start_classify.m

最新的模式识别分类工具箱,希望对朋友们有用!

% Main function for the GUI single-algorithm (main) screen
% --------------------------------------------------------
% Modified 03/19/02 - Vittorio Castelli


Npoints = 100; %Number of points on each axis of the decision region

hm = findobj('Tag', 'Messages'); 
hParent = get(hm,'Parent'); %Get calling window tag
set(hm,'String',''); 

h =  findobj('Tag', 'TestSetError');
set(h, 'String', '');
h =  findobj('Tag', 'TrainSetError');
set(h, 'String', '');


error_method_val = get(findobj('Tag', 'popErrorEstimationMethod'),'Value');
error_method_str = get(findobj('Tag', 'popErrorEstimationMethod'),'String');
error_method 	 = char(error_method_str(error_method_val(1)));

framework_val = get(findobj('Tag', 'frameworkToBeUsed'),'Value');
framework_str = get(findobj('Tag', 'frameworkToBeUsed'),'String');
framework_str = cellstr(framework_str);
framework     = char (framework_str(framework_val(1)));

% retrieves the parameters from the interface

if (isempty(whos('targets')))    
  set(hm,'String','No targets on workspace. Please load targets.')   
  break
end

if (isempty(whos('features')))
  set(hm,'String','No features on workspace. Please load features.')
  break
end 

h = findobj('Tag', 'Redraws');   
redraws = str2num(get(h, 'String'));   
if isempty(redraws), 
  set(hm,'String','Please select how many redraws are needed.')      
  break
else     
  if (redraws < 1), 
    set(hm,'String','Number of redraws must be larger than 0.')     
    break    
  end   
end      

h = findobj('Tag', 'PercentTraining'); 
percent = str2num(get(h, 'String'));   

if strcmp(error_method, 'Holdout'),
  if isempty(percent), 
    set(hm,'String','Please select the percentage of training vectors.')     
    break;
  else     
    if (floor(percent/100*length(targets)) < 1),     
      set(hm,'String','Number of training vectors must be larger than 0.')     
      break;
    end   
    if (percent >= 100),     
      set(hm,'String','Number of training vectors must be smaller than 100%.')     
      break;
    end   
    if (floor((1-percent/100)*length(targets)) < 1),     
      set(hm,'String','Number of test vectors must be larger than 0.')     
      break    
    end   
  end
end      


% Some variable definitions
Nclasses  = find_classes(targets); %Number of classes in targets
test_err  = zeros(Nclasses+1,redraws);   
train_err = zeros(Nclasses+1,redraws);  


% Find the region for the grid
[region,x,y]  = calculate_region(features, [zeros(1,4) Npoints]);    

h	      = findobj('Tag', 'Algorithm'); 
val	      = get(h,'Value');     
algorithm     = get(h,'String'); 
algorithm     = deblank(char(algorithm(val,:)));
h	      = findobj('Tag', 'Preprocessing'); 
val           = get(h,'Value');     
preprocessing = get(h,'String'); 
preprocessing = deblank(char(preprocessing(val,:)));

PreprocessingParameters = str2num(char(get(findobj('Tag', 'txtPreprocessingParameters'),'String')));

% Get algorithm parameters. 
% If they don't contain a string, turn them into a number

if strcmp(get(findobj('Tag', 'txtAlgorithmParameters'),'Visible'),'on'),
  AlgorithmParameters = char(get(findobj('Tag', 'txtAlgorithmParameters'),'String'));
else
  AlgorithmParameters = char(get(findobj('Tag', 'txtAlgorithmParametersLong'),'String'));
end

if (~isempty(str2num(AlgorithmParameters))),
  AlgorithmParameters = str2num(AlgorithmParameters);
end

plot_on = strcmp(get(findobj(gcbf,'Label','&Show center of partitions during training'),'Checked'),'on');

%Now that the data is OK, start working

set(gcf,'pointer','watch');
hold off
plot_scatter(features, targets, hParent)
axis(region(1:4))
hold on
drawnow

%
% 
%

if (strcmp(framework, 'Framework_for_Class'))
  new_start_classify;
else
  for i = 1: redraws,  
    set(hm, 'String', ['Processing iteration ' num2str(i) ' of ' num2str(redraws) ' iterations...']);
    
    %  Make a draw according to the error method chosen
    L = length(targets);

    switch error_method

    case cellstr('Resubstitution')
      test_indices  = 1:L;
      train_indices = 1:L;
      
    case cellstr('Holdout')
      [test_indices, train_indices] = make_a_draw(floor(percent/100*L), L); ...

    case cellstr('Cross-Validation')
      chunk          = floor(L/redraws);
      test_indices   = 1 + (i-1)*chunk : i * chunk;
      train_indices  = [1:(i-1)*chunk, i * chunk + 1:L];
    end

    train_features = features(:, train_indices);    
    train_targets  = targets (:, train_indices);    
    test_features  = features(:, test_indices);     
    test_targets   = targets (:, test_indices);     
    
    fprintf('have %d training and %d test samples, fraction = %f\n', ...
	length(train_targets), length(test_targets), percent);
    

    %  Preprocess and then find decision region
    switch preprocessing

    case cellstr('None')
      fprintf('Generating decision region\n')
      fprintf('algorithm = %s,\n',algorithm);
      D = feval(algorithm, train_features, train_targets, AlgorithmParameters, region); 
      fprintf('Calculating the error\n')
      [train_err(:,i), test_err(:,i)] = ...
	  calculate_error (D, train_features, train_targets, ...
	  test_features, test_targets, region, Nclasses);
      
    case cellstr('PCA')
      % preprocessing with PCA
      disp('Performing preprocessing\n')
      [reduced_features, reduced_targets, uw] = ...
	  feval(preprocessing, train_features, train_targets, PreprocessingParameters, region); 
      disp('Generating decision region')
      [region, x, y] = calculate_region(reduced_features, region);    
      D = feval(algorithm, reduced_features, reduced_targets, AlgorithmParameters, region); 
      disp('Calculating the error')
      [train_err(:,i), test_err(:,i)] = calculate_error (D, reduced_features, reduced_targets, uw*test_features, test_targets, region, Nclasses);      
      
    case cellstr('Fishers_linear_discriminant')
      disp('Performing preprocessing')
      [reduced_features, reduced_targets, w] = feval(preprocessing, train_features, train_targets, [], region); 
      [region, x, y] = calculate_region(reduced_features, region);    
      disp('Generating decision region')
      D = feval(algorithm, reduced_features, reduced_targets, AlgorithmParameters, region); 
      disp('Calculating the error')
      [train_err(:,i), test_err(:,i)] = calculate_error (D, reduced_features, reduced_targets, [w'*test_features; zeros(1,length(test_targets))], test_targets, region, Nclasses);      
      % Replot the data
      hold off
      plot_scatter([w'*features; zeros(1,length(targets))], targets, hParent)
      hold on
      
    otherwise
      disp('Performing preprocessing')
      if strcmp(get(findobj(gcbf, 'Tag', '&Options&SeparatePreprocessing'),'Checked'),'on'),
	disp('Perform seperate preprocessing for each class.')
	in0 = find(train_targets == 0);
	in1 = find(train_targets == 1);
	[reduced_features0, reduced_targets0] = ...
	    feval(preprocessing, train_features(:,in0), ...
	    train_targets(in0), PreprocessingParameters, region, ...
	    plot_on);  

	[reduced_features1, reduced_targets1] = ...
	    feval(preprocessing, train_features(:,in1), ...
	    train_targets(in1), PreprocessingParameters, region, ...
	    plot_on); 
	
	reduced_features = [reduced_features0, reduced_features1];
	reduced_targets  = [reduced_targets0,  reduced_targets1];
      else
	[reduced_features, reduced_targets] = ...
	    feval(preprocessing, train_features, train_targets, ...
	    PreprocessingParameters, region, plot_on); 
      end
      pause(1);
      plot_process([]);
      indices = find(sum(isfinite(reduced_features)) > 0);
      reduced_features = reduced_features(:,indices);
      reduced_targets  = reduced_targets(:,indices);

      if (i == redraws)
	%Plot only during the last iteration
	plot_scatter(reduced_features, reduced_targets, hParent, 1)
	axis(region(1:4))
      end
      %Show Voronoi diagram
      if ~isempty(findobj('Tag','Voronoi diagram')),
	%Voronoi diagram figure exists
	figure(findobj('Tag','Voronoi diagram'))
	clf;
      else
	figure;
	set(gcf,'Tag','Voronoi diagram');
      end
      hold on
      contour(x,y,voronoi_regions(reduced_features, region),length(reduced_targets))
      plot_scatter(reduced_features, reduced_targets)
      hold off
      axis(region(1:4));
      grid on;
      title('Voronoi regions')
      figure(hParent)
      if ((sum(reduced_targets) <= 1) & (sum(~reduced_targets) <= 1) & (~strcmp(algorithm,'None')))
	error('Too few reduced points (This program needs at least two points of each class). Please restart.')
      else
	if strcmp(algorithm,'None'),
	  %No classification was asked for
	  set(gcf,'pointer','arrow');     
	  return
	end
      end
      disp('Generating decision region')
      D = feval(algorithm, reduced_features, reduced_targets, AlgorithmParameters, region); 
      disp('Calculating the error')
      [train_err(:,i), test_err(:,i)] = calculate_error (D, train_features, train_targets, test_features, test_targets, region, Nclasses);
      
    end      
  end      
  
  %Display error
  h =  findobj('Tag', 'TestSetError');
  s = 'Test set errors: ';
  for j = 1:Nclasses,
    s = [s 'Class ' num2str(j) ': ' num2str(mean(test_err(j,:)),2) '. '];
  end
  s = [s 'Total: ' num2str(mean(test_err(Nclasses+1,:)),2)];
  set(h, 'String', s);

  h =  findobj('Tag', 'TrainSetError');

  h =  findobj('Tag', 'TrainSetError');
  s = 'Train set errors: ';
  for j = 1:Nclasses,
    s = [s 'Class ' num2str(j) ': ' num2str(mean(train_err(j,:)),2) '. '];
  end
  s = [s 'Total: ' num2str(mean(train_err(Nclasses+1,:)),2)];
  set(h, 'String', s);

  %Show decision region
  [s,h]=contour(x,y,D,1);set(h,'LineWidth',2,'EdgeColor','k');
  if strcmp(get(findobj(gcbf,'Label','Shade &Decision Regions'),'Checked'),'on'),
    contourf(x,y,D,1);colormap([198 198 255; 240 255 240]/255);
  end

  %Show Bayes decision region and error (if possible)
  hBayes = findobj('Tag','chkBayes');
  if ((get(hBayes, 'Value')) & (~isempty(whos('p0')))),
    Dbayes = decision_region(m0, s0.^2, w0, m1, s1.^2, w1, p0, region);
    [s,h]	 = contour(x,y,Dbayes,1);set(h,'LineWidth',1,'EdgeColor','r');
    [classify, Bayes_err]  = classification_error(Dbayes, features, targets, region);
    h =  findobj('Tag', 'BayesError');
    s = 'Bayes errors: ';
    for j = 1:Nclasses,
      s = [s 'Class ' num2str(j) ': ' num2str(1 - classify(j,j),2) '. '];
    end
    s = [s 'Total: ' num2str(mean(1 - diag(classify)),2)];
    set(h, 'String', s);
  end   
  grid on

  %Replot training points if necessary
  if strcmp(get(findobj(gcbf,'Label','Show &Training points'),'Checked'),'on'),
    plot_scatter(train_features, train_targets, hParent, 2)
  end

  hold off

  %That's all folks!
  s = 'Finished!';
  if (redraws > 1),
    s = [s ' (Note that only the last decision region is shown)'];
  end
  set(hm, 'String', s);   
  set(gcf,'pointer','arrow');
end