hierarchical_clustering_index.h

/***********************************************************************
 * Software License Agreement (BSD License)
 *
 * Copyright 2008-2011  Marius Muja (mariusm@cs.ubc.ca). All rights reserved.
 * Copyright 2008-2011  David G. Lowe (lowe@cs.ubc.ca). All rights reserved.
 *
 * THE BSD LICENSE
 *
 * 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 AUTHOR ``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 AUTHOR 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.
 *************************************************************************/
 
#ifndef OPENCV_FLANN_HIERARCHICAL_CLUSTERING_INDEX_H_
#define OPENCV_FLANN_HIERARCHICAL_CLUSTERING_INDEX_H_
 
 
#include <algorithm>
#include <map>
#include <limits>
#include <cmath>
 
#include "general.h"
#include "nn_index.h"
#include "dist.h"
#include "matrix.h"
#include "result_set.h"
#include "heap.h"
#include "allocator.h"
#include "random.h"
#include "saving.h"
 
 
namespace cvflann
{
 
struct HierarchicalClusteringIndexParams : public IndexParams
{
    HierarchicalClusteringIndexParams(int branching = 32,
                                      flann_centers_init_t centers_init = FLANN_CENTERS_RANDOM,
                                      int trees = 4, int leaf_size = 100)
    {
        (*this)["algorithm"] = FLANN_INDEX_HIERARCHICAL;
        // The branching factor used in the hierarchical clustering
        (*this)["branching"] = branching;
        // Algorithm used for picking the initial cluster centers
        (*this)["centers_init"] = centers_init;
        // number of parallel trees to build
        (*this)["trees"] = trees;
        // maximum leaf size
        (*this)["leaf_size"] = leaf_size;
    }
};
 
 
template <typename Distance>
class HierarchicalClusteringIndex : public NNIndex<Distance>
{
public:
    typedef typename Distance::ElementType ElementType;
    typedef typename Distance::ResultType DistanceType;
 
private:
 
 
    typedef void (HierarchicalClusteringIndex::* centersAlgFunction)(int, int*, int, int*, int&);
 
    centersAlgFunction chooseCenters;
 
 
 
    void chooseCentersRandom(int k, int* dsindices, int indices_length, int* centers, int& centers_length)
    {
        UniqueRandom r(indices_length);
 
        int index;
        for (index=0; index<k; ++index) {
            bool duplicate = true;
            int rnd;
            while (duplicate) {
                duplicate = false;
                rnd = r.next();
                if (rnd<0) {
                    centers_length = index;
                    return;
                }
 
                centers[index] = dsindices[rnd];
 
                for (int j=0; j<index; ++j) {
                    DistanceType sq = distance(dataset[centers[index]], dataset[centers[j]], dataset.cols);
                    if (sq<1e-16) {
                        duplicate = true;
                    }
                }
            }
        }
 
        centers_length = index;
    }
 
 
    void chooseCentersGonzales(int k, int* dsindices, int indices_length, int* centers, int& centers_length)
    {
        int n = indices_length;
 
        int rnd = rand_int(n);
        CV_DbgAssert(rnd >=0 && rnd < n);
 
        centers[0] = dsindices[rnd];
 
        int index;
        for (index=1; index<k; ++index) {
 
            int best_index = -1;
            DistanceType best_val = 0;
            for (int j=0; j<n; ++j) {
                DistanceType dist = distance(dataset[centers[0]],dataset[dsindices[j]],dataset.cols);
                for (int i=1; i<index; ++i) {
                    DistanceType tmp_dist = distance(dataset[centers[i]],dataset[dsindices[j]],dataset.cols);
                    if (tmp_dist<dist) {
                        dist = tmp_dist;
                    }
                }
                if (dist>best_val) {
                    best_val = dist;
                    best_index = j;
                }
            }
            if (best_index!=-1) {
                centers[index] = dsindices[best_index];
            }
            else {
                break;
            }
        }
        centers_length = index;
    }
 
 
    void chooseCentersKMeanspp(int k, int* dsindices, int indices_length, int* centers, int& centers_length)
    {
        int n = indices_length;
 
        double currentPot = 0;
        DistanceType* closestDistSq = new DistanceType[n];
 
        // Choose one random center and set the closestDistSq values
        int index = rand_int(n);
        CV_DbgAssert(index >=0 && index < n);
        centers[0] = dsindices[index];
 
        // Computing distance^2 will have the advantage of even higher probability further to pick new centers
        // far from previous centers (and this complies to "k-means++: the advantages of careful seeding" article)
        for (int i = 0; i < n; i++) {
            closestDistSq[i] = distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols);
            closestDistSq[i] = ensureSquareDistance<Distance>( closestDistSq[i] );
            currentPot += closestDistSq[i];
        }
 
 
        const int numLocalTries = 1;
 
        // Choose each center
        int centerCount;
        for (centerCount = 1; centerCount < k; centerCount++) {
 
            // Repeat several trials
            double bestNewPot = -1;
            int bestNewIndex = 0;
            for (int localTrial = 0; localTrial < numLocalTries; localTrial++) {
 
                // Choose our center - have to be slightly careful to return a valid answer even accounting
                // for possible rounding errors
                double randVal = rand_double(currentPot);
                for (index = 0; index < n-1; index++) {
                    if (randVal <= closestDistSq[index]) break;
                    else randVal -= closestDistSq[index];
                }
 
                // Compute the new potential
                double newPot = 0;
                for (int i = 0; i < n; i++) {
                    DistanceType dist = distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols);
                    newPot += std::min( ensureSquareDistance<Distance>(dist), closestDistSq[i] );
                }
 
                // Store the best result
                if ((bestNewPot < 0)||(newPot < bestNewPot)) {
                    bestNewPot = newPot;
                    bestNewIndex = index;
                }
            }
 
            // Add the appropriate center
            centers[centerCount] = dsindices[bestNewIndex];
            currentPot = bestNewPot;
            for (int i = 0; i < n; i++) {
                DistanceType dist = distance(dataset[dsindices[i]], dataset[dsindices[bestNewIndex]], dataset.cols);
                closestDistSq[i] = std::min( ensureSquareDistance<Distance>(dist), closestDistSq[i] );
            }
        }
 
        centers_length = centerCount;
 
        delete[] closestDistSq;
    }
 
 
    void GroupWiseCenterChooser(int k, int* dsindices, int indices_length, int* centers, int& centers_length)
    {
        const float kSpeedUpFactor = 1.3f;
 
        int n = indices_length;
 
        DistanceType* closestDistSq = new DistanceType[n];
 
        // Choose one random center and set the closestDistSq values
        int index = rand_int(n);
        CV_DbgAssert(index >=0 && index < n);
        centers[0] = dsindices[index];
 
        for (int i = 0; i < n; i++) {
            closestDistSq[i] = distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols);
        }
 
 
        // Choose each center
        int centerCount;
        for (centerCount = 1; centerCount < k; centerCount++) {
 
            // Repeat several trials
            double bestNewPot = -1;
            int bestNewIndex = 0;
            DistanceType furthest = 0;
            for (index = 0; index < n; index++) {
 
                // We will test only the potential of the points further than current candidate
                if( closestDistSq[index] > kSpeedUpFactor * (float)furthest ) {
 
                    // Compute the new potential
                    double newPot = 0;
                    for (int i = 0; i < n; i++) {
                        newPot += std::min( distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols)
                                            , closestDistSq[i] );
                    }
 
                    // Store the best result
                    if ((bestNewPot < 0)||(newPot <= bestNewPot)) {
                        bestNewPot = newPot;
                        bestNewIndex = index;
                        furthest = closestDistSq[index];
                    }
                }
            }
 
            // Add the appropriate center
            centers[centerCount] = dsindices[bestNewIndex];
            for (int i = 0; i < n; i++) {
                closestDistSq[i] = std::min( distance(dataset[dsindices[i]], dataset[dsindices[bestNewIndex]], dataset.cols)
                                             , closestDistSq[i] );
            }
        }
 
        centers_length = centerCount;
 
        delete[] closestDistSq;
    }
 
 
public:
 
 
    HierarchicalClusteringIndex(const Matrix<ElementType>& inputData, const IndexParams& index_params = HierarchicalClusteringIndexParams(),
                                Distance d = Distance())
        : dataset(inputData), params(index_params), root(NULL), indices(NULL), distance(d)
    {
        memoryCounter = 0;
 
        size_ = dataset.rows;
        veclen_ = dataset.cols;
 
        branching_ = get_param(params,"branching",32);
        centers_init_ = get_param(params,"centers_init", FLANN_CENTERS_RANDOM);
        trees_ = get_param(params,"trees",4);
        leaf_size_ = get_param(params,"leaf_size",100);
 
        if (centers_init_==FLANN_CENTERS_RANDOM) {
            chooseCenters = &HierarchicalClusteringIndex::chooseCentersRandom;
        }
        else if (centers_init_==FLANN_CENTERS_GONZALES) {
            chooseCenters = &HierarchicalClusteringIndex::chooseCentersGonzales;
        }
        else if (centers_init_==FLANN_CENTERS_KMEANSPP) {
            chooseCenters = &HierarchicalClusteringIndex::chooseCentersKMeanspp;
        }
        else if (centers_init_==FLANN_CENTERS_GROUPWISE) {
            chooseCenters = &HierarchicalClusteringIndex::GroupWiseCenterChooser;
        }
        else {
            FLANN_THROW(cv::Error::StsError, "Unknown algorithm for choosing initial centers.");
        }
 
        root = new NodePtr[trees_];
        indices = new int*[trees_];
 
        for (int i=0; i<trees_; ++i) {
            root[i] = NULL;
            indices[i] = NULL;
        }
    }
 
    HierarchicalClusteringIndex(const HierarchicalClusteringIndex&);
    HierarchicalClusteringIndex& operator=(const HierarchicalClusteringIndex&);
 
    virtual ~HierarchicalClusteringIndex()
    {
        if (root!=NULL) {
            delete[] root;
        }
 
        if (indices!=NULL) {
            free_indices();
            delete[] indices;
        }
    }
 
    size_t size() const CV_OVERRIDE
    {
        return size_;
    }
 
    size_t veclen() const CV_OVERRIDE
    {
        return veclen_;
    }
 
 
    int usedMemory() const CV_OVERRIDE
    {
        return pool.usedMemory+pool.wastedMemory+memoryCounter;
    }
 
    void buildIndex() CV_OVERRIDE
    {
        if (branching_<2) {
            FLANN_THROW(cv::Error::StsError, "Branching factor must be at least 2");
        }
 
        free_indices();
 
        for (int i=0; i<trees_; ++i) {
            indices[i] = new int[size_];
            for (size_t j=0; j<size_; ++j) {
                indices[i][j] = (int)j;
            }
            root[i] = pool.allocate<Node>();
            computeClustering(root[i], indices[i], (int)size_, branching_,0);
        }
    }
 
 
    flann_algorithm_t getType() const CV_OVERRIDE
    {
        return FLANN_INDEX_HIERARCHICAL;
    }
 
 
    void saveIndex(FILE* stream) CV_OVERRIDE
    {
        save_value(stream, branching_);
        save_value(stream, trees_);
        save_value(stream, centers_init_);
        save_value(stream, leaf_size_);
        save_value(stream, memoryCounter);
        for (int i=0; i<trees_; ++i) {
            save_value(stream, *indices[i], size_);
            save_tree(stream, root[i], i);
        }
 
    }
 
 
    void loadIndex(FILE* stream) CV_OVERRIDE
    {
        if (root!=NULL) {
            delete[] root;
        }
 
        if (indices!=NULL) {
            free_indices();
            delete[] indices;
        }
 
        load_value(stream, branching_);
        load_value(stream, trees_);
        load_value(stream, centers_init_);
        load_value(stream, leaf_size_);
        load_value(stream, memoryCounter);
 
        indices = new int*[trees_];
        root = new NodePtr[trees_];
        for (int i=0; i<trees_; ++i) {
            indices[i] = new int[size_];
            load_value(stream, *indices[i], size_);
            load_tree(stream, root[i], i);
        }
 
        params["algorithm"] = getType();
        params["branching"] = branching_;
        params["trees"] = trees_;
        params["centers_init"] = centers_init_;
        params["leaf_size"] = leaf_size_;
    }
 
 
    void findNeighbors(ResultSet<DistanceType>& result, const ElementType* vec, const SearchParams& searchParams) CV_OVERRIDE
    {
 
        const int maxChecks = get_param(searchParams,"checks",32);
        const bool explore_all_trees = get_param(searchParams,"explore_all_trees",false);
 
        // Priority queue storing intermediate branches in the best-bin-first search
        const cv::Ptr<Heap<BranchSt>>& heap = Heap<BranchSt>::getPooledInstance(cv::utils::getThreadID(), (int)size_);
 
        std::vector<bool> checked(size_,false);
        int checks = 0;
        for (int i=0; i<trees_; ++i) {
            findNN(root[i], result, vec, checks, maxChecks, heap, checked, explore_all_trees);
            if (!explore_all_trees && (checks >= maxChecks) && result.full())
                break;
        }
 
        BranchSt branch;
        while (heap->popMin(branch) && (checks<maxChecks || !result.full())) {
            NodePtr node = branch.node;
            findNN(node, result, vec, checks, maxChecks, heap, checked, false);
        }
 
        CV_Assert(result.full());
    }
 
    IndexParams getParameters() const CV_OVERRIDE
    {
        return params;
    }
 
 
private:
 
    struct Node
    {
        int pivot;
        int size;
        Node** childs;
        int* indices;
        int level;
    };
    typedef Node* NodePtr;
 
 
 
    typedef BranchStruct<NodePtr, DistanceType> BranchSt;
 
 
 
    void save_tree(FILE* stream, NodePtr node, int num)
    {
        save_value(stream, *node);
        if (node->childs==NULL) {
            int indices_offset = (int)(node->indices - indices[num]);
            save_value(stream, indices_offset);
        }
        else {
            for(int i=0; i<branching_; ++i) {
                save_tree(stream, node->childs[i], num);
            }
        }
    }
 
 
    void load_tree(FILE* stream, NodePtr& node, int num)
    {
        node = pool.allocate<Node>();
        load_value(stream, *node);
        if (node->childs==NULL) {
            int indices_offset;
            load_value(stream, indices_offset);
            node->indices = indices[num] + indices_offset;
        }
        else {
            node->childs = pool.allocate<NodePtr>(branching_);
            for(int i=0; i<branching_; ++i) {
                load_tree(stream, node->childs[i], num);
            }
        }
    }
 
 
    void free_indices()
    {
        if (indices!=NULL) {
            for(int i=0; i<trees_; ++i) {
                if (indices[i]!=NULL) {
                    delete[] indices[i];
                    indices[i] = NULL;
                }
            }
        }
    }
 
 
    void computeLabels(int* dsindices, int indices_length,  int* centers, int centers_length, int* labels, DistanceType& cost)
    {
        cost = 0;
        for (int i=0; i<indices_length; ++i) {
            ElementType* point = dataset[dsindices[i]];
            DistanceType dist = distance(point, dataset[centers[0]], veclen_);
            labels[i] = 0;
            for (int j=1; j<centers_length; ++j) {
                DistanceType new_dist = distance(point, dataset[centers[j]], veclen_);
                if (dist>new_dist) {
                    labels[i] = j;
                    dist = new_dist;
                }
            }
            cost += dist;
        }
    }
 
    void computeClustering(NodePtr node, int* dsindices, int indices_length, int branching, int level)
    {
        node->size = indices_length;
        node->level = level;
 
        if (indices_length < leaf_size_) { // leaf node
            node->indices = dsindices;
            std::sort(node->indices,node->indices+indices_length);
            node->childs = NULL;
            return;
        }
 
        std::vector<int> centers(branching);
        std::vector<int> labels(indices_length);
 
        int centers_length;
        (this->*chooseCenters)(branching, dsindices, indices_length, &centers[0], centers_length);
 
        if (centers_length<branching) {
            node->indices = dsindices;
            std::sort(node->indices,node->indices+indices_length);
            node->childs = NULL;
            return;
        }
 
 
        //  assign points to clusters
        DistanceType cost;
        computeLabels(dsindices, indices_length, &centers[0], centers_length, &labels[0], cost);
 
        node->childs = pool.allocate<NodePtr>(branching);
        int start = 0;
        int end = start;
        for (int i=0; i<branching; ++i) {
            for (int j=0; j<indices_length; ++j) {
                if (labels[j]==i) {
                    std::swap(dsindices[j],dsindices[end]);
                    std::swap(labels[j],labels[end]);
                    end++;
                }
            }
 
            node->childs[i] = pool.allocate<Node>();
            node->childs[i]->pivot = centers[i];
            node->childs[i]->indices = NULL;
            computeClustering(node->childs[i],dsindices+start, end-start, branching, level+1);
            start=end;
        }
    }
 
 
 
    void findNN(NodePtr node, ResultSet<DistanceType>& result, const ElementType* vec, int& checks, int maxChecks,
                const cv::Ptr<Heap<BranchSt>>& heap, std::vector<bool>& checked, bool explore_all_trees = false)
    {
        if (node->childs==NULL) {
            if (!explore_all_trees && (checks>=maxChecks) && result.full()) {
                return;
            }
            for (int i=0; i<node->size; ++i) {
                int index = node->indices[i];
                if (!checked[index]) {
                    DistanceType dist = distance(dataset[index], vec, veclen_);
                    result.addPoint(dist, index);
                    checked[index] = true;
                    ++checks;
                }
            }
        }
        else {
            DistanceType* domain_distances = new DistanceType[branching_];
            int best_index = 0;
            domain_distances[best_index] = distance(vec, dataset[node->childs[best_index]->pivot], veclen_);
            for (int i=1; i<branching_; ++i) {
                domain_distances[i] = distance(vec, dataset[node->childs[i]->pivot], veclen_);
                if (domain_distances[i]<domain_distances[best_index]) {
                    best_index = i;
                }
            }
            for (int i=0; i<branching_; ++i) {
                if (i!=best_index) {
                    heap->insert(BranchSt(node->childs[i],domain_distances[i]));
                }
            }
            delete[] domain_distances;
            findNN(node->childs[best_index],result,vec, checks, maxChecks, heap, checked, explore_all_trees);
        }
    }
 
private:
 
 
    const Matrix<ElementType> dataset;
 
    IndexParams params;
 
 
    size_t size_;
 
    size_t veclen_;
 
    NodePtr* root;
 
    int** indices;
 
 
    Distance distance;
 
    PooledAllocator pool;
 
    int memoryCounter;
 
    int branching_;
    int trees_;
    flann_centers_init_t centers_init_;
    int leaf_size_;
 
 
};
 
}
 
 
#endif /* OPENCV_FLANN_HIERARCHICAL_CLUSTERING_INDEX_H_ */