prepareData.m~

function [trdata,tedata] = prepareData()
%
% Downloads, preprocess and extracts features from Caltech 101
% return trdata and tedata each of which is struct where
%     .data contains the data matrix
%     .labels contains the corresponding labels.
%

% the directory for downloading Caltech101
c101dir = '../data/c101';
ntr = 30;
nte = 30;

% load parameters
% bs : stride for average downsampling in layer1 and layer1
% bw : kernel size for average downsampling layer1 layer2
% ct : connection table between layer1 and layer2 features
% ker: weighting kernel for Local Contrast Normalization
%      fixed gaussian weights.
%      Not the convolutional kernels they are randomly generated
%      below.
params = load('../data/params.mat');

% randomize convolution parameters (they are all zeros initially)
params.kc.layer1 = -0.11 + 0.22 * rand(size(params.ct.layer1,1),9,9);
params.kc.layer2 = -0.11 + 0.22 * rand(size(params.ct.layer2,1),9,9);

% BH: added this
params.c101 = 0;
params.subSamplePercent = 0.5;
params.split = 4;

% first get caltech 101 data
catdir = getC101(c101dir);

fprintf('\n\n\n\n%s\n\n\n\n\n','Extracting Features with Random Kernels... (This might take long depending on your system)');

if params.c101
    % perorms random feature extraction and returns
    % two matrices, data and corresponding labels
    [data,labels] = processImages(catdir,params,ntr,nte);
    
    % prepare training and testing sets
    % trdata and tedata and structs
    % trdata.data contains a large training data
    % matrix
    [trdata,tedata] = prepareTrainTestData(data,labels,ntr,nte);
    
else
    [trdata,trlabels,tedata,telabels] = processImages(catdir,params,ntr,nte);
end