MultitaskKernelNormalizer.h

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 Christian Widmer
 * Copyright (C) 2009 Max-Planck-Society
 */

#ifndef _MULTITASKKERNELNORMALIZER_H___
#define _MULTITASKKERNELNORMALIZER_H___

#include <shogun/kernel/normalizer/KernelNormalizer.h>
#include <shogun/kernel/Kernel.h>
#include <algorithm>
#include <vector>



namespace shogun
{
class CMultitaskKernelNormalizer: public CKernelNormalizer
{

public:

    CMultitaskKernelNormalizer() : CKernelNormalizer(), scale(1.0)
    {
    }

    CMultitaskKernelNormalizer(std::vector<int32_t> task_vector)
        : CKernelNormalizer(), scale(1.0)
    {

        num_tasks = get_num_unique_tasks(task_vector);

        // set both sides equally
        set_task_vector(task_vector);

        // init similarity matrix
        similarity_matrix = std::vector<float64_t>(num_tasks * num_tasks);

    }

    virtual ~CMultitaskKernelNormalizer()
    {
    }

    virtual bool init(CKernel* k)
    {

        //same as first-element normalizer
        CFeatures* old_lhs=k->lhs;
        CFeatures* old_rhs=k->rhs;
        k->lhs=old_lhs;
        k->rhs=old_lhs;

        if (strcmp(k->get_name(), "WeightedDegree") == 0) {
            SG_INFO("using first-element normalization\n");
            scale=k->compute(0, 0);
        } else {
            SG_INFO("no inner normalization for non-WDK kernel\n");
            scale=1.0;
        }

        k->lhs=old_lhs;
        k->rhs=old_rhs;

        ASSERT(k);
        int32_t num_lhs = k->get_num_vec_lhs();
        int32_t num_rhs = k->get_num_vec_rhs();
        ASSERT(num_lhs>0);
        ASSERT(num_rhs>0);

        //std::cout << "scale: " << scale << std::endl;

        return true;
    }

    int32_t get_num_unique_tasks(std::vector<int32_t> vec) {

        //sort
        std::sort(vec.begin(), vec.end());

        //reorder tasks with unique prefix
        std::vector<int32_t>::iterator endLocation = std::unique(vec.begin(), vec.end());

        //count unique tasks
        int32_t num_vec = std::distance(vec.begin(), endLocation);

        return num_vec;

    }

    virtual float64_t normalize(float64_t value, int32_t idx_lhs,
            int32_t idx_rhs)
    {

        //lookup tasks
        int32_t task_idx_lhs = task_vector_lhs[idx_lhs];
        int32_t task_idx_rhs = task_vector_rhs[idx_rhs];

        //lookup similarity
        float64_t task_similarity = get_task_similarity(task_idx_lhs,
                task_idx_rhs);

        //take task similarity into account
        float64_t similarity = (value/scale) * task_similarity;


        return similarity;

    }

    virtual float64_t normalize_lhs(float64_t value, int32_t idx_lhs)
    {
        SG_ERROR("normalize_lhs not implemented");
        return 0;
    }

    virtual float64_t normalize_rhs(float64_t value, int32_t idx_rhs)
    {
        SG_ERROR("normalize_rhs not implemented");
        return 0;
    }

public:

    std::vector<int32_t> get_task_vector_lhs() const
    {
        return task_vector_lhs;
    }

    void set_task_vector_lhs(std::vector<int32_t> vec)
    {
        task_vector_lhs = vec;
    }

    std::vector<int32_t> get_task_vector_rhs() const
    {
        return task_vector_rhs;
    }

    void set_task_vector_rhs(std::vector<int32_t> vec)
    {
        task_vector_rhs = vec;
    }

    void set_task_vector(std::vector<int32_t> vec)
    {
        task_vector_lhs = vec;
        task_vector_rhs = vec;
    }

    float64_t get_task_similarity(int32_t task_lhs, int32_t task_rhs)
    {

        ASSERT(task_lhs < num_tasks && task_lhs >= 0);
        ASSERT(task_rhs < num_tasks && task_rhs >= 0);

        return similarity_matrix[task_lhs * num_tasks + task_rhs];

    }

    void set_task_similarity(int32_t task_lhs, int32_t task_rhs,
            float64_t similarity)
    {

        ASSERT(task_lhs < num_tasks && task_lhs >= 0);
        ASSERT(task_rhs < num_tasks && task_rhs >= 0);

        similarity_matrix[task_lhs * num_tasks + task_rhs] = similarity;

    }

    virtual const char* get_name() const
    {
        return "MultitaskKernelNormalizer";
    }

    inline CMultitaskKernelNormalizer* KernelNormalizerToMultitaskKernelNormalizer(CKernelNormalizer* n)
    {
        return dynamic_cast<CMultitaskKernelNormalizer*>(n);
    }


protected:

    std::vector<float64_t> similarity_matrix;

    int32_t num_tasks;

    std::vector<int32_t> task_vector_lhs;

    std::vector<int32_t> task_vector_rhs;

    float64_t scale;

};
}
#endif