% ensure everything ends in slash!
addpath('./helper/');
workdir = './matfiles/';
heat_savedir = [workdir 'heat_kernels_q/'];
exp_savedir = [workdir 'experiments/svm/allnew/'];

C1  = {{'M11'}};	% no regex here
C2  = {{'.*'}};
%C1  = {{'C11'} {'E11'} {'G154'} {'M11'}};
%C2  = {{'.*'}  {'.*'}  {'.*'}   {'.*'}};
%C2  = {{'C.*'} {'E.*'} {'G.*'}  {'M.*'}};



% --- Build CV Pool -----------------------------------------------------------
load([workdir 'topics.mat']);
load([workdir 'member.7k.mat']);
maxNumPerClass   = 2500;
numFolds         = 40;
numTrainPerClass = 150;
numTestPerClass  = 300;

if ~exist(exp_savedir,'dir'), mkdir(exp_savedir); end
for ii=1:length(C1)
	[ids1 ids2 topids1 topids2] = formBinaryPool( ...
						C1{ii}, C2{ii}, topics, member, maxNumPerClass);
	c1labels = topics(topids1);
	c1labels = [c1labels;repmat({'_'},1,length(c1labels)-1) {''}];
	c1labels = cell2mat(reshape(c1labels,1,numel(c1labels)));

	save([exp_savedir c1labels 'vsAll.pool.mat'], ...
			'ids1', 'ids2', 'topids1', 'topids2',...
			'maxNumPerClass','numFolds','numTrainPerClass','numTestPerClass');
end


% --- Perform experiments -----------------------------------------------------
load([workdir 'tf.7k.mat']);
load([workdir 'doclen.7k.mat']);

cc=[0.50 0.25];
tt=[1.28 1.44 1.60 1.76];
trbfgamma = [1 2 3 4];
rbfgamma = [.5 1 2 3 4];
cut = 2000;

K_file = mktemp(['/noback/lowbow/translation/tmp/K.XXXXXX.mtx']);
svmflags = [' -t 4 -u "' K_file '" '];

% --- Loop through paramters to test translation kernel -----------------------
filelist = dir(exp_savedir);
ptrn = '(.*)vs(.*)\.pool\.mat';
fprintf('CV RBF Kernel under translation experiment.\n');
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([exp_savedir file],'file')
        continue;
    end
	c1labels = char(match{1}(1));
	c2labels = char(match{1}(2));
	
	load([exp_savedir file]);
	ids = [ids1 ids2];

	fprintf('%s vs. %s\n',c1labels,c2labels);
	
	% --- Standard RBF kernel -------------------------------------------------
	fprintf('\tCV standard RBF kernel (with cut=%d):\n',cut);
	acc = zeros(length(rbfgamma),numFolds);
	for ii=1:length(rbfgamma)
		fprintf('\t\trbfgamma = %f\n',rbfgamma(ii));
		svmflags = [' -t 2 -g ' num2str(rbfgamma(ii)) ' '];

		% --- Cross-Validation ------------------------------------------------
		for ff=1:numFolds
			[trn tst]=sampleFromBinaryPool(maxNumPerClass,numTrainPerClass,numTestPerClass);
			trn_vects = tf(:,ids(trn.samps));
			tst_vects = tf(:,ids(tst.samps));
			acc(ii,ff) = svmWrapper( trn_vects, trn.labels, trn.ids,...
							tst_vects, tst.labels, tst.ids, svmflags, 0 );
		end

		avgacc(ii) = mean(acc(ii,:))/(2*numTestPerClass);
		stdacc(ii) = std(acc(ii,:))/(2*numTestPerClass);
		fprintf('\t\t\tmean: %f stddev: %f\n', avgacc(ii), stdacc(ii));
	end


	% --- Translated RBF kernel -------------------------------------------------
	fprintf('\tCV translated RBF kernel (with cut=%d):\n',cut);
	% prepare for feature selection
	if cut>0
		[df I] = sort(full(sum(spones(tf),2)));
		Jsub = I((end-cut+1):end); % jvd: i.e., the words with highest doc freq get removed
		Jind = sub2ind([size(tf,1) size(tf,1)],Jsub,Jsub);
	end


	tacc = zeros(length(cc),length(tt),length(trbfgamma),numFolds);
	for ii=1:length(cc)
		c = cc(ii);
		for jj=1:length(tt)
			t = tt(jj);
			H_file = [heat_savedir 'Hc' num2str(100*c) 't' num2str(100*t) '.7k.mat'];
			fprintf('\t\t%s\n',H_file);
			load(H_file);

			
			% --- Optional Feature selection (on T) ---------------------------
			if cut>0
				H(Jsub,:)=0;H(:,Jsub)=0;H(Jind)=1;
			end
			T = normalizeRows(H);	% jvd: make sure normalize last!

			
			% --- Prepare for kernel ------------------------------------------
			[D Dnull] = transD2tf(tf(:,ids),doclen(ids),T);		% rbf kernel, D
			%D=tf'*T*T'*tf; Dnull = sparse(size(D,1),1);		% linear kernel, D
			Dcache = [[0 Dnull'];[Dnull D]];


			% --- RBF Kernel --------------------------------------------------
			for kk=1:length(trbfgamma)
				fprintf('\t\t\ttrbfgamma = %f\n',trbfgamma(kk));
				
				% --- Cross-Validation ----------------------------------------
				for ff=1:numFolds
					[trn tst] = sampleFromBinaryPool(maxNumPerClass,numTrainPerClass,numTestPerClass);
					% necessary to key into our augmented kernel
					I = [0 trn.samps tst.samps]+1;
					K = exp(-trbfgamma(ii) * Dcache(I,I));	% rbf kernel, K
					%K = Dcache(I,I);						% linear kernel, K
					jvdmatrixWrite( K_file, K );
					trn_vects = tf(:,ids(trn.samps));
					tst_vects = tf(:,ids(tst.samps));
					tacc(ii,jj,kk,ff) = svmWrapper( trn_vects, trn.labels, trn.ids,...
										tst_vects, tst.labels, tst.ids, svmflags, 0 );
				end
				
				avgtacc(ii,jj,kk) = mean(tacc(ii,jj,kk,:))/(2*numTestPerClass);
				stdtacc(ii,jj,kk) = std(tacc(ii,jj,kk,:))/(2*numTestPerClass);
				fprintf('\t\t\t\tmean: %f stddev: %f\n', avgtacc(ii,jj,kk), stdtacc(ii,jj,kk));
			end
			
			% checkpoint save!
			save([exp_savedir c1labels 'vsAll' num2str(cut)/1000 'k.svmrbf.mat'],...
					'cut','cc','tt','trbfgamma','tacc','rbfgamma','acc',...
					'avgtacc','stdtacc','avgacc','stdacc');
		end
	end

	save([exp_savedir c1labels 'vsAll' num2str(cut)/1000 'k.svmrbf.mat'],...
			'cut','cc','tt','trbfgamma','tacc','rbfgamma','acc',...
			'avgtacc','stdtacc','avgacc','stdacc');
end


if exist(K_file,'file')
	system(['rm ' K_file]);
end