train_test.asv

the matlab code of ginisvm - an svm implementation good at probability regression.

%----------------------------------------
% First Read in the input data
% trainx -> input data
% Ytrain -> input label
% crossx -> cross-validation data
% Ycross -> cross-validation label
% The training and test labels should be binary inputs
% or prior probabilities.
% An example for a three class label is [0 0 1] to indicate
% that the training label belongs to class 3. Or it could
% be [0.1 0.3 0.6] to indicate prior confidence.
% KTYPE and KSCALE are the kernel parameters. For example
% KTYPE = 6 is for gaussian and KSCALE is the sigma parameter
% of the gaussian.
% Cgini is the C parameter in SVM
% B is the gamma parameter (see JMLR paper).
% Niter is the number of iterations of the randomized SMO
% algorithm. 
%---------------------------------------
[N,D] = size(trainx);
[Ny,M] = size(Ytrain);
if Ny ~= N,
   error('Training Data size neq labels');
end;

[Ncross,D] = size(crossx);
[Nycross,M] = size(Ycross);
if Nycross ~= Ncross,
   error('Cross-validation Data size neq labels');
end;


global KTYPE;
global KSCALE;

%--------------------------------
% Parameters to tune
%--------------------------------

B = 0.01;         % gamma parameter
Niter = 100000;   % For better convergence
Cgini = 1;       % To prevent overfitting
%Cgini = 2.471/B;       % To prevent overfitting
KTYPE = 1;       % Kernel type 
KSCALE = 10000;      % Parameter
%--------------------------------



plotflag = 1;

for i = 1:N,
   yindex(i) = find(Ytrain(i,:) > 0.5);
end;

for i = 1:Ncross,
   ycrossindex(i) = find(Ycross(i,:) > 0.5);
end;

%---------------------------------------
% Train the ginisvm
%---------------------------------------
fprintf('Starting SVM Training....');
[testcoeff,testbias] = ginitrain(trainx,Ytrain,Cgini*ones(N,1),Niter,B*ones(N,1));
fprintf('....Done\n');

%---------------------------------------
% Compute the sparsity index.
%---------------------------------------
spind = find(sum(abs(testcoeff),2) < 1e-5);
nsv = length(spind);
for k = 1:nsv,
   testcoeff(spind(k),:) = zeros(1,M);
end;
fprintf('Sparsity Index = %d\n',nsv/N*100);
   
%---------------------------------------
% Performance on training set
%---------------------------------------

fprintf('Evaluating Performance on Training set\n');
errordist = zeros(1,M);
error = 0;
eflag = zeros(N,1);

for k = 1:N,
   mvalue = kernel(trainx(k,:),trainx)*testcoeff+testbias;
   [result, resultmargin] = ginitest(mvalue,B);
   [maxval,ind] = max(result);
   if ind ~= yindex(k),
      error = error + 1;
      eflag(k) = 1;
      errordist(yindex(k)) = errordist(yindex(k)) + 1;
   end;
end;
fprintf('Multi-class Train Error = %d percent \n',(error/N)*100);
for i = 1:M,
   fprintf('Class %d Error = %d percent \n',i,(errordist(i)/N)*100);
end;
clear result resultmargin;

%---------------------------------------
% Performance on test set
%---------------------------------------
fprintf('Evaluating Performance on Cross-validation set\n');
errordist = zeros(1,M);
error = 0;
eflagcross = zeros(Ncross,1);

for k = 1:Ncross,
   mvalue = kernel(crossx(k,:),trainx)*testcoeff+testbias;
   [result, resultmargin] = ginitest(mvalue,B);
   [maxval,ind] = max(result);
   if ind ~= ycrossindex(k),
      error = error + 1;
      eflagcross(k) = 1;
      errordist(ycrossindex(k)) = errordist(ycrossindex(k)) + 1;
   end;
end;
fprintf('Multi-class Cross-validation Error = %d percent \n',(error/Ncross)*100);
for i = 1:M,
   fprintf('Class %d Error = %d percent \n',i,(errordist(i)/Ncross)*100);
end;
clear result resultmargin;

%---------------------------------------
% Plot the probability Contour
%---------------------------------------
if plotflag == 1,
   fprintf('Plotting Contour ....');
   figure;
   giniplot(trainx,Ytrain,testcoeff,testbias',B);
   fprintf('....done\n');
end;