computeInstanceWideKernel.m

function  [instanceStarts, instanceEnds, instanceWideKernel] = computeInstanceWideKernel(allSeqsAsBags, thisInstances, sparseComputation, debugLevel, debugMsgLocation)

% COMPUTEINSTANCEWIDEKERNEL 
% Computes the instance-wide kernel matrix given a bag of instances
%
% INPUT PARAMS
% Param 'allSeqsAsBags' (cell array)
% Collection of all segments per sequence
%
% Param 'thisInstances' (vector)
% #instances in each bag (in other words, #segments per sequence) 
%
% Param 'sparseComputation' (0/1)
% Flag specifying whether sparseComputation is to be performed (0/1)
%
% Param 'debugLevel' (0/1/2)
%
% Param 'debugMsgLocation' (1/fileID)
%
%
% OUTPUT PARAMS
% Param 'instanceStarts' (vector)
% Indexes in the vector where instances (segments) of any sequence begin
%
% Param 'instanceEnds' (vector)
% Indexes in the vector where istances (segments) of any sequence end
%
% Param 'instanceWideKernel' (matrix)
% The instance-wide kernel matrix of dimension #Instances x #Instances
%
% ADDITIONAL NOTES
%
% Author: snikumbh@mpi-inf.mpg.de

if nargin < 3
    debugMsgLocation = 1;
    debugLevel = 2;
    sparseComputation = 0;
end
logMessages(debugMsgLocation, sprintf('Onto compute the Std MI kernel\n'), debugLevel);
nBags = length(allSeqsAsBags);

tic;
instanceEnds = cumsum(thisInstances);
instanceStarts = cumsum(thisInstances) - thisInstances + 1;
nInstances = instanceEnds(end);
logMessages(debugMsgLocation, sprintf('instances starts and ends collected in %.4f seconds\n', toc), debugLevel);


if sparseComputation
    logMessages(debugMsgLocation, sprintf('Performing sparse computation for kernel on instances\n'), debugLevel);
    tic;
    allSeqsAsMat = zeros( size( allSeqsAsBags{1}, 1) , nInstances);
    for i=1:nBags
	allSeqsAsMat(:, instanceStarts(i): instanceEnds(i)) = allSeqsAsBags{i};
    end
    logMessages(debugMsgLocation, sprintf('Loop converting cells to matrix completed in %.4f seconds\n', toc), debugLevel);
    % Above is a looping version instead of cell2mat (see next line). This takes 
    % nearly half the time for really long sequences and too many instances.
    % tic; allSeqsAsMat = cell2mat(allSeqsAsBags); toc;
    tic;
    instanceWideKernel = full(allSeqsAsMat' * allSeqsAsMat);
    logMessages(debugMsgLocation, sprintf('full kernel computed in %.4f seconds\n', toc), debugLevel);
    % nBags x nBags kernel actually need not be computed
    clear allSeqsAsMat;
    %
else
    % direct (not sparseComputation)
    % This is tested neither for correctness nor efficiency. We recommend using sparseComputation.
    logMessages(debugMsgLocation, sprintf('Computing directly, %d rows.\n', nBags), debugLevel);
    stdMultiInstanceKernel = zeros(nBags);
    for i=1:nBags
        if mod(i, 10) == 0
            logMessages(debugMsgLocation, sprintf('%d, ', i), debugLevel);
        end
        iInstances = size(allSeqsAsBags{i}, 2);
        for j=i:nBags
            term = transpose(allSeqsAsBags{i}) * allSeqsAsBags{j};
	    jInstances = size(allSeqsAsBags{j}, 2);
            stdMultiInstanceKernel(i,j) = sum(term(:)) * (1/iInstances) * (1/jInstances);
        end
    end
    logMessages(debugMsgLocation, sprintf('\n'), debugLevel);
    % Fill-up the other half
    stdMultiInstanceKernel = triu(stdMultiInstanceKernel, 1) + stdMultiInstanceKernel';
    % Normalize the kernel matrix -- done by calling normalizeKernel
    stdMultiInstanceKernel = normalizeKernel(stdMultiInstanceKernel);
    %
    %
end %if-else ends

end % function ends