% DCSC Dynamic Classifier Selection Combiner
%
% V = DCSC(A,W,K,TYPE)
% V = A*DCSC([],W,K,TYPE)
% V = A*(W*DCSC(K,TYPE))
% D = B*V
%
% INPUT
% A Dataset used for training base classifiers as well as combiner
% B Dataset used for testing (executing) the combiner
% W Set of trained or untrained base classifiers
% K Integer, number of neighbors
% TYPE 'soft' (default) or 'crisp'
%
% OUTPUT
% V Trained Dynamic Classifier Selection
% D Dataset with prob. products (over base classifiers) per class
%
% DESCRIPTION
% This dynamic classifier selection takes for every object to be
% classified the K nearest neighbors of an evaluation set (training set) A
% and determines which classifier performs best over this set of objects.
% If the base classifiers (STACKED or PARALLEL) are untrained, A is used to
% train them as well.
%
% The selection of the best classifier can be made in a soft or in a crisp
% way. If TYPE is 'soft' (default) classifier confidences are averaged (see
% CLASSC), otherwise the best classifier is selected by voting.
%
% REFERENCE
% G. Giacinto and F. Roli, Methods for Dynamic Classifier Selection
% 10th Int. Conf. on Image Anal. and Proc., Venice, Italy (1999), 659-664.
%
% SEE ALSO (<a href="http://37steps.com/prtools">PRTools Guide</a>)
% DATASETS, MAPPINGS, STACKED, NAIVEBC, CLASSC, TESTD, 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
function OUT = dcsc(varargin)
name = 'Dynamic Classifier Selection';
DefaultNumNeigh = 20;
DefaultType = 'soft';
argin = shiftargin(varargin,{'prmapping','scalar'});
argin = setdefaults(argin,[],[],[],[]);
[par1,par2,par3,par4] = deal(argin{:});
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Empty Call DCSC, DCSC([],K,TYPE)
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
if isempty(par1)
if isempty(par3), par3 = DefaultType; end
if isempty(par2), par2 = DefaultNumNeigh; end
% If there are no inputs, return an untrained mapping.
% (PRTools transfers the latter into the first)
OUT = prmapping(mfilename,'combiner',{par2,par3});
OUT = setname(OUT,name);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Storing Base Classifiers: W*DCSC, DCSC(W,K,TYPE)
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
elseif ismapping(par1)
if isempty(par3), par3 = DefaultType; end
if isempty(par2), par2 = DefaultNumNeigh; end
% call like OUT = DCSC(W,k) or OUT = W*DCSC([],k)
% store trained or untrained base classifiers,
% ready for later training of the combiner
BaseClassf = par1;
if ~isparallel(BaseClassf) & ~isstacked(BaseClassf)
error('Parallel or stacked set of base classifiers expected');
end
if ~(isa(par2,'double') & length(par2)==1 & isint(par2))
error('Number of neighbors should be integer scalar')
end
OUT = prmapping(mfilename,'untrained',{BaseClassf,par2,par3});
OUT = setname(OUT,name);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Training Combiner (and base classifiers if needed):
% A*(W*DCSC), A*(W*DCSC([],K,TYPE)), DCSC(A,W,K,TYPE)
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
elseif isdataset(par1) & ismapping(par2) & ...
(isuntrained(par2) | (istrained(par2) & (isstacked(par2) | isparallel(par2))))
% call like OUT = DCSC(TrainSet,W,k,type) or TrainSet*(W*DCSC([],k,type))
% (PRTools transfers the latter into the first)
% W is a set of trained or untrained set classifiers.
if isempty(par4), par4 = DefaultType; end
if isempty(par3), par3 = DefaultNumNeigh; end
TrainSet = par1;
BaseClassf = par2;
islabtype(TrainSet,'crisp'); % allow crisp labels only
isvaldfile(TrainSet,1,2); % at least one object per class, 2 classes
TrainSet = setprior(TrainSet,getprior(TrainSet,0)); % avoid many warnings
if ~(isstacked(BaseClassf) | isparallel(BaseClassf))
if iscombiner(BaseClassf)
error('No base classifiers found')
end
Data = getdata(BaseClassf);
BaseClassf = Data.BaseClassf; % base classifiers were already stored
end
if isuntrained(BaseClassf) % base classifiers untrained, so train them!
BaseClassf = TrainSet*BaseClassf;
n = length(getdata(BaseClassf));
else % base classifiers are trained, just check label lists
BaseClassifiers = getdata(BaseClassf);
n = length(BaseClassifiers);
for j=1:n
if ~isequal(getlabels(BaseClassifiers{j}),getlablist(TrainSet))
error('Training set and base classifiers should deal with same labels')
end
end
end
Data.BaseClassf = BaseClassf;
if ~isempty(par3) || ~isempty(par4) % overrules previously defined k (NumNeigh)
Data.NumNeigh = par3;
end
if ~isempty(par4) % overrules previously defined type
if strcmp(par4,'soft')
Data.SoftVote = 1;
else
Data.SoftVote = 0;
end
end
% Let us determine for every object in the trainingset how good it is
% classified by every base classifier
[m,p,c] = getsize(TrainSet);
%n = size(BaseClassf,2)/c; % nr. of base classifiers
ClassTrain1 = reshape(+(TrainSet*BaseClassf),m,c,n);
ClassTrain2 = zeros(m,n);
for j=1:n
ct = ClassTrain1(:,:,j)*normm;
ct = ct(sub2ind([m,c],[1:m]',[getnlab(TrainSet)]));
ClassTrain2(:,j) = ct;
end
if ~Data.SoftVote % find crisp outcomes
[cmax,L] = max(ClassTrain2,[],2);
ClassTrain2 = zeros(m,n);
ClassTrain2(sub2ind([m,n],[1:m]',L)) = ones(length(L),1);
end
Data.ClassTrain = ClassTrain2;
Data.TrainSet = TrainSet;
OUT = prmapping(mfilename,'trained',Data,getlablist(TrainSet),size(TrainSet,2),getsize(TrainSet,3));
OUT = setname(OUT,name);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Evaluation of the trained combiner V: B*V, DCSC(B,V,K)
% TYPE cannot be changed anymore
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
elseif isdataset(par1) & ismapping(par2) & istrained(par2)
if isempty(par3), par3 = DefaultNumNeigh; end
TestSet = par1;
BaseClassf = getdata(par2,'BaseClassf');
TrainSet = getdata(par2,'TrainSet');
[m,p,c] = getsize(TrainSet);
n = size(BaseClassf,2)/c; % nr. of base classifiers
if ~isempty(par3) || ~isempty(par4)
k = par3;
else
k = getdata(par2,'NumNeigh');
end
D = distm(TestSet,TrainSet); % disputable procedure for parallel base classifiers
[dd,J] = sort(+D,2); % J stores the numeric labels of the nearest training objects
ClassPerf = zeros(size(TestSet,1),n); % base classifier performances per testobject
TestSize = size(TestSet,1);
N = [1:TestSize];
kk = k;
ClassTrain = getdata(par2,'ClassTrain');
while ~isempty(N) & kk > 0
for j=1:n % find for every testobject and for all classifiers the mean
% performance (i.e. the mean of the correct class assignments) over
% its neighborhood
ClassPerf(:,j) = mean(reshape(ClassTrain(J(:,1:kk),j),TestSize,kk),2);
end
[cc,L(N,:)] = sort(-ClassPerf(N,:),2);
NN = find(cc(:,1) == cc(:,2)); % solve ties
N = N(NN); % select objects that suffer from ties
kk = kk-1; % try once more with smaller set of neighbors
end
U = getdata(BaseClassf);
d = zeros(TestSize,c);
feats = 0; % counter to find feature numbers for parallel classifiers
for j=1:n
featsize = size(U{j},1);
Lj = find(L(:,1)==j);
if ~isempty(Lj)
if isparallel(BaseClassf) % retrieve features for parallel classifiers
d(Lj,:) = +(TestSet(Lj,feats+1:feats+featsize)*U{j});
else
d(Lj,:) = +(TestSet(Lj,:)*U{j});
end
end
feats = feats+featsize;
end
OUT = setdat(TestSet,d,par2);
else
error('Illegal input');
end