%CLASSC Convert classifier to normalized classifier (yielding confidences)
%
% V = CLASSC(W)
% V = W*CLASSC
% D = CLASSC(A*W)
% D = A*W*CLASSC
% D = CLASSC(A,W)
%
% INPUT
% W Trained or untrained classifier
% A Dataset
%
% OUTPUT
% V Normalized classifier producing confidences instead of
% densities or distances (after training if W is untrained)
%
% DESCRIPTION
% The trained or untrained classifier W may yield densities or unnormalised
% confidences. The latter holds for two-class discriminants like FISHERC
% and SVC as well as for neural networks. Such classifiers use or should
% use CNORMC to convert distances to confidences. In multi-class problems
% as well as in combining schemes they do not produce normalises
% confidences. These outcomes, like the density outcomes of classifiers
% liek QDC, LDC and PARZENC, can be converted by CLASSC into confidences:
% the sum of the outcomes will be one for every object.
%
% In case W is a one-dimensional mapping, it is converted into a two-class
% classifier, provided that during the construction a class label was
% supplied. If not, the mapping cannot be converted and an error is
% generated.
%
% CLASSC lists the outcomes on the screen in case no output argument is
% supplied. Also true and estimated labels are supplied.
%
% SEE ALSO (<a href="http://37steps.com/prtools">PRTools Guide</a>)
% MAPPINGS, DATASETS, CNORMC, LABELD
% Copyright: R.P.W. Duin, r.p.w.duin@37steps.com
% Faculty EWI, Delft University of Technology
% P.O. Box 5031, 2600 GA Delft, The Netherlands
% $Id: classc.m,v 1.5 2010/02/23 15:21:54 duin Exp $
function w = classc(w,flag)
if nargin < 2, flag = 0; end % flag forces non=combiner behavior avoiding recursion
if (nargin == 0)
% Untrained mapping.
w = prmapping('classc','combiner',flag);
elseif (ismapping(w))
% If mapping is stacked or parallel, recurse over the individual
% sub-mappings and call CLASSC for each of them.
if ((isstacked(w)) | (isparallel(w))) & (flag == 0)
v = cell(1,length(w.data));
for j = 1:length(w.data)
if ismapping(w.data{j}) % the parallel combiner may have nonmapping data
v{j} = feval(mfilename,w.data{j});
else
v{j} = w.data{j};
end
end
w = setdata(w,v);
w = feval(mfilename,w,1); % and here CLASSC is called for the combiner avoiding recursion
else
conv = get(w,'out_conv');
if (conv < 1)
% Set the "normalization" bit in the mapping's output conversion flag
w = set(w,'out_conv',conv+2);
end;
end
elseif (isdataset(w))
if ismapping(flag)
if nargout == 1
w = feval(mfilename,w*flag);
else
feval(mfilename,w*flag);
clear w;
end
return
end
w = w*normm;
w = w*costm;
if nargout == 0 % list outcomes on the screen
ww = +w;
ss = repmat('-',1,9*size(ww,2));
fprintf('\n True Estimated Class \nLabels Labels Confidences\n');
fprintf('------------------%s\n',ss);
nlab = getnlab(w);
[wmax,K] = max(ww,[],2);
lablist = getlablist(w);
if ~isempty(lablist) & ~ischar(lablist)
nlab = lablist(nlab);
K = lablist(K);
end
for j=1:size(ww,1)
if (nlab(j) ~= K(j))
fprintf(' %3.0f ->%3.0f ',nlab(j),K(j));
else
fprintf(' %3.0f %3.0f ',nlab(j),K(j));
end
fprintf(' %7.4f',ww(j,:));
fprintf('\n');
end
lablist = getlablist(w);
if ischar(lablist)
fprintf('\n');
for j=1:size(lablist,1)
fprintf(' %i %s\n',j,lablist(j,:));
end
end
clear w;
end
else
error('input should be mapping or dataset');
end
return