release / genetic-gpu

  #include <iostream>
  
  //stim libraries
  #include <stim/envi/envi.h>
  #include <stim/image/image.h>
  #include <stim/ui/progressbar.h>
  #include <stim/parser/filename.h>
  #include <stim/parser/table.h>
  #include <stim/parser/arguments.h>
  //input arguments
  stim::arglist args;
  #include <fstream>
  #include <thread>
  #include <random>
  #include <vector>
  #include <math.h>
  #include <limits>
  
  #define NOMINMAX
  
  
  
  //GA
  #include "ga_gpu.h"
  #include "enviload.h"
  
  
  //envi input file and associated parameters
  stim::envi E;										//ENVI binary file object
  unsigned int B;										//shortcuts storing the spatial and spectral size of the ENVI image
  //mask and class information used for training
  //std::vector< stim::image<unsigned char> > C;		//2D array used to access each mask C[m][p], where m = mask# and p = pixel#
  std::vector<unsigned int> nP;						//array holds the number of pixels in each mask: nP[m] is the number of pixels in mask m
  size_t nC = 0;										//number of classes
  size_t tP = 0;										//total number of pixels in all masks: tP = nP[0] + nP[1] + ... + nP[nC]
  float* fea;
  
  //ga_gpu class object
  ga_gpu ga;
  bool debug;
  bool binaryClass;
  int binClassOne;
  
  //creating struct to pass to thread functions as it limits number of arguments to 3
  typedef struct {
  	float* S;
  	float* Sb;
  	float* Sw;
  	float* lda;
  }gnome;
  gnome gnom;
  
  
  void gpuComputeEignS( size_t g, size_t fea){	
  	//eigen value computation will return  r = (nC-1) eigen vectors so new projected data will have dimension of r rather than f
  	//  std::thread::id this_id = std::this_thread::get_id();
  	//	std::cout<<"thread id is "<< this_id<<std::endl;  
  	size_t f = fea;
  	//std::thread::id g = std::this_thread::get_id();	
  	float* LeftEigVectors_a = (float*) malloc(f * f * sizeof(float));			
  	float* gSw_a = (float*) malloc(f * f * sizeof(float));						//copy of between class scatter		
  	std::memcpy(gSw_a, &gnom.Sw[g * f * f], f * f *sizeof(float));
  	if(debug){
  		std::cout<<"From Eigen function: Sb and Sw "<<std::endl;
  		displayS(gSw_a, f);									//display Sb				
  		displayS(&gnom.Sb[g * f * f], f);									//display Sw
  		std::cout<<std::endl;
  	}	
  	
  	std::vector<unsigned int> features = ga.getGnome(g);
  	std::vector<unsigned int> featuresunique;
  	int flag = 0;
  	std::sort(features.begin(), features.end()); // 1 1 2 2 3 3 3 4 4 5 5 6 7 
  	std::unique_copy(features.begin(), features.end(), std::back_inserter(featuresunique));
  	if(featuresunique.size()< features.size()){
  		f = featuresunique.size();
  	}
  				
  	size_t r = nC-1;                                         //LDA projected dimension (limited to number of classes - 1  by rank)
  	if(r > f){
  		r = f;
  	}
  
  	int info;
  	float* EigenvaluesI_a = (float*)malloc(f * sizeof(float));
  	float* Eigenvalues_a = (float*)malloc(f * sizeof(float));
  	int *IPIV = (int*) malloc(sizeof(int) * f);
  	//computing inverse of matrix Sw
  	memset(IPIV, 0, f * sizeof(int));
  	LAPACKE_sgetrf(LAPACK_COL_MAJOR, (int)f, (int)f, gSw_a, (int)f, IPIV);						
  	// DGETRI computes the inverse of a matrix using the LU factorization computed by DGETRF.
  	LAPACKE_sgetri(LAPACK_COL_MAJOR, (int)f, gSw_a, (int)f, IPIV);
  	
  	float* gSbSw_a  = (float*)calloc(f * f, sizeof(float));
  	//mtxMul(gSbSw_a, gSw_a, &gnom.Sb[g * f * f * sizeof(float)], f, f, f,f);
  	mtxMul(gSbSw_a, gSw_a, &gnom.Sb[g * f * f], f, f, f,f);
  	if(debug){
  		std::cout<<"From Eigen function: inverse of sw and ratio of sb and sw (Sb/Sw)";
  		displayS(gSw_a, f);			//display inverse of Sw (1/Sw)				
  		displayS(gSbSw_a, f);		//display ratio of Sb and Sw (Sb/Sw)
  	}
     
     //compute left eigenvectors for current gnome from ratio of between class scatter and within class scatter:  Sb/Sw 
  	info = LAPACKE_sgeev(LAPACK_COL_MAJOR, 'V', 'N', (int)f, gSbSw_a, (int)f, Eigenvalues_a, EigenvaluesI_a, LeftEigVectors_a, (int)f, 0, (int)f);
  	//sort eignevalue indices in descending order							
  	size_t* sortedindx = sortIndx(Eigenvalues_a, f);
  	//displayS(LeftEigVectors_a, f);					//display Eignevectors (Note these are -1 * matlab eigenvectors does not change fitness score results but keep in mind while projecting data on it)		
  	//sorting left eigenvectors  (building forward transformation matrix As)
  	for (size_t rowE = 0; rowE < r; rowE++){
  		for (size_t colE = 0; colE < f; colE++){
  			size_t ind1 = g * r * f + rowE * f + colE;			
  			//size_t ind1 =  rowE * f + colE;
  			size_t ind2 = sortedindx[rowE] * f + colE;				//eigenvector as row vector			
  			gnom.lda[ind1] = LeftEigVectors_a[ind2];
  		}
  	}
  		
  	if(debug){
  		std::cout<<"Eigenvalues are"<<std::endl;
  		for(size_t n = 0 ; n < f; n ++){
  			std::cout << Eigenvalues_a[n] << ", " ;
  		}
  		std::cout<< std::endl;
  		std::cout<<"From Eigen function: Eignevector"<<std::endl;
  	  				
  		std::cout<<"LDA basis is "<<std::endl;
  		std::cout << "r is " << r << std::endl;
  		for(size_t l = 0 ; l < r; l++){
  			for(size_t n = 0 ; n < f; n ++){
  				std::cout << gnom.lda[g * l * f + l * f + n] << ", " ;
  			}
  			std::cout<<std::endl;
  		}
  			
  	}
  	//Extract only r eigne vectors as a LDA projection basis
  	float* tempgSb = (float*)calloc(r * f, sizeof(float));
  	//mtxMul(tempgSb, &gnom.lda[g * r * f * sizeof(float)], &gnom.Sb[g * f * f * sizeof(float)], r, f, f,f);
  	//mtxMul(tempgSb, &lda[g * r * f ], gSb, r, f, f,f);
  	mtxMul(tempgSb, &gnom.lda[g * r * f], &gnom.Sb[g * f * f], r, f, f,f);
  	float* nSb = (float*)calloc(r * r, sizeof(float));
  	mtxMultranspose(nSb, tempgSb, &gnom.lda[g * r * f], r, f, r, f);
  	
  	float* tempgSw = (float*)calloc(r * f, sizeof(float));
  	//mtxMul(tempgSw, &gnom.lda[g * r * f * sizeof(float)], &gnom.Sw[g * f * f * sizeof(float)], r, f, f,f);
  	mtxMul(tempgSw, &gnom.lda[g * r * f], &gnom.Sw[g * f * f], r, f, f,f);
  	float* nSw = (float*)calloc(r * r, sizeof(float));
  	mtxMultranspose(nSw, tempgSw, &gnom.lda[g * r * f], r, f, r, f);
  	if(debug){
  		std::cout<<"From Eigen function: projected Sb sn Sw"<<std::endl;
  		displayS(nSb, r);									//display Sb				
  		displayS(nSw, r);									//display Sw
  		std::cout<<std::endl;
  	}
  	
  	cv::Mat newSw = cv::Mat((int)r, (int)r, CV_32FC1, nSw);		//within scatter of gnome g in the population 
  	cv::Mat newSb = cv::Mat((int)r, (int)r, CV_32FC1, nSb);		//within scatter of gnome g in the population 
  			
  	//fisher's ratio from ratio of projected sb and sw
  	float fisherRatio = cv::determinant(newSb) /cv::determinant(newSw);
  	gnom.S[g] = 1/fisherRatio;
  	if (debug) {
  		std::cout<<"Score["<<g<<"]: "<< gnom.S[g]<<std::endl;
  	
  		std::cout << "best gnoem is " << std::endl;
  		for (size_t i = 0; i < f; i++)
  			std::cout << ga.P[ga.f * g + i] << ", ";
  		std::cout << std::endl;
  	}
  //	if(!isfinite(gnom.S[g])){
  //      std::cout<<"-----------------------------------------------"<<std::endl;
  //      std::cout<<"Displaying intermediate values of gnome for which score is non finite"<<std::endl;
  //     	std::cout<<"population "<<std::endl;
  //			for(int i = 0; i < ga.f; i++){
  //				std::cout<<"\t"<<ga.P[g * ga.f + i];
  //			}
  //      std::cout<<std::endl;
  //      std::cout<<"Sb determinant is "<<cv::determinant(newSb)<<"\t Sw determinant is "<<cv::determinant(newSb)<<std::endl;
  //     	std::cout<<"fisher ratio is "<<fisherRatio<<std::endl;
  //			std::cout<<"Score["<<g<<"]: "<< gnom.S[g]<<std::endl;	
  //      std::cout<<"------------------------------------------------"<<std::endl;
  //
  //	}
  		
  	
  	if(IPIV!= NULL) std::free(IPIV);	
  	if(gSw_a!= NULL) std::free(gSw_a);
  	if(gSbSw_a!= NULL) std::free(gSbSw_a);
  	if(Eigenvalues_a!= NULL) std::free(Eigenvalues_a);
  	if(EigenvaluesI_a!= NULL) std::free(EigenvaluesI_a);
  	if(tempgSb!= NULL) std::free(tempgSb);
  	if(tempgSw!= NULL) std::free(tempgSw);
  	
  }	
  
  	
  void fitnessFunction( float* sb, float* sw, float* lda, float* M, float* cM, size_t f, cudaDeviceProp props, size_t gen, std::ofstream& profilefile){  
  		
  		size_t tP = 0;																	//total number of pixels
  		std::for_each(nP.begin(), nP.end(), [&] (size_t n) {
  			tP += n;
  		});		
   		size_t nC = nP.size();													//total number of classes
  		
  		//--------------Compute between class scatter 		
  	//	ga.cpu_computeSbSw(sb, sw,  M,  cM, nC,  tP, nP, gen);		
  		//
  		//if(debug){
  		//	std::cout<<"cpu results of Sb sn Sw"<<std::endl;
  		//	displayS(sb, ga.f);									//display Sb				
  		//	displayS(sw, ga.f);									//display Sw
  		// }
  		
  		//ga.callingGpuComputeSbSw(sb, sw,  nC, tP, nP, props, gen);			
  		ga.gpu_computeSbSw(sb, sw, nC, tP, props, gen, debug, profilefile);
  		//ga.cpu_computeSbSw(sb, sw, M, cM, nC, tP, nP);
  
  		if(debug){
  			std::cout<<"From fitness function: gpu results of Sb sn Sw"<<std::endl;
  			displayS(sb, ga.f);									//display Sb				
  			displayS(sw, ga.f);									//display Sw
  			std::cout<<std::endl;
  		}				
  		
  	// -----------------------   Linear discriminant Analysis   --------------------------------------
  		//timer.start();	
  		//structure is created to pass variable to thread function as it accepts only 3 arguments 
  		gnom.S = ga.S;
  		gnom.Sw = sw;
  		gnom.Sb = sb;
  		gnom.lda = lda;
  
  		//calling function without using threads
  		for (size_t i = 0; i<ga.p; i++){			
  			//calling function for eigencomputation
  			gpuComputeEignS(i, f);
  			//std::cout<<"Score["<<i<<"]: "<< ga.S[i]<<std::endl;
  		}	
  	
  		//std::vector<std::thread> threads;
  		//for (size_t g = 0; g<ga.p; g++){
  		//	//creating thread to do eigen computation 
  		//	threads.push_back(std::thread(gpuComputeEignS, g, f));	
  		//}	
  		//
  		//// loop again to join the threads
  		//for (auto& t : threads)
  		//	t.join();
  
  		const auto elapsed1 = timer.time_elapsed();
  		if(gen > ga.gnrtn - 2){
  			std::cout << "gpu_eigen time "<<std::chrono::duration_cast<std::chrono::microseconds>(elapsed1).count() << "us" << std::endl;
        profilefile<< "gpu_eigen time "<<std::chrono::duration_cast<std::chrono::microseconds>(elapsed1).count() << "us" << std::endl;
  		}
  		//ga.S = gnom.score;
  		//size_t bestGnomeIdx = ga.sortSIndx()[0];
  
  }//end of fitness function
  
  void binaryclassifier(int classnum){
  	unsigned int* target = (unsigned int*) calloc(tP, sizeof(unsigned int));
  	memcpy(target, ga.T, tP * sizeof(unsigned int));
  	for(int i = 0 ; i < tP; i++){
  		if(target[i]==classnum){
  			ga.T[i] = 1;
  			
  		}else
  			ga.T[i] = 0;
  	}
  }
  
  
  
  void advertisement() {
  	std::cout << std::endl;
  	std::cout << "=========================================================================" << std::endl;
  	std::cout << "Thank you for using the GA-GPU features selection for spectroscopic image!" << std::endl;
  	std::cout << "=========================================================================" << std::endl << std::endl;
  //	std::cout << args.str();
  }
  
  int main(int argc, char* argv[]){
  	
  //Add the argument options and set some of the default parameters
  	args.add("help", "print this help");
  	args.section("Genetic Algorithm");
  	args.add("features", "select features selection algorithm parameters","10", "number of features to be selected");
  	args.add("classes", "image masks used to specify classes", "", "class1.bmp class2.bmp class3.bmp");
  	args.add("population", "total number of feature subsets in puplation matrix", "1000");
  	args.add("generations", "number of generationsr", "50");
  	args.add("initial_guess", "initial guess of featues", "");
  	args.add("debug", "display intermediate data for debugging");
  	args.add("binary", "Select features for binary classes", "");
  	args.add("trim", "this gives wavenumber to use in trim option of siproc which trims all bands from envi file except gagpu selected bands");
  
  	args.parse(argc,argv);                  //parse the command line arguments
  
  //Print the help text if set
  	if(args["help"].is_set()){				//display the help text if requested
  		advertisement();
  		std::cout<<std::endl<<"usage: ga-gpu input output --option [A B C ...]"<<std::endl;
  		std::cout<<std::endl<<std::endl;
  		std::cout<<args.str()<<std::endl;
  		exit(1);
  	}
  	if (args.nargs() < 2) {										//if the user doesn't provide input and output files
  		std::cout << "ERROR: GA-GPU requires an input (ENVI) file and an output (features, text) file." << std::endl;
  		return 1;
  	}
  	if (args["classes"].nargs() < 2) {							//if the user doesn't specify at least two class images
  		std::cout << "ERROR: GA-GPU requires at least two class images to be specified using the --classes option" << std::endl;
  		return 1;
  	}
  
  	std::string outfile = args.arg(1);							//outfile is text file where bnad index, LDA-basis, wavelength and if --trim option is set then trim wavelengths are set respectively 
    std::string profile_file = "profile_" + outfile ;
    std::ofstream profilefile(profile_file.c_str(), std::ios::out);			//open outfstream for outfile
  	
  	time_t t_start = time(NULL);								//start a timer for file reading 	
  	E.open(args.arg(0), std::string(args.arg(0)) + ".hdr");		//open header file 		
  	size_t X = E.header.samples;								//total number of pixels in X dimension
  	size_t Y = E.header.lines;									//total number of pixels in Y dimension
  	B = (unsigned int)E.header.bands;							//total number of bands (features)
  	std::vector<double> wavelengths = E.header.wavelength;		//wavelengths of each band
  
  	if(E.header.interleave != stim::envi_header::BIP){			//this code can only load bip files and hence check that in header file
  		std::cout<<"this code works for only bip files. please convert file to bip file"<<std::endl;
  		exit(1);												//if file is not bip file exit code execution
  	}	
  	
  ///--------------------------Load features--------------------------------------------- 
  	nP = ga_load_class_images(argc, args, &nC, &tP);			//load supervised class images
  	ga.F = load_features( nC, tP, B, E, nP);					//generate the feature matrix
  	ga.T = ga_load_responses(tP, nC, nP);						//load the responses for RF training
  	E.close();													//close the hyperspectral file
  	time_t t_end = time(NULL);
  	std::cout<<"Total time: "<<t_end - t_start<<" s"<<std::endl;	
  	//display_PixelfeatureNclass(ga.F, ga.T , B , 1);			//Print one value from feature matrix to debug feature loading
  	//std::cout << "pixel target is " << ga.T[0] << "  " << ga.T[1] << "  " << ga.T[tP - 2] << "  " << ga.T[tP - 1]<<std::endl;
  ///--------------------------Genetic algorith configurations with defult paramets and from argument values---------------------
  	ga.f = args["features"].as_int(0);					//number of features to be selected by user default value is 10
  	ga.p = args["population"].as_int(0);				//population size to be selected by user default value is 1000
  	ga.gnrtn = args["generations"].as_int(0);			//number of generations to be selected by user default value is 50
  	if(args["binary"]) {						//set this option when features are to be selected as binary clas features (class vs stroma) 
  		binClassOne = args["binary"].as_int(0);			//sel class number here, if 2 then features are selected for (class-2 vs stroma) 
  		//feture selection for class selected by user with user arguments (make it binary class data by making chosen class label as 1 and al other class labels 0 from multiclass data )
  		//to select feature for all classes in joint class data using binary class system need to write a script with loop covering all classes
  		binaryclassifier(binClassOne);
  	}  ///not fully implemented yet
  
  	ga.ub = B;						//upper bound is number of bands (i.e. size of z dimension)  Note: for this particular application and way code is written lower bound is 0 and upper bound is size of z dimension
  	ga.uniformRate = 0.5;			//uniform rate is used in crossover
  	ga.mutationRate = 0.5f;		//in percentage for mutation operation on gnome
  	ga.tournamentSize = 5;			//for crossover best parents are selected from tournament of gnomes  
  	ga.elitism = true;				// if it is true then best gnome of current generation is passed to next generation
  	//initial guess of population
  	ga.i_guess = (unsigned int*) calloc(ga.f, sizeof(unsigned int));		
  	debug = args["debug"];				
  
  //==================Generate intial population =================================================
  	std::vector<unsigned int> i_guess(ga.f);
  	for (size_t b = 0; b < ga.f; b++)     //generate default initial guess 
  		i_guess[b] = rand() % B + 0;
  
  	if (args["initial_guess"].is_set()) {//if the user specifies the --initialguess option & provides feature indices as initial guess
  		size_t nf = args["initial_guess"].nargs();				//get the number of arguments after initial_guess
  		if (nf == 1 || nf == ga.f) {								//check if file with initial guessed indices is given or direct indices are given as argument
  			if (nf == 1) {		 //if initial guessed feature indices are given in file										
  				std::ifstream in;		//open the file containing the baseline points
  				in.open(args["initial_guess"].as_string().c_str());
  				if (in.is_open()){	//if file is present and can be opened then read it 
  					unsigned int b_ind;
  					while (in >> b_ind)  //get each band index and push it into the vector
  						i_guess.push_back(b_ind);					
  				}
  				else 
  					std::cout << "cannot open file of initial_guess indices" << std::endl;
  			}
  			else if (nf == ga.f) {	 //if direct indices are given as argument
  				for (size_t b = 0; b < nf; b++)								//for each band given by the user
  					i_guess[b] = args["initial_guess"].as_int(b);			//store that band in the i_guess array
  			}
  		}		
  	}
  
  	ga.initialize_population(i_guess, debug);      //initialize first population set for first generation, user can pass manually preferred features from command line 
  	//display_gnome(0);	
  
  //------------------Calculate class means and total mean of features----------------------------
  	float* M = (float*)calloc( B , sizeof(float));			//total mean of entire feature martix for all features (bands B)
  	ga.ttlMean(M, tP, B);									//calculate total mean, ga.F is entire feature matrix, M is mean for all bands B(features)
  	if(debug) ga.dispalymean(M);							//if option --debug is used display all bands mean
  	
  	//display band index of bands with mean zero, this indicates that band has all zero values 
  	std::cout<<"Display features indices with zero mean "<<std::endl;
  	for(unsigned int i = 0; i < B; i++){
  		if(M[i]== 0)
  			std::cout<<"\t"<<i;
  	}
  	std::cout<<std::endl;
  //	std::cout << "pixel target is " << ga.T[0] << "  " << ga.T[1] << "  " << ga.T[tP - 2] << "  " << ga.T[tP - 1]<<std::endl;
  	float* cM = (float*)calloc(nC * B , sizeof(float));		//cM is nC X B matrix with each row as mean of all samples in one class for all features (bands B)
  	ga.classMean(cM, tP, nC, B, nP);						//calculate class mean, ga.F is entire feature matrix, M is mean for all bands B(features)
  	if(debug) ga.dispalyClassmean(cM, nC);
  
  //------------------------------------GPU init----------------------------------------------------
  	//checking for cuda device 
  	int count;
  	HANDLE_ERROR(cudaGetDeviceCount(&count));
  	if(count < 1){
  		std::cout<<"no cuda device is available"<<std::endl;
  		return 1;
  	}
  	cudaDeviceProp props;
  	HANDLE_ERROR(cudaGetDeviceProperties(&props, 0));
  	ga.gpuInitializationfrommain(M, cM, nP, tP, nC);
  
  //feture selection for class selected by user with user arguments (make it binary class data by making chosen class label as 1 and al other class labels 0 from multiclass data )
  //to select feature for all classes in joint class data using binary class system need to write a script with loop covering all classes
  	//if(binaryClass){									
  	//	binaryclassifier(binClassOne);
  	//}
  
  //============================= GA evolution by generations ====================================================
  	std::vector<unsigned int> bestgnome;												//holds best gnome after each generation evaluation 
  	size_t bestG_Indx;																	//This gives index of best gnome in the current population to get best gnome and its fitness value
  	unsigned int* newPop = (unsigned int*) calloc(ga.p * ga.f, sizeof(unsigned int));   //temprory storage of new population
  	double* best_S = (double*) calloc (ga.gnrtn, sizeof(double));						//stores fitness value of best gnome at each iteration
  	float* lda = (float*) calloc (ga.p * (nC-1) * ga.f, sizeof(float));					//stores LDA basis for each gnome so that we can have best gnome's LDA basis 																	
  	float* sb = (float*) calloc( ga.p * ga.f * ga.f , sizeof(float)) ;					//3d matrix for between class scatter (each 2d matrix between class scatter for one gnome) 
  	float* sw = (float*) calloc( ga.p * ga.f * ga.f , sizeof(float)) ;					//3d matrix for within class scatter (each 2d matrix within class scatter for one gnome) 
  	ga.zerobandcheck(M, true);										//checking bands with all zeros and duplicated bands in a gnome replacing them with other bands avoiding duplication and zero mean 
  	ga.zerobandcheck(M, true);										//Repeating zeroband cheack as some of these bands are not replaced in previous run and gave random results
  	time_t gpu_timestart = time(NULL);					//start a timer for total evoluation 
  
  	for (size_t gen = 0; gen < ga.gnrtn; gen++){		//for each generation find fitness value of all gnomes in population matrix and generate population for next generation
  		//std::cout<<"Generation: "<<gen<<std::endl;
  		fitnessFunction(sb, sw, lda, M , cM, ga.f, props, gen, profilefile);		//Evaluate phe(feature matrix for current population) for fitness of all gnomes in current population 
  		timer.start();												//start timer for new population generation
  		bestG_Indx = ga.evolvePopulation(newPop, M, debug);			//evolve population to generate new generation population
  		const auto pop_generation = timer.time_elapsed(); 			// end timer for new population generation
  		if(gen >ga.gnrtn -2){
  			std::cout << "population evolution time "<<std::chrono::duration_cast<std::chrono::microseconds>(pop_generation).count() << "us" << std::endl;
    			profilefile<<"population evolution time "<<std::chrono::duration_cast<std::chrono::microseconds>(pop_generation).count() << "us" <<std::endl;
  		}
  						
  		best_S[gen] = ga.S[bestG_Indx];								//score of best gnome in current generation 
  		bestgnome = ga.getGnome(bestG_Indx);						//Best gnome of current populaation																				
  		ga.generateNewP(newPop);									//replace current population with new populaiton in the ga classs object
  		ga.zerobandcheck(M, false);										//checking bands with all zeros and duplicated bands in a gnome replacing them with other bands avoiding duplication and zero mean 
  		ga.zerobandcheck(M, false);										//Repeating zeroband cheack as some of these bands are not replaced in previous run and gave random results
  	}//end generation	  
  	
  	time_t gpu_timeend = time(NULL);					//end a timer for total evoluation 
  	std::cout<<"Total gpu time: "<<gpu_timeend - gpu_timestart<<" s"<<std::endl;
   	profilefile<<"Total gpu time: "<<gpu_timeend - gpu_timestart<<" s"<<std::endl;
  
  //================================ Results of GA ===============================================================
  	std::cout<<"best gnome's fitness value is "<<best_S[ga.gnrtn-1]<<std::endl;	
  	std::cout<<"best gnome is:	";
  	for(size_t i = 0; i < ga.f; i++){
  		std::cout<<" "<<(bestgnome.at(i));
     	}
  	std::cout<<std::endl;	
  
  	//create a text file to store the LDA stats (features subset and  LDA-basis)
  	////format of CSV file is: 1st row - band index, 2nd LDA basis depending on number of classes, 3rd - wavenumber corresponding to band index and it --trim is selected then trim wavnumbersare also given
  	std::ofstream csv(outfile.c_str(), std::ios::out);			//open outfstream for outfile
  	size_t ldaindx = bestG_Indx * (nC-1) * ga.f ;				//Compute LDA basis index of best gnome
  	
  	//if(binaryClass){											//this option is for binary class feature selection from joint classes but not fully implemented
  	//	csv<<binClassOne<<std::endl;
  	//}
  
   //fitness values of best gnome is 
  	csv<<"best gnome's fitness value is "<<best_S[ga.gnrtn-1]<<std::endl;									//output fitness value of best gnome in last generation
  	//output gnome i.e. band index of selected featurs
  	csv<<(bestgnome.at(0));									//output feature subset
  	for(size_t i = 1; i < ga.f; i++)
  		csv<<","<<(bestgnome.at(i));
  	csv<<std::endl;
  
  	//output LDA basis of size r X f, r is nC - 1 as LDA projection is rank limited by number of classes - 1
  	for (size_t i = 0; i < nC-1; i++){						
  		csv<<lda[ldaindx + i * ga.f ];
  		for (size_t j = 1; j < ga.f; j++){
  			csv<<","<<lda[ldaindx + i * ga.f +j];
  		}
  		csv << std::endl;
  	}
  	//output actual wavelenths corresponding to those band indices 
  	csv << (wavelengths[bestgnome.at(0)]);
  	for (size_t i = 1; i < ga.f; i++)
  		csv << "," << (wavelengths[bestgnome.at(i)]);
  	csv << std::endl;
  
  
  	if (args["trim"].is_set()) {
  		csv << "trim info" << std::endl;
  		std::sort(bestgnome.begin(), bestgnome.end());			//sort features index in best gnome 
  
  		std::vector<unsigned int> trimindex(ga.f);				//create a vector to store temprory trim index bounds 
  		std::vector<unsigned int> finaltrim_ind;				//create a vector to store final trim index bounds 
  		std::vector<unsigned int> trim_wv;						//create a vector to store final trim wavelength bounds 
  
  		//trim index 
  		trimindex.push_back(1);									//1st trimming band index is 1
  		for (size_t i = 0; i < ga.f; i++) {						// for each feature find its bound indexes 
  			trimindex[i * 2] = bestgnome.at(i) - 1;		
  			trimindex[i * 2 + 1] = bestgnome.at(i) + 1;			
  		}
  		trimindex.push_back(B);								    //last bound index is B 		
  
  		//organize trim index
  		int k = 0;
  		for (size_t i = 0; i < ga.f + 1; i++) {						// find valid pair of trim indices bound excluding adjacent trim indices
  			if (trimindex[2 * i] < trimindex[2 * i + 1]) {
  				finaltrim_ind.push_back(trimindex[2 * i]);			//this is left bound
  				finaltrim_ind.push_back(trimindex[2 * i + 1]);
  				k = k + 2;
  			}
  		}
  		//add duplicated trim indices as single index to final trim index
  		for (size_t i = 0; i < ga.f + 1; i++) {						//check each pair of trim indices for duplications	
  			if (trimindex[2 * i] == trimindex[2 * i + 1]) {			// (duplication caused due to adjacent features)
  				finaltrim_ind.push_back(trimindex[2 * i]);			// remove duplicated trim indices replace by one
  				k = k + 1;
  			}
  		}
  
  
  		////output actual wavelenths corresponding to those trim indices 
  		////these wavenumber are grouped in pairs, check each pair, if duplicated numbers are there in pair delete one and keep other as band to trim, if 2nd wavnumber is smaller than 1st in a pair ignore that pair
  		////e.g [1, 228, 230, 230, 232, 350,352, 351, 353, 1200] pairas [1(start)-228,230-230, 232-350, 352-351, 353-1200(end)], trimming wavenumbers are [1-228, 230, 233-350, 353-1200]
  		csv << (wavelengths[finaltrim_ind.at(0)]);
  		for (size_t i = 1; i < ga.f * + 2 ; i++)
  			csv << "," << (wavelengths[finaltrim_ind.at(i)]);
  		csv << std::endl;
  	} //end trim option
  
  
    //free gpu pointers
    ga.gpu_Destroy();
  
  	//free pointers
  	std::free(sb);
  	std::free(sw);
  	std::free(M);
  	std::free(cM);
  	std::free(best_S);
  	std::free(lda);
  	std::free(newPop);	
  	
  }//end main