learn_filters_load_save.m

Standard model object recognition matlab code

function f = learn_filters_load_save (prev_dir,savefile,layer)

% FUNCTION f = learn_filters_load_save (prev_dir,savefile,layer)
% 
% This function returns the learned features, learned from
% prev_dir, and save them in savefile.  The input
% "param" determines which set of learn parameters to choose in
% init_learn_param.m

global ON_CON;
ON_CON = 10;

% "layer" is a switch that determines the following parameters.
init_learn_param; 
%learn_num_patch     = 10;   % number of patches per training image
%learn_patch_siz     = [3];  % size of patches (spatial dimension)
%learn_num_afferents = 10;   % number of afferents within the patch
%learn_scales        = [3];  % [1:8];
%shift               = 1;

D = dir([prev_dir '/r_i*.mat']);
f = [];

disp(['Learning on ' num2str(length(D)) ' images: ' prev_dir]);

for i = 1:length(D)
  fprintf('.'); % Display the progress
  if mod(i,50) == 0
    fprintf('\n');
  end
  
  % Find the response and retrieve the last layer.
  load ([prev_dir D(i).name]);
  X = r;
  
  % Tag with the first learn_scale.
  scale_tag = ['s' num2str(learn_scales(1))];

  % Learn the snapshots and save.
  for which_scale = learn_scales
    Xs = getfield(X, scale_tag);
    
    if max(max(max(abs(Xs))))>0 % For non-blank images only
      [f1, f2] = learn_snapshots (Xs, learn_patch_siz(which_scale), ...
	  learn_num_patch, learn_num_afferents);
      f2 = (double(f1>=ON_CON)); % Do f2 first.
      f1 = (double(f1-ON_CON)>=0).*(f1-ON_CON);

      % Check the dimension of learned filters, before returning the results.
      % This is for the case when we had learned only 1 feature.
      if size(f1,4) <= 1
	f0 = zeros([size(f1) 1]);
	f0(:,:,:,1) = f1;
	f1 = f0;
      end
      if size(f2,4) <= 1
	f0 = zeros([size(f2) 1]);
	f0(:,:,:,1) = f2;
	f2 = f0;
      end
      f_tmp = filt_package([],f1,f2,shift(which_scale), ...
	  ['s' num2str(which_scale)]);

      % Merge with the previous filters
      if isstruct(f) % & isfield(f, ['f1_' scale_tag])
	%f = aux_merge_filters(f,which_scale,f_tmp,which_scale,scale_tag);
	f = aux_merge_filters(f,learn_scales(1),f_tmp,which_scale,scale_tag);
      else
	f = f_tmp;
      end
    end
  end
end

fprintf('\n');
save(savefile, 'f', 'learn_num_patch', 'learn_patch_siz', ...
    'learn_num_afferents', 'learn_scales');