opencv/3rdparty/flann/algorithms/autotuned_index.h

572 lines
19 KiB
C++

/***********************************************************************
* 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 AUTOTUNEDINDEX_H_
#define AUTOTUNEDINDEX_H_
#include "constants.h"
#include "nn_index.h"
#include "ground_truth.h"
#include "index_testing.h"
namespace cvflann
{
class AutotunedIndex : public NNIndex
{
NNIndex* bestIndex;
IndexParams* bestParams;
SearchParams bestSearchParams;
Matrix<float>* sampledDataset;
Matrix<float>* testDataset;
Matrix<int>* gt_matches;
float speedup;
/**
* The dataset used by this index
*/
const Matrix<float> dataset;
/**
* Index parameters
*/
const AutotunedIndexParams& params;
/**
* Number of features in the dataset.
*/
int size_;
/**
* Length of each feature.
*/
int veclen_;
public:
AutotunedIndex(const Matrix<float>& inputData, const AutotunedIndexParams& params_ = AutotunedIndexParams() ) :
dataset(inputData), params(params_)
{
size_ = dataset.rows;
veclen_ = dataset.cols;
bestIndex = NULL;
}
virtual ~AutotunedIndex()
{
delete bestIndex;
delete bestParams;
};
/**
Method responsible with building the index.
*/
virtual void buildIndex()
{
bestParams = estimateBuildParams();
bestIndex = bestParams->createIndex(dataset);
bestIndex->buildIndex();
speedup = estimateSearchParams(bestSearchParams);
}
/**
Saves the index to a stream
*/
virtual void saveIndex(FILE* stream)
{
bestIndex->saveIndex(stream);
}
/**
Loads the index from a stream
*/
virtual void loadIndex(FILE* stream)
{
bestIndex->loadIndex(stream);
}
/**
Method that searches for nearest-neighbors
*/
virtual void findNeighbors(ResultSet& result, const float* vec, const SearchParams& /*searchParams*/)
{
bestIndex->findNeighbors(result, vec, bestSearchParams);
}
/**
Number of features in this index.
*/
virtual int size() const
{
return bestIndex->size();
}
/**
The length of each vector in this index.
*/
virtual int veclen() const
{
return bestIndex->veclen();
}
/**
The amount of memory (in bytes) this index uses.
*/
virtual int usedMemory() const
{
return bestIndex->usedMemory();
}
/**
* Algorithm name
*/
virtual flann_algorithm_t getType() const
{
return bestIndex->getType();
}
/**
Estimates the search parameters required in order to get a certain precision.
If testset is not given it uses cross-validation.
*/
// virtual Params estimateSearchParams(float precision, Dataset<float>* testset = NULL)
// {
// Params params;
//
// return params;
// }
private:
struct CostData {
float searchTimeCost;
float buildTimeCost;
float timeCost;
float memoryCost;
float totalCost;
};
typedef pair<CostData, KDTreeIndexParams> KDTreeCostData;
typedef pair<CostData, KMeansIndexParams> KMeansCostData;
void evaluate_kmeans(CostData& cost, const KMeansIndexParams& kmeans_params)
{
StartStopTimer t;
int checks;
const int nn = 1;
logger.info("KMeansTree using params: max_iterations=%d, branching=%d\n", kmeans_params.iterations, kmeans_params.branching);
KMeansIndex kmeans(*sampledDataset, kmeans_params);
// 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, params.target_precision, checks, nn);;
float datasetMemory = (float)(sampledDataset->rows*sampledDataset->cols*sizeof(float));
cost.memoryCost = (kmeans.usedMemory()+datasetMemory)/datasetMemory;
cost.searchTimeCost = searchTime;
cost.buildTimeCost = buildTime;
cost.timeCost = (buildTime*params.build_weight+searchTime);
logger.info("KMeansTree buildTime=%g, searchTime=%g, timeCost=%g, buildTimeFactor=%g\n",buildTime, searchTime, cost.timeCost, params.build_weight);
}
void evaluate_kdtree(CostData& cost, const KDTreeIndexParams& kdtree_params)
{
StartStopTimer t;
int checks;
const int nn = 1;
logger.info("KDTree using params: trees=%d\n",kdtree_params.trees);
KDTreeIndex kdtree(*sampledDataset, kdtree_params);
t.start();
kdtree.buildIndex();
t.stop();
float buildTime = (float)t.value;
//measure search time
float searchTime = test_index_precision(kdtree, *sampledDataset, *testDataset, *gt_matches, params.target_precision, checks, nn);
float datasetMemory = (float)(sampledDataset->rows*sampledDataset->cols*sizeof(float));
cost.memoryCost = (kdtree.usedMemory()+datasetMemory)/datasetMemory;
cost.searchTimeCost = searchTime;
cost.buildTimeCost = buildTime;
cost.timeCost = (buildTime*params.build_weight+searchTime);
logger.info("KDTree buildTime=%g, searchTime=%g, timeCost=%g\n",buildTime, searchTime, cost.timeCost);
}
// 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;
//
// }
// };
KMeansCostData optimizeKMeans()
{
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 = ARRAY_LEN(maxIterations)*ARRAY_LEN(branchingFactors);
vector<KMeansCostData> kmeansCosts(kmeansParamSpaceSize);
// CostData* kmeansCosts = new CostData[kmeansParamSpaceSize];
// evaluate kmeans for all parameter combinations
int cnt = 0;
for (size_t i=0; i<ARRAY_LEN(maxIterations); ++i) {
for (size_t j=0; j<ARRAY_LEN(branchingFactors); ++j) {
kmeansCosts[cnt].second.centers_init = CENTERS_RANDOM;
kmeansCosts[cnt].second.branching = branchingFactors[j];
kmeansCosts[cnt].second.iterations = maxIterations[j];
evaluate_kmeans(kmeansCosts[cnt].first, kmeansCosts[cnt].second);
int k = cnt;
// order by time cost
while (k>0 && kmeansCosts[k].first.timeCost < kmeansCosts[k-1].first.timeCost) {
swap(kmeansCosts[k],kmeansCosts[k-1]);
--k;
}
++cnt;
}
}
// 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];
// }
float optTimeCost = kmeansCosts[0].first.timeCost;
// recompute total costs factoring in the memory costs
for (int i=0;i<kmeansParamSpaceSize;++i) {
kmeansCosts[i].first.totalCost = (kmeansCosts[i].first.timeCost/optTimeCost + params.memory_weight * kmeansCosts[i].first.memoryCost);
int k = i;
while (k>0 && kmeansCosts[k].first.totalCost < kmeansCosts[k-1].first.totalCost) {
swap(kmeansCosts[k],kmeansCosts[k-1]);
k--;
}
}
// display the costs obtained
for (int i=0;i<kmeansParamSpaceSize;++i) {
logger.info("KMeans, branching=%d, iterations=%d, time_cost=%g[%g] (build=%g, search=%g), memory_cost=%g, cost=%g\n",
kmeansCosts[i].second.branching, kmeansCosts[i].second.iterations,
kmeansCosts[i].first.timeCost,kmeansCosts[i].first.timeCost/optTimeCost,
kmeansCosts[i].first.buildTimeCost, kmeansCosts[i].first.searchTimeCost,
kmeansCosts[i].first.memoryCost,kmeansCosts[i].first.totalCost);
}
return kmeansCosts[0];
}
KDTreeCostData optimizeKDTree()
{
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 };
size_t kdtreeParamSpaceSize = ARRAY_LEN(testTrees);
vector<KDTreeCostData> kdtreeCosts(kdtreeParamSpaceSize);
// evaluate kdtree for all parameter combinations
int cnt = 0;
for (size_t i=0; i<ARRAY_LEN(testTrees); ++i) {
kdtreeCosts[cnt].second.trees = testTrees[i];
evaluate_kdtree(kdtreeCosts[cnt].first, kdtreeCosts[cnt].second);
int k = cnt;
// order by time cost
while (k>0 && kdtreeCosts[k].first.timeCost < kdtreeCosts[k-1].first.timeCost) {
swap(kdtreeCosts[k],kdtreeCosts[k-1]);
--k;
}
++cnt;
}
// 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];
// }
float optTimeCost = kdtreeCosts[0].first.timeCost;
// recompute costs for kd-tree factoring in memory cost
for (size_t i=0;i<kdtreeParamSpaceSize;++i) {
kdtreeCosts[i].first.totalCost = (kdtreeCosts[i].first.timeCost/optTimeCost + params.memory_weight * kdtreeCosts[i].first.memoryCost);
int k = i;
while (k>0 && kdtreeCosts[k].first.totalCost < kdtreeCosts[k-1].first.totalCost) {
swap(kdtreeCosts[k],kdtreeCosts[k-1]);
k--;
}
}
// display costs obtained
for (size_t i=0;i<kdtreeParamSpaceSize;++i) {
logger.info("kd-tree, trees=%d, time_cost=%g[%g] (build=%g, search=%g), memory_cost=%g, cost=%g\n",
kdtreeCosts[i].second.trees,kdtreeCosts[i].first.timeCost,kdtreeCosts[i].first.timeCost/optTimeCost,
kdtreeCosts[i].first.buildTimeCost, kdtreeCosts[i].first.searchTimeCost,
kdtreeCosts[i].first.memoryCost,kdtreeCosts[i].first.totalCost);
}
return kdtreeCosts[0];
}
/**
Chooses the best nearest-neighbor algorithm and estimates the optimal
parameters to use when building the index (for a given precision).
Returns a dictionary with the optimal parameters.
*/
IndexParams* estimateBuildParams()
{
int sampleSize = int(params.sample_fraction*dataset.rows);
int testSampleSize = min(sampleSize/10, 1000);
logger.info("Entering autotuning, dataset size: %d, sampleSize: %d, testSampleSize: %d\n",dataset.rows, sampleSize, testSampleSize);
// 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 new LinearIndexParams();
}
// We use a fraction of the original dataset to speedup the autotune algorithm
sampledDataset = dataset.sample(sampleSize);
// We use a cross-validation approach, first we sample a testset from the dataset
testDataset = sampledDataset->sample(testSampleSize,true);
// We compute the ground truth using linear search
logger.info("Computing ground truth... \n");
gt_matches = new Matrix<int>(testDataset->rows, 1);
StartStopTimer t;
t.start();
compute_ground_truth(*sampledDataset, *testDataset, *gt_matches, 0);
t.stop();
float bestCost = (float)t.value;
IndexParams* bestParams = new LinearIndexParams();
// Start parameter autotune process
logger.info("Autotuning parameters...\n");
KMeansCostData kmeansCost = optimizeKMeans();
if (kmeansCost.first.totalCost<bestCost) {
bestParams = new KMeansIndexParams(kmeansCost.second);
bestCost = kmeansCost.first.totalCost;
}
KDTreeCostData kdtreeCost = optimizeKDTree();
if (kdtreeCost.first.totalCost<bestCost) {
bestParams = new KDTreeIndexParams(kdtreeCost.second);
bestCost = kdtreeCost.first.totalCost;
}
// free the memory used by the datasets we sampled
delete sampledDataset;
delete testDataset;
delete gt_matches;
return bestParams;
}
/**
Estimates the search time parameters needed to get the desired precision.
Precondition: the index is built
Postcondition: the searchParams will have the optimum params set, also the speedup obtained over linear search.
*/
float estimateSearchParams(SearchParams& searchParams)
{
const int nn = 1;
const long SAMPLE_COUNT = 1000;
assert(bestIndex!=NULL); // must have a valid index
float speedup = 0;
int samples = min(dataset.rows/10, SAMPLE_COUNT);
if (samples>0) {
Matrix<float>* testDataset = dataset.sample(samples);
logger.info("Computing ground truth\n");
// we need to compute the ground truth first
Matrix<int> gt_matches(testDataset->rows,1);
StartStopTimer t;
t.start();
compute_ground_truth(dataset, *testDataset, gt_matches,1);
t.stop();
float linear = (float)t.value;
int checks;
logger.info("Estimating number of checks\n");
float searchTime;
float cb_index;
if (bestIndex->getType() == KMEANS) {
logger.info("KMeans algorithm, estimating cluster border factor\n");
KMeansIndex* kmeans = (KMeansIndex*)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, params.target_precision, checks, 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);
((KMeansIndexParams*)bestParams)->cb_index = cb_index;
}
else {
searchTime = test_index_precision(*bestIndex, dataset, *testDataset, gt_matches, params.target_precision, checks, nn, 1);
}
logger.info("Required number of checks: %d \n",checks);;
searchParams.checks = checks;
delete testDataset;
speedup = linear/searchTime;
}
return speedup;
}
};
}
#endif /* AUTOTUNEDINDEX_H_ */