NbodyTree.cpp

浏览该文件的文档.

/*
 * Copyright (c) The Shogun Machine Learning Toolbox
 * Written (w) 2014 Parijat Mazumdar
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice, this
 * list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 * this list of conditions and the following disclaimer in the documentation
 * and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 * The views and conclusions contained in the software and documentation are those
 * of the authors and should not be interpreted as representing official policies,
 * either expressed or implied, of the Shogun Development Team.
 */

#include <shogun/multiclass/tree/NbodyTree.h>
#include <shogun/distributions/KernelDensity.h>

using namespace shogun;

CNbodyTree::CNbodyTree(int32_t leaf_size, EDistanceType d)
: CTreeMachine<NbodyTreeNodeData>()
{
    init();

    m_leaf_size=leaf_size;
    m_dist=d;
}

void CNbodyTree::build_tree(CDenseFeatures<float64_t>* data)
{
    REQUIRE(data,"data not set\n");
    REQUIRE(m_leaf_size>0,"Leaf size should be greater than 0\n");

    m_knn_done=false;
    m_data=data->get_feature_matrix();

    m_vec_id=SGVector<index_t>(m_data.num_cols);
    m_vec_id.range_fill(0);

    set_root(recursive_build(0,m_data.num_cols-1));
}

void CNbodyTree::query_knn(CDenseFeatures<float64_t>* data, int32_t k)
{
    REQUIRE(data,"Query data not supplied\n")
    REQUIRE(data->get_num_features()==m_data.num_rows,"query data dimension should be same as training data dimension\n")

    m_knn_done=true;
    SGMatrix<float64_t> qfeats=data->get_feature_matrix();
    m_knn_dists=SGMatrix<float64_t>(k,qfeats.num_cols);
    m_knn_indices=SGMatrix<index_t>(k,qfeats.num_cols);
    int32_t dim=qfeats.num_rows;

    for (int32_t i=0;i<qfeats.num_cols;i++)
    {
        CKNNHeap* heap=new CKNNHeap(k);
        bnode_t* root=NULL;
        if (m_root)
            root=dynamic_cast<bnode_t*>(m_root);

        float64_t mdist=min_dist(root,qfeats.matrix+i*dim,dim);
        query_knn_single(heap,mdist,root,qfeats.matrix+i*dim,dim);
        memcpy(m_knn_dists.matrix+i*k,heap->get_dists(),k*sizeof(float64_t));
        memcpy(m_knn_indices.matrix+i*k,heap->get_indices(),k*sizeof(index_t));

        delete(heap);
    }
}

SGVector<float64_t> CNbodyTree::log_kernel_density(SGMatrix<float64_t> test, EKernelType kernel, float64_t h, float64_t atol, float64_t rtol)
{
    int32_t dim=m_data.num_rows;
    REQUIRE(test.num_rows==dim,"dimensions of training data and test data should be the same\n")

    float64_t log_atol=CMath::log(atol*m_data.num_cols);
    float64_t log_rtol=CMath::log(rtol);
    float64_t log_kernel_norm=CKernelDensity::log_norm(kernel,h,dim);
    SGVector<float64_t> log_density(test.num_cols);
    for (int32_t i=0;i<test.num_cols;i++)
    {
        bnode_t* root=NULL;
        if (m_root)
            root=dynamic_cast<bnode_t*>(m_root);

        float64_t lower_dist=0;
        float64_t upper_dist=0;
        min_max_dist(test.matrix+i*dim,root,lower_dist,upper_dist,dim);

        float64_t min_bound=CMath::log(m_data.num_cols)+CKernelDensity::log_kernel(kernel,upper_dist,h);
        float64_t max_bound=CMath::log(m_data.num_cols)+CKernelDensity::log_kernel(kernel,lower_dist,h);
        float64_t spread=logdiffexp(max_bound,min_bound);

        get_kde_single(root,test.matrix+i*dim,kernel,h,log_atol,log_rtol,log_kernel_norm,min_bound,spread,min_bound,spread);
        log_density[i]=logsumexp(min_bound,spread-CMath::log(2))+log_kernel_norm-CMath::log(m_data.num_cols);
    }

    return log_density;
}

SGVector<float64_t> CNbodyTree::log_kernel_density_dual(SGMatrix<float64_t> test, SGVector<index_t> qid, bnode_t* qroot, EKernelType kernel, float64_t h, float64_t atol, float64_t rtol)
{
    int32_t dim=m_data.num_rows;
    REQUIRE(test.num_rows==dim,"dimensions of training data and test data should be the same\n")

    float64_t log_atol=CMath::log(atol*m_data.num_cols*test.num_cols);
    float64_t log_rtol=CMath::log(rtol);
    float64_t log_kernel_norm=CKernelDensity::log_norm(kernel,h,dim);
    SGVector<float64_t> log_density(test.num_cols);
    log_density.fill_vector(log_density.vector,log_density.vlen,-CMath::INFTY);

    bnode_t* rroot=NULL;
    if (m_root)
        rroot=dynamic_cast<bnode_t*>(m_root);

    float64_t upper_dist=max_dist_dual(rroot,qroot);
    float64_t lower_dist=min_dist_dual(rroot,qroot);
    float64_t min_bound=CMath::log(test.num_cols)+CMath::log(m_data.num_cols)+CKernelDensity::log_kernel(kernel,upper_dist,h);
    float64_t max_bound=CMath::log(test.num_cols)+CMath::log(m_data.num_cols)+CKernelDensity::log_kernel(kernel,lower_dist,h);
    float64_t spread=logdiffexp(max_bound,min_bound);

    kde_dual(rroot,qroot,qid,test,log_density,kernel,h,log_atol,log_rtol,log_kernel_norm,min_bound,spread,min_bound,spread);

    float64_t log_n=CMath::log(m_data.num_cols);
    for (int32_t i=0;i<test.num_cols;i++)
        log_density[i]=log_density[i]+log_kernel_norm-log_n;

    return log_density;
}

SGMatrix<float64_t> CNbodyTree::get_knn_dists()
{
    if (m_knn_done)
        return m_knn_dists;

    SG_ERROR("knn query has not been executed yet\n");
    return SGMatrix<float64_t>();
}

SGMatrix<index_t> CNbodyTree::get_knn_indices()
{
    if (m_knn_done)
        return m_knn_indices;

    SG_ERROR("knn query has not been executed yet\n");
    return SGMatrix<index_t>();
}

void CNbodyTree::query_knn_single(CKNNHeap* heap, float64_t mdist, bnode_t* node, float64_t* arr, int32_t dim)
{
    if (mdist>heap->get_max_dist())
        return;

    if (node->data.is_leaf)
    {
        index_t start=node->data.start_idx;
        index_t end=node->data.end_idx;

        for (int32_t i=start;i<=end;i++)
            heap->push(m_vec_id[i],distance(m_vec_id[i],arr,dim));

        return;
    }

    bnode_t* cleft=node->left();
    bnode_t* cright=node->right();

    float64_t min_dist_left=min_dist(cleft,arr,dim);
    float64_t min_dist_right=min_dist(cright,arr,dim);

    if (min_dist_left<=min_dist_right)
    {
        query_knn_single(heap,min_dist_left,cleft,arr,dim);
        query_knn_single(heap,min_dist_right,cright,arr,dim);
    }
    else
    {
        query_knn_single(heap,min_dist_right,cright,arr,dim);
        query_knn_single(heap,min_dist_left,cleft,arr,dim);
    }

    SG_UNREF(cleft);
    SG_UNREF(cright);
}

float64_t CNbodyTree::distance(index_t vec, float64_t* arr, int32_t dim)
{
    float64_t ret=0;
    for (int32_t i=0;i<dim;i++)
        ret+=add_dim_dist(m_data(i,vec)-arr[i]);

    return actual_dists(ret);
}

CBinaryTreeMachineNode<NbodyTreeNodeData>* CNbodyTree::recursive_build(index_t start, index_t end)
{
    bnode_t* node=new bnode_t();
    init_node(node,start,end);

    // stopping critertia
    if (end-start+1<m_leaf_size*2)
    {
        node->data.is_leaf=true;
        return node;
    }

    node->data.is_leaf=false;
    index_t dim=find_split_dim(node);
    index_t mid=(end+start)/2;
    partition(dim,start,end,mid);

    bnode_t* child_left=recursive_build(start,mid);
    bnode_t* child_right=recursive_build(mid+1,end);

    node->left(child_left);
    node->right(child_right);

    return node;
}

void CNbodyTree::get_kde_single(bnode_t* node,float64_t* data, EKernelType kernel, float64_t h, float64_t log_atol, float64_t log_rtol,
    float64_t log_norm, float64_t min_bound_node, float64_t spread_node, float64_t &min_bound_global, float64_t &spread_global)
{
    int32_t n_node=CMath::log(node->data.end_idx-node->data.start_idx+1);
    int32_t n_total=CMath::log(m_data.num_cols);

    // local bound criterion met
    if ((log_norm+spread_node+n_total-n_node)<=logsumexp(log_atol,log_rtol+log_norm+min_bound_node))
        return;

    // global bound criterion met
    if ((log_norm+spread_global)<=logsumexp(log_atol,log_rtol+log_norm+min_bound_global))
        return;

    // node is leaf
    if (node->data.is_leaf)
    {
        min_bound_global=logdiffexp(min_bound_global,min_bound_node);
        spread_global=logdiffexp(spread_global,spread_node);

        for (int32_t i=node->data.start_idx;i<=node->data.end_idx;i++)
        {
            float64_t pt_eval=CKernelDensity::log_kernel(kernel,distance(m_vec_id[i],data,m_data.num_rows),h);
            min_bound_global=logsumexp(pt_eval,min_bound_global);
        }

        return;
    }

    bnode_t* lchild=node->left();
    bnode_t* rchild=node->right();

    float64_t lower_dist=0;
    float64_t upper_dist=0;
    min_max_dist(data,lchild,lower_dist,upper_dist,m_data.num_rows);

    int32_t n_l=lchild->data.end_idx-lchild->data.start_idx+1;
    float64_t lower_bound_childl=CMath::log(n_l)+CKernelDensity::log_kernel(kernel,upper_dist,h);
    float64_t spread_childl=logdiffexp(log(n_l)+CKernelDensity::log_kernel(kernel,lower_dist,h),lower_bound_childl);

    min_max_dist(data,rchild,lower_dist,upper_dist,m_data.num_rows);
    int32_t n_r=rchild->data.end_idx-rchild->data.start_idx+1;
    float64_t lower_bound_childr=CMath::log(n_r)+CKernelDensity::log_kernel(kernel,upper_dist,h);
    float64_t spread_childr=logdiffexp(log(n_r)+CKernelDensity::log_kernel(kernel,lower_dist,h),lower_bound_childr);

    // update global bounds
    min_bound_global=logdiffexp(min_bound_global,min_bound_node);
    min_bound_global=logsumexp(min_bound_global,lower_bound_childl);
    min_bound_global=logsumexp(min_bound_global,lower_bound_childr);

    spread_global=logdiffexp(spread_global,spread_node);
    spread_global=logsumexp(spread_global,spread_childl);
    spread_global=logsumexp(spread_global,spread_childr);

    get_kde_single(lchild,data,kernel,h,log_atol,log_rtol,log_norm,lower_bound_childl,spread_childl,min_bound_global,spread_global);
    get_kde_single(rchild,data,kernel,h,log_atol,log_rtol,log_norm,lower_bound_childr,spread_childr,min_bound_global,spread_global);

    SG_UNREF(lchild);
    SG_UNREF(rchild);
}

void CNbodyTree::kde_dual(bnode_t* refnode, bnode_t* querynode, SGVector<index_t> qid, SGMatrix<float64_t> qdata, SGVector<float64_t> log_density, EKernelType kernel_type, float64_t h, float64_t log_atol, float64_t log_rtol, float64_t log_norm, float64_t min_bound_node, float64_t spread_node, float64_t &min_bound_global, float64_t &spread_global)
{
    int32_t dim=m_data.num_rows;
    float64_t n_node=CMath::log(refnode->data.end_idx-refnode->data.start_idx+1)+CMath::log(querynode->data.end_idx-querynode->data.start_idx+1);
    float64_t n_total=CMath::log(m_data.num_cols*qdata.num_cols);

    bool global_criterion=(log_norm+spread_global)<=logsumexp(log_atol,log_rtol+log_norm+min_bound_global);
    bool local_criterion=(log_norm+spread_node+n_total-n_node)<=logsumexp(log_atol,log_rtol+log_norm+min_bound_node);

    // global bound criterion met || local bound criterion met
    if (global_criterion || local_criterion)
    {
        // log density of all query points in the node is increased by K(mean + spread/2)
        float64_t center_density=logsumexp(min_bound_node,spread_node-CMath::log(2))-CMath::log(querynode->data.end_idx-querynode->data.start_idx+1);
        for (int32_t i=querynode->data.start_idx;i<=querynode->data.end_idx;i++)
            log_density[qid[i]]=logsumexp(log_density[qid[i]],center_density);

        return;
    }

    // both are leaves
    if (refnode->data.is_leaf && querynode->data.is_leaf)
    {
        min_bound_global=logdiffexp(min_bound_global,min_bound_node);
        spread_global=logdiffexp(spread_global,spread_node);

        // point by point evavuation of density
        for (int32_t i=querynode->data.start_idx;i<=querynode->data.end_idx;i++)
        {
            float64_t q=-CMath::INFTY;
            for (int32_t j=refnode->data.start_idx;j<=refnode->data.end_idx;j++)
            {
                float64_t pt_eval=CKernelDensity::log_kernel(kernel_type,distance(m_vec_id[j],qdata.matrix+dim*qid[i],dim),h);
                q=logsumexp(q,pt_eval);
            }

            min_bound_global=logsumexp(min_bound_global,q);
            log_density[qid[i]]=logsumexp(log_density[qid[i]],q);
        }

        return;
    }

    // if query node is leaf - just recurse on the reference tree
    if (querynode->data.is_leaf)
    {
        bnode_t* lchild=refnode->left();
        bnode_t* rchild=refnode->right();
        int32_t queryn=querynode->data.end_idx-querynode->data.start_idx+1;

        // compute bounds for query node and left child of ref node
        float64_t lower_dist=min_dist_dual(querynode,lchild);
        float64_t upper_dist=max_dist_dual(querynode,lchild);
        int32_t refn_l=lchild->data.end_idx-lchild->data.start_idx+1;
        float64_t lower_bound_childl=CMath::log(queryn)+CMath::log(refn_l)+CKernelDensity::log_kernel(kernel_type,upper_dist,h);
        float64_t spread_childl=logdiffexp(CMath::log(queryn)+CMath::log(refn_l)+CKernelDensity::log_kernel(kernel_type,lower_dist,h),lower_bound_childl);

        // compute bounds for query node and right child of ref node
        lower_dist=min_dist_dual(querynode,rchild);
        upper_dist=max_dist_dual(querynode,rchild);
        int32_t refn_r=rchild->data.end_idx-rchild->data.start_idx+1;
        float64_t lower_bound_childr=CMath::log(queryn)+CMath::log(refn_r)+CKernelDensity::log_kernel(kernel_type,upper_dist,h);
        float64_t spread_childr=logdiffexp(CMath::log(queryn)+CMath::log(refn_r)+CKernelDensity::log_kernel(kernel_type,lower_dist,h),lower_bound_childr);

        // update global bounds
        min_bound_global=logdiffexp(min_bound_global,min_bound_node);
        min_bound_global=logsumexp(min_bound_global,lower_bound_childl);
        min_bound_global=logsumexp(min_bound_global,lower_bound_childr);

        spread_global=logdiffexp(spread_global,spread_node);
        spread_global=logsumexp(spread_global,spread_childl);
        spread_global=logsumexp(spread_global,spread_childr);

        kde_dual(lchild,querynode,qid,qdata,log_density,kernel_type,h,log_atol,log_rtol,log_norm,lower_bound_childl,spread_childl, min_bound_global,spread_global);
        kde_dual(rchild,querynode,qid,qdata,log_density,kernel_type,h,log_atol,log_rtol,log_norm,lower_bound_childr,spread_childr, min_bound_global,spread_global);

        SG_UNREF(lchild);
        SG_UNREF(rchild);
        return;
    }

    // if reference node is leaf - just recurse on the query tree
    if (refnode->data.is_leaf)
    {
        int32_t ref_n=refnode->data.end_idx-refnode->data.start_idx+1;
        bnode_t* lchild=querynode->left();
        bnode_t* rchild=querynode->right();

        int32_t query_nl=lchild->data.end_idx-lchild->data.start_idx+1;
        int32_t query_nr=rchild->data.end_idx-rchild->data.start_idx+1;

        // compute bounds for left child of query node and ref node
        float64_t lower_dist=min_dist_dual(refnode,lchild);
        float64_t upper_dist=max_dist_dual(refnode,lchild);
        float64_t lower_bound_childl=CMath::log(query_nl)+CMath::log(ref_n)+CKernelDensity::log_kernel(kernel_type,upper_dist,h);
        float64_t spread_childl=logdiffexp(CMath::log(query_nl)+CMath::log(ref_n)+CKernelDensity::log_kernel(kernel_type,lower_dist,h),lower_bound_childl);

        // compute bounds for right child of query node and ref node
        lower_dist=min_dist_dual(querynode,rchild);
        upper_dist=max_dist_dual(querynode,rchild);
        float64_t lower_bound_childr=CMath::log(query_nr)+CMath::log(ref_n)+CKernelDensity::log_kernel(kernel_type,upper_dist,h);
        float64_t spread_childr=logdiffexp(CMath::log(query_nr)+CMath::log(ref_n)+CKernelDensity::log_kernel(kernel_type,lower_dist,h),lower_bound_childr);

        // update global bounds
        min_bound_global=logdiffexp(min_bound_global,min_bound_node);
        min_bound_global=logsumexp(min_bound_global,lower_bound_childl);
        min_bound_global=logsumexp(min_bound_global,lower_bound_childr);

        spread_global=logdiffexp(spread_global,spread_node);
        spread_global=logsumexp(spread_global,spread_childl);
        spread_global=logsumexp(spread_global,spread_childr);

        kde_dual(refnode,lchild,qid,qdata,log_density,kernel_type,h,log_atol,log_rtol,log_norm,lower_bound_childl,spread_childl,min_bound_global,spread_global);
        kde_dual(refnode,rchild,qid,qdata,log_density,kernel_type,h,log_atol,log_rtol,log_norm,lower_bound_childr,spread_childr,min_bound_global,spread_global);

        SG_UNREF(lchild);
        SG_UNREF(rchild);
        return;
    }

    // if none of above -  apply 4 way recursion in both trees: left-left, left-right, right-left, right-right
    bnode_t* refchildl=refnode->left();
    bnode_t* refchildr=refnode->right();
    bnode_t* querychildl=querynode->left();
    bnode_t* querychildr=querynode->right();

    float64_t refn_l=refchildl->data.end_idx-refchildl->data.start_idx+1;
    float64_t refn_r=refchildr->data.end_idx-refchildr->data.start_idx+1;
    float64_t queryn_l=querychildl->data.end_idx-querychildl->data.start_idx+1;
    float64_t queryn_r=querychildr->data.end_idx-querychildr->data.start_idx+1;

    // left child-left child bounds
    float64_t lower_dist=min_dist_dual(querychildl,refchildl);
    float64_t upper_dist=max_dist_dual(querychildl,refchildl);
    float64_t lower_bound_ll=CMath::log(queryn_l)+CMath::log(refn_l)+CKernelDensity::log_kernel(kernel_type,upper_dist,h);
    float64_t spread_ll=logdiffexp(CMath::log(queryn_l)+CMath::log(refn_l)+CKernelDensity::log_kernel(kernel_type,lower_dist,h),lower_bound_ll);

    // left-right bounds
    lower_dist=min_dist_dual(querychildl,refchildr);
    upper_dist=max_dist_dual(querychildl,refchildr);
    float64_t lower_bound_lr=CMath::log(queryn_l)+CMath::log(refn_r)+CKernelDensity::log_kernel(kernel_type,upper_dist,h);
    float64_t spread_lr=logdiffexp(CMath::log(queryn_l)+CMath::log(refn_r)+CKernelDensity::log_kernel(kernel_type,lower_dist,h),lower_bound_lr);

    // right-left bounds
    lower_dist=min_dist_dual(querychildr,refchildl);
    upper_dist=max_dist_dual(querychildr,refchildl);
    float64_t lower_bound_rl=CMath::log(queryn_r)+CMath::log(refn_l)+CKernelDensity::log_kernel(kernel_type,upper_dist,h);
    float64_t spread_rl=logdiffexp(CMath::log(queryn_r)+CMath::log(refn_l)+CKernelDensity::log_kernel(kernel_type,lower_dist,h),lower_bound_rl);

    // right-right bounds
    lower_dist=min_dist_dual(querychildr,refchildr);
    upper_dist=max_dist_dual(querychildr,refchildr);
    float64_t lower_bound_rr=CMath::log(queryn_r)+CMath::log(refn_r)+CKernelDensity::log_kernel(kernel_type,upper_dist,h);
    float64_t spread_rr=logdiffexp(CMath::log(queryn_r)+CMath::log(refn_r)+CKernelDensity::log_kernel(kernel_type,lower_dist,h),lower_bound_rr);

    // update global bound and spread
    min_bound_global=logdiffexp(min_bound_global,min_bound_node);
    min_bound_global=logsumexp(min_bound_global,lower_bound_ll);
    min_bound_global=logsumexp(min_bound_global,lower_bound_lr);
    min_bound_global=logsumexp(min_bound_global,lower_bound_rl);
    min_bound_global=logsumexp(min_bound_global,lower_bound_rr);

    spread_global=logdiffexp(spread_global,spread_node);
    spread_global=logsumexp(spread_global,spread_ll);
    spread_global=logsumexp(spread_global,spread_lr);
    spread_global=logsumexp(spread_global,spread_rl);
    spread_global=logsumexp(spread_global,spread_rr);

    // left-left and left-right recursions
    kde_dual(refchildl,querychildl,qid,qdata,log_density,kernel_type,h,log_atol,log_rtol,log_norm,lower_bound_ll,spread_ll, min_bound_global,spread_global);
    kde_dual(refchildr,querychildl,qid,qdata,log_density,kernel_type,h,log_atol,log_rtol,log_norm,lower_bound_lr,spread_lr, min_bound_global,spread_global);

    // right-left and right-right recursions
    kde_dual(refchildl,querychildr,qid,qdata,log_density,kernel_type,h,log_atol,log_rtol,log_norm,lower_bound_rl,spread_rl, min_bound_global,spread_global);
    kde_dual(refchildr,querychildr,qid,qdata,log_density,kernel_type,h,log_atol,log_rtol,log_norm,lower_bound_rr,spread_rr, min_bound_global, spread_global);

    SG_UNREF(refchildl);
    SG_UNREF(refchildr);
    SG_UNREF(querychildl);
    SG_UNREF(querychildr);
}

void CNbodyTree::partition(index_t dim, index_t start, index_t end, index_t mid)
{
    // in-place partial quick-sort
    index_t left=start;
    index_t right=end;
    while (true)
    {
        index_t midindex=left;
        for (int32_t i=left;i<right;i++)
        {
            if (m_data(dim,m_vec_id[i])<m_data(dim,m_vec_id[right]))
            {
                CMath::swap(*(m_vec_id.vector+i),*(m_vec_id.vector+midindex));
                midindex+=1;
            }
        }

        CMath::swap(*(m_vec_id.vector+midindex),*(m_vec_id.vector+right));
        if (midindex==mid)
            break;
        else if (midindex<mid)
            left=midindex+1;
        else
            right=midindex-1;
    }
}

index_t CNbodyTree::find_split_dim(bnode_t* node)
{
    SGVector<float64_t> upper_bounds=node->data.bbox_upper;
    SGVector<float64_t> lower_bounds=node->data.bbox_lower;

    index_t max_dim=0;
    float64_t max_spread=-1;
    for (int32_t i=0;i<m_data.num_rows;i++)
    {
        float64_t spread=upper_bounds[i]-lower_bounds[i];
        if (spread>max_spread)
        {
            max_spread=spread;
            max_dim=i;
        }
    }

    return max_dim;
}

void CNbodyTree::init()
{
    m_data=SGMatrix<float64_t>();
    m_leaf_size=1;
    m_vec_id=SGVector<index_t>();
    m_dist=D_EUCLIDEAN;
    m_knn_done=false;
    m_knn_dists=SGMatrix<float64_t>();
    m_knn_indices=SGMatrix<index_t>();

    SG_ADD(&m_data,"m_data","data matrix",MS_NOT_AVAILABLE);
    SG_ADD(&m_leaf_size,"m_leaf_size","leaf size",MS_NOT_AVAILABLE);
    SG_ADD(&m_vec_id,"m_vec_id","id of vectors",MS_NOT_AVAILABLE);
    SG_ADD(&m_knn_done,"knn_done","knn done or not",MS_NOT_AVAILABLE);
    SG_ADD(&m_knn_dists,"m_knn_dists","knn distances",MS_NOT_AVAILABLE);
    SG_ADD(&m_knn_indices,"knn_indices","knn indices",MS_NOT_AVAILABLE);
}