function [ids trn_samps tst_samps trn_zs tst_zs] = prepOneVsRest( ...
				c1labels, topics, member, ...
				maxNumCachedPerClass, numFolds, numTrainPerClass, numTestPerClass )
%PREPONEVSREST   Prepares a sampling experiment.
%   Prepares a cross-validation experiment for RCV1 corpus as we've formatted
%   it.
%
%   Arguments:
%	   c1labels           regex of things to search in topics (cell)
%	   topics             list of possible topics (cell)
%	   member             binary membership matrix (length(topics) rows by
%	                          number of docs)
%      maxNumCachedPerClass   max num of docs in pool from which we sample
%      numFolds           number of times to sample from pool
%      numTrainPerClass   number of training per class per fold
%      numTestPerClass    number of testing per class per fold
%
%   Returns:
%      ids         corpus docs used in sample pool
%      trn_samps   randperms used to index into pool
%      tst_samps   randperms used to index into pool
%      trn_zs      corresponding (training) labels
%      tst_zs      corresponding (testing) labels
%
%   Authors:
%       Joshua V. Dillon; Feb. 21, 2007; jvdillon AAT purdue DDOT edu


% find the documents that belong to c1labels
topic_ids = find(regexplist(c1labels,topics));
% if any doc has 1 or more c1labels, z=1, else z=0
z = any(member(topic_ids,:),1);	

% determine class representaion 
sumz = full(sum(z));	% how many c1 docs are there?
N = min([sumz,length(z)-sumz,maxNumCachedPerClass]);
% (doing this in case future version supports asymmetric validation)
N1=N;
N2=N;

% randomly find N1 class1 indecies (cols in the corpus)
I   = randperm(sumz);
z1  = full(find(z));
ids1 = z1( I(1:N1) );

% randomly find N2 class2 indecies (cols in the corpus)
I   = randperm(length(z)-sumz);
z2  = full(find(~z));
ids2 = z2( I(1:N2) );
% ids2: offset not req'd since id1 is mutex with id2 (see comment below)

ids = [ids1 ids2];	% ids1 is mutex with ids2

% set CV paramters based on function input
%(doing this in case future version supports asymmetric validation)
N1trn = numTrainPerClass;  N1tst = numTestPerClass;
N2trn = numTrainPerClass;  N2tst = numTestPerClass;

% following are pre-grown
trn_samps = zeros(N1trn+N2trn,numFolds);
trn_zs    = zeros(N1trn+N2trn,numFolds);
tst_samps = zeros(N1tst+N2tst,numFolds);
tst_zs    = zeros(N1tst+N2tst,numFolds);

% for each "fold" form some sample indecies (which will index into the
% length(ids) number of corpus indecies
for ii=1:numFolds
	I = randperm(N1);
	trn1 = I(1:N1trn);
	tst1 = I((1+N1trn):(N1tst+N1trn));	% tst1 is disjoint from trn1

	I = randperm(N2);	
	trn2 = I(1:N2trn);
	tst2 = I((1+N2trn):(N2tst+N2trn));	% tst2 is disjoint from trn2

	% offsetting since cached kernel is constructed from I=[id1 id2].  this
	% will "key" into the cached matrix to extract a training kernel and a
	% testing kernel
	trn_samps(:,ii) = [trn1 trn2+N1]';	
	tst_samps(:,ii) = [tst1 tst2+N1]';

	% labels are unpermuted
	trn_zs(:,ii) = [repmat(1,N1trn,1) ; repmat(-1,N2trn,1)];
	tst_zs(:,ii) = [repmat(1,N1tst,1) ; repmat(-1,N2tst,1)];
end