train_lin_denois.m~

很好的matlab模式识别工具箱

% Training of linear PCA model to tuned for image restoration.
%

% Modifications:
% 5-may-2004, VF
% 14-apr-2004, VF
% 13-apr-2004, VF
% 17-mar-2004, VF

% == Setting ==========================================
Num_Folds = 1;

%New_Dim_Range = [5 10 20 30 40 60 80 100];
New_Dim_Range = [1 2];

input_data_file = 'circle_data';
output_data_file = [];
%input_data_file = '/home.dokt/xfrancv/data/usps_noisy';
%output_data_file = 'LinearPCAModelUSPS';

% ========================================================

% Loads training and testing data.
load(input_data_file);
[Dim,Num_Data] = size( trn.X );

% Data partitioning
[itrn,itst]=crossval(Num_Data,Num_Folds);

% Tuning linear PCA model
Mse = [];

for new_dim = New_Dim_Range,
    
 fprintf('\nnew_dim = %d\n', new_dim);
   
 cv_mse = 0;  
 for i=1:Num_Folds,
    
   fprintf('\n');
    
   trn_X = trn.gnd_X(:,itrn{i});
   val_gnd_X = trn.gnd_X(:,itst{i});
   val_corr_X = trn.X(:,itst{i});

   
   fprintf('Computing Linear PCA...');
   lin_model = pca(trn_X, new_dim);
   fprintf('done\n');

   fprintf('Projecting validation data...');
   val_reconst_X = lpimage( val_corr_X, lin_model );
   fprintf('done.\n');

   dummy = (val_reconst_X - val_gnd_X).^2;
  
   mse = sum(dummy(:))/size(val_gnd_X,2);
  
   fprintf('folder %d/%d: validation errors mse=%f\n', ...
      i, Num_Folds, mse);
   
   cv_mse = cv_mse + mse;
 end
 
 cv_mse = cv_mse/Num_Folds;
 
 Mse(find(new_dim==New_Dim_Range)) = cv_mse;
 
 fprintf('new_dim = %d: mse = %f\n', new_dim, cv_mse);
end
   
[dummy,inx] = min(Mse);
fprintf('\nMin(mse) = %f, dim = %f\n', ...
   Mse(inx), New_Dim_Range(inx) );

fprintf('Computing optimal Kernel PCA...');
lpca_model = pca( trn.X, New_Dim_Range(inx) );
fprintf('done.\n');

if isempty(output_data_file),
  figure; hold on;
  xlabel('dim'); ylabel('mse');

  plot(New_Dim_Range,Mse);
else
   save(output_data_file,'New_Dim_Range',...
      'Mse','Num_Folds','input_data_file',...
      'output_data_file','lpca_model');
end

% plots 2D data
if Dim == 2 & isempty(output_data_file),
  X = lpimage(tst.X, lin_model );

  mse = sum(sum((X-tst.gnd_X).^2 ));
  fprintf('\ntest mse=%f\n', mse);
  
  figure; hold on;
  h0=ppatterns(tst.gnd_X,'r+');
  h1=ppatterns(tst.X,'gx');
  h2=ppatterns(X,'bo');
  legend([h0.points h1.points h2.points],'Ground truth','Noisy','Reconst');
end


% EOF