SVM_setup.m

function [train_data,train_class,test_data,test_class,fig_handle] = SVM_setup(design_opt,train_N,test_N,rand_seed)
% function [train_data,test_data,train_class,test_class,fig_handle] = SVM_setup(design_opt,train_N,test_N)
% This function sets up the SVM examples
% Inputs: design_opt   - 1 for linearly separable, 2 for nonseparable
%         train_N      - number of training points per population
%         test_N       - number of test points per population
%         rand_seed    - <optional> random seed (default=0)
% Outputs: train_data  - the training set
%          train_class - the classifications at the training points
%          test_data   - the test set
%          test_class  - the classifications at the test points
%          fig_handle  - <optional> return the handle of the figure
%                        of a scatter plot of the data
%
% design_opt = 1 has class=1  centered at (1,0)
%                    class=-1 centered at (0,1)
% design_opt = 2 has class=1  centered at (1,0), (0,1), (2,1)
%                    class=-1 centered at (0,0), (1,1), (2,0)
global GAUSSQR_PARAMETERS
if ~isstruct(GAUSSQR_PARAMETERS)
    error('GAUSSQR_PARAMETERS does not exist ... did you forget to call rbfsetup?')
end
statsOn = GAUSSQR_PARAMETERS.STATISTICS_TOOLBOX_AVAILABLE;
if not(statsOn)
    error('Sorry, but for the moment you need the stats package to run this')
end

if nargin==3
    rand_seed = 0;
elseif nargin~=4
    error('Unacceptable arguments, nargin=%d',nargin)
end

plot_on = 0;
if nargout==5
    plot_on = 1;
end

% This allows us to control the results
% Some older versions of Matlab are handled with this block
if exist('rng','builtin')
    rng(rand_seed);
else
    rand('state',rand_seed);
    randn('state',rand_seed);
    if not(exist('randi','builtin'))
        randi = @(n,r,c) ceil(n*rand(r,c));
    end
end
% This shouldn't be necessary, but seems to be.
randi = @(n,r,c) ceil(n*rand(r,c));

% Set up the data points based on what the user requests
switch design_opt
    case 1
        grnmean = [1,0];
        redmean = [0,1];
        covmat = eye(2);
    case 2
        grnmean = [1,0;0,1;2,1];
        redmean = [0,0;1,1;2,0];
        covmat = .5*eye(2);
    otherwise
        error('design_opt must be either 1 or 2')
end
grnmean_N = size(grnmean,1);
redmean_N = size(redmean,1);

% How much randomness do we want in our training set
buffer = .2;
        
% Generate some manufactured data and attempt to classify it
% The data will be generated by normal distributions with different means
grnpop = mvnrnd(grnmean(randi(grnmean_N,test_N,1),:),repmat(covmat,[1,1,test_N]),test_N);
redpop = mvnrnd(redmean(randi(redmean_N,test_N,1),:),repmat(covmat,[1,1,test_N]),test_N);

% Generate a training set from which to learn the classifier
grnpts = mvnrnd(grnpop(randi(test_N,train_N,1),:),repmat(covmat*buffer,[1,1,train_N]),train_N);
redpts = mvnrnd(redpop(randi(test_N,train_N,1),:),repmat(covmat*buffer,[1,1,train_N]),train_N);

% Create a vector of data and associated classifications
% Green label 1, red label -1
train_data = [grnpts;redpts];
train_class = ones(2*train_N,1);
train_class(train_N+1:2*train_N) = -1;
test_data = [grnpop;redpop];
test_class = ones(2*test_N,1);
test_class(test_N+1:2*test_N) = -1;

% Scatter plot of the input data
if plot_on
    fig_handle = figure;
    hold on
    plot(grnpop(:,1),grnpop(:,2),'g+','markersize',12)
    plot(redpop(:,1),redpop(:,2),'rx','markersize',12)
    plot(grnpop(:,1),grnpop(:,2),'bs','markersize',12)
    plot(redpop(:,1),redpop(:,2),'bo','markersize',12)
    plot(grnmean(:,1),grnmean(:,2),'gs','markersize',12,'MarkerFaceColor','g')
    plot(redmean(:,1),redmean(:,2),'ro','markersize',12,'MarkerFaceColor','r')
    plot(grnpts(:,1),grnpts(:,2),'g+','markersize',7)
    plot(redpts(:,1),redpts(:,2),'rx','markersize',7)
    hold off
end