//------------------------------------------------------------------------
// Print summary of execution
//------------------------------------------------------------------------
void printSummary(
const KMlocal& theAlg, // the algorithm
const KMdata& dataPts, // the points
KMfilterCenters& ctrs) // the centers
{
cout << "Number of stages: " << theAlg.getTotalStages() << "\n";
cout << "Average distortion: " <<
ctrs.getDist(false)/double(ctrs.getNPts()) << "\n";
// print final center points
cout << "(Final Center Points:\n";
ctrs.print();
cout << ")\n";
// get/print final cluster assignments
KMctrIdxArray closeCtr = new KMctrIdx[dataPts.getNPts()];
double* sqDist = new double[dataPts.getNPts()];
ctrs.getAssignments(closeCtr, sqDist);
*kmOut << "(Cluster assignments:\n"
<< " Point Center Squared Dist\n"
<< " ----- ------ ------------\n";
for (int i = 0; i < dataPts.getNPts(); i++) {
*kmOut << " " << setw(5) << i
<< " " << setw(5) << closeCtr[i]
<< " " << setw(10) << sqDist[i]
<< "\n";
}
*kmOut << ")\n";
delete [] closeCtr;
delete [] sqDist;
}
// standard constructor
KMfilterCenters::KMfilterCenters(int k, KMdata& p, double df)
: KMcenters(k, p) {
if (p.getKcTree() == NULL) { // kc-tree not yet built?
kmError("Building kc-tree", KMwarn);
p.buildKcTree(); // build it now
}
sums = kmAllocPts(kCtrs, getDim());
sumSqs = new double[kCtrs];
weights = new int[kCtrs];
dists = new double[kCtrs];
currDist = KM_HUGE;
dampFactor = df;
currDBIndex = KM_HUGE;
currXBIndex = KM_HUGE;
invalidate(); // distortions are initially invalid
}
static std::vector<double> computeKMeans(
KMdata & datapoints,
uint64_t const k,
uint64_t const runs,
bool const debug
)
{
KMterm const term(runs, 0, 0, 0, 0.10, 0.10, 3, 0.50, 10, 0.95);
datapoints.buildKcTree();
KMfilterCenters ctrs(k, datapoints);
KMlocalHybrid kmAlg(ctrs, term);
ctrs = kmAlg.execute();
std::vector<double> centrevector;
for ( uint64_t i = 0; i < k; ++i )
{
centrevector.push_back(ctrs[i][0]);
if ( debug )
std::cerr << "centre[" << i << "]=" << ctrs[i][0] << std::endl;
}
std::sort(centrevector.begin(),centrevector.end());
return centrevector;
}
//------------------------------------------------------------------------
// Print summary of execution
//------------------------------------------------------------------------
void printSummary(
const KMlocal& theAlg, // the algorithm
const KMdata& dataPts, // the points
KMfilterCenters& ctrs) // the centers
{
ofstream out; // output data file stream
out.open("output3.txt",ios::out);
if(!out){ // si l'ouverture a échoué
cerr << "Impossible d'ouvrir le fichier erreur2" << endl;
exit(0);
}
out << "Number of stages: " << theAlg.getTotalStages() << "\n";
out << "Average distortion: " <<
ctrs.getDist(false)/double(ctrs.getNPts()) << "\n";
// print final center points
out << "(Final Center Points:\n";
ctrs.print();
out << ")\n";
// get/print final cluster assignments
KMctrIdxArray closeCtr = new KMctrIdx[dataPts.getNPts()];
double* sqDist = new double[dataPts.getNPts()];
ctrs.getAssignments(closeCtr, sqDist); // obligé d'avoir la distance ?
out << "(Cluster assignments:\n"
<< "Point Center :\n"
<< "----- ------ \n";
for (int i = 0; i < dataPts.getNPts(); i++) {
out << setw(5) << i
<< setw(5) << closeCtr[i]
<< "\n";
}
out << ")\n";
delete [] closeCtr;
delete [] sqDist;
out.close(); // on ferme le fichier
}
/**
* \fn void exportSTIPs(std::string stip, int dim, const KMdata& dataPts)
* \brief STIPs exportation function in the format 1 point = 1 line.
* Each dimension are separated from one space (" ").
*
* \param[in] stip Name of the file containing the STIPs.
* \param[in] dim The STIPs dimension.
* \param[in] dataPts The KMlocal object which will be containing STIPs.
*/
void exportSTIPs(std::string stip, int dim, const KMdata& dataPts){
int nPts = dataPts.getNPts(); // actual number of points
// open fiouverture en écriture avec effacement du fichier ouvert
ofstream sSTIPs(stip.c_str(), ios::out | ios::trunc);
if(!sSTIPs){
std::cerr << "Impossible d'ouvrir le fichier !" << std::endl;
exit(EXIT_FAILURE);
}
for(int i=0; i<nPts ;i++){
for(int d = 0; d<dim ; d++){
sSTIPs << dataPts[i][d] << " ";
}
sSTIPs << std::endl;
}
sSTIPs.close();
}
// standard constructor
KMcenters::KMcenters(int k, KMdata& p)
: kCtrs(k), pts(&p) {
ctrs = kmAllocPts(kCtrs, p.getDim());
}
//------------------------------------------------------------------------
// Print summary of execution
//------------------------------------------------------------------------
void printSummary(const KMlocal& theAlg, // the algorithm
const KMdata& dataPts, // the points
KMfilterCenters& ctrs) // the centers
{
double dbval, xbval;
dbval = ctrs.getDBIndex();
xbval = ctrs.getXBIndex();
FILE* logfp = fopen("kmean.log", "wt");
fprintf(logfp, "Number of stages: %d\n", theAlg.getTotalStages());
fprintf(logfp, "Average distortion: %g\n", ctrs.getDist(false)/double(ctrs.getNPts()));
fprintf(logfp, "DB-index: %g\n", dbval);
fprintf(logfp, "XB-index: %g\n", xbval);
cout << "Number of stages: " << theAlg.getTotalStages() << "\n";
cout << "Average distortion: " <<
ctrs.getDist(false)/double(ctrs.getNPts()) << "\n";
cout << "DB-index = " << dbval << "\n";
cout << "XB-index = " << xbval << "\n";
KMctrIdxArray closeCtr = new KMctrIdx[dataPts.getNPts()];
double* sqDist = new double[dataPts.getNPts()];
ctrs.getAssignments(closeCtr, sqDist);
int* hist = new int[ctrs.getK()];
memset(hist, 0, sizeof(int) * ctrs.getK() );
for(int i=0; i < dataPts.getNPts(); i++ )
{
int k = closeCtr[i];
hist[k] ++;
}
KMpoint kpt;
for(int i=0; i< ctrs.getK(); i++ )
{
fprintf(logfp, "%3d-th, #%5d ", i, hist[i]);
// print final center points
kpt = ctrs[i];
fprintf(logfp, " [");
for (int j = 0; j < ctrs.getDim(); j++)
{
fprintf(logfp, "%7g ", kpt[j]);
}
fprintf(logfp, " ]\n");
}
fclose(logfp);
// write to fea-file
CFeaFileWriter theFeaWriter;
char saveName[256];
sprintf(saveName, "kmain_%s", infname.c_str() );
theFeaWriter.openFile(saveName);
FEA_FILE_HEADER feaHeader;
feaHeader.nVersion = FEA_FILE_VERSION;
feaHeader.nRecords = ctrs.getK();
feaHeader.nFeaDim = ctrs.getDim();
feaHeader.nElemType = ELEM_TYPE_FLOAT;
feaHeader.nElemSize = sizeof(float);
feaHeader.bIndexTab = 0;
feaHeader.bVarLen = 0;
sprintf(feaHeader.szFeaName, "kmean codebook");
theFeaWriter.setFileHeader(feaHeader);
float* pFea = new float[ctrs.getDim()];
for(int i=0; i< ctrs.getK(); i++ )
{
// save final center points
kpt = ctrs[i];
for (int j = 0; j < ctrs.getDim(); j++)
pFea[j] = kpt[j];
theFeaWriter.writeRecordAt(pFea, i);
}
theFeaWriter.flush2Disk();
theFeaWriter.closeFile();
theFeaWriter.releaseMemory();
delete [] pFea;
delete [] closeCtr;
delete [] sqDist;
delete [] hist;
}
/**
* \fn void kmIvanAlgorithm(int ic, int dim, const KMdata& dataPts, int k, KMfilterCenters& ctrs)
* \brief This is an optimized KMeans algorithm. Ivan's algorithm uses
* basic KMeans algorithm (here the Lloyd's one) and the idea was to
* initialize centers intelligently.
*
* \param[in] ic The iteration coefficient will determine the number of iterations in each phases.
* \param[in] dim Points and centers's dimension.
* \param[in] dataPts The data we want to compute the centers.
* \param[in] k The number of centers.
* \param[out] ctrs The centers.
*
* The Ivan's algorithm is divided into 3 phases. The first phase is executed on
* 25 per cent of the data (randomly sampled). To begin, the centers are randomly generated.
* Then ic * 4 iterations of a KMeans algorithm are executed.
* During the second part we cluster 50 per cent of the data using the older centroids.
* This step is computed ic * 2 times.
* Finally, we make ic * 1 iteration on all the data.
*
*/
void kmIvanAlgorithm(int ic, int dim, const KMdata& dataPts, int k, KMfilterCenters& ctrs){
int nPts = dataPts.getNPts();
KMdata subDataPts(dim,nPts); // maxPts = nPts since subDataPts is a sample of dataPts
int* randomVector = NULL;
double** centersBuffer = NULL;
centersBuffer = (double **) malloc(k*sizeof(double*));
for(int c=0 ; c<k ; c++){
centersBuffer[c] = (double*) malloc(dim*sizeof(double));
}
// ic : iteration coefficient
int nrPhases = 3;
for(int i=0 ; i<nrPhases ; i++){
std::cerr<<"phase:"<<i<<std::endl;
int maxIter = (int) pow(2,(nrPhases-1-i));
int sampleSize = floor(nPts/maxIter);
std::cout << "Applying k-means: " << endl;
std::cout << "Clustering " << sampleSize << " vectors (ie. " << 100/maxIter << " percent of the data)";
std::cout << " in " << k << " clusters";
if (i>0){
std::cout << " using older centroids..." << std::endl;
}
else{
std::cout << "..." << endl;
std::cout << "Initializing centroids by sampling..." << std::endl;
}
if (i == nrPhases-1){
// Filling subDataPts
for(int s=0; s<nPts ; s++){
for(int d=0 ; d<dim ; d++){
subDataPts[s][d] = dataPts[s][d];
}
}
subDataPts.setNPts(nPts);
}
else{
// Filling the random vector permiting to sampling "uniformly" (as more as we can) the data
std::cerr<<"ok++"<<std::endl;
randomVector = (int*) malloc(sampleSize * sizeof(int));
srand(time(NULL)); // initialisation of rand
for(int s=0; s<sampleSize ;s++){
int r = (int) rand()%(nPts);
int index = 0;
while(index<s && randomVector[index] != r){
index++;
}
if(s==0 || randomVector[index] != r)
randomVector[s] = r;
else{
s--;
}
}
std::cerr<<"ok--"<<std::endl;
// Filling subDataPts
for(int s=0; s<sampleSize ; s++){
for(int d=0 ; d<dim ; d++){
subDataPts[s][d] = dataPts[randomVector[s]][d];
}
}
subDataPts.setNPts(sampleSize);
}
subDataPts.buildKcTree();
// Allocate centers with subData
KMfilterCenters newCtrs(k, subDataPts);
// Initializing the centers (randomly for the first iteration)
if(i==0){
(newCtrs).genRandom();
}
else{
for(int c = 0; c < k ; c++){
for(int d=0 ; d<dim ; d++){
(newCtrs)[c][d] = centersBuffer[c][d];
}
}
}
for(int iteration = 0 ; iteration < ic*maxIter ; iteration++){ // ic : iteration coefficient
(newCtrs).lloyd1Stage();
}
// Saving the old centers in centersBuffer
for(int c = 0; c < k ; c++){
for(int d=0 ; d<dim ; d++){
centersBuffer[c][d] = (newCtrs)[c][d];
}
}
if(i==nrPhases-1){
for(int c = 0; c < k ; c++){
for(int d=0 ; d<dim ; d++){
ctrs[c][d] = centersBuffer[c][d];
}
}
}
free(randomVector);
randomVector = NULL;
}
for(int c=0 ; c<k ; c++){
free(centersBuffer[c]);
}
free(centersBuffer);
}
