gistfile1.matlab

% syn_sofm
%
% This function implements a Kohonen SOFM for clustering of a synthetic data
% set of 6 Gaussien distributions.
%
% March 2010, Mattias Ohlsson 
% Email: [email protected]

% Clear 
clear all;
close all;

% Default parameters
def_size = [5 5];
def_ndata = 100;
def_epoch = 100;

% Load the synthetic cluster data
ndata = input(sprintf('How many data? [%d] ', def_ndata));
if ndata > 0
  [P,T] = loadclust1(ndata);
else
  [P,T] = loadfacesORL(0.5);
end

% Initialize the random number generator
% tmp = input('Seed to the random number generator (Default current time): ');
% if tmp > 0
%   rand('state',tmp);
%   randn('state',tmp);
% else
%   rand('state',sum(100*clock));
%   randn('state',sum(100*clock));
% end

% Give the size of the output grid
tmp = input(sprintf('Give the size of the output grid [%d %d] ', ...
		    def_size(1), def_size(2)));
if tmp > 0
  gsize = tmp;
else
  gsize = def_size;
end

% Create the SOFM network
net = newsom(P,gsize,'gridtop','dist',0.9,1000,0.02);

% Hom many epochs
tmp = input(sprintf('Number of epochs? [%d] ', def_epoch));
if tmp > 0
  epoch = tmp;
else
  epoch = def_epoch;
end
noepoch = 0;

if ndata > 0
    %% PLOTTING PART
    % Plot of initial setup
    w = net.iw{1,1};
    plot(P(1,:),P(2,:),'.g','markersize',10)
    hold on;
    plotsom(net.iw{1,1},net.layers{1}.distances);
    text = sprintf('Self Organizing Feature Map %dx%d nodex', gsize(1),gsize(2)); 
    title(text);
    drawnow;
    hold off;
    %% END OF PLOTTING PART
end

% Start to train the network
net.trainParam.epochs = epoch;
net = train(net,P);

% Some statistics
Y = sim(net,P);
Yc = vec2ind(Y);
for i=1:gsize(2)
  for j=1:gsize(1)
    nowins(i,j) = length(find(Yc == (i-1)*gsize(1)+j));
  end
end

if ndata > 0
    %% PLOTTING PART
    plot(P(1,:),P(2,:),'.g','markersize',10)
    hold on;
    plotsom(net.iw{1,1},net.layers{1}.distances);
    text = sprintf('Result after %d epochs', epoch);
    title(text);
    drawnow;
    hold off;
    %% END OF PLOTTING PART
else
    for i = 1:gsize(2)
        for j = 1:gsize(1)
            showfacecluster(net, 'sofm', P, T, nowins(i, j));
        end
    end
end

disp('Matrix of the number of "wins" for the output nodes');
disp(nowins);