SGSparseVector.cpp

浏览该文件的文档.

/*
 * 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) 2013,2014 Thoralf Klein
 * Written (W) 2012 Fernando José Iglesias García
 * Written (W) 2010,2012 Soeren Sonnenburg
 * Copyright (C) 2010 Berlin Institute of Technology
 * Copyright (C) 2012 Soeren Sonnenburg
 */

#include <shogun/lib/SGSparseVector.h>
#include <shogun/lib/SGVector.h>
#include <shogun/mathematics/Math.h>
#include <shogun/io/File.h>

namespace shogun
{

template <class T>
SGSparseVector<T>::SGSparseVector() : SGReferencedData()
{
    init_data();
}

template <class T>
SGSparseVector<T>::SGSparseVector(SGSparseVectorEntry<T> * feats, index_t num_entries,
                                  bool ref_counting) :
    SGReferencedData(ref_counting),
    num_feat_entries(num_entries), features(feats)
{
}

template <class T>
SGSparseVector<T>::SGSparseVector(index_t num_entries, bool ref_counting) :
    SGReferencedData(ref_counting),
    num_feat_entries(num_entries)
{
    features = SG_MALLOC(SGSparseVectorEntry<T>, num_feat_entries);
}

template <class T>
SGSparseVector<T>::SGSparseVector(const SGSparseVector &orig) :
    SGReferencedData(orig)
{
    copy_data(orig);
}

template <class T>
SGSparseVector<T>::~SGSparseVector()
{
    unref();
}

template <class T>
T SGSparseVector<T>::dense_dot(T alpha, T * vec, int32_t dim, T b)
{
    ASSERT(vec)
    T result = b;

    if (features)
    {
        for (int32_t i = 0; i < num_feat_entries; i++)
        {
            if (features[i].feat_index < dim)
            {
                result += alpha * vec[features[i].feat_index] * features[i].entry;
            }
        }
    }

    return result;
}

template <class T>
void SGSparseVector<T>::add_to_dense(T alpha, T * vec, int32_t dim, bool abs_val)
{
    REQUIRE(vec, "vec must not be NULL\n");

    if (abs_val)
    {
        for (int32_t i = 0; i < num_feat_entries; i++)
        {
            vec[features[i].feat_index] += alpha*CMath::abs(features[i].entry);
        }
    }
    else
    {
        for (int32_t i = 0; i < num_feat_entries; i++)
        {
            vec[features[i].feat_index] += alpha*features[i].entry;
        }
    }
}

template <class T>
template <typename ST>
T SGSparseVector<T>::dense_dot(SGVector<ST> vec)
{
    ASSERT(vec)
    T result(0.0);

    if (features)
    {
        for (int32_t i = 0; i < num_feat_entries; i++)
        {
            if (features[i].feat_index < vec.vlen)
                result += static_cast<T>(vec[features[i].feat_index])
                          * features[i].entry;
        }
    }

    return result;
}

template complex128_t SGSparseVector<complex128_t>::dense_dot<float64_t>(SGVector<float64_t>);
template complex128_t SGSparseVector<complex128_t>::dense_dot<int32_t>(SGVector<int32_t> vec);
template float64_t SGSparseVector<float64_t>::dense_dot<int32_t>(SGVector<int32_t> vec);

template <class T>
T SGSparseVector<T>::sparse_dot(const SGSparseVector<T> &v)
{
    return sparse_dot(*this, v);
}

template <class T>
T SGSparseVector<T>::sparse_dot(const SGSparseVector<T> &a, const SGSparseVector<T> &b)
{
    if (a.num_feat_entries == 0 || b.num_feat_entries == 0)
    {
        return 0;
    }

    if (!a.is_sorted() || !b.is_sorted())
    {
        return dot_prod_expensive_unsorted(a, b);
    }

    T dot_prod = 0;
    index_t a_idx = 0, b_idx = 0;

    while (a_idx < a.num_feat_entries && b_idx < b.num_feat_entries)
    {
        if (a.features[a_idx].feat_index < b.features[b_idx].feat_index)
        {
            a_idx++;
        }
        else if (a.features[a_idx].feat_index > b.features[b_idx].feat_index)
        {
            b_idx++;
        }
        else
        {
            // a.features[a_idx].feat_index == b.features[b_idx].feat_index
            dot_prod += a.features[a_idx].entry * b.features[b_idx].entry;
            a_idx++;
            b_idx++;
        }
    }

    return dot_prod;
}

template<class T>
int32_t SGSparseVector<T>::get_num_dimensions()
{
    if (!features)
    {
        return 0;
    }

    int32_t dimensions = -1;

    for (index_t i = 0; i < num_feat_entries; i++)
    {
        if (features[i].feat_index > dimensions)
        {
            dimensions = features[i].feat_index;
        }
    }

    return dimensions + 1;
}

template<class T>
void SGSparseVector<T>::sort_features(bool stable_pointer)
{
    if (!num_feat_entries)
    {
        return;
    }

    // remember old pointer to enforce quarantee
    const SGSparseVectorEntry<T> * old_features_ptr = features;

    int32_t * feat_idx = SG_MALLOC(int32_t, num_feat_entries);

    for (index_t j = 0; j < num_feat_entries; j++)
    {
        feat_idx[j] = features[j].feat_index;
    }

    CMath::qsort_index(feat_idx, features, num_feat_entries);
    SG_FREE(feat_idx);

    for (index_t j = 1; j < num_feat_entries; j++)
    {
        REQUIRE(features[j - 1].feat_index <= features[j].feat_index,
                "sort_features(): failed sanity check %d <= %d after sorting (comparing indices features[%d] <= features[%d], features=%d)\n",
                features[j - 1].feat_index, features[j].feat_index, j - 1, j, num_feat_entries);
    }

    // compression: removing zero-entries and merging features with same index
    int32_t last_index = 0;

    for (index_t j = 1; j < num_feat_entries; j++)
    {
        // always true, but kept for future changes
        REQUIRE(last_index < j,
                "sort_features(): target index %d must not exceed source index j=%d",
                last_index, j);
        REQUIRE(features[last_index].feat_index <= features[j].feat_index,
                "sort_features(): failed sanity check %d = features[%d].feat_index <= features[%d].feat_index = %d\n",
                features[last_index].feat_index, last_index, j, features[j].feat_index);

        // merging of features with same index
        if (features[last_index].feat_index == features[j].feat_index)
        {
            features[last_index].entry += features[j].entry;
            continue;
        }

        // only skip to next element if current one is not zero
        if (features[last_index].entry != 0.0)
        {
            last_index++;
        }

        features[last_index] = features[j];
    }

    // remove single first element if zero (not caught by loop)
    if (features[last_index].entry == 0.0)
    {
        last_index--;
    }

    int32_t new_feat_count = last_index + 1;
    ASSERT(new_feat_count <= num_feat_entries);

    // shrinking vector
    if (!stable_pointer)
    {
        SG_SINFO("shrinking vector from %d to %d\n", num_feat_entries, new_feat_count);
        features = SG_REALLOC(SGSparseVectorEntry<T>, features, num_feat_entries, new_feat_count);
    }

    num_feat_entries = new_feat_count;

    for (index_t j = 1; j < num_feat_entries; j++)
    {
        REQUIRE(features[j - 1].feat_index < features[j].feat_index,
                "sort_features(): failed sanity check %d < %d after sorting (comparing indices features[%d] < features[%d], features=%d)\n",
                features[j - 1].feat_index, features[j].feat_index, j - 1, j, num_feat_entries);
    }

    // compare with old pointer to enforce quarantee
    if (stable_pointer)
    {
        ASSERT(old_features_ptr == features);
    }

    return;
}

template<class T>
bool SGSparseVector<T>::is_sorted() const
{
    if (num_feat_entries == 0 || num_feat_entries == 1)
    {
        return true;
    }

    for (index_t j = 1; j < num_feat_entries; j++)
    {
        if (features[j - 1].feat_index >= features[j].feat_index)
        {
            return false;
        }
    }

    return true;
}

template<class T>
T SGSparseVector<T>::get_feature(int32_t index)
{
    T ret = 0;

    if (features)
    {
        for (index_t i = 0; i < num_feat_entries; i++)
            if (features[i].feat_index == index)
            {
                ret += features[i].entry ;
            }
    }

    return ret ;
}

template<class T>
SGVector<T> SGSparseVector<T>::get_dense()
{
    SGVector<T> dense;

    if (features)
    {
        dense.resize_vector(get_num_dimensions());
        dense.zero();

        for (index_t i = 0; i < num_feat_entries; i++)
        {
            dense.vector[features[i].feat_index] += features[i].entry;
        }
    }

    return dense;
}

template<class T>
SGVector<T> SGSparseVector<T>::get_dense(int32_t dimension)
{
    SGVector<T> dense(dimension);
    dense.zero();

    if (features)
    {
        REQUIRE(get_num_dimensions() <= dimension, "get_dense(dimension=%d): sparse dimension %d exceeds requested dimension\n",
                dimension, get_num_dimensions());

        for (index_t i = 0; i < num_feat_entries; i++)
        {
            dense.vector[features[i].feat_index] += features[i].entry;
        }
    }

    return dense;
}

template<class T>
SGSparseVector<T> SGSparseVector<T>::clone() const
{
    SGSparseVectorEntry <T> * copy = SG_MALLOC(SGSparseVectorEntry <T>, num_feat_entries);
    memcpy(copy, features, num_feat_entries * sizeof(SGSparseVectorEntry <T>));
    return SGSparseVector<T>(copy, num_feat_entries);
}

template<class T> void SGSparseVector<T>::load(CFile * loader)
{
    ASSERT(loader)
    unref();

    SG_SET_LOCALE_C;
    loader->get_sparse_vector(features, num_feat_entries);
    SG_RESET_LOCALE;
}

template<class T> void SGSparseVector<T>::save(CFile * saver)
{
    ASSERT(saver)

    SG_SET_LOCALE_C;
    saver->set_sparse_vector(features, num_feat_entries);
    SG_RESET_LOCALE;
}

template <>
void SGSparseVector<complex128_t>::load(CFile * loader)
{
    SG_SERROR("SGSparseVector::load():: Not supported for complex128_t\n");
}

template <>
void SGSparseVector<complex128_t>::save(CFile * saver)
{
    SG_SERROR("SGSparseVector::save():: Not supported for complex128_t\n");
}

template <class T>
void SGSparseVector<T>::copy_data(const SGReferencedData &orig)
{
    num_feat_entries = ((SGSparseVector *)(&orig))->num_feat_entries;
    features = ((SGSparseVector *)(&orig))->features;
}

template <class T>
void SGSparseVector<T>::init_data()
{
    num_feat_entries = 0;
    features = NULL;
}

template <class T>
void SGSparseVector<T>::free_data()
{
    num_feat_entries = 0;
    SG_FREE(features);
}

template <class T>
T SGSparseVector<T>::dot_prod_expensive_unsorted(const SGSparseVector<T> &a, const SGSparseVector<T> &b)
{
    SG_SWARNING("Computing sparse_dot(a,b) on unsorted vectors is very expensive: O(n^2)\n");
    SG_SWARNING("Using fallback to give correct results because upstream code does not sort.\n");

    T dot_prod = 0;

    for (index_t b_idx = 0; b_idx < b.num_feat_entries; ++b_idx)
    {
        const T tmp = b.features[b_idx].entry;

        for (index_t a_idx = 0; a_idx < a.num_feat_entries; ++a_idx)
        {
            if (a.features[a_idx].feat_index == b.features[b_idx].feat_index)
            {
                dot_prod += tmp * a.features[a_idx].entry;
            }
        }
    }

    return dot_prod;
}

template <>
void SGSparseVector<bool>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%d", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry ? 1 : 0);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<char>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%c", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<int8_t>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%d", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<uint8_t>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%u", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<int16_t>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%d", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<uint16_t>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%u", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<int32_t>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%d", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<uint32_t>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%u", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<int64_t>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%lld", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<uint64_t>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%llu ", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<float32_t>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%g", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<float64_t>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%.18g", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<floatmax_t>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
    {
        SG_SPRINT("%s%s%d:%.36Lg", prefix, i == 0 ? "" : " ", features[i].feat_index, features[i].entry);
    }

    SG_SPRINT("%s]\n", prefix);
}

template <>
void SGSparseVector<complex128_t>::display_vector(const char * name, const char * prefix)
{
    SG_SPRINT("%s%s=[", prefix, name);

    for (int32_t i = 0; i < num_feat_entries; i++)
        SG_SPRINT("%s%s%d:(%.18lg+i%.18lg)", prefix, i == 0 ? "" : " ", features[i].feat_index,
                  features[i].entry.real(), features[i].entry.imag());

    SG_SPRINT("%s]\n", prefix);
}

template class SGSparseVector<bool>;
template class SGSparseVector<char>;
template class SGSparseVector<int8_t>;
template class SGSparseVector<uint8_t>;
template class SGSparseVector<int16_t>;
template class SGSparseVector<uint16_t>;
template class SGSparseVector<int32_t>;
template class SGSparseVector<uint32_t>;
template class SGSparseVector<int64_t>;
template class SGSparseVector<uint64_t>;
template class SGSparseVector<float32_t>;
template class SGSparseVector<float64_t>;
template class SGSparseVector<floatmax_t>;
template class SGSparseVector<complex128_t>;
}