plotsvm2d.m

CU SVM Classifier Matlab Toolbox

function PlotSVM2D(data,params,options)
% PlotSVM2D trains a binary SVM classifier (SVM-type: C-SVC) using 
% specified data and plots decision boundary of the SVM classifier in 2D.
%
% Synopsis:
%  PlotSVM(data)
%  PlotSVM(data,params)
%  PlotSVM(data,params,options)
% Inputs:
% data - training data
%  data.X a row of column vectors, input data.
%  data.y a row vector, the labels of data.X.
% params - the parameters used to train the SVM
%  params.kernel_type the kernel function type (default RBF).
%       0 linear kernel
%       1 polynomial
%       2 RBF
%       3 Sigmoid function
%  params.C cost coeficient (default 1).
%  params.gamma constant 'gamma' in the kernel function (default 0.5).
%  params.degree the degree of the polynomial kernel (default 3).
%  params.coef0 constant 'coef0' in the kernel function (default 0).
% options - optional_inputs
%  options.background [int] If 1 then backgroud is colored 
%    according to the value of decision function; (default 1).
%  options.sv [int] If 1 then the support vectors are marked; 
%    (defaultly 1).
%  options.sv_size [int] Size of marker used to denote SVs.
%  options.margin [int] If 1 then margin is displayed; (default 1).
%  options.gridx [int] Density of grid in y-axis; (default 50).
%  options.gridy [int] Density of grid in y-axis; (default 50).
%  options.color [int] Color of decision boundary; (default 'k').
%
% Output:
%  H [struct] Handles of used graphical objects.
%

% Modifications:
% 09-24-2004, first created by Xi Long

% -- Process input arguments -----------------
if (nargin < 1) || (nargin > 3)
   disp(' Incorrect number of input variables.\n');
   help PlotSVM2D;
   return;
end

if ~isfield(data, 'X')
    disp(' Error: invalid data structure.\n');
    return; 
end
if ~isfield(data, 'y')
    disp(' Error: invalid data structure.\n');
    return; 
end
if isempty(data.X)
    disp(' Error: sample field is empty.\n');
    return; 
end
if size(data.X,2) ~= size(data.y,2)
    disp(' Error: sample and label do not match.\n');
    return; 
end

if nargin < 2, params=[]; end
if ~isfield(params,'kernel_type'), params.kernel_type = 2; end
if ~isfield(params,'C'), params.C = 1; end
if ~isfield(params,'gamma'), params.gamma = 0.5; end
if ~isfield(params,'degree'), params.degree = 3; end
if ~isfield(params,'coef0'), params.coef0 = 0; end

if nargin < 3, options=[]; end
if ~isfield(options,'background'), options.background = 1; end
if ~isfield(options,'sv'), options.sv = 1; end
if ~isfield(options,'margin'), options.margin = 1; end
if ~isfield(options,'gridx'), options.gridx = 50; end
if ~isfield(options,'gridy'), options.gridy = 50; end
if ~isfield(options,'sv_size'), options.sv_size = 12; end
if ~isfield(options,'color'), options.color = 'k'; end

switch params.kernel_type
    case 0
        model=LinearSVC_Train(data, 0, params.C);
    case 1
        model=PolySVC_Train(data, 0, params.C, params.gamma, params.degree, params.coef0);
    case 2
        model=RbfSVC_Train(data, 0, params.C, params.gamma);
    case 3
        model=SigmSVC_Train(data, 0, params.C, params.gamma, params.coef0);
    otherwise
        disp('Unknown kernel type!')
        return;
end 


% get axis
a = axis;
old_hold = ishold;
hold on;

% plot Support Vectos
if options.sv,
%  inx = find(model.sv.y==1);
%  ppatterns(model.sv.X(:,inx),['o' marker_color(1)],options.sv_size);
%  inx = find(model.sv.y==2);
%  ppatterns(model.sv.X(:,inx),['o' marker_color(2)],options.sv_size);
  tmp=PlotPtns(model.SV, 'ok' ,options.sv_size);
  H.sv = tmp.points;
end

% limits of current figure
xmin=a(1);
xmax=a(2);
ymin=a(3);
ymax=a(4);
  
% makes grid 
[X,Y] = meshgrid(xmin:(xmax-xmin)/options.gridx:xmax,...
                 ymin:(ymax-ymin)/options.gridy:ymax);

% make testing patterns covering whole grid
tst_data.X=[reshape(X',1,prod(size(X)));reshape(Y',1,prod(size(Y)))];

% classify points
tst_data.y=ones(size(tst_data.X,2));


%[ClassRate, dec_fun, Ns, ConfMatrix, pred_labels]= SVMTest(tst_data, Labels, model.AlphaY, model.sv.X, model.Bias,model.Parameters, model.nSV, model.nLabel);
output=SVM_Predict(0, tst_data, model, 0);

dec_fun=output.predvalue';

% compute color limits
l=(-min(dec_fun)+max(dec_fun))/2;

% reshape dec_fun
Z = reshape(dec_fun,size(X,1),size(X,2))';

% colors background 
if options.background,
  H.background = pcolor(X,Y,Z);
end

% smooth shading
shading interp;

% plots decision boundary
[dummy,H.boundary] = contour(X,Y,Z,[0,0],options.color);

% plots margins
if options.margin,
   [dummy,H.margin_plus] = contour(X,Y,Z,[1,1],[options.color,'--']);
   [dummy,H.margin_minus] = contour(X,Y,Z,[-1,-1],[options.color,'--']);
end

% set color limits and colormap
if options.background,
  set(H.background, 'LineStyle','none' );
  set(gca,'Clim',[-l l]);
  
  % creates colormap and sets it up
  g=gray(64);
  cmp=[g(33:end,:);flipud(g(33:end,:))];
  cmp(1:32,1)=cmp(1:32,1)/2;
  cmp(1:32,3)=cmp(1:32,3)/2;
  cmp(33:end,3)=cmp(33:end,3)/2;
%  cmp(33:end,2)=cmp(33:end,2)/2;
  colormap(cmp)
end

if ~old_hold,
   hold off;
end

if nargout >= 1, varargout{1} = H; end

return;