GaussianKernel.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) 1999-2010 Soeren Sonnenburg
 * Written (W) 2011 Abhinav Maurya
 * Written (W) 2012 Heiko Strathmann
 * Copyright (C) 1999-2009 Fraunhofer Institute FIRST and Max-Planck-Society
 * Copyright (C) 2010 Berlin Institute of Technology
 */

#include <shogun/lib/common.h>
#include <shogun/base/Parameter.h>
#include <shogun/kernel/GaussianKernel.h>
#include <shogun/features/DotFeatures.h>
#include <shogun/io/SGIO.h>
#include <shogun/mathematics/Math.h>

using namespace shogun;

void CGaussianKernel::set_width(float64_t w)
{
    REQUIRE(w>0, "width (%f) must be positive\n",w);
    m_log_width=CMath::log(w/2.0)/2.0;
}

float64_t CGaussianKernel::get_width() const
{
    return CMath::exp(m_log_width*2.0)*2.0;
}

CGaussianKernel::CGaussianKernel() : CDotKernel()
{
    init();
}

CGaussianKernel::CGaussianKernel(int32_t size, float64_t w) : CDotKernel(size)
{
    init();
    set_width(w);
}

CGaussianKernel::CGaussianKernel(CDotFeatures* l, CDotFeatures* r,
        float64_t w, int32_t size) : CDotKernel(size)
{
    init();
    set_width(w);
    init(l,r);
}

CGaussianKernel::~CGaussianKernel()
{
    cleanup();
}

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

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

#include <typeinfo>
CSGObject *CGaussianKernel::shallow_copy() const
{
    // TODO: remove this after all the classes get shallow_copy properly implemented
    // this assert is to avoid any subclass of CGaussianKernel accidentally called
    // with the implement here
    ASSERT(typeid(*this) == typeid(CGaussianKernel))
    CGaussianKernel *ker = new CGaussianKernel(cache_size, get_width());
    if (lhs)
    {
        ker->init(lhs, rhs);
    }
    return ker;
}

void CGaussianKernel::cleanup()
{
    if (sq_lhs != sq_rhs)
        SG_FREE(sq_rhs);
    sq_rhs = NULL;

    SG_FREE(sq_lhs);
    sq_lhs = NULL;

    CKernel::cleanup();
}

void CGaussianKernel::precompute_squared_helper(float64_t* &buf, CDotFeatures* df)
{
    ASSERT(df)
    int32_t num_vec=df->get_num_vectors();
    buf=SG_MALLOC(float64_t, num_vec);

    for (int32_t i=0; i<num_vec; i++)
        buf[i]=df->dot(i,df, i);
}

bool CGaussianKernel::init(CFeatures* l, CFeatures* r)
{
    cleanup();

    CDotKernel::init(l, r);
    precompute_squared();
    return init_normalizer();
}

float64_t CGaussianKernel::compute(int32_t idx_a, int32_t idx_b)
{
    if (!m_compact)
    {
        float64_t result=distance(idx_a,idx_b);
        return CMath::exp(-result);
    }

    int32_t len_features, power;
    len_features=((CDotFeatures*) lhs)->get_dim_feature_space();
    power=(len_features%2==0) ? (len_features+1):len_features;

    float64_t result=distance(idx_a,idx_b);
    float64_t result_multiplier=1-(sqrt(result))/3;

    if (result_multiplier<=0)
        result_multiplier=0;
    else
        result_multiplier=pow(result_multiplier, power);

    return result_multiplier*exp(-result);
}

void CGaussianKernel::load_serializable_post() throw (ShogunException)
{
    CKernel::load_serializable_post();
    precompute_squared();
}

void CGaussianKernel::precompute_squared()
{
    if (!lhs || !rhs)
        return;

    precompute_squared_helper(sq_lhs, (CDotFeatures*) lhs);

    if (lhs==rhs)
        sq_rhs=sq_lhs;
    else
        precompute_squared_helper(sq_rhs, (CDotFeatures*) rhs);
}

SGMatrix<float64_t> CGaussianKernel::get_parameter_gradient(
        const TParameter* param, index_t index)
{
    REQUIRE(lhs && rhs, "Features not set!\n")

    if (!strcmp(param->m_name, "log_width"))
    {
        SGMatrix<float64_t> derivative=SGMatrix<float64_t>(num_lhs, num_rhs);

        for (int j=0; j<num_lhs; j++)
            for (int k=0; k<num_rhs; k++)
            {
                float64_t element=distance(j,k);
                derivative(j,k)=exp(-element)*element*2.0;
            }

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

void CGaussianKernel::init()
{
    set_width(1.0);
    set_compact_enabled(false);
    sq_lhs=NULL;
    sq_rhs=NULL;
    SG_ADD(&m_log_width, "log_width", "Kernel width in log domain", MS_AVAILABLE, GRADIENT_AVAILABLE);
    SG_ADD(&m_compact, "compact", "Compact enabled option", MS_AVAILABLE);
}

float64_t CGaussianKernel::distance(int32_t idx_a, int32_t idx_b)
{
    return (sq_lhs[idx_a]+sq_rhs[idx_b]-2*CDotKernel::compute(idx_a,idx_b))/get_width();
}