MulticlassOneVsRestStrategy.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) 2012 Chiyuan Zhang
 * Copyright (C) 2012 Chiyuan Zhang
 */

#include <shogun/multiclass/MulticlassOneVsRestStrategy.h>
#include <shogun/labels/BinaryLabels.h>
#include <shogun/labels/MulticlassLabels.h>
#include <shogun/mathematics/Math.h> 

using namespace shogun;

CMulticlassOneVsRestStrategy::CMulticlassOneVsRestStrategy()
    : CMulticlassStrategy()
{
}

CMulticlassOneVsRestStrategy::CMulticlassOneVsRestStrategy(EProbHeuristicType prob_heuris)
    : CMulticlassStrategy(prob_heuris)
{
}

SGVector<int32_t> CMulticlassOneVsRestStrategy::train_prepare_next()
{
    for (int32_t i=0; i < m_orig_labels->get_num_labels(); ++i)
    {
        if (((CMulticlassLabels*) m_orig_labels)->get_int_label(i)==m_train_iter)
            ((CBinaryLabels*) m_train_labels)->set_label(i, +1.0);
        else
            ((CBinaryLabels*) m_train_labels)->set_label(i, -1.0);
    }

    // increase m_train_iter *after* setting labels
    CMulticlassStrategy::train_prepare_next();

    return SGVector<int32_t>();
}

int32_t CMulticlassOneVsRestStrategy::decide_label(SGVector<float64_t> outputs)
{
    if (m_rejection_strategy && m_rejection_strategy->reject(outputs))
        return CDenseLabels::REJECTION_LABEL;

    return SGVector<float64_t>::arg_max(outputs.vector, 1, outputs.vlen);
}

SGVector<index_t> CMulticlassOneVsRestStrategy::decide_label_multiple_output(SGVector<float64_t> outputs, int32_t n_outputs)
{
    float64_t* outputs_ = SG_MALLOC(float64_t, outputs.vlen);
    int32_t* indices_ = SG_MALLOC(int32_t, outputs.vlen);
    for (int32_t i=0; i<outputs.vlen; i++)
    {
        outputs_[i] = outputs[i];
        indices_[i] = i;
    }
    CMath::qsort_backward_index(outputs_,indices_,outputs.vlen);
    SGVector<index_t> result(n_outputs);
    for (int32_t i=0; i<n_outputs; i++)
        result[i] = indices_[i];
    SG_FREE(outputs_);
    SG_FREE(indices_);
    return result;
}

void CMulticlassOneVsRestStrategy::rescale_outputs(SGVector<float64_t> outputs)
{
    switch(get_prob_heuris_type())
    {
        case OVA_NORM:
            rescale_heuris_norm(outputs);
            break;
        case OVA_SOFTMAX:
            SG_ERROR("%s::rescale_outputs(): Need to specify sigmoid parameters!\n", get_name());
            break;
        case PROB_HEURIS_NONE:
            break;
        default:
            SG_ERROR("%s::rescale_outputs(): Unknown OVA probability heuristic type!\n", get_name());
            break;
    }
}

void CMulticlassOneVsRestStrategy::rescale_outputs(SGVector<float64_t> outputs,
        const SGVector<float64_t> As, const SGVector<float64_t> Bs)
{
    if (get_prob_heuris_type()==OVA_SOFTMAX)
        rescale_heuris_softmax(outputs,As,Bs);
    else
        rescale_outputs(outputs);
}

void CMulticlassOneVsRestStrategy::rescale_heuris_norm(SGVector<float64_t> outputs)
{
    if (m_num_classes != outputs.vlen)
    {
        SG_ERROR("%s::rescale_heuris_norm(): size(outputs) = %d != m_num_classes = %d\n", 
                get_name(), outputs.vlen, m_num_classes);
    }

    float64_t norm = SGVector<float64_t>::sum(outputs);
    norm += 1E-10;
    for (int32_t i=0; i<outputs.vlen; i++) 
        outputs[i] /= norm;
}

void CMulticlassOneVsRestStrategy::rescale_heuris_softmax(SGVector<float64_t> outputs,
        const SGVector<float64_t> As, const SGVector<float64_t> Bs)
{
    if (m_num_classes != outputs.vlen)
    {
        SG_ERROR("%s::rescale_heuris_softmax(): size(outputs) = %d != m_num_classes = %d\n", 
                get_name(), outputs.vlen, m_num_classes);
    }

    for (int32_t i=0; i<outputs.vlen; i++) 
        outputs[i] = CMath::exp(-As[i]*outputs[i]-Bs[i]);

    float64_t norm = SGVector<float64_t>::sum(outputs);
    norm += 1E-10;
    for (int32_t i=0; i<outputs.vlen; i++) 
        outputs[i] /= norm;
}