RescaleFeatures.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) 20013 Viktor Gal
 * Copyright (C) 2013 Viktor Gal
 */

#include <shogun/preprocessor/RescaleFeatures.h>

using namespace shogun;

CRescaleFeatures::CRescaleFeatures()
 : CDensePreprocessor<float64_t>(),
 m_initialized(false)
{
    register_parameters();
}

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

bool CRescaleFeatures::init(CFeatures* features)
{
    if (!m_initialized)
    {
        ASSERT(features->get_feature_class()==C_DENSE);
        ASSERT(features->get_feature_type()==F_DREAL);

        CDenseFeatures<float64_t>* simple_features=(CDenseFeatures<float64_t>*) features;
        int32_t num_examples = simple_features->get_num_vectors();
        int32_t num_features = simple_features->get_num_features();
        REQUIRE(num_examples > 1,
                        "number of feature vectors should be at least 2!\n");

        SG_INFO("Extracting min and range values for each feature\n")

        m_min = SGVector<float64_t>(num_features);
        m_range = SGVector<float64_t>(num_features);
        SGMatrix<float64_t> feature_matrix=((CDenseFeatures<float64_t>*)features)->get_feature_matrix();
        for (index_t i = 0; i < num_features; i++)
        {
            SGVector<float64_t> vec = feature_matrix.get_row_vector(i);
            float64_t cur_min = vec[0];
            float64_t cur_max = vec[0];

            /* find the max and min values in one loop */
            for (index_t j = 1; j < vec.vlen; j++)
            {
                cur_min = CMath::min(vec[j], cur_min);
                cur_max = CMath::max(vec[j], cur_max);
            }
    
            /* only rescale if range > 0 */
            if ((cur_max - cur_min) > 0) {
                m_min[i] = cur_min;
                m_range[i] = 1.0/(cur_max - cur_min);
            }
            else {
                m_min[i] = 0.0;
                m_range[i] = 1.0;
            }
        }

        m_initialized = true;

        return true;
    }

    return false;
}

void CRescaleFeatures::cleanup()
{
    m_initialized = false;
}

SGMatrix<float64_t> CRescaleFeatures::apply_to_feature_matrix(CFeatures* features)
{
    ASSERT(m_initialized);

    SGMatrix<float64_t> feature_matrix=((CDenseFeatures<float64_t>*)features)->get_feature_matrix();
    ASSERT(feature_matrix.num_rows == m_min.vlen);

    for (index_t i = 0; i < feature_matrix.num_cols; i++)
    {
        float64_t* vec = feature_matrix.get_column_vector(i);
        SGVector<float64_t>::vec1_plus_scalar_times_vec2(vec, -1.0, m_min.vector, feature_matrix.num_rows);
        for (index_t j = 0; j < feature_matrix.num_rows; j++) {
            vec[j] *= m_range[j];
        }
    }

    return feature_matrix;
}

SGVector<float64_t> CRescaleFeatures::apply_to_feature_vector(SGVector<float64_t> vector)
{
    ASSERT(m_initialized);
    ASSERT(m_min.vlen == vector.vlen);

    float64_t* ret = SG_MALLOC(float64_t, vector.vlen);
    SGVector<float64_t>::add(ret, 1.0, vector.vector, -1.0, m_min.vector, vector.vlen);
    for (index_t i = 0; i < vector.vlen; i++) {
        ret[i] *= m_range[i];
    }

    return SGVector<float64_t>(ret,vector.vlen);
}

void CRescaleFeatures::register_parameters()
{
    SG_ADD(&m_min, "min", "minimum values of each feature", MS_NOT_AVAILABLE);
    SG_ADD(&m_range, "range", "Reciprocal of the range of each feature", MS_NOT_AVAILABLE);
    SG_ADD(&m_initialized, "initialized", "Indicator of the state of the preprocessor.", MS_NOT_AVAILABLE);
}