/*
Creates a new R-near neighbor data structure (PRNearNeighborStructT)
from the parameters <thresholdR> and <successProbability> for the
data set <dataSet>. <nPoints> is the number of points in the data
set and <dimension> is the dimension of the points.
The set <sampleQueries> is a set with query sample points
(R-NN DS's parameters are optimized for query points from the set
<sampleQueries>). <sampleQueries> could be a sample of points from the
actual query set or from the data set. When computing the estimated
number of collisions of a sample query point <q> with the data set
points, if there is a point in the data set with the same pointer
with <q> (that is when <q> is a data set point), then the
corresponding point (<q>) is not considered in the data set (for the
purpose of computing the respective #collisions estimation).
*/
PRNearNeighborStructT initSelfTunedRNearNeighborWithDataSet(RealT thresholdR,
RealT successProbability,
Int32T nPoints,
IntT dimension,
PPointT *dataSet,
IntT nSampleQueries,
PPointT *sampleQueries,
MemVarT memoryUpperBound){
initializeLSHGlobal();
PRNearNeighborStructT nnStruct = NULL;
RNNParametersT optParameters = computeOptimalParameters(thresholdR, successProbability, nPoints, dimension, dataSet, nSampleQueries, sampleQueries, memoryUpperBound);
if (!optParameters.useUfunctions) {
DPRINTF("Used L=%d\n", optParameters.parameterL);
}else{
DPRINTF("Used m = %d\n", optParameters.parameterM);
DPRINTF("Used L = %d\n", optParameters.parameterL);
}
TimeVarT timeInit = 0;
TIMEV_START(timeInit);
// Init the R-NN data structure.
if (optParameters.typeHT != HT_HYBRID_CHAINS){
nnStruct = initLSH(optParameters, nPoints);
}else{
printRNNParameters(DEBUG_OUTPUT, optParameters);
nnStruct = initLSH_WithDataSet(optParameters, nPoints, dataSet);
}
TIMEV_END(timeInit);
DPRINTF("Time for initializing: %0.6lf\n", timeInit);
DPRINTF("Allocated memory: %lld\n", totalAllocatedMemory);
TimeVarT timeAdding = 0;
if (optParameters.typeHT != HT_HYBRID_CHAINS){
// Add the points to the LSH buckets.
TIMEV_START(timeAdding);
for(IntT i = 0; i < nPoints; i++){
addNewPointToPRNearNeighborStruct(nnStruct, dataSet[i]);
}
TIMEV_END(timeAdding);
printf("Time for adding points: %0.6lf\n", timeAdding);
DPRINTF("Allocated memory: %lld\n", totalAllocatedMemory);
}
DPRINTF("Time for creating buckets: %0.6lf\n", timeBucketCreation);
DPRINTF("Time for putting buckets into UH: %0.6lf\n", timeBucketIntoUH);
DPRINTF("Time for computing GLSH: %0.6lf\n", timeComputeULSH);
DPRINTF("NGBuckets: %d\n", nGBuckets);
return nnStruct;
}
// Determines the run-time coefficients of the different parts of the //确定查询算法不同部分的运行时间
// query algorithm. Values that are computed and returned are
// <lshPrecomp>, <uhashOver>, <distComp>. <lshPrecomp> is the time for
// pre-computing one function from the LSH family. <uhashOver> is the
// time for getting a bucket from a hash table (of buckets).<distComp>
// is the time to compute one distance between two points. These times
// are computed by constructing a R-NN DS on a sample data set and
// running a sample query set on it.
void determineRTCoefficients(RealT thresholdR,
RealT successProbability,
BooleanT useUfunctions,
IntT typeHT, //建立hash表的类型
IntT dimension,
Int32T nPoints,
PPointT *realData,
RealT &lshPrecomp,
RealT &uhashOver,
RealT &distComp){
// use a subset of the original data set. 使用原始数据集的一个子集
// there is not much theory behind the formula below. //减小运算规模
IntT n = nPoints / 50; //最多生成n各点，缩小50倍
if (n < 100) { //如果生成的点的个数小于100，则使桶的数量与数据集点的数量一样多
n = nPoints;
}
if (n > 10000) {
n = 10000;
}
// Initialize the data set to use.
PPointT *dataSet;
FAILIF(NULL == (dataSet = (PPointT*)MALLOC(n * sizeof(PPointT))));
for(IntT i = 0; i < n; i++){ //从真实数据集中随机取n个点 （最多10000个）
dataSet[i] = realData[genRandomInt(0, nPoints - 1)];
}
IntT hashTableSize = n; //哈希表大小也初始化为n，是指hashTableSize放的点的个数，还是放的桶的个数？
RNNParametersT algParameters;
algParameters.parameterR = thresholdR; //半径
algParameters.successProbability = successProbability;
algParameters.dimension = dimension;
#ifdef USE_L1_DISTANCE
algParameters.parameterR2 = thresholdR; //使用L1距离，R2=R
#else
algParameters.parameterR2 = SQR(thresholdR); //使用L2 R2=R^2
#endif
algParameters.useUfunctions = useUfunctions;
algParameters.parameterK = 16; //k 设定为16，只是测试，估算运算时间，可能是先随机设置一个时间，之后再在代码中改成16，因为16是bestK.
algParameters.parameterW = PARAMETER_W_DEFAULT; //W=4，manuel中说经过多次测试，4是最好的值
algParameters.parameterT = n; //点的个数
algParameters.typeHT = typeHT; //桶的类型HT_HYBRID_CHAINS，在line405里面定义的。
if (algParameters.useUfunctions){
algParameters.parameterM = computeMForULSH(algParameters.parameterK, algParameters.successProbability); //经过改进的L和M
algParameters.parameterL = algParameters.parameterM * (algParameters.parameterM - 1) / 2;
}else{
algParameters.parameterM = computeLfromKP(algParameters.parameterK, algParameters.successProbability); //论文里面的M=L
algParameters.parameterL = algParameters.parameterM;
}
// FAILIF(NULL == (dataSet = (PPointT*)MALLOC(n * sizeof(PPointT))));
// for(IntT i = 0; i < n; i++){
// FAILIF(NULL == (dataSet[i] = (PPointT)MALLOC(sizeof(PointT))));
// FAILIF(NULL == (dataSet[i]->coordinates = (RealT*)MALLOC(dimension * sizeof(RealT))));
// dataSet[i]->index = i;
// sqrLength = 0;
// for(IntT d = 0; d < dimension; d++){
// if (i == 0) {
// dataSet[i]->coordinates[d] = genUniformRandom(-100, 100);
// }else{
// dataSet[i]->coordinates[d] = dataSet[0]->coordinates[d];
// }
// sqrLength += SQR(dataSet[i]->coordinates[d]);
// }
// dataSet[i]->sqrLength = sqrLength;
// }
// switch on timing
BooleanT tempTimingOn = timingOn; //初始化为True
timingOn = TRUE;
// initialize result arrays
PPointT *result = NULL; //结果集以及其初始化
IntT resultSize = 0;
IntT nNNs;
IntT nSucReps;
do{
// create the test structure
PRNearNeighborStructT nnStruct;
switch(algParameters.typeHT){
case HT_LINKED_LIST:
nnStruct = initLSH(algParameters, n);
// add points to the test structure
for(IntT i = 0; i < n; i++){
addNewPointToPRNearNeighborStruct(nnStruct, realData[i]);
}
break;
case HT_HYBRID_CHAINS:
nnStruct = initLSH_WithDataSet(algParameters, n, dataSet); //初始化数据结构，参数集，点的个数，数据集，对点进行映射转换，桶进行映射转换，点存入桶中
break;
default:
ASSERT(FALSE);
}
// query point
PPointT queryPoint;
// FAILIF(NULL == (queryPoint = (PPointT)MALLOC(sizeof(PointT))));
// FAILIF(NULL == (queryPoint->coordinates = (RealT*)MALLOC(dimension * sizeof(RealT))));
// RealT sqrLength = 0;
// for(IntT i = 0; i < dimension; i++){
// queryPoint->coordinates[i] = dataSet[0]->coordinates[i];
// //queryPoint->coordinates[i] = 0.1;
// sqrLength += SQR(queryPoint->coordinates[i]);
// }
//queryPoint->coordinates[0] = dataPoint->coordinates[0] + 0.0001;
//queryPoint->sqrLength = sqrLength;
// reset the R parameter so that there are no NN neighbors.
setResultReporting(nnStruct, FALSE);
//DPRINTF1("X\n");
lshPrecomp = 0;
uhashOver = 0;
distComp = 0;
IntT nReps = 20;
nSucReps = 0;
for(IntT rep = 0; rep < nReps; rep++){
queryPoint = realData[genRandomInt(0, nPoints - 1)]; //查询点为数据集中随机抽取出来的一个点
timeComputeULSH = 0;
timeGetBucket = 0;
timeCycleBucket = 0;
nOfDistComps = 0; //点与点比较的次数
//返回查找到的近邻点数，并将查询到的近邻点存入result中。
nNNs = getNearNeighborsFromPRNearNeighborStruct(nnStruct, queryPoint, result, resultSize);
//DPRINTF("Time to compute LSH: %0.6lf\n", timeComputeULSH);
//DPRINTF("Time to get bucket: %0.6lf\n", timeGetBucket);
//DPRINTF("Time to cycle through buckets: %0.9lf\n", timeCycleBucket);
//DPRINTF("N of dist comp: %d\n", nOfDistComps);
ASSERT(nNNs == 0); //若一个点都没有找到，将发生中断。
if (nOfDistComps >= MIN(n / 10, 100)){ //与足够的点比较过，才将时间计入
nSucReps++;
lshPrecomp += timeComputeULSH / algParameters.parameterK / algParameters.parameterM; //一个点对一个哈希函数的处理时间。共有k*L个哈希函数
uhashOver += timeGetBucket / algParameters.parameterL; //找到一个链表中桶的时间
distComp += timeCycleBucket / nOfDistComps; //遍历链表中桶，并与桶里面的点比较的时间
}
}
if (nSucReps >= 5){
lshPrecomp /= nSucReps;
uhashOver /= nSucReps;
distComp /= nSucReps;
DPRINTF1("RT coeffs computed.\n");
}else{
algParameters.parameterR *= 2; // double the radius and repeat //比较的点数不够，将半径扩大，重复比较
DPRINTF1("Could not determine the RT coeffs. Repeating.\n");
}
freePRNearNeighborStruct(nnStruct);
}while(nSucReps < 5); //做一个有效值的判断，要获得5次有效值
FREE(dataSet);
FREE(result);
timingOn = tempTimingOn;
}
// Determines the run-time coefficients of the different parts of the
// query algorithm. Values that are computed and returned are
// <lshPrecomp>, <uhashOver>, <distComp>. <lshPrecomp> is the time for
// pre-computing one function from the LSH family. <uhashOver> is the
// time for getting a bucket from a hash table (of buckets).<distComp>
// is the time to compute one distance between two points. These times
// are computed by constructing a R-NN DS on a sample data set and
// running a sample query set on it.
void determineRTCoefficients(RealT thresholdR,
RealT successProbability,
BooleanT useUfunctions,
IntT typeHT,
IntT dimension,
Int32T nPoints,
PPointT *realData,
RealT &lshPrecomp,
RealT &uhashOver,
RealT &distComp){
// use a subset of the original data set.
// there is not much theory behind the formula below.
IntT n = nPoints / 50;
if (n < 100) {
n = nPoints;
}
if (n > 10000) {
n = 10000;
}
// Initialize the data set to use.
PPointT *dataSet;
FAILIF(NULL == (dataSet = (PPointT*)MALLOC(n * sizeof(PPointT))));
for(IntT i = 0; i < n; i++){
dataSet[i] = realData[genRandomInt(0, nPoints - 1)];
}
IntT hashTableSize = n;
RNNParametersT algParameters;
algParameters.parameterR = thresholdR;
algParameters.successProbability = successProbability;
algParameters.dimension = dimension;
#ifdef USE_L1_DISTANCE
algParameters.parameterR2 = thresholdR;
#else
algParameters.parameterR2 = SQR(thresholdR);
#endif
algParameters.useUfunctions = useUfunctions;
algParameters.parameterK = 16;
algParameters.parameterW = PARAMETER_W_DEFAULT;
algParameters.parameterT = n;
algParameters.typeHT = typeHT;
if (algParameters.useUfunctions){
algParameters.parameterM = computeMForULSH(algParameters.parameterK, algParameters.successProbability);
algParameters.parameterL = algParameters.parameterM * (algParameters.parameterM - 1) / 2;
}else{
algParameters.parameterM = computeLfromKP(algParameters.parameterK, algParameters.successProbability);
algParameters.parameterL = algParameters.parameterM;
}
// FAILIF(NULL == (dataSet = (PPointT*)MALLOC(n * sizeof(PPointT))));
// for(IntT i = 0; i < n; i++){
// FAILIF(NULL == (dataSet[i] = (PPointT)MALLOC(sizeof(PointT))));
// FAILIF(NULL == (dataSet[i]->coordinates = (RealT*)MALLOC(dimension * sizeof(RealT))));
// dataSet[i]->index = i;
// sqrLength = 0;
// for(IntT d = 0; d < dimension; d++){
// if (i == 0) {
// dataSet[i]->coordinates[d] = genUniformRandom(-100, 100);
// }else{
// dataSet[i]->coordinates[d] = dataSet[0]->coordinates[d];
// }
// sqrLength += SQR(dataSet[i]->coordinates[d]);
// }
// dataSet[i]->sqrLength = sqrLength;
// }
// switch on timing
BooleanT tempTimingOn = timingOn;
timingOn = TRUE;
// initialize result arrays
PPointT *result = NULL;
IntT resultSize = 0;
IntT nNNs;
IntT nSucReps;
do{
// create the test structure
PRNearNeighborStructT nnStruct;
switch(algParameters.typeHT){
case HT_LINKED_LIST:
nnStruct = initLSH(algParameters, n);
// add points to the test structure
for(IntT i = 0; i < n; i++){
addNewPointToPRNearNeighborStruct(nnStruct, realData[i]);
}
break;
case HT_HYBRID_CHAINS:
nnStruct = initLSH_WithDataSet(algParameters, n, dataSet);
break;
default:
ASSERT(FALSE);
}
// query point
PPointT queryPoint;
// FAILIF(NULL == (queryPoint = (PPointT)MALLOC(sizeof(PointT))));
// FAILIF(NULL == (queryPoint->coordinates = (RealT*)MALLOC(dimension * sizeof(RealT))));
// RealT sqrLength = 0;
// for(IntT i = 0; i < dimension; i++){
// queryPoint->coordinates[i] = dataSet[0]->coordinates[i];
// //queryPoint->coordinates[i] = 0.1;
// sqrLength += SQR(queryPoint->coordinates[i]);
// }
//queryPoint->coordinates[0] = dataPoint->coordinates[0] + 0.0001;
//queryPoint->sqrLength = sqrLength;
// reset the R parameter so that there are no NN neighbors.
setResultReporting(nnStruct, FALSE);
//DPRINTF1("X\n");
lshPrecomp = 0;
uhashOver = 0;
distComp = 0;
IntT nReps = 20;
nSucReps = 0;
for(IntT rep = 0; rep < nReps; rep++){
queryPoint = realData[genRandomInt(0, nPoints - 1)];
timeComputeULSH = 0;
timeGetBucket = 0;
timeCycleBucket = 0;
nOfDistComps = 0;
nNNs = getNearNeighborsFromPRNearNeighborStruct(nnStruct, queryPoint, result, resultSize);
//DPRINTF("Time to compute LSH: %0.6lf\n", timeComputeULSH);
//DPRINTF("Time to get bucket: %0.6lf\n", timeGetBucket);
//DPRINTF("Time to cycle through buckets: %0.9lf\n", timeCycleBucket);
//DPRINTF("N of dist comp: %d\n", nOfDistComps);
ASSERT(nNNs == 0);
if (nOfDistComps >= MIN(n / 10, 100)){
nSucReps++;
lshPrecomp += timeComputeULSH / algParameters.parameterK / algParameters.parameterM;
uhashOver += timeGetBucket / algParameters.parameterL;
distComp += timeCycleBucket / nOfDistComps;
}
}
if (nSucReps >= 5){
lshPrecomp /= nSucReps;
uhashOver /= nSucReps;
distComp /= nSucReps;
DPRINTF1("RT coeffs computed.\n");
}else{
algParameters.parameterR *= 2; // double the radius and repeat
DPRINTF1("Could not determine the RT coeffs. Repeating.\n");
}
freePRNearNeighborStruct(nnStruct);
}while(nSucReps < 5);
FREE(dataSet);
FREE(result);
timingOn = tempTimingOn;
}