ANOVAKernel.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) 2011 Andrew Tereskin
 * Copyright (C) 2011 Berlin Institute of Technology and Max-Planck-Society
 */

#include <math.h>
#include <shogun/kernel/ANOVAKernel.h>
#include <shogun/mathematics/Math.h>

using namespace shogun;

CANOVAKernel::CANOVAKernel(): CDotKernel(0), cardinality(1.0)
{
    register_params();
}

CANOVAKernel::CANOVAKernel(int32_t cache, int32_t d)
: CDotKernel(cache), cardinality(d)
{
    register_params();
}

CANOVAKernel::CANOVAKernel(
    CDenseFeatures<float64_t>* l, CDenseFeatures<float64_t>* r, int32_t d, int32_t cache)
  : CDotKernel(cache), cardinality(d)
{
    register_params();
    init(l, r);
}

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

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

    bool result = CDotKernel::init(l,r);

    init_normalizer();
    return result;
}

float64_t CANOVAKernel::compute(int32_t idx_a, int32_t idx_b)
{
    return compute_rec1(idx_a, idx_b);
}

float64_t CANOVAKernel::compute_rec1(int32_t idx_a, int32_t idx_b)
{
    int32_t alen, blen;
    bool afree, bfree;

    float64_t* avec=
        ((CDenseFeatures<float64_t>*) lhs)->get_feature_vector(idx_a, alen, afree);
    float64_t* bvec=
        ((CDenseFeatures<float64_t>*) rhs)->get_feature_vector(idx_b, blen, bfree);
    ASSERT(alen==blen)

    float64_t result = compute_recursive1(avec, bvec, alen);

    ((CDenseFeatures<float64_t>*) lhs)->free_feature_vector(avec, idx_a, afree);
    ((CDenseFeatures<float64_t>*) rhs)->free_feature_vector(bvec, idx_b, bfree);

    return result;
}

float64_t CANOVAKernel::compute_rec2(int32_t idx_a, int32_t idx_b)
{
    int32_t alen, blen;
    bool afree, bfree;

    float64_t* avec=
        ((CDenseFeatures<float64_t>*) lhs)->get_feature_vector(idx_a, alen, afree);
    float64_t* bvec=
        ((CDenseFeatures<float64_t>*) rhs)->get_feature_vector(idx_b, blen, bfree);
    ASSERT(alen==blen)

    float64_t result = compute_recursive2(avec, bvec, alen);

    ((CDenseFeatures<float64_t>*) lhs)->free_feature_vector(avec, idx_a, afree);
    ((CDenseFeatures<float64_t>*) rhs)->free_feature_vector(bvec, idx_b, bfree);

    return result;
}

void CANOVAKernel::register_params()
{
    SG_ADD(&cardinality, "cardinality", "Kernel cardinality.", MS_AVAILABLE);
}


float64_t CANOVAKernel::compute_recursive1(float64_t* avec, float64_t* bvec, int32_t len)
{
    int32_t DP_len=(cardinality+1)*(len+1);
    float64_t* DP = SG_MALLOC(float64_t, DP_len);

    ASSERT(DP)
    int32_t d=cardinality;
    int32_t offs=cardinality+1;

    ASSERT(DP_len==(len+1)*offs)

    for (int32_t j=0; j < len+1; j++)
        DP[j] = 1.0;

    for (int32_t k=1; k < d+1; k++)
    {
        // TRAP d>len case
        if (k-1>=len)
            return 0.0;

        DP[k*offs+k-1] = 0;
        for (int32_t j=k; j < len+1; j++)
            DP[k*offs+j]=DP[k*offs+j-1]+avec[j-1]*bvec[j-1]*DP[(k-1)*offs+j-1];
    }

    float64_t result=DP[d*offs+len];

    SG_FREE(DP);

    return result;
}

float64_t CANOVAKernel::compute_recursive2(float64_t* avec, float64_t* bvec, int32_t len)
{
    float64_t* KD = SG_MALLOC(float64_t, cardinality+1);
    float64_t* KS = SG_MALLOC(float64_t, cardinality+1);
    float64_t* vec_pow = SG_MALLOC(float64_t, len);

    ASSERT(vec_pow)
    ASSERT(KS)
    ASSERT(KD)

    int32_t d=cardinality;
    for (int32_t i=0; i < len; i++)
        vec_pow[i] = 1;

    for (int32_t k=1; k < d+1; k++)
    {
        KS[k] = 0;
        for (int32_t i=0; i < len; i++)
        {
            vec_pow[i] *= avec[i]*bvec[i];
            KS[k] += vec_pow[i];
        }
    }

    KD[0] = 1;
    for (int32_t k=1; k < d+1; k++)
    {
        float64_t sum = 0;
        for (int32_t s=1; s < k+1; s++)
        {
            float64_t sign = 1.0;
            if (s % 2 == 0)
                sign = -1.0;

            sum += sign*KD[k-s]*KS[s];
        }

        KD[k] = sum / k;
    }
    float64_t result=KD[d];
    SG_FREE(vec_pow);
    SG_FREE(KS);
    SG_FREE(KD);

    return result;
}