fix hmmViterbi llh

chapter13/HMM/hmmViterbi.m

@@ -1,11 +1,12 @@
-function z = hmmViterbi(x, model)
+function [z, llh] = hmmViterbi(x, model)
 % Viterbi algorithm calculated in log scale to improve numerical stability.
 % This is a wrapper function which transform input and call underlying algorithm
 % Input:
 %   x: 1 x n integer vector which is the sequence of observations
 %   model:  model structure
 % Output:
 %   z: 1 x n latent state
+%   llh:  loglikelihood
 % Written by Mo Chen ([email protected]).
 A = model.A;
 E = model.E;
@@ -15,4 +16,4 @@
 d = max(x);
 X = sparse(x,1:n,1,d,n);
 M = E*X;
-z = hmmViterbi_(M, A, s);
+[z,llh] = hmmViterbi_(M, A, s);

chapter13/HMM/hmmViterbi_.m

@@ -1,12 +1,12 @@
-function [z, p] = hmmViterbi_(M, A, s)
+function [z, llh] = hmmViterbi_(M, A, s)
 % Implmentation function of Viterbi algorithm. 
 % Input:
 %   M: k x n emmision data matrix M=E*X
 %   A: k x k transition matrix
 %   s: k x 1 starting probability (prior)
 % Output:
 %   z: 1 x n latent state
-%   p: 1 x n probability
+%   llh:  loglikelihood
 % Written by Mo Chen ([email protected]).
 [k,n] = size(M);
 Z = zeros(k,n);
@@ -20,6 +20,6 @@
     Z = Z(idx,:);
     Z(:,t) = 1:k;
 end
-[v,idx] = max(v);
+[llh,idx] = max(v);
 z = Z(idx,:);
-p = exp(v);
+

demo/ch13/hmm_demo.m

@@ -5,7 +5,7 @@
 n = 10000;
 [x, model] = hmmRnd(d, k, n);
 %%
-z = hmmViterbi(x,model);
+[z,p] = hmmViterbi(x,model);
 %%
 [alpha,llh] = hmmFilter(x,model);
 %%