MulticlassOVREvaluation.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.
 *
 * Copyright (C) 2012 Sergey Lisitsyn
 */

#include <shogun/evaluation/MulticlassOVREvaluation.h>
#include <shogun/evaluation/ROCEvaluation.h>
#include <shogun/evaluation/PRCEvaluation.h>
#include <shogun/labels/MulticlassLabels.h>
#include <shogun/mathematics/Statistics.h>

using namespace shogun;

CMulticlassOVREvaluation::CMulticlassOVREvaluation() :
    CEvaluation(), m_binary_evaluation(NULL), m_graph_results(NULL), m_num_graph_results(0)
{
}

CMulticlassOVREvaluation::CMulticlassOVREvaluation(CBinaryClassEvaluation* binary_evaluation) :
    CEvaluation(), m_binary_evaluation(NULL), m_graph_results(NULL), m_num_graph_results(0)
{
    set_binary_evaluation(binary_evaluation);
}

CMulticlassOVREvaluation::~CMulticlassOVREvaluation()
{
    SG_UNREF(m_binary_evaluation);
    if (m_graph_results)
    {
        for (int32_t i=0; i<m_num_graph_results; i++)
            m_graph_results[i].~SGMatrix<float64_t>();
        SG_FREE(m_graph_results);
    }
}

float64_t CMulticlassOVREvaluation::evaluate(CLabels* predicted, CLabels* ground_truth)
{
    ASSERT(m_binary_evaluation)
    ASSERT(predicted)
    ASSERT(ground_truth)
    int32_t n_labels = predicted->get_num_labels();
    ASSERT(n_labels)
    CMulticlassLabels* predicted_mc = (CMulticlassLabels*)predicted;
    CMulticlassLabels* ground_truth_mc = (CMulticlassLabels*)ground_truth;
    int32_t n_classes = predicted_mc->get_multiclass_confidences(0).size();
    ASSERT(n_classes>0)
    m_last_results = SGVector<float64_t>(n_classes);
    
    SGMatrix<float64_t> all(n_labels,n_classes);
    for (int32_t i=0; i<n_labels; i++)
    {
        SGVector<float64_t> confs = predicted_mc->get_multiclass_confidences(i);
        for (int32_t j=0; j<n_classes; j++)
        {
            all(i,j) = confs[j];
        }
    }
    if (dynamic_cast<CROCEvaluation*>(m_binary_evaluation) || dynamic_cast<CPRCEvaluation*>(m_binary_evaluation))
    {
        for (int32_t i=0; i<m_num_graph_results; i++)
            m_graph_results[i].~SGMatrix<float64_t>();
        SG_FREE(m_graph_results);
        m_graph_results = SG_MALLOC(SGMatrix<float64_t>, n_classes);
        m_num_graph_results = n_classes;
    }
    for (int32_t c=0; c<n_classes; c++)
    {
        CLabels* pred = new CBinaryLabels(SGVector<float64_t>(all.get_column_vector(c),n_labels,false));
        SGVector<float64_t> gt_vec(n_labels);
        for (int32_t i=0; i<n_labels; i++)
        {
            if (ground_truth_mc->get_label(i)==c)
                gt_vec[i] = +1.0;
            else
                gt_vec[i] = -1.0;
        }
        CLabels* gt = new CBinaryLabels(gt_vec);
        m_last_results[c] = m_binary_evaluation->evaluate(pred, gt);

        if (dynamic_cast<CROCEvaluation*>(m_binary_evaluation))
        {
            new (&m_graph_results[c]) SGMatrix<float64_t>();
            m_graph_results[c] = ((CROCEvaluation*)m_binary_evaluation)->get_ROC();
        }
        if (dynamic_cast<CPRCEvaluation*>(m_binary_evaluation))
        {
            new (&m_graph_results[c]) SGMatrix<float64_t>();
            m_graph_results[c] = ((CPRCEvaluation*)m_binary_evaluation)->get_PRC();
        }
    }
    return CStatistics::mean(m_last_results);
}