%GENDAT Random sampling of datasets for training and testing
%
% [A,B,IA,IB] = GENDAT(X,N,SEED)
% [A,B,IA,IB] = X*GENDAT([],N,SEED)
% [A,B,IA,IB] = X*GENDAT(N,SEED)
% [A,B,IA,IB] = GENDAT(X,ALF,SEED)
% [A,B,IA,IB] = X*GENDAT([],ALF,SEED)
% [A,B,IA,IB] = X*GENDAT(ALF,SEED)
%
% INPUT
% X Dataset.
% N,ALF Number/fraction of objects to be selected (def: bootstrapping).
% Alternatively a vector of numbers of objects for each class.
% SEED A state of the random number generation according to RANDRESET
%
% OUTPUT
% A,B Datasets
% IA,IB Original indices from the dataset X
%
% DESCRIPTION
% Generation of N objects from dataset X. They are stored in dataset A,
% the remaining objects in dataset B. IA and IB are the indices of the
% objects selected from X for A and B. The random object generation follows
% the class prior probabilities. So is the prior probability of a class is
% PA, then in expectation PA*N objects are selected from that class. If N
% is large or if one of the classes has too few objects in A, the number of
% generated objects might be less than N.
%
% If N is a vector of sizes, exactly N(i) objects are generated for class i.
% Classes are ordered as given by GETLABLIST(A).
%
% If the function is called without specifying N, the data set X is
% bootstrapped and stored in A. Not selected samples are stored in B.
%
% ALF should be a scalar < 1. For each class a fraction ALF of the objects
% is selected for A and the not selected objects are stored in B.
%
% If X is a cell array of datasets the command is executed for each
% dataset separately. Results are stored in cell arrays. For each dataset
% the random seed is reset, resulting in aligned sets for the generated
% datasets if the sets in X were aligned.
%
% EXAMPLES
% See PREX_PLOTC.
%
% SEE ALSO (<a href="http://37steps.com/prtools">PRTools Guide</a>)
% DATASETS, MAPPINGS, GENSUBSETS, RANDRESET
% Copyright: R.P.W. Duin, r.p.w.duin@37steps.com
function [A,B,IA,IB] = gendat(varargin)
argin = shiftargin(varargin,'vector');
argin = setdefaults(argin,[],[],[]);
if mapping_task(argin,'definition')
A = define_mapping(argin,'generator','Data sampling');
return
end
% execution
[X,N,seed] = deal(argin{:});
if isempty(seed)
seed = randreset;
end
randreset(seed);
% If the input is a cell array of datasets, apply this procedure
% to the individual datasets.
if (iscell(X))
[A,B,IA,IB] = X*feval(mfilename,[],N,seed);
return
end
% When required, get the right number of objects from the given
% fraction ALF.
if isdouble(X), Xdouble = true; else, Xdouble = false; end
if ~isdatafile(X), X = prdataset(X); end
X = setlablist(X); % remove empty classes first
[m,k,c] = getsize(X);
% we need at least one class below:
unlabeled = 0;
if c==0,
X=cdats(X,1);
c=1;
unlabeled = 1; % we need to correct for labeling at the end
end
R = classsizes(X);
if ~isempty(N) && length(N) ~= 1 && length(N) ~= c
error('Data size should be scalar or a vector matching the number of classes')
end
if ~islabtype(X,'crisp')
if numel(N) > 1
prwarning(1,'Specification of numbers of objects per class not possible for given label type')
N = sum(N);
end
if N < 1, N = ceil(N*m); end
end
if ~isempty(N) && all(N < 1) && islabtype(X,'crisp')
%DXD it should also be possible to have a fraction for each of the
%classes, I think...
if length(N)==1
N = ceil(N*R);
else
N = ceil(N(:).*R(:));
end
end
% Depending if N (or ALF) is given, the objects are created using
% subsampling or bootstrapping.
IA = [];
if (nargin < 2) || (isempty(N)) % Bootstrap
for i=1:c
J = findnlab(X,i);
K = ceil(rand(R(i),1)*R(i));
IA = [IA; J(K)];
end
else % Subsampling
if ~islabtype(X,'crisp')
K = randperm(m);
if (N > m)
%DXD: I would like to have just a warning:
%error('More objects requested than available.')
prwarning(4,'More objects requested than available.')
%N = m;
K = repmat(K,1,ceil(N/m));
end
IA = K(1:N);
else
%p = X.prior; % avoid warning
if isempty(X,'prior')
p = classsizes(X);
p = p/sum(p);
else
p = getprior(X);
end
%p = getprior(X);
N = genclass(N,p);
for i=1:c
J = findnlab(X,i);
K = randperm(R(i));
if (N(i) > R(i))
%DXD: I would like to have just a warning:
%error('More objects requested than available.')
prwarning(1,'More objects requested than available in class %d.',i)
%N(i) = R(i);
end
K = repmat(K,1,ceil(N(i)/R(i)));
IA = [IA; J(K(1:N(i)))];
end
end
end
% Finally, extract the datasets:
IB = [1:m]';
IB(IA) = [];
if unlabeled
X = setlabels(X,[]); % reset unlabeling
end
A = X(IA,:);
B = X(IB,:);
if Xdouble
A = +A; B = +B;
end
return;