Demo.rar_DEMO_KNN kddcup99_kddcup99_kddcup99/knn

%DEMO clear clc close all %% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %STEP 1 %Load data and calculate the mean face %load data display('loading data...'); [test_ids,test_imgs,train_ids,train_imgs,nonface_imgs]=load_data(); %calculate the average face display('calculate the average face...'); avg_face=getAvgFace(train_imgs); %get the size of the original images (all datasets have images with the %same dimensions) img_size=[size(train_imgs,2),size(train_imgs,3)]; %get the number of samples in the training data n=size(train_imgs,1); %Too much figures..ask for closing figures of step1 choice = questdlg('would you like to close all figures of step1?', ... 'Step1', ... 'Yes','No','No'); % Handle response switch choice case 'Yes' close all end %% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %STEP 2 %get principle components W and the N latent coordinates %resize data to be (height*width)xN matrix train_imgs=(reshape(train_imgs,[size(train_imgs,1),size(train_imgs,2)*... size(train_imgs,3)]))'; nonface_imgs=(reshape(nonface_imgs,[size(nonface_imgs,1),size(nonface_imgs,2)*... size(nonface_imgs,3)]))'; test_imgs=(reshape(test_imgs,[size(test_imgs,1),size(test_imgs,2)*... size(test_imgs,3)]))'; avg_face=avg_face(:); %resize the mean to be (height*width)x1 vector display('calculate the principal components...'); [W,X,L]=PCA_(train_imgs,avg_face); %get the principle components %get the bound of x axis in plots m=length(L); %Scree Plot figure; plot(L); title('Scree Plot'); ylim([-200,max(L(:))]) xlim([-50,m]) xlabel('Index of Eigenvalues') % x-axis label set(gca,'XTickLabel',num2str(get(gca,'XTick').')) %to avoid x10^k ylabel('Eigenvalues') % y-axis label set(gca,'YTickLabel',num2str(get(gca,'YTick').')) %to avoid x10^k print('visualization\\ScreePlot','-dpng'); %save it %Fraction of Variance (FVE) FVE=cumsum(L(:))/sum(L(:)); %calc FVE %plot FVE figure; plot(FVE); title('Fraction of Variance Explained'); xlabel('c') % x-axis label ylabel('FVE') % y-axis label ylim([min(FVE(:)),1]) print('visualization\\FVE','-dpng'); %save it %Plot first 10 principle components (Eigenvectors) figure('units','normalized','outerposition',[0 0 1 1]) %full screen for i=1:10 subplot(2,5,i); imshow(reshape(W(:,i),img_size(1),img_size(2)),[]); colormap(jet); title(sprintf('Eigenvector %d',i)); end print('visualization\\10EigenVectors','-dpng'); %save it %reconstruct the data %pick random image num=round(rand*n-1)+1; %reconstruct the original images using few number of basis vectors figure('units','normalized','outerposition',[0 0 1 1]) %full screen j=2; subplot(2,5,1); imshow(reshape(train_imgs(:,num),img_size(1),img_size(2)),[]); title('original image'); for i=5:10:90 display(sprintf('reconstructing the data using only the first %d basis vectors...',i)); X=W(:,1:i)'*(train_imgs(:,num)-avg_face); %recalculate X Y=(W(:,1:i)*X)+avg_face; %reconstruct the image using (i)th basis vectors subplot(2,5,j); j=j+1; imshow(reshape(Y,img_size(1),img_size(2)),[]); title(sprintf('Using %d principal components',i)); end print('visualization\\reconst_low','-dpng'); %save it %find the smallest dimensionality such that at least 90% of the variance is %explained alpha=0.9; indices=find(FVE>=alpha); %pick all c values (number of dimentions) that %satisfy the condition >= alpha index=indices(1); %get the first one display(sprintf('The smallest dimensionality that explain at least 90%s of the variance=%d','%',index)); %Too much figures..ask for closing figures of step2 choice = questdlg('would you like to close all figures of step2?', ... 'Step2', ... 'Yes','No','No'); % Handle response switch choice case 'Yes' close all end %% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %STEP 3 %Reconstruction using low dimensionality data %calculate the latent coordinates using only the first 115 principle %components. %we can compute the first c(th) principle components using: %[W,X,L]=PCA_(train_imgs,avg_face,index); %get the first (index=value) principle components %or.. W=W(:,1:index); %discard principle components starting from the basis vector after (index) X=W'*(train_imgs-repmat(avg_face,[1,n])); %recalculate X %reconstruct face images (training set) display('reconstructing face images...'); %Y=sqrt(n)*(W*X)+repmat(avg_face,[1,n]); Y=(W*X)+repmat(avg_face,[1,n]); %reconstruct nonface images display('reconstructing non-face images...'); %avg_nonface(:,:)=mean(nonface_imgs,2); %get mean of non face images X_nonfaces=W'*(nonface_imgs-repmat(avg_face,[1,size(nonface_imgs,2)])); %recalculate X Y_nonface=(W*X_nonfaces)+repmat(avg_face,[1,size(nonface_imgs,2)]); %reconstruct %compute the absolute difference images of training face images diff_face=abs(Y-train_imgs); %compute the squared reconstruction error of training face images error_face=(sum((Y-train_imgs).^2,1)); %compute the absolute difference images of non-face images diff_nonface=abs(Y_nonface-nonface_imgs); %compute the squared reconstruction error of non-face images error_nonface=(sum((Y_nonface-nonface_imgs).^2,1)); %show 5 random samples from reconstructed faces and reconstructed non-faces %images %pick 5 reconstructed face images for i=1:5 figure('units','normalized','outerposition',[0 0 1 1]) %full screen num=round(rand*n-1)+1; %original image ax1 =subplot(1,3,1); imshow(reshape(train_imgs(:,num),img_size(1),img_size(2)),[]); title(sprintf('original image# %d',num)); colormap(ax1,gray); %reconstructed image ax2=subplot(1,3,2); imshow(reshape(Y(:,num),img_size(1),img_size(2)),[]); title('reconstructed image'); colormap(ax2,gray); %abs error ax3=subplot(1,3,3); imshow(reshape(diff_face(:,num),img_size(1),img_size(2)),[]); title('abs error'); colormap(ax3,jet); axes('Position',[0 0 1 1],'Xlim',[0 1],'Ylim',[0 1],'Box','off',... 'Visible','off','Units','normalized', 'clipping' , 'off'); %report squared reconstruction error t=text(0.5, 1,sprintf('Squared Reconstruction Error: %f',... error_face(num)),'HorizontalAlignment',... 'center','VerticalAlignment', 'top'); t.FontSize = 12; %save it print(fullfile('visualization',sprintf('reconst_face_%d',i)),'-dpng'); end %pick 5 reconstructed non-face images for i=1:5 figure('units','normalized','outerposition',[0 0 1 1]) %full screen num=round(rand*size(nonface_imgs,2)-1)+1; %original image ax1=subplot(1,3,1); imshow(reshape(nonface_imgs(:,num),img_size(1),img_size(2)),[]); colormap gray %restore the gray colormap title(sprintf('original image# %d',num)); colormap(ax1,gray) %reconstructed image ax2=subplot(1,3,2); colormap(ax2,gray); imshow(reshape(Y_nonface(:,num),img_size(1),img_size(2)),[]); title('reconstructed image'); %abs error ax3=subplot(1,3,3); imshow(reshape(diff_nonface(:,num),img_size(1),img_size(2)),[]); colormap(ax3,jet); title('abs error'); axes('Position',[0 0 1 1],'Xlim',[0 1],'Ylim',[0 1],'Box','off',... 'Visible','off','Units','normalized', 'clipping' , 'off'); %report squared reconstruction error t=text(0.5, 1,sprintf('Squared Reconstruction Error: %f',... error_nonface(num)),'HorizontalAlignment',... 'center','VerticalAlignment', 'top'); t.FontSize = 12; %save it print(fullfile('visualization',sprintf('reconst_nonface_%d',i)),'-dpng'); end %picking a threshold value to distinguish between face and non-face images T=(mean(error_face)+(max(error_face)-mean(error_face))/2+... mean(error_nonface)-(mean(error_nonface)-mi