analyze_data_snn.m

来自「神经网络的工具箱, 神经网络的工具箱,」· M 代码 · 共 107 行

function analysis = analyze_data_snn(P)
%ANALYZE_DATA_SNN classify inputs in bool/boolgroup/cycle/continous variables.
%
%  Syntax
%
%   analysis = analyze_data_snn(P)
%
%  Description
%
%   ANALYZE_DATA_SNN takes
%    P   - [N x MU] matrix of inputs/outputs
%   and returns 
%    analysis - a struct containing information about P.
%
%   (N = #inputs/outputs; MU = #patterns)
%

if isnumeric(P)
   readdata = P';
   [MU, NM] = size(readdata);
   
   %check for booleans.
   bool_ind = find ...
      (  ( sum((readdata == repmat(min(readdata),MU,1)) + ...
               (readdata == repmat(max(readdata),MU,1))) + ...
           sum(isnan(readdata)) ) == MU  );
     
% BUGFIX: use function double, because in Matlab 6.5 bools 
% form a logical array, while it was a double array in former
% Matlab versions.  
%<    bools = ...
%<      (readdata(:,bool_ind) == repmat(max(readdata(:,bool_ind)), MU, 1));
%>
   bools = double(...
       (readdata(:,bool_ind) == repmat(max(readdata(:,bool_ind)), MU, 1)));     
%>>>
   bools(find(isnan(readdata(:,bool_ind)))) = NaN;

   %check for 'boolgroups' % boolgroup: sum(group) = 1, with group = [i:k];
   NBOOL = size(bool_ind,2);
   ngroups = 0;
   bool_groups = [];
   group_begin = 1;
   while(group_begin<=NBOOL)
	for group_end = NBOOL:-1:group_begin
	    score = sum(bools(:, group_begin:group_end),2); 
	    if (sum(score == 1 | isnan(score)) == MU) 
               ngroups = ngroups + 1;
	       bool_groups = [bool_groups; zeros(1,NM)];
	       bool_groups(ngroups, bool_ind([group_begin:group_end])) = 1;
               group_begin = group_end + 1;
	       break;
            end
        end 
        group_begin = group_begin + 1;
   end

   %continuous variables
   cont_ind = setdiff([1:NM], bool_ind);
   conts = readdata(:, cont_ind);
   
   %check for cycles: x^2 + y^2 = 1
   NCONT = size(cont_ind, 2);
   ncycles = 0; 
   cycles = [];
   cycle_begin = 1;
   while (cycle_begin < NCONT)
      score = abs(sum((conts(:,[cycle_begin, cycle_begin+1])).^2, 2) - 1);
      if (score < 0.001) 
	 ncycles = ncycles + 1;
         cycles = [cycles; zeros(1,NM)];
	 cycles(ncycles, cont_ind([cycle_begin:(cycle_begin+1)])) = 1;
         cycle_begin = cycle_begin + 1;
      end
      cycle_begin = cycle_begin + 1;
   end


   for nm = 1:NM
       analysis(nm).min = min(P(nm,:), [], 2);
       analysis(nm).max = max(P(nm,:), [], 2);
       analysis(nm).group = nm;
   end
   for ind = bool_ind
       analysis(ind).type = 'bool';
   end
   for ind = cont_ind
       analysis(ind).type = 'continuous';
   end
   for k = 1:size(cycles,1)
       cycle = find(cycles(k,:));
       for ind = cycle
           analysis(ind).group = cycle; 
       end
   end
   for k = 1:size(bool_groups,1)
       group = find(bool_groups(k,:));
       for ind = group
           analysis(ind).group = group; 
       end
   end
  
else
   error('ANALYZE_DATA_SNN: input must be (numeric) matrix');
end