gethmmlikeforz.m

来自「Continuous Profile Models (CPM) Matlab T」· M 代码 · 共 48 行

function [hmmLikeForZU uMat]= getHMMlikeForZ(z,G,scalesExpRep,...
    gammaSum5,gammaSum6,u,u2,sig2inv,binNum);

% helper function to calculate the HMM portion of the expected complete LOG
% likelihood.  two terms drop out 1) normalization constant of the
% gaussian, and 2) the x-related term in the quadratic expansion
%
% used only for z updates, as u updates has a more efficient version
%
% note too it is only doing it for 1 bin, (the input params should only be
% given for this bin
%
% called by z2Function, and uFunction
%
% sig2inv = G.sigmas(binNum,:).^(-2);
%
% if want it for kk only, then specify which kk

%if ~exist('kk')
%    kk = 1:G.numSamples;
%end

uMat=0;

zRep = zeros(G.numSamples,G.numStates);
for cc=1:G.numClass
  classInd = G.class{cc};
  numInClass = length(classInd);
  zTauS = z(G.stateToScaleTau(:,2),cc);
  zRep(classInd,:)=repmat(zTauS,[1 numInClass])';
end
zPhi = scalesExpRep.*zRep;
%zPhi = squeeze(zPhi(kk,:));

if G.USE_CPM2         
    %% remember, this is just for one bin
    term1 = sum(sig2inv.*sum(u.*zPhi.*gammaSum6,2)');
    term2 = sum(sig2inv.*sum(u2.*zPhi.^2.*gammaSum5,2)')/2;
else   
    term1 = sum(u.*sig2inv.*sum(zPhi.*gammaSum6,2)');
    term2 = sum(u2.*sig2inv.*sum(zPhi.^2.*gammaSum5,2)')/2;
end

hmmLikeForZU = term1 - term2;

%keyboard;

return;