autotuned_index.h

/***********************************************************************
 * Software License Agreement (BSD License)
 *
 * Copyright 2008-2009  Marius Muja (mariusm@cs.ubc.ca). All rights reserved.
 * Copyright 2008-2009  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_AUTOTUNED_INDEX_H_
#define OPENCV_FLANN_AUTOTUNED_INDEX_H_
 
 
#include <sstream>
 
#include "nn_index.h"
#include "ground_truth.h"
#include "index_testing.h"
#include "sampling.h"
#include "kdtree_index.h"
#include "kdtree_single_index.h"
#include "kmeans_index.h"
#include "composite_index.h"
#include "linear_index.h"
#include "logger.h"
 
namespace cvflann
{
 
template<typename Distance>
NNIndex<Distance>* create_index_by_type(const Matrix<typename Distance::ElementType>& dataset, const IndexParams& params, const Distance& distance);
 
 
struct AutotunedIndexParams : public IndexParams
{
    AutotunedIndexParams(float target_precision = 0.8, float build_weight = 0.01, float memory_weight = 0, float sample_fraction = 0.1)
    {
        (*this)["algorithm"] = FLANN_INDEX_AUTOTUNED;
        // precision desired (used for autotuning, -1 otherwise)
        (*this)["target_precision"] = target_precision;
        // build tree time weighting factor
        (*this)["build_weight"] = build_weight;
        // index memory weighting factor
        (*this)["memory_weight"] = memory_weight;
        // what fraction of the dataset to use for autotuning
        (*this)["sample_fraction"] = sample_fraction;
    }
};
 
 
template <typename Distance>
class AutotunedIndex : public NNIndex<Distance>
{
public:
    typedef typename Distance::ElementType ElementType;
    typedef typename Distance::ResultType DistanceType;
 
    AutotunedIndex(const Matrix<ElementType>& inputData, const IndexParams& params = AutotunedIndexParams(), Distance d = Distance()) :
        dataset_(inputData), distance_(d)
    {
        target_precision_ = get_param(params, "target_precision",0.8f);
        build_weight_ =  get_param(params,"build_weight", 0.01f);
        memory_weight_ = get_param(params, "memory_weight", 0.0f);
        sample_fraction_ = get_param(params,"sample_fraction", 0.1f);
        bestIndex_ = NULL;
        speedup_ = 0;
    }
 
    AutotunedIndex(const AutotunedIndex&);
    AutotunedIndex& operator=(const AutotunedIndex&);
 
    virtual ~AutotunedIndex()
    {
        if (bestIndex_ != NULL) {
            delete bestIndex_;
            bestIndex_ = NULL;
        }
    }
 
    virtual void buildIndex() CV_OVERRIDE
    {
        std::ostringstream stream;
        bestParams_ = estimateBuildParams();
        print_params(bestParams_, stream);
        Logger::info("----------------------------------------------------\n");
        Logger::info("Autotuned parameters:\n");
        Logger::info("%s", stream.str().c_str());
        Logger::info("----------------------------------------------------\n");
 
        bestIndex_ = create_index_by_type(dataset_, bestParams_, distance_);
        bestIndex_->buildIndex();
        speedup_ = estimateSearchParams(bestSearchParams_);
        stream.str(std::string());
        print_params(bestSearchParams_, stream);
        Logger::info("----------------------------------------------------\n");
        Logger::info("Search parameters:\n");
        Logger::info("%s", stream.str().c_str());
        Logger::info("----------------------------------------------------\n");
    }
 
    virtual void saveIndex(FILE* stream) CV_OVERRIDE
    {
        save_value(stream, (int)bestIndex_->getType());
        bestIndex_->saveIndex(stream);
        save_value(stream, get_param<int>(bestSearchParams_, "checks"));
    }
 
    virtual void loadIndex(FILE* stream) CV_OVERRIDE
    {
        int index_type;
 
        load_value(stream, index_type);
        IndexParams params;
        params["algorithm"] = (flann_algorithm_t)index_type;
        bestIndex_ = create_index_by_type<Distance>(dataset_, params, distance_);
        bestIndex_->loadIndex(stream);
        int checks;
        load_value(stream, checks);
        bestSearchParams_["checks"] = checks;
    }
 
    virtual void findNeighbors(ResultSet<DistanceType>& result, const ElementType* vec, const SearchParams& searchParams) CV_OVERRIDE
    {
        int checks = get_param<int>(searchParams,"checks",FLANN_CHECKS_AUTOTUNED);
        if (checks == FLANN_CHECKS_AUTOTUNED) {
            bestIndex_->findNeighbors(result, vec, bestSearchParams_);
        }
        else {
            bestIndex_->findNeighbors(result, vec, searchParams);
        }
    }
 
 
    IndexParams getParameters() const CV_OVERRIDE
    {
        return bestIndex_->getParameters();
    }
 
    SearchParams getSearchParameters() const
    {
        return bestSearchParams_;
    }
 
    float getSpeedup() const
    {
        return speedup_;
    }
 
 
    virtual size_t size() const CV_OVERRIDE
    {
        return bestIndex_->size();
    }
 
    virtual size_t veclen() const CV_OVERRIDE
    {
        return bestIndex_->veclen();
    }
 
    virtual int usedMemory() const CV_OVERRIDE
    {
        return bestIndex_->usedMemory();
    }
 
    virtual flann_algorithm_t getType() const CV_OVERRIDE
    {
        return FLANN_INDEX_AUTOTUNED;
    }
 
private:
 
    struct CostData
    {
        float searchTimeCost;
        float buildTimeCost;
        float memoryCost;
        float totalCost;
        IndexParams params;
    };
 
    void evaluate_kmeans(CostData& cost)
    {
        StartStopTimer t;
        int checks;
        const int nn = 1;
 
        Logger::info("KMeansTree using params: max_iterations=%d, branching=%d\n",
                     get_param<int>(cost.params,"iterations"),
                     get_param<int>(cost.params,"branching"));
        KMeansIndex<Distance> kmeans(sampledDataset_, cost.params, distance_);
        // measure index build time
        t.start();
        kmeans.buildIndex();
        t.stop();
        float buildTime = (float)t.value;
 
        // measure search time
        float searchTime = test_index_precision(kmeans, sampledDataset_, testDataset_, gt_matches_, target_precision_, checks, distance_, nn);
 
        float datasetMemory = float(sampledDataset_.rows * sampledDataset_.cols * sizeof(float));
        cost.memoryCost = (kmeans.usedMemory() + datasetMemory) / datasetMemory;
        cost.searchTimeCost = searchTime;
        cost.buildTimeCost = buildTime;
        Logger::info("KMeansTree buildTime=%g, searchTime=%g, build_weight=%g\n", buildTime, searchTime, build_weight_);
    }
 
 
    void evaluate_kdtree(CostData& cost)
    {
        StartStopTimer t;
        int checks;
        const int nn = 1;
 
        Logger::info("KDTree using params: trees=%d\n", get_param<int>(cost.params,"trees"));
        KDTreeIndex<Distance> kdtree(sampledDataset_, cost.params, distance_);
 
        t.start();
        kdtree.buildIndex();
        t.stop();
        float buildTime = (float)t.value;
 
        //measure search time
        float searchTime = test_index_precision(kdtree, sampledDataset_, testDataset_, gt_matches_, target_precision_, checks, distance_, nn);
 
        float datasetMemory = float(sampledDataset_.rows * sampledDataset_.cols * sizeof(float));
        cost.memoryCost = (kdtree.usedMemory() + datasetMemory) / datasetMemory;
        cost.searchTimeCost = searchTime;
        cost.buildTimeCost = buildTime;
        Logger::info("KDTree buildTime=%g, searchTime=%g\n", buildTime, searchTime);
    }
 
 
    //    struct KMeansSimpleDownhillFunctor {
    //
    //        Autotune& autotuner;
    //        KMeansSimpleDownhillFunctor(Autotune& autotuner_) : autotuner(autotuner_) {}
    //
    //        float operator()(int* params) {
    //
    //            float maxFloat = numeric_limits<float>::max();
    //
    //            if (params[0]<2) return maxFloat;
    //            if (params[1]<0) return maxFloat;
    //
    //            CostData c;
    //            c.params["algorithm"] = KMEANS;
    //            c.params["centers-init"] = CENTERS_RANDOM;
    //            c.params["branching"] = params[0];
    //            c.params["max-iterations"] = params[1];
    //
    //            autotuner.evaluate_kmeans(c);
    //
    //            return c.timeCost;
    //
    //        }
    //    };
    //
    //    struct KDTreeSimpleDownhillFunctor {
    //
    //        Autotune& autotuner;
    //        KDTreeSimpleDownhillFunctor(Autotune& autotuner_) : autotuner(autotuner_) {}
    //
    //        float operator()(int* params) {
    //            float maxFloat = numeric_limits<float>::max();
    //
    //            if (params[0]<1) return maxFloat;
    //
    //            CostData c;
    //            c.params["algorithm"] = KDTREE;
    //            c.params["trees"] = params[0];
    //
    //            autotuner.evaluate_kdtree(c);
    //
    //            return c.timeCost;
    //
    //        }
    //    };
 
 
 
    void optimizeKMeans(std::vector<CostData>& costs)
    {
        Logger::info("KMEANS, Step 1: Exploring parameter space\n");
 
        // explore kmeans parameters space using combinations of the parameters below
        int maxIterations[] = { 1, 5, 10, 15 };
        int branchingFactors[] = { 16, 32, 64, 128, 256 };
 
        int kmeansParamSpaceSize = FLANN_ARRAY_LEN(maxIterations) * FLANN_ARRAY_LEN(branchingFactors);
        costs.reserve(costs.size() + kmeansParamSpaceSize);
 
        // evaluate kmeans for all parameter combinations
        for (size_t i = 0; i < FLANN_ARRAY_LEN(maxIterations); ++i) {
            for (size_t j = 0; j < FLANN_ARRAY_LEN(branchingFactors); ++j) {
                CostData cost;
                cost.params["algorithm"] = FLANN_INDEX_KMEANS;
                cost.params["centers_init"] = FLANN_CENTERS_RANDOM;
                cost.params["iterations"] = maxIterations[i];
                cost.params["branching"] = branchingFactors[j];
 
                evaluate_kmeans(cost);
                costs.push_back(cost);
            }
        }
 
        //         Logger::info("KMEANS, Step 2: simplex-downhill optimization\n");
        //
        //         const int n = 2;
        //         // choose initial simplex points as the best parameters so far
        //         int kmeansNMPoints[n*(n+1)];
        //         float kmeansVals[n+1];
        //         for (int i=0;i<n+1;++i) {
        //             kmeansNMPoints[i*n] = (int)kmeansCosts[i].params["branching"];
        //             kmeansNMPoints[i*n+1] = (int)kmeansCosts[i].params["max-iterations"];
        //             kmeansVals[i] = kmeansCosts[i].timeCost;
        //         }
        //         KMeansSimpleDownhillFunctor kmeans_cost_func(*this);
        //         // run optimization
        //         optimizeSimplexDownhill(kmeansNMPoints,n,kmeans_cost_func,kmeansVals);
        //         // store results
        //         for (int i=0;i<n+1;++i) {
        //             kmeansCosts[i].params["branching"] = kmeansNMPoints[i*2];
        //             kmeansCosts[i].params["max-iterations"] = kmeansNMPoints[i*2+1];
        //             kmeansCosts[i].timeCost = kmeansVals[i];
        //         }
    }
 
 
    void optimizeKDTree(std::vector<CostData>& costs)
    {
        Logger::info("KD-TREE, Step 1: Exploring parameter space\n");
 
        // explore kd-tree parameters space using the parameters below
        int testTrees[] = { 1, 4, 8, 16, 32 };
 
        // evaluate kdtree for all parameter combinations
        for (size_t i = 0; i < FLANN_ARRAY_LEN(testTrees); ++i) {
            CostData cost;
            cost.params["algorithm"] = FLANN_INDEX_KDTREE;
            cost.params["trees"] = testTrees[i];
 
            evaluate_kdtree(cost);
            costs.push_back(cost);
        }
 
        //         Logger::info("KD-TREE, Step 2: simplex-downhill optimization\n");
        //
        //         const int n = 1;
        //         // choose initial simplex points as the best parameters so far
        //         int kdtreeNMPoints[n*(n+1)];
        //         float kdtreeVals[n+1];
        //         for (int i=0;i<n+1;++i) {
        //             kdtreeNMPoints[i] = (int)kdtreeCosts[i].params["trees"];
        //             kdtreeVals[i] = kdtreeCosts[i].timeCost;
        //         }
        //         KDTreeSimpleDownhillFunctor kdtree_cost_func(*this);
        //         // run optimization
        //         optimizeSimplexDownhill(kdtreeNMPoints,n,kdtree_cost_func,kdtreeVals);
        //         // store results
        //         for (int i=0;i<n+1;++i) {
        //             kdtreeCosts[i].params["trees"] = kdtreeNMPoints[i];
        //             kdtreeCosts[i].timeCost = kdtreeVals[i];
        //         }
    }
 
    IndexParams estimateBuildParams()
    {
        std::vector<CostData> costs;
 
        int sampleSize = int(sample_fraction_ * dataset_.rows);
        int testSampleSize = std::min(sampleSize / 10, 1000);
 
        Logger::info("Entering autotuning, dataset size: %d, sampleSize: %d, testSampleSize: %d, target precision: %g\n", dataset_.rows, sampleSize, testSampleSize, target_precision_);
 
        // For a very small dataset, it makes no sense to build any fancy index, just
        // use linear search
        if (testSampleSize < 10) {
            Logger::info("Choosing linear, dataset too small\n");
            return LinearIndexParams();
        }
 
        // We use a fraction of the original dataset to speedup the autotune algorithm
        sampledDataset_ = random_sample(dataset_, sampleSize);
        // We use a cross-validation approach, first we sample a testset from the dataset
        testDataset_ = random_sample(sampledDataset_, testSampleSize, true);
 
        // We compute the ground truth using linear search
        Logger::info("Computing ground truth... \n");
        gt_matches_ = Matrix<int>(new int[testDataset_.rows], testDataset_.rows, 1);
        StartStopTimer t;
        t.start();
        compute_ground_truth<Distance>(sampledDataset_, testDataset_, gt_matches_, 0, distance_);
        t.stop();
 
        CostData linear_cost;
        linear_cost.searchTimeCost = (float)t.value;
        linear_cost.buildTimeCost = 0;
        linear_cost.memoryCost = 0;
        linear_cost.params["algorithm"] = FLANN_INDEX_LINEAR;
 
        costs.push_back(linear_cost);
 
        // Start parameter autotune process
        Logger::info("Autotuning parameters...\n");
 
        optimizeKMeans(costs);
        optimizeKDTree(costs);
 
        float bestTimeCost = costs[0].searchTimeCost;
        for (size_t i = 0; i < costs.size(); ++i) {
            float timeCost = costs[i].buildTimeCost * build_weight_ + costs[i].searchTimeCost;
            if (timeCost < bestTimeCost) {
                bestTimeCost = timeCost;
            }
        }
 
        float bestCost = costs[0].searchTimeCost / bestTimeCost;
        IndexParams bestParams = costs[0].params;
        if (bestTimeCost > 0) {
            for (size_t i = 0; i < costs.size(); ++i) {
                float crtCost = (costs[i].buildTimeCost * build_weight_ + costs[i].searchTimeCost) / bestTimeCost +
                                memory_weight_ * costs[i].memoryCost;
                if (crtCost < bestCost) {
                    bestCost = crtCost;
                    bestParams = costs[i].params;
                }
            }
        }
 
        delete[] gt_matches_.data;
        delete[] testDataset_.data;
        delete[] sampledDataset_.data;
 
        return bestParams;
    }
 
 
 
    float estimateSearchParams(SearchParams& searchParams)
    {
        const int nn = 1;
        const size_t SAMPLE_COUNT = 1000;
 
        CV_Assert(bestIndex_ != NULL && "Requires a valid index"); // must have a valid index
 
        float speedup = 0;
 
        int samples = (int)std::min(dataset_.rows / 10, SAMPLE_COUNT);
        if (samples > 0) {
            Matrix<ElementType> testDataset = random_sample(dataset_, samples);
 
            Logger::info("Computing ground truth\n");
 
            // we need to compute the ground truth first
            Matrix<int> gt_matches(new int[testDataset.rows], testDataset.rows, 1);
            StartStopTimer t;
            t.start();
            compute_ground_truth<Distance>(dataset_, testDataset, gt_matches, 1, distance_);
            t.stop();
            float linear = (float)t.value;
 
            int checks;
            Logger::info("Estimating number of checks\n");
 
            float searchTime;
            float cb_index;
            if (bestIndex_->getType() == FLANN_INDEX_KMEANS) {
                Logger::info("KMeans algorithm, estimating cluster border factor\n");
                KMeansIndex<Distance>* kmeans = (KMeansIndex<Distance>*)bestIndex_;
                float bestSearchTime = -1;
                float best_cb_index = -1;
                int best_checks = -1;
                for (cb_index = 0; cb_index < 1.1f; cb_index += 0.2f) {
                    kmeans->set_cb_index(cb_index);
                    searchTime = test_index_precision(*kmeans, dataset_, testDataset, gt_matches, target_precision_, checks, distance_, nn, 1);
                    if ((searchTime < bestSearchTime) || (bestSearchTime == -1)) {
                        bestSearchTime = searchTime;
                        best_cb_index = cb_index;
                        best_checks = checks;
                    }
                }
                searchTime = bestSearchTime;
                cb_index = best_cb_index;
                checks = best_checks;
 
                kmeans->set_cb_index(best_cb_index);
                Logger::info("Optimum cb_index: %g\n", cb_index);
                bestParams_["cb_index"] = cb_index;
            }
            else {
                searchTime = test_index_precision(*bestIndex_, dataset_, testDataset, gt_matches, target_precision_, checks, distance_, nn, 1);
            }
 
            Logger::info("Required number of checks: %d \n", checks);
            searchParams["checks"] = checks;
 
            speedup = linear / searchTime;
 
            delete[] gt_matches.data;
            delete[] testDataset.data;
        }
 
        return speedup;
    }
 
private:
    NNIndex<Distance>* bestIndex_;
 
    IndexParams bestParams_;
    SearchParams bestSearchParams_;
 
    Matrix<ElementType> sampledDataset_;
    Matrix<ElementType> testDataset_;
    Matrix<int> gt_matches_;
 
    float speedup_;
 
    const Matrix<ElementType> dataset_;
 
    float target_precision_;
    float build_weight_;
    float memory_weight_;
    float sample_fraction_;
 
    Distance distance_;
 
 
};
}
 
 
#endif /* OPENCV_FLANN_AUTOTUNED_INDEX_H_ */