%PARALLEL Combining classifiers in different feature spaces
%
% WC = PARALLEL(W1,W2,W3, ....) or WC = [W1;W2;W3; ...]
% WC = PARALLEL({W1;W2;W3; ...}) or WC = [{W1;W2;W3; ...}]
% WC = PARALLEL(WC,W1,W2, ....) or WC = [WC;W2;W3; ...]
% WC = PARALELL(C);
% WC = PARALLEL(WC,N);
%
% INPUT
% W1,W2,... Base classifiers to be combined.
% WC Parallel combined classifier
% C Cell array of classifiers
% N Integer array
%
% OUTPUT
% WC Combined classifier.
%
% DESCRIPTION
% The base classifiers (or mappings) W1, W2, W3, ... defined in different
% feature spaces are combined in WC. This is a classifier defined for the
% total number of features and with the combined set of outputs. So, for three
% two-class classifiers defined for the classes 'c1' and 'c2', a dataset A is
% mapped by D = A*WC on the outputs 'c1','c2','c1','c2','c1','c2' which are
% the feature labels of D. Note that classification by LABELD(D) finds for
% each vector in D the feature label of the column with the maximum value.
% This is equivalent to using the maximum combiner MAXC.
%
% Other fixed combining rules like PRODC, MEANC, and VOTEC can be applied
% by D = A*WC*PRODC etc. A trained combiner like FISHERC has to be supplied
% with the appropriate training set by AC = A*WC; VC = AC*FISHERC. So the
% expression VC = A*WC*FISHERC yields a classifier and not a dataset as with
% fixed combining rules. This classifier operates in the intermediate feature
% space, the output space of the set of base classifiers. A new dataset B has
% to be mapped to this intermediate space first by BC = B*WC before it can be
% classified by D = BC*VC. As this is equivalent to D = B*WC*VC, the total
% trained combiner is WTC = WC*VC = WC*A*WC*FISHERC. To simplify this procedure
% PRTools executes the training of a combined classifier by
% WTC = A*(WC*FISHERC) as WTC = WC*A*WC*FISHERC.
%
% In order to allow for training an untrained parallel combined classifier by
% A*WC the subsets of the features of A that apply for the individual base
% classifiers of WC should be known to WC. Subset sizes are stored in the
% dataset description if constructed by horizontal concatenation:
% A = [A1 A2 A3}; This can be overruled by storing them in the classifier
% by the call WC = PARALLEL(WC,N), in which N is an array of integers, such
% that sum(N) equals the feature size of A.
%
% SEE ALSO (<a href="http://37steps.com/prtools">PRTools Guide</a>)
% MAPPINGS, DATASETS, MAXC, MINC, MEANC, MEDIANC, PRODC, FISHERC, STACKED
%
% EXAMPLES
% See PREX_COMBINING.
% Copyright: R.P.W. Duin, duin@ph.tn.tudelft.nl
% Faculty of Applied Sciences, Delft University of Technology
% P.O. Box 5046, 2600 GA Delft, The Netherlands
% $Id: parallel.m,v 1.4 2009/03/18 16:17:59 duin Exp $
function out = parallel (varargin)
% If there are no arguments, just return an empty map.
if (nargin == 0)
out = prmapping(mfilename,'combiner');
return
end
% If there is one argument, assume it is a mapping or cell array of these.
if (nargin == 1)
v = varargin{1};
if (~iscell(v))
ismapping(v); % Assert that V is a single mapping.
out = prmapping('parallel',getmapping_type(v),{v},getlabels(v));
out = set(out,'size',getsize(v));
else
% If V is a cell array of mappings, unpack it and call this function
% for each cell.
out = feval(mfilename,v{:});
end
return
end
if ( nargin == 2 & ( isparallel(varargin{1}) | iscell(varargin{1}) ) ...
& ~ismapping(varargin{2}) )
% special case, store dataset sizes
if iscell(varargin{1})
w = parallel(varargin{1});
else
w = varargin{1};
end
n = varargin{2};
if length(w.data) ~= length(n)
error('Wrong number of classsizes')
end
w.data = [w.data {n}];
out = w;
% If there are multiple arguments and the first is not a dataset,
% combine the supplied mappings.
% elseif ((nargin > 2) | ~isa(varargin{1},'prdataset'))
elseif (~(isa(varargin{1},'prdataset')))
v1 = varargin{1};
if (isempty(v1) & ~ismapping(v1))
start = 3; v1 = varargin{2};
else
start = 2;
end
ismapping(v1); % Assert that V1 is a mapping.
[k,n] = size(v1); % Extract V1's characteristics.
labels = getlabels(v1);
type = getmapping_type(v1);
if (~strcmp(getmapping_file(v1),mfilename))
% If V1 is not already a parallel combiner, make it into a cell array
% of mappings.
v = {v1};
else
% V1 is already a parallel combiner: get the mapping data.
v = getdata(v1);
end
% Now concatenate all base classifiers as cells in V.
for j = start:nargin
v2 = varargin{j};
if j==nargin & ~ismapping(v2)
v = [v {v2}];
else
if (~strcmp(type,getmapping_type(v2)))
error('mappings should be of equal type')
end
v = [v {v2}];
k = k + size(v2,1);
n = n + size(v2,2);
labels = [labels; getlabels(v2)];
end
end
if length(v) == 1
out = v{1}; % just one mapping left: return it
else
out = prmapping('parallel',type,v,labels,k,n); % Construct the combined mapping.
end
else
% Execution: dataset * parallel_mapping.
a = varargin{1};
if nargin > 2 & ~ismapping(varargin{end})
v = parallel(varargin(2:end-1),varargin{end});
elseif nargin > 2
v = parallel(varargin(2:end));
else
v = varargin{2};
end
ismapping(v); % Assert that V is a mapping.
out = []; n = 0;
if isuntrained(v)
% get feature sizes of datasets to be supplied to the classifiers
s = v.data{end};
if ismapping(s)
s = getuser(a,'featsets');
if isempty(s)
error(['Training of parallel combined untrained classifier not possible.' ...
newline 'Feature sizes should be stored in the classifier first.'])
end
else
v.data = v.data(1:end-1);
end
if sum(s) ~= size(a,2)
error('Feature size of dataset does not match with classifier')
end
r = length(s);
if length(v.data) ~= r
error('Number of classifiers does not match number of feature subsets')
end
t = sprintf('Training %i base classifiers: ',r);
prwaitbar(r,t)
for j=1:r
prwaitbar(r,j,[t getname(v{j})]);
N = [n+1:n+s(j)]; % to features indexed by N.
n = n + s(j);
w = a(:,N)*v{j};
out = [out; w];
end
prwaitbar(0)
else
for j = 1:length(v.data)
sz = size(v{j},1); % Classifier V{J} is applied
N = [n+1:n+sz]; % to features indexed by N.
b = a(:,N)*v{j};
if (size(v{j},2) == 1) % Restore 2D outputs for k->1 classifiers.
b = b(:,1);
end
n = n + sz;
out = [out b]; % Concatenate mapped data.
end
end
end
return