performPCA.m

% EE 660 Project
% Manoj Kumar P.A  (manojpamk@gmail.com)

% Performs PCA to reduce the images from N (typically 100) dimension to M dimension

% Execution path : ~/Acads/Fall2015/EE660/ProjectStuff

% NAIVE Method for now : Convert DxDx3 image into 1x3D^2 row vector


skipSuperVectorCreation = 0;
if (skipSuperVectorCreation ~= 1)
    clc;
    % clear all;
    close all;
    fprintf('Loading data...');
    load matlabData/Rawdata/rawData_rgb_smooth.mat;
%     load matlabData/Rawdata/rawData_grayscale_normal.mat;
%     load matlabData/Rawdata/rawData_grayscale_smooth.mat;
    
    numChannels = 3;        % Change to 1 for grayscale
    fprintf('Done\n');

    myclasses = {'flowers','dogs','houses','aeroplane','ship','car','motorcycle','bus','beach','mountain'};

    %==========================================================================
    % First, vectorize and normalize
    superVector = [];

    for classIndex = 1:length(myclasses)   


       fprintf('Vectorising for class %s..\n',cell2mat(myclasses(classIndex)));
       eval(sprintf('classVar = class_%s;',cell2mat(myclasses(classIndex))));

       temp = zeros(size(classVar,1),size(classVar,2)*size(classVar,3)*numChannels);
       for imageIndex = 1:size(temp,1)
           myImage = squeeze(classVar(imageIndex,:,:,:));
           temp(imageIndex,:) = reshape(reshape(permute(myImage,[1 3 2]),[size(classVar,2)*numChannels size(classVar,2)])',[1 size(temp,2)]);
       end

       superVector = [superVector; [temp classIndex*ones(size(temp,1),1)]];

    end

    % Normalising to zero-mean unit-variance
    % Probably the variance normalisation shouldn't be done now. Only afte reducing dimension?
    m = mean(superVector(:,1:end-1));
    % v = var(superVector(:,1:end-1));
    superVector(:,1:end-1) = superVector(:,1:end-1) - repmat(m,[size(superVector,1) 1]);
    % superVector(:,1:end-1) = superVector(:,1:end-1)./sqrt(repmat(v,[size(superVector,1) 1]));

    clear m v temp classVar myImage class_*;
    fprintf('Saving into superVector_grayscale.mat...\n');    
    save('matlabData/Supervectors/superVector_rgb_smooth.mat');
    %==========================================================================
end
fprintf('Completed superVector preparation. Commencing PCA... \n');
clear;

% Temporary shortcut 
% load matlabData/superVector.mat;

skipPCA = 1;
if (skipPCA ~= 1)
    %==========================================================================
    % Now, get to the PCA part - This block requires a lot of memory and time.
    % Best to execute as a standalone code, and hence skipping over by default.

    % Q is 5.8G for 8240x30,000 
    load('matlabData/Supervectors/superVector_rgb_smooth.mat');

    X = superVector(:,1:end-1); labels = superVector(:,end);
    clear superVector;

    fprintf('Creating covariance matrix...\n');
    Q = X'*X; clear X;    % Takes 5-10 min
      
    
    % This is THE MOST resource-consuming step. Please be wary about it!    
    fprintf('Computing the eigen vectors and eigen values...\n');
    [e,l] = eig(Q); clear Q; l = diag(l);
    l = fliplr(l');
    save('matlabData/AfterEigAnalysis/afterEigAnalysis_grayscale_smooth.mat','-v7.3');

else
    load('matlabData/afterEigAnalysis.mat','l'); 
    % Now, 'e' is the eigenmatrix, 'l' is the eigenvalue matrix
end

% Choosing a reduced dimension
% Command to obtain this : 
% l = fliplr(l');
% c = cumsum(l)./sum(l);
% min(find(c>0.99)), etc..

% Grayscale, Smoothening (max 10000)
% Relative Strength (%)     First n values
% 97*                        443
% 99                         878
% 99.9                       2018
% 99.99                      3617
% 99.999                     5412
 
% Grayscale, Normal (max 10000)
% Relative Strength (%)     First n values
% 92*                        531   
% 99                         2824
% 99.9                       5282
% 99.99                      6863
% 99.999                     7727

% RGB, Smoothening (max 30000)
% Relative Strength (%)     First n values
% 96*                       467
% 99                        1208
% 99.9                      2858
% 99.99                     4837

% RGB, Normal (max 30000)
% Relative Strength (%)     First n values
% 91*                       511
% 99                        3114
% 99.9                      5755
% 99.99                     7347



m = 511;  % The reduced dimension
l = l(1:m);
load('matlabData/afterEigAnalysis.mat','e');     % This will take some time
e = e(:,end-m+1:end); e = fliplr(e);            % Since e-vectors are ascending by default
save('temp.mat','e','l');
clear;
load('matlabData/Supervectors/superVector_grayscale_smooth.mat');
labels = superVector(:,end); superVector = superVector(:,1:end-1);
load('temp.mat','e');
delete temp.mat
superVector = superVector*e;
clear e;

% Now, do zero-mean unit variance normalisation
m = mean(superVector);
v = var(superVector);
superVector = superVector - repmat(m,[size(superVector,1) 1]);
superVector = superVector./sqrt(repmat(v,[size(superVector,1) 1]));


% Creating a bag-of-words representation for the labels
a = eye(10); bag_of_words_labels = zeros(size(superVector,1),length(myclasses));
for classIndex = 1:length(myclasses)
    bag_of_words_labels(824*(classIndex-1)+1:824*(classIndex),:) = repmat(a(classIndex,:),[824 1]);
end
clear a classIndex;

save('matlabData/AfterDimReduction/afterDimReduction_rgb_normal_91_511.mat');
superVector = [superVector labels];
arffwrite('matlabData/ArffFiles/afterDimReduction_rgb_normal_91_511',superVector);

%==========================================================================