addpath('./helper/');
workdir = './matfiles/';
heat_savedir = [workdir 'heat_kernels_q/'];


% --- Prep For CV -------------------------------------------------------------
%maxNumCachedPerClass = 4000;
%numFolds             = 20;
%numTrainPerClass     = 250;
%numTestPerClass      = 100;
%c1labels = {'M11\d*','M12\d*','M13\d*','M14\d*'};
%load([workdir 'topics.mat']);
%load([workdir 'member.7k.mat']);
%
%% class1 will have label '1' while class2 will have label '-1'
%[ids trn_samps tst_samps trn_zs tst_zs] = prepOneVsRest( c1labels, topics, member, ...
%		maxNumCachedPerClass, numFolds, numTrainPerClass, numTestPerClass );
%
%clear topics member;
%% save all variables here
%save([workdir 'OneVsRest.7k.mat']); 
load([workdir 'OneVsRest.7k.mat']); 


% --- Initialization ----------------------------------------------------------
load([workdir 'tf.7k.mat']);
tf = tf(:,ids);	% prune unused documents (cols)

%% remove unused words (rows)
%% this is not necessary since it would only affect the transD2 computation,
%%   which never divides by the doc-freq of words
%inactivewords = find(~any(tf,2));
%tf(inactivewords,:) = [];   % (rows)

% jvd: comment out the next four lines to disable feature selection
cut = 2000;
[df I] = sort(sum(tf,2));
Jsub = I((end-cut+1):end);	% jvd: i.e., the words with highest doc freq
Jind = sub2ind([size(tf,1) size(tf,1)],Jsub,Jsub);


filelist = dir(heat_savedir);
ptrn = 'Hc(\d+)t(\d+)';


% --- Loop through paramters to test translation kernel -----------------------
fprintf('CV Linear Kernel under translation\n');
ii = 0;
for fln = {filelist.name}
	file = char(fln);
	match = regexp(file,ptrn,'tokens');
	% jvd: by re-checking file existence, we can control execution by removing,
	%      but not adding kernels
	if numel(match)==0 | ~exist([heat_savedir file],'file')
		continue;
	end
	ii = ii+1;
	fprintf('%s\n',file);

	c = str2num(match{1}{1})/100;
	t = str2num(match{1}{2})/100;
	load([heat_savedir file]);
	params(:,ii) = [c;t];	% consolidating

	% --- Optional Feature selection (on T) -----------------------------------
	% H(inactivewords,inactivewords) = [];	% jvd: not needed, see above

	% jvd: do not translate the top 2000 words set all rows and cols of words
	%         that occur too frequently to zero, then fixup the diag those rows
	%         to be one.  in this way the most occuring words are not
	%         translated.
	if exist('Jsub','var') & exist('Jind','var') & exist('cut','var')
		H(Jsub,:)=0;H(:,Jsub)=0;H(Jind)=1;
	end
	
	% jvd: make sure normalize last!
	T = normalizeRows(H);

	% jvd: the linear kernel is simply:
	B   = tf' * T*T' * tf;

	% --- Translated Linear Kernel --------------------------------------------
	TAcc = svmCVOneVsRest( trn_samps, tst_samps, trn_zs, tst_zs, tf, ...
				numFolds, '', 'tranlin', B);
	tacc(:,ii) = TAcc;
end


% --- Standard Linear Kernel --------------------------------------------------
fprintf('CV Linear Kernel\n');
acc = svmCVOneVsRest( trn_samps, tst_samps, trn_zs, tst_zs, tf, ...
			numFolds, '', 'stndlin');

if exist('Jsub','var') & exist('Jind','var') & exist('cut','var')
	clear_except tacc acc params Jsub Jind workdir;
else
	clear_except tacc acc params workdir;
end
save([workdir 'acc_svm_lin_feat_q.7k.mat']);