PluginEstimate.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) 1999-2009 Soeren Sonnenburg
 * Copyright (C) 1999-2009 Fraunhofer Institute FIRST and Max-Planck-Society
 */

#include <shogun/lib/common.h>
#include <shogun/io/SGIO.h>
#include <shogun/features/StringFeatures.h>
#include <shogun/labels/Labels.h>
#include <shogun/labels/BinaryLabels.h>
#include <shogun/labels/RegressionLabels.h>
#include <shogun/distributions/LinearHMM.h>
#include <shogun/classifier/PluginEstimate.h>

using namespace shogun;

CPluginEstimate::CPluginEstimate(float64_t pos_pseudo, float64_t neg_pseudo)
: CMachine(), m_pos_pseudo(1e-10), m_neg_pseudo(1e-10),
    pos_model(NULL), neg_model(NULL), features(NULL)
{
    m_parameters->add(&m_pos_pseudo,
            "pos_pseudo","pseudo count for positive class");
    m_parameters->add(&m_neg_pseudo,
            "neg_pseudo", "pseudo count for negative class");

    m_parameters->add((CSGObject**) &pos_model,
            "pos_model", "LinearHMM modelling positive class.");
    m_parameters->add((CSGObject**) &neg_model,
            "neg_model", "LinearHMM modelling negative class.");

    m_parameters->add((CSGObject**) &features,
            "features", "String Features.");
}

CPluginEstimate::~CPluginEstimate()
{
    SG_UNREF(pos_model);
    SG_UNREF(neg_model);

    SG_UNREF(features);
}

bool CPluginEstimate::train_machine(CFeatures* data)
{
    ASSERT(m_labels);
    ASSERT(m_labels->get_label_type() == LT_BINARY);
    if (data)
    {
        if (data->get_feature_class() != C_STRING ||
                data->get_feature_type() != F_WORD)
        {
            SG_ERROR("Features not of class string type word\n");
        }

        set_features((CStringFeatures<uint16_t>*) data);
    }
    ASSERT(features);

    SG_UNREF(pos_model);
    SG_UNREF(neg_model);

    pos_model=new CLinearHMM(features);
    neg_model=new CLinearHMM(features);

    SG_REF(pos_model);
    SG_REF(neg_model);

    int32_t* pos_indizes=SG_MALLOC(int32_t, ((CStringFeatures<uint16_t>*) features)->get_num_vectors());
    int32_t* neg_indizes=SG_MALLOC(int32_t, ((CStringFeatures<uint16_t>*) features)->get_num_vectors());

    ASSERT(m_labels->get_num_labels()==features->get_num_vectors());

    int32_t pos_idx=0;
    int32_t neg_idx=0;

    for (int32_t i=0; i<m_labels->get_num_labels(); i++)
    {
        if (((CBinaryLabels*) m_labels)->get_label(i) > 0)
            pos_indizes[pos_idx++]=i;
        else
            neg_indizes[neg_idx++]=i;
    }

    SG_INFO( "training using pseudos %f and %f\n", m_pos_pseudo, m_neg_pseudo);
    pos_model->train(pos_indizes, pos_idx, m_pos_pseudo);
    neg_model->train(neg_indizes, neg_idx, m_neg_pseudo);

    SG_FREE(pos_indizes);
    SG_FREE(neg_indizes);

    return true;
}

CBinaryLabels* CPluginEstimate::apply_binary(CFeatures* data)
{
    if (data)
    {
        if (data->get_feature_class() != C_STRING ||
            data->get_feature_type() != F_WORD)
        {
            SG_ERROR("Features not of class string type word\n");
        }

        set_features((CStringFeatures<uint16_t>*) data);
    }

    ASSERT(features);
    SGVector<float64_t> result(features->get_num_vectors());

    for (int32_t vec=0; vec<features->get_num_vectors(); vec++)
        result[vec] = apply_one(vec);

    return new CBinaryLabels(result);
}

float64_t CPluginEstimate::apply_one(int32_t vec_idx)
{
    ASSERT(features);

    int32_t len;
    bool free_vec;
    uint16_t* vector=features->get_feature_vector(vec_idx, len, free_vec);

    if ((!pos_model) || (!neg_model))
        SG_ERROR( "model(s) not assigned\n");

    float64_t result=pos_model->get_log_likelihood_example(vector, len) - neg_model->get_log_likelihood_example(vector, len);
    features->free_feature_vector(vector, vec_idx, free_vec);
    return result;
}