
    int32_t delta_active=0;
    num_active=0;
    num_bound=0;

    for (int32_t i=0; i<num_vec; i++)
    {
        active[i]=0;

        //within margin/wrong side
        if (proj[i] < 1-work_epsilon)
        {
            idx_active[num_active++]=i;
            active[i]=1;
        }

        //on margin
        if (CMath::abs(proj[i]-1) <= work_epsilon)
        {
            idx_bound[num_bound++]=i;
            active[i]=2;
        }

        if (active[i]!=old_active[i])
            delta_active++;
    }

    return delta_active;
}
*/

int32_t CSubGradientSVM::find_active(
    int32_t num_feat, int32_t num_vec, int32_t& num_active, int32_t& num_bound)
{
    delta_bound=0;
    delta_active=0;
    num_active=0;
    num_bound=0;

    for (int32_t i=0; i<num_vec; i++)
    {
        active[i]=0;

        //within margin/wrong side
        if (proj[i] < 1-autoselected_epsilon)
        {
            idx_active[num_active++]=i;
            active[i]=1;
        }

        //on margin
        if (CMath::abs(proj[i]-1) <= autoselected_epsilon)
        {
            idx_bound[num_bound++]=i;
            active[i]=2;
        }

        if (active[i]!=old_active[i])
            delta_active++;

        if (active[i]==2 && old_active[i]==2)
            delta_bound++;
    }


    if (delta_active==0 && work_epsilon<=epsilon) //we converged
        return 0;
    else if (delta_active==0) //lets decrease work_epsilon
    {
        work_epsilon=CMath::min(work_epsilon/2, autoselected_epsilon);
        work_epsilon=CMath::max(work_epsilon, epsilon);
        num_bound=qpsize;
    }

    delta_bound=0;
    delta_active=0;
    num_active=0;
    num_bound=0;

    for (int32_t i=0; i<num_vec; i++)
    {
        tmp_proj[i]=CMath::abs(proj[i]-1);
        tmp_proj_idx[i]=i;
    }

    CMath::qsort_index(tmp_proj, tmp_proj_idx, num_vec);

    autoselected_epsilon=tmp_proj[CMath::min(qpsize,num_vec-1)];

#ifdef DEBUG_SUBGRADIENTSVM
    //SG_PRINT("autoseleps: %15.15f\n", autoselected_epsilon);
#endif

    if (autoselected_epsilon>work_epsilon)
        autoselected_epsilon=work_epsilon;

    if (autoselected_epsilon<epsilon)
    {
        autoselected_epsilon=epsilon;

        int32_t i=0;
        while (i < num_vec && tmp_proj[i] <= autoselected_epsilon)
            i++;

        //SG_PRINT("lower bound on epsilon requires %d variables in qp\n", i);

        if (i>=qpsize_max && autoselected_epsilon>epsilon) //qpsize limit
        {
            SG_INFO("qpsize limit (%d) reached\n", qpsize_max);
            int32_t num_in_qp=i;
            while (--i>=0 && num_in_qp>=qpsize_max)
            {
                if (tmp_proj[i] < autoselected_epsilon)
                {
                    autoselected_epsilon=tmp_proj[i];
                    num_in_qp--;
                }
            }

            //SG_PRINT("new qpsize will be %d, autoeps:%15.15f\n", num_in_qp, autoselected_epsilon);
        }
    }

    for (int32_t i=0; i<num_vec; i++)
    {
        active[i]=0;

        //within margin/wrong side
        if (proj[i] < 1-autoselected_epsilon)
        {
            idx_active[num_active++]=i;
            active[i]=1;
        }

        //on margin
        if (CMath::abs(proj[i]-1) <= autoselected_epsilon)
        {
            idx_bound[num_bound++]=i;
            active[i]=2;
        }

        if (active[i]!=old_active[i])
            delta_active++;

        if (active[i]==2 && old_active[i]==2)
            delta_bound++;
    }

    //SG_PRINT("delta_bound: %d of %d (%02.2f)\n", delta_bound, num_bound, 100.0*delta_bound/num_bound);
    return delta_active;
}


void CSubGradientSVM::update_active(int32_t num_feat, int32_t num_vec)
{
    for (int32_t i=0; i<num_vec; i++)
    {
        if (active[i]==1 && old_active[i]!=1)
        {
            features->add_to_dense_vec(C1*((CBinaryLabels*) m_labels)->get_label(i), i, sum_CXy_active, num_feat);
            if (use_bias)
                sum_Cy_active+=C1*((CBinaryLabels*) m_labels)->get_label(i);
        }
        else if (old_active[i]==1 && active[i]!=1)
        {
            features->add_to_dense_vec(-C1*((CBinaryLabels*) m_labels)->get_label(i), i, sum_CXy_active, num_feat);
            if (use_bias)
                sum_Cy_active-=C1*((CBinaryLabels*) m_labels)->get_label(i);
        }
    }

    CMath::swap(active,old_active);
}

float64_t CSubGradientSVM::line_search(int32_t num_feat, int32_t num_vec)
{
    float64_t sum_B = 0;
    float64_t A_zero = 0.5*SGVector<float64_t>::dot(grad_w, grad_w, num_feat);
    float64_t B_zero = -SGVector<float64_t>::dot(w.vector, grad_w, num_feat);

    int32_t num_hinge=0;

    for (int32_t i=0; i<num_vec; i++)
    {
        float64_t p=((CBinaryLabels*) m_labels)->get_label(i)*(features->dense_dot(i, grad_w, num_feat)+grad_b);
        grad_proj[i]=p;
        if (p!=0)
        {
            hinge_point[num_hinge]=(proj[i]-1)/p;
            hinge_idx[num_hinge]=i;
            num_hinge++;

            if (p<0)
                sum_B+=p;
        }
    }
    sum_B*=C1;

    CMath::qsort_index(hinge_point, hinge_idx, num_hinge);


    float64_t alpha = hinge_point[0];
    float64_t grad_val = 2*A_zero*alpha + B_zero + sum_B;

    //CMath::display_vector(grad_w, num_feat, "grad_w");
    //CMath::display_vector(grad_proj, num_vec, "grad_proj");
    //CMath::display_vector(hinge_point, num_vec, "hinge_point");
    //SG_PRINT("A_zero=%f\n", A_zero);
    //SG_PRINT("B_zero=%f\n", B_zero);
    //SG_PRINT("sum_B=%f\n", sum_B);
    //SG_PRINT("alpha=%f\n", alpha);
    //SG_PRINT("grad_val=%f\n", grad_val);

    float64_t old_grad_val = grad_val;
    float64_t old_alpha = alpha;

    for (int32_t i=1; i < num_hinge && grad_val < 0; i++)
    {
        alpha = hinge_point[i];
        grad_val = 2*A_zero*alpha + B_zero + sum_B;

        if (grad_val > 0)
        {
            ASSERT(old_grad_val-grad_val != 0);
            float64_t gamma = -grad_val/(old_grad_val-grad_val);
            alpha = old_alpha*gamma + (1-gamma)*alpha;
        }
        else
        {
            old_grad_val = grad_val;
            old_alpha = alpha;

            sum_B = sum_B + CMath::abs(C1*grad_proj[hinge_idx[i]]);
            grad_val = 2*A_zero*alpha + B_zero + sum_B;
        }
    }

    return alpha;
}

float64_t CSubGradientSVM::compute_min_subgradient(
    int32_t num_feat, int32_t num_vec, int32_t num_active, int32_t num_bound)
{
    float64_t dir_deriv=0;

    if (num_bound > 0)
    {

        CTime t2;
        SGVector<float64_t>::add(v, 1.0, w.vector, -1.0, sum_CXy_active, num_feat);

        if (num_bound>=qpsize_max && num_it_noimprovement!=10) // if qp gets to large, lets just choose a random beta
        {
            //SG_PRINT("qpsize too large  (%d>=%d) choosing random subgradient/beta\n", num_bound, qpsize_max);
            for (int32_t i=0; i<num_bound; i++)
                beta[i]=CMath::random(0.0,1.0);
        }
        else
        {
            memset(beta, 0, sizeof(float64_t)*num_bound);

            float64_t bias_const=0;

            if (use_bias)
                bias_const=1;

            for (int32_t i=0; i<num_bound; i++)
            {
                for (int32_t j=i; j<num_bound; j++)
                {
                    Z[i*num_bound+j]= 2.0*C1*C1*((CBinaryLabels*) m_labels)->get_label(idx_bound[i])*((CBinaryLabels*) m_labels)->get_label(idx_bound[j])*
                        (features->dot(idx_bound[i], features, idx_bound[j]) + bias_const);

                    Z[j*num_bound+i]=Z[i*num_bound+j];

                }

                Zv[i]=-2.0*C1*((CBinaryLabels*) m_labels)->get_label(idx_bound[i])*
                    (features->dense_dot(idx_bound[i], v, num_feat)-sum_Cy_active);
            }

            //CMath::display_matrix(Z, num_bound, num_bound, "Z");
            //CMath::display_vector(Zv, num_bound, "Zv");
            t2.stop();
#ifdef DEBUG_SUBGRADIENTSVM
            t2.time_diff_sec(true);
#endif

            CTime t;
            CQPBSVMLib solver(Z,num_bound, Zv,num_bound, 1.0);
            //solver.set_solver(QPB_SOLVER_GRADDESC);
            //solver.set_solver(QPB_SOLVER_GS);
#ifdef USE_CPLEX
            solver.set_solver(QPB_SOLVER_CPLEX);
#else
            solver.set_solver(QPB_SOLVER_SCAS);
#endif

            solver.solve_qp(beta, num_bound);

            t.stop();
#ifdef DEBUG_SUBGRADIENTSVM
            tim+=t.time_diff_sec(true);
#else
            tim+=t.time_diff_sec(false);
#endif

            //CMath::display_vector(beta, num_bound, "beta gs");
            //solver.set_solver(QPB_SOLVER_CPLEX);
            //solver.solve_qp(beta, num_bound);
            //CMath::display_vector(beta, num_bound, "beta cplex");

            //CMath::display_vector(grad_w, num_feat, "grad_w");
            //SG_PRINT("grad_b:%f\n", grad_b);
        }

        SGVector<float64_t>::add(grad_w, 1.0, w.vector, -1.0, sum_CXy_active, num_feat);
        grad_b = -sum_Cy_active;

        for (int32_t i=0; i<num_bound; i++)
        {
            features->add_to_dense_vec(-C1*beta[i]*((CBinaryLabels*) m_labels)->get_label(idx_bound[i]), idx_bound[i], grad_w, num_feat);
            if (use_bias)
                grad_b -=  C1 * ((CBinaryLabels*) m_labels)->get_label(idx_bound[i])*beta[i];
        }

        dir_deriv = SGVector<float64_t>::dot(grad_w, v, num_feat) - grad_b*sum_Cy_active;
        for (int32_t i=0; i<num_bound; i++)
        {
            float64_t val= C1*((CBinaryLabels*) m_labels)->get_label(idx_bound[i])*(features->dense_dot(idx_bound[i], grad_w, num_feat)+grad_b);
            dir_deriv += CMath::max(0.0, val);
        }
    }
    else
    {
        SGVector<float64_t>::add(grad_w, 1.0, w.vector, -1.0, sum_CXy_active, num_feat);
        grad_b = -sum_Cy_active;

        dir_deriv = SGVector<float64_t>::dot(grad_w, grad_w, num_feat)+ grad_b*grad_b;
    }

    return dir_deriv;
}

float64_t CSubGradientSVM::compute_objective(int32_t num_feat, int32_t num_vec)
{
    float64_t result= 0.5 * SGVector<float64_t>::dot(w.vector,w.vector, num_feat);

    for (int32_t i=0; i<num_vec; i++)
    {
        if (proj[i]<1.0)
            result += C1 * (1.0-proj[i]);
    }

    return result;
}

void CSubGradientSVM::compute_projection(int32_t num_feat, int32_t num_vec)
{
    for (int32_t i=0; i<num_vec; i++)
        proj[i]=((CBinaryLabels*) m_labels)->get_label(i)*(features->dense_dot(i, w.vector, num_feat)+bias);
}

void CSubGradientSVM::update_projection(float64_t alpha, int32_t num_vec)
{
    SGVector<float64_t>::vec1_plus_scalar_times_vec2(proj,-alpha, grad_proj, num_vec);
}

void CSubGradientSVM::init(int32_t num_vec, int32_t num_feat)
{
    // alloc normal and bias inited with 0
    w=SGVector<float64_t>(num_feat);
    w.zero();
    //CMath::random_vector(w, num_feat, -1.0, 1.0);
    bias=0;
    num_it_noimprovement=0;
    grad_b=0;
    qpsize_limit=5000;

    grad_w=SG_MALLOC(float64_t, num_feat);
    memset(grad_w,0,sizeof(float64_t)*num_feat);

    sum_CXy_active=SG_MALLOC(float64_t, num_feat);
    memset(sum_CXy_active,0,sizeof(float64_t)*num_feat);

    v=SG_MALLOC(float64_t, num_feat);
    memset(v,0,sizeof(float64_t)*num_feat);

    old_v=SG_MALLOC(float64_t, num_feat);
    memset(old_v,0,sizeof(float64_t)*num_feat);

    sum_Cy_active=0;

    proj= SG_MALLOC(float64_t, num_vec);
    memset(proj,0,sizeof(float64_t)*num_vec);

    tmp_proj=SG_MALLOC(float64_t, num_vec);
    memset(proj,0,sizeof(float64_t)*num_vec);

    tmp_proj_idx= SG_MALLOC(int32_t, num_vec);
    memset(tmp_proj_idx,0,sizeof(int32_t)*num_vec);

    grad_proj= SG_MALLOC(float64_t, num_vec);
    memset(grad_proj,0,sizeof(float64_t)*num_vec);

    hinge_point= SG_MALLOC(float64_t, num_vec);
    memset(hinge_point,0,sizeof(float64_t)*num_vec);

    hinge_idx= SG_MALLOC(int32_t, num_vec);
    memset(hinge_idx,0,sizeof(int32_t)*num_vec);

    active=SG_MALLOC(uint8_t, num_vec);
    memset(active,0,sizeof(uint8_t)*num_vec);

    old_active=SG_MALLOC(uint8_t, num_vec);
    memset(old_active,0,sizeof(uint8_t)*num_vec);

    idx_bound=SG_MALLOC(int32_t, num_vec);
    memset(idx_bound,0,sizeof(int32_t)*num_vec);

    idx_active=SG_MALLOC(int32_t, num_vec);
    memset(idx_active,0,sizeof(int32_t)*num_vec);

    Z=SG_MALLOC(float64_t, qpsize_limit*qpsize_limit);
    memset(Z,0,sizeof(float64_t)*qpsize_limit*qpsize_limit);

    Zv=SG_MALLOC(float64_t, qpsize_limit);
    memset(Zv,0,sizeof(float64_t)*qpsize_limit);

    beta=SG_MALLOC(float64_t, qpsize_limit);
    memset(beta,0,sizeof(float64_t)*qpsize_limit);

    old_Z=SG_MALLOC(float64_t, qpsize_limit*qpsize_limit);
    memset(old_Z,0,sizeof(float64_t)*qpsize_limit*qpsize_limit);

    old_Zv=SG_MALLOC(float64_t, qpsize_limit);
    memset(old_Zv,0,sizeof(float64_t)*qpsize_limit);

    old_beta=SG_MALLOC(float64_t, qpsize_limit);
    memset(old_beta,0,sizeof(float64_t)*qpsize_limit);

}

void CSubGradientSVM::cleanup()
{
    SG_FREE(hinge_idx);
    SG_FREE(hinge_point);
    SG_FREE(grad_proj);
    SG_FREE(proj);
    SG_FREE(tmp_proj);
    SG_FREE(tmp_proj_idx);
    SG_FREE(active);
    SG_FREE(old_active);
    SG_FREE(idx_bound);
    SG_FREE(idx_active);
    SG_FREE(sum_CXy_active);
    SG_FREE(grad_w);
    SG_FREE(v);
    SG_FREE(Z);
    SG_FREE(Zv);
    SG_FREE(beta);
    SG_FREE(old_v);
    SG_FREE(old_Z);
    SG_FREE(old_Zv);
    SG_FREE(old_beta);

    hinge_idx=NULL;
    proj=NULL;
    active=NULL;
    old_active=NULL;
    idx_bound=NULL;
    idx_active=NULL;
    sum_CXy_active=NULL;
    grad_w=NULL;
    v=NULL;
    Z=NULL;
    Zv=NULL;
    beta=NULL;
}

bool CSubGradientSVM::train_machine(CFeatures* data)
{
    tim=0;
    SG_INFO("C=%f epsilon=%f\n", C1, epsilon);
    ASSERT(m_labels);

    if (data)
    {
        if (!data->has_property(FP_DOT))
            SG_ERROR("Specified features are not of type CDotFeatures\n");
        set_features((CDotFeatures*) data);
    }
    ASSERT(get_features());

    int32_t num_iterations=0;
    int32_t num_train_labels=m_labels->get_num_labels();
    int32_t num_feat=features->get_dim_feature_space();
    int32_t num_vec=features->get_num_vectors();

    ASSERT(num_vec==num_train_labels);

    init(num_vec, num_feat);

    int32_t num_active=0;
    int32_t num_bound=0;
    float64_t alpha=0;
    float64_t dir_deriv=0;
    float64_t obj=0;
    delta_active=num_vec;
    last_it_noimprovement=-1;

    work_epsilon=0.99;
    autoselected_epsilon=work_epsilon;

    compute_projection(num_feat, num_vec);

    CTime time;
#ifdef DEBUG_SUBGRADIENTSVM
    float64_t loop_time=0;
#endif
    while (!(CSignal::cancel_computations()))
    {
        CTime t;
        delta_active=find_active(num_feat, num_vec, num_active, num_bound);

        update_active(num_feat, num_vec);

#ifdef DEBUG_SUBGRADIENTSVM
        SG_PRINT("==================================================\niteration: %d ", num_iterations);
        obj=compute_objective(num_feat, num_vec);
        SG_PRINT("objective:%.10f alpha: %.10f dir_deriv: %f num_bound: %d num_active: %d work_eps: %10.10f eps: %10.10f auto_eps: %10.10f time:%f\n",
                obj, alpha, dir_deriv, num_bound, num_active, work_epsilon, epsilon, autoselected_epsilon, loop_time);
#else
      SG_ABS_PROGRESS(work_epsilon, -CMath::log10(work_epsilon), -CMath::log10(0.99999999), -CMath::log10(epsilon), 6);
#endif
        //CMath::display_vector(w, w_dim, "w");
        //SG_PRINT("bias: %f\n", bias);
        //CMath::display_vector(proj, num_vec, "proj");
        //CMath::display_vector(idx_active, num_active, "idx_active");
        //SG_PRINT("num_active: %d\n", num_active);
        //CMath::display_vector(idx_bound, num_bound, "idx_bound");
        //SG_PRINT("num_bound: %d\n", num_bound);
        //CMath::display_vector(sum_CXy_active, num_feat, "sum_CXy_active");
        //SG_PRINT("sum_Cy_active: %f\n", sum_Cy_active);
        //CMath::display_vector(grad_w, num_feat, "grad_w");
        //SG_PRINT("grad_b:%f\n", grad_b);

        dir_deriv=compute_min_subgradient(num_feat, num_vec, num_active, num_bound);

        alpha=line_search(num_feat, num_vec);

        if (num_it_noimprovement==10 || num_bound<qpsize_max)
        {
            float64_t norm_grad=SGVector<float64_t>::dot(grad_w, grad_w, num_feat) +
                grad_b*grad_b;

#ifdef DEBUG_SUBGRADIENTSVM
            SG_PRINT("CHECKING OPTIMALITY CONDITIONS: "
                    "work_epsilon: %10.10f delta_active:%d alpha: %10.10f norm_grad: %10.10f a*norm_grad:%10.16f\n",
                    work_epsilon, delta_active, alpha, norm_grad, CMath::abs(alpha*norm_grad));
#else
            SG_ABS_PROGRESS(work_epsilon, -CMath::log10(work_epsilon), -CMath::log10(0.99999999), -CMath::log10(epsilon), 6);
#endif

            if (work_epsilon<=epsilon && delta_active==0 && CMath::abs(alpha*norm_grad)<1e-6)
                break;
            else
                num_it_noimprovement=0;
        }

        if ((dir_deriv<0 || alpha==0) && (work_epsilon<=epsilon && delta_active==0))
        {
            if (last_it_noimprovement==num_iterations-1)
            {
                SG_PRINT("no improvement...\n");
                num_it_noimprovement++;
            }
            else
                num_it_noimprovement=0;

            last_it_noimprovement=num_iterations;
        }

        SGVector<float64_t>::vec1_plus_scalar_times_vec2(w.vector, -alpha, grad_w, num_feat);
        bias-=alpha*grad_b;

        update_projection(alpha, num_vec);
        //compute_projection(num_feat, num_vec);
        //CMath::display_vector(w.vector, w_dim, "w");
        //SG_PRINT("bias: %f\n", bias);
        //CMath::display_vector(proj, num_vec, "proj");

        t.stop();
#ifdef DEBUG_SUBGRADIENTSVM
        loop_time=t.time_diff_sec();
#endif
        num_iterations++;

        if (get_max_train_time()>0 && time.cur_time_diff()>get_max_train_time())
            break;
    }
    SG_DONE();

    SG_INFO("converged after %d iterations\n", num_iterations);

    obj=compute_objective(num_feat, num_vec);
    SG_INFO("objective: %f alpha: %f dir_deriv: %f num_bound: %d num_active: %d\n",
            obj, alpha, dir_deriv, num_bound, num_active);

#ifdef DEBUG_SUBGRADIENTSVM
    CMath::display_vector(w.vector, w.vlen, "w");
    SG_PRINT("bias: %f\n", bias);
#endif
    SG_INFO("solver time:%f s\n", tim);

    cleanup();

    return true;
}