|
wolffd@0
|
1 % Make an HMM with Gaussian observations
|
|
wolffd@0
|
2 % X1 -> X2
|
|
wolffd@0
|
3 % | |
|
|
wolffd@0
|
4 % v v
|
|
wolffd@0
|
5 % Y1 Y2
|
|
wolffd@0
|
6
|
|
wolffd@0
|
7 intra = zeros(2);
|
|
wolffd@0
|
8 intra(1,2) = 1;
|
|
wolffd@0
|
9 inter = zeros(2);
|
|
wolffd@0
|
10 inter(1,1) = 1;
|
|
wolffd@0
|
11 n = 2;
|
|
wolffd@0
|
12
|
|
wolffd@0
|
13 Q = 2; % num hidden states
|
|
wolffd@0
|
14 O = 2; % size of observed vector
|
|
wolffd@0
|
15 ns = [Q O];
|
|
wolffd@0
|
16 bnet = mk_dbn(intra, inter, ns, 'discrete', 1, 'observed', 2);
|
|
wolffd@0
|
17
|
|
wolffd@0
|
18 prior0 = normalise(rand(Q,1));
|
|
wolffd@0
|
19 transmat0 = mk_stochastic(rand(Q,Q));
|
|
wolffd@0
|
20 mu0 = rand(O,Q);
|
|
wolffd@0
|
21 Sigma0 = repmat(eye(O), [1 1 Q]);
|
|
wolffd@0
|
22 bnet.CPD{1} = tabular_CPD(bnet, 1, prior0);
|
|
wolffd@0
|
23 %% we set the cov prior to 0 to give same results as HMM toolbox
|
|
wolffd@0
|
24 %bnet.CPD{2} = gaussian_CPD(bnet, 2, 'mean', mu0, 'cov', Sigma0, 'cov_prior_weight', 0);
|
|
wolffd@0
|
25 bnet.CPD{2} = gaussian_CPD(bnet, 2, 'mean', mu0, 'cov', Sigma0);
|
|
wolffd@0
|
26 bnet.CPD{3} = tabular_CPD(bnet, 3, transmat0);
|
|
wolffd@0
|
27
|
|
wolffd@0
|
28
|
|
wolffd@0
|
29 T = 5; % fixed length sequences
|
|
wolffd@0
|
30
|
|
wolffd@0
|
31 engine = {};
|
|
wolffd@0
|
32 engine{end+1} = smoother_engine(jtree_2TBN_inf_engine(bnet));
|
|
wolffd@0
|
33 engine{end+1} = smoother_engine(hmm_2TBN_inf_engine(bnet));
|
|
wolffd@0
|
34 engine{end+1} = hmm_inf_engine(bnet);
|
|
wolffd@0
|
35 engine{end+1} = jtree_unrolled_dbn_inf_engine(bnet, T);
|
|
wolffd@0
|
36 %engine{end+1} = frontier_inf_engine(bnet);
|
|
wolffd@0
|
37 engine{end+1} = bk_inf_engine(bnet, 'clusters', {[1]});
|
|
wolffd@0
|
38 engine{end+1} = jtree_dbn_inf_engine(bnet);
|
|
wolffd@0
|
39
|
|
wolffd@0
|
40
|
|
wolffd@0
|
41 inf_time = cmp_inference_dbn(bnet, engine, T);
|
|
wolffd@0
|
42
|
|
wolffd@0
|
43 ncases = 2;
|
|
wolffd@0
|
44 max_iter = 2;
|
|
wolffd@0
|
45 [learning_time, CPD, LL, cases] = cmp_learning_dbn(bnet, engine, T, 'ncases', ncases, 'max_iter', max_iter);
|
|
wolffd@0
|
46
|
|
wolffd@0
|
47 % Compare to HMM toolbox
|
|
wolffd@0
|
48
|
|
wolffd@0
|
49 data = zeros(O, T, ncases);
|
|
wolffd@0
|
50 for i=1:ncases
|
|
wolffd@0
|
51 data(:,:,i) = cell2num(cases{i}(bnet.observed, :));
|
|
wolffd@0
|
52 end
|
|
wolffd@0
|
53
|
|
wolffd@0
|
54 tic
|
|
wolffd@0
|
55 [LL2, prior2, transmat2, mu2, Sigma2] = mhmm_em(data, prior0, transmat0, mu0, Sigma0, [], 'max_iter', max_iter);
|
|
wolffd@0
|
56 t=toc;
|
|
wolffd@0
|
57 disp(['HMM toolbox took ' num2str(t) ' seconds '])
|
|
wolffd@0
|
58
|
|
wolffd@0
|
59 e = 1;
|
|
wolffd@0
|
60 assert(approxeq(prior2, CPD{e,1}.CPT))
|
|
wolffd@0
|
61 assert(approxeq(mu2, CPD{e,2}.mean))
|
|
wolffd@0
|
62 assert(approxeq(Sigma2, CPD{e,2}.cov))
|
|
wolffd@0
|
63 assert(approxeq(transmat2, CPD{e,3}.CPT))
|
|
wolffd@0
|
64 assert(approxeq(LL2, LL{e}))
|
