GaussianARDKernel.cpp

浏览该文件的文档.

/*
 * 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) 2015 Wu Lin
 * Written (W) 2012 Jacob Walker
 *
 * Adapted from WeightedDegreeRBFKernel.cpp
 */

#include <shogun/kernel/GaussianARDKernel.h>
#include <shogun/mathematics/Math.h>

#ifdef HAVE_LINALG_LIB
#include <shogun/mathematics/linalg/linalg.h>
#endif

using namespace shogun;

CGaussianARDKernel::CGaussianARDKernel() : CExponentialARDKernel()
{
    init();
}

CGaussianARDKernel::~CGaussianARDKernel()
{
}

void CGaussianARDKernel::init()
{

#ifdef HAVE_LINALG_LIB
    m_sq_lhs=SGVector<float64_t>();
    m_sq_rhs=SGVector<float64_t>();
    SG_ADD(&m_sq_lhs, "sq_lhs", "squared left-hand side", MS_NOT_AVAILABLE);
    SG_ADD(&m_sq_rhs, "sq_rhs", "squared right-hand side", MS_NOT_AVAILABLE);
#endif
}

float64_t CGaussianARDKernel::distance(int32_t idx_a, int32_t idx_b)
{
    float64_t result=0.0;
#ifdef HAVE_LINALG_LIB
    REQUIRE(lhs, "Left features (lhs) not set!\n")
    REQUIRE(rhs, "Right features (rhs) not set!\n")

    if (lhs==rhs && idx_a==idx_b)
        return result;

    if (m_ARD_type==KT_SCALAR)
    {
        result=(m_sq_lhs[idx_a]+m_sq_rhs[idx_b]-2.0*CDotKernel::compute(idx_a,idx_b));
        result*=CMath::exp(2.0*m_log_weights[0]);
    }
    else
    {
        SGVector<float64_t> avec=get_feature_vector(idx_a, lhs);
        SGVector<float64_t> bvec=get_feature_vector(idx_b, rhs);
        avec=linalg::add(avec, bvec, 1.0, -1.0);
        result=compute_helper(avec, avec);
    }
#endif /* HAVE_LINALG_LIB */
    return result/2.0;
}

#ifdef HAVE_LINALG_LIB
CGaussianARDKernel::CGaussianARDKernel(int32_t size)
        : CExponentialARDKernel(size)
{
    init();
}

CGaussianARDKernel::CGaussianARDKernel(CDotFeatures* l,
        CDotFeatures* r, int32_t size)
        : CExponentialARDKernel(size)
{
    init();
}

bool CGaussianARDKernel::init(CFeatures* l, CFeatures* r)
{
    bool status=CExponentialARDKernel::init(l,r);

    if (m_ARD_type==KT_SCALAR)
        precompute_squared();

    return status;
}

SGVector<float64_t> CGaussianARDKernel::precompute_squared_helper(CDotFeatures* df)
{
    REQUIRE(df, "Features not set\n")
    int32_t num_vec=df->get_num_vectors();
    SGVector<float64_t> sq(num_vec);
    for (int32_t i=0; i<num_vec; i++)
        sq[i]=df->dot(i,df, i);
    return sq;
}

void CGaussianARDKernel::precompute_squared()
{
    if (!lhs || !rhs)
        return;
    SG_SPRINT("called\n");
    m_sq_lhs=precompute_squared_helper((CDotFeatures*) lhs);

    if (lhs==rhs)
        m_sq_rhs=m_sq_lhs;
    else
        m_sq_rhs=precompute_squared_helper((CDotFeatures*) rhs);
}


CGaussianARDKernel* CGaussianARDKernel::obtain_from_generic(CKernel* kernel)
{
    if (kernel->get_kernel_type()!=K_GAUSSIANARD)
    {
        SG_SERROR("Provided kernel is not of type CGaussianARDKernel!\n");
    }

    /* since an additional reference is returned */
    SG_REF(kernel);
    return (CGaussianARDKernel*)kernel;
}

float64_t CGaussianARDKernel::compute_helper(SGVector<float64_t> avec, SGVector<float64_t>bvec)
{
    SGMatrix<float64_t> left;
    SGMatrix<float64_t> left_transpose;
    float64_t scalar_weight=1.0;
    if (m_ARD_type==KT_SCALAR)
    {
        left=SGMatrix<float64_t>(avec.vector,1,avec.vlen,false);
        scalar_weight=CMath::exp(m_log_weights[0]);
    }
    else if(m_ARD_type==KT_FULL || m_ARD_type==KT_DIAG)
    {
        left_transpose=get_weighted_vector(avec);
        left=SGMatrix<float64_t>(left_transpose.matrix,1,left_transpose.num_rows,false);
    }
    else
        SG_ERROR("Unsupported ARD type\n");
    SGMatrix<float64_t> right=compute_right_product(bvec, scalar_weight);
    SGMatrix<float64_t> res=linalg::matrix_product(left, right);
    return res[0]*scalar_weight;
}

float64_t CGaussianARDKernel::compute_gradient_helper(SGVector<float64_t> avec,
    SGVector<float64_t> bvec, float64_t scale, index_t index)
{
    float64_t result=0.0;

    if(m_ARD_type==KT_DIAG)
        result=2.0*avec[index]*bvec[index]*CMath::exp(2.0*m_log_weights[index]);
    else
    {
        SGMatrix<float64_t> res;

        if (m_ARD_type==KT_SCALAR)
        {
            SGMatrix<float64_t> left(avec.vector,1,avec.vlen,false);
            SGMatrix<float64_t> right(bvec.vector,bvec.vlen,1,false);
            res=linalg::matrix_product(left, right);
            result=2.0*res[0]*CMath::exp(2.0*m_log_weights[0]);
        }
        else if(m_ARD_type==KT_FULL)
        {
            int32_t row_index=0;
            int32_t col_index=index;
            int32_t offset=m_weights_rows;
            int32_t total_offset=0;
            while(col_index>=offset && offset>0)
            {
                col_index-=offset;
                total_offset+=offset;
                offset--;
                row_index++;
            }
            col_index+=row_index;

            SGVector<float64_t> row_vec=SGVector<float64_t>(m_log_weights.vector+total_offset,m_weights_rows-row_index,false);   
            row_vec[0]=CMath::exp(row_vec[0]);

            SGMatrix<float64_t> row_vec_r(row_vec.vector,row_vec.vlen,1,false);
            SGMatrix<float64_t> left(avec.vector+row_index,1,avec.vlen-row_index,false);

            res=linalg::matrix_product(left, row_vec_r);
            result=res[0]*bvec[col_index];

            SGMatrix<float64_t> row_vec_l(row_vec.vector,1,row_vec.vlen,false);
            SGMatrix<float64_t> right(bvec.vector+row_index,bvec.vlen-row_index,1,false);

            res=linalg::matrix_product(row_vec_l, right);
            result+=res[0]*avec[col_index];

            if(row_index==col_index)
                result*=row_vec[0];
            row_vec[0]=CMath::log(row_vec[0]);
        }
        else
        {
            SG_ERROR("Unsupported ARD type\n");
        }

    }
    return result*scale;
}


SGVector<float64_t> CGaussianARDKernel::get_parameter_gradient_diagonal(
        const TParameter* param, index_t index)
{
    REQUIRE(param, "Param not set\n");
    REQUIRE(lhs , "Left features not set!\n");
    REQUIRE(rhs, "Right features not set!\n");

    if (lhs==rhs)
    {
        if (!strcmp(param->m_name, "log_weights"))
        {
            SGVector<float64_t> derivative(num_lhs);
            derivative.zero();
            return derivative;
        }
    }
    else
    {
        int32_t length=CMath::min(num_lhs, num_rhs);
        SGVector<float64_t> derivative(length);
        check_weight_gradient_index(index);
        for (index_t j=0; j<length; j++)
        {
            if (!strcmp(param->m_name, "log_weights") )
            {
                if (m_ARD_type==KT_SCALAR)
                {
                    float64_t dist=distance(j,j);
                    derivative[j]=CMath::exp(-dist)*(-dist*2.0);
                }
                else
                {
                    SGVector<float64_t> avec=get_feature_vector(j, lhs);
                    SGVector<float64_t> bvec=get_feature_vector(j, rhs);
                    derivative[j]=get_parameter_gradient_helper(param,index,j,j,avec,bvec);
                }

            }
        }
        return derivative;
    }

    SG_ERROR("Can't compute derivative wrt %s parameter\n", param->m_name);
    return SGVector<float64_t>();
}


float64_t CGaussianARDKernel::get_parameter_gradient_helper(
    const TParameter* param, index_t index, int32_t idx_a,
    int32_t idx_b, SGVector<float64_t> avec, SGVector<float64_t> bvec)
{
    REQUIRE(param, "Param not set\n");

    if (!strcmp(param->m_name, "log_weights"))
    {
        bvec=linalg::add(avec, bvec, 1.0, -1.0);
        float64_t scale=-kernel(idx_a,idx_b)/2.0;
        return  compute_gradient_helper(bvec, bvec, scale, index);
    }
    else
    {
        SG_ERROR("Can't compute derivative wrt %s parameter\n", param->m_name);
        return 0.0;
    }
}

SGMatrix<float64_t> CGaussianARDKernel::get_parameter_gradient(
        const TParameter* param, index_t index)
{
    REQUIRE(param, "Param not set\n");
    REQUIRE(lhs , "Left features not set!\n");
    REQUIRE(rhs, "Right features not set!\n");

    if (!strcmp(param->m_name, "log_weights"))
    {
        SGMatrix<float64_t> derivative(num_lhs, num_rhs);
        check_weight_gradient_index(index);
        for (index_t j=0; j<num_lhs; j++)
        {
            SGVector<float64_t> avec=get_feature_vector(j, lhs);
            for (index_t k=0; k<num_rhs; k++)
            {
                if (m_ARD_type==KT_SCALAR)
                {
                    float64_t dist=distance(j,k);
                    derivative(j,k)=CMath::exp(-dist)*(-dist*2.0);
                }
                else
                {
                    SGVector<float64_t> bvec=get_feature_vector(k, rhs);
                    derivative(j,k)=get_parameter_gradient_helper(param,index,j,k,avec,bvec);
                }
            }
        }
        return derivative;
    }
    else
    {
        SG_ERROR("Can't compute derivative wrt %s parameter\n", param->m_name);
        return SGMatrix<float64_t>();
    }
}
#endif /* HAVE_LINALG_LIB */