Mercurial > hg > camir-aes2014

annotate toolboxes/FullBNT-1.0.7/bnt/examples/dynamic/HHMM/Mgram/mgram1.m @ 0:e9a9cd732c1e tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
rev   line source
wolffd@0 1 % a multigram is a degenerate 2HHMM where the bottom level HMMs emit deterministic strings
wolffd@0 2 % and the the top level abstract states are independent of each other
wolffd@0 3 % cf. HSMM/test_mgram2
wolffd@0 4
wolffd@0 5 words = {'the', 't', 'h', 'e'};
wolffd@0 6 data = 'the';
wolffd@0 7 nwords = length(words);
wolffd@0 8 word_len = zeros(1, nwords);
wolffd@0 9 word_prob = normalise(ones(1,nwords));
wolffd@0 10 word_logprob = log(word_prob);
wolffd@0 11 for wi=1:nwords
wolffd@0 12 word_len(wi)=length(words{wi});
wolffd@0 13 end
wolffd@0 14 D = max(word_len);
wolffd@0 15
wolffd@0 16 alphasize = 26;
wolffd@0 17 data = letter2num(data);
wolffd@0 18 T = length(data);
wolffd@0 19
wolffd@0 20 % node numbers
wolffd@0 21 W = 1; % top level state = word id
wolffd@0 22 L = 2; % bottom level state = letter position within word
wolffd@0 23 F = 3;
wolffd@0 24 O = 4;
wolffd@0 25
wolffd@0 26 ss = 4;
wolffd@0 27 intra = zeros(ss,ss);
wolffd@0 28 intra(W,[F L O])=1;
wolffd@0 29 intra(L,[O F])=1;
wolffd@0 30
wolffd@0 31 inter = zeros(ss,ss);
wolffd@0 32 inter(W,W)=1;
wolffd@0 33 inter(L,L)=1;
wolffd@0 34 inter(F,[W L])=1;
wolffd@0 35
wolffd@0 36 % node sizes
wolffd@0 37 ns = zeros(1,ss);
wolffd@0 38 ns(W) = nwords;
wolffd@0 39 ns(L) = D;
wolffd@0 40 ns(F) = 2;
wolffd@0 41 ns(O) = alphasize;
wolffd@0 42
wolffd@0 43
wolffd@0 44 % Make the DBN
wolffd@0 45 bnet = mk_dbn(intra, inter, ns, 'observed', O);
wolffd@0 46 eclass = bnet.equiv_class;
wolffd@0 47
wolffd@0 48
wolffd@0 49
wolffd@0 50 % uniform start distrib over words, uniform trans mat
wolffd@0 51 Wstart = normalise(ones(1,nwords));
wolffd@0 52 Wtrans = mk_stochastic(ones(nwords,nwords));
wolffd@0 53
wolffd@0 54 % always start in state 1 for each bottom level HMM
wolffd@0 55 delta1_start = zeros(1, D);
wolffd@0 56 delta1_start(1) = 1;
wolffd@0 57 Lstart = repmat(delta1_start, nwords, 1);
wolffd@0 58 LRtrans = mk_leftright_transmat(D, 0); % 0 self loop prob
wolffd@0 59 Ltrans = repmat(LRtrans, [1 1 nwords]);
wolffd@0 60
wolffd@0 61 % Finish in the last letter of each word
wolffd@0 62 Fprob = zeros(nwords, D, 2);
wolffd@0 63 Fprob(:,:,1)=1;
wolffd@0 64 for i=1:nwords
wolffd@0 65 Fprob(i,length(words{i}),2)=1;
wolffd@0 66 Fprob(i,length(words{i}),1)=0;
wolffd@0 67 end
wolffd@0 68
wolffd@0 69 % Each state uniquely emits a letter
wolffd@0 70 Oprob = zeros(nwords, D, alphasize);
wolffd@0 71 for i=1:nwords
wolffd@0 72 for l=1:length(words{i})
wolffd@0 73 a = double(words{i}(l))-96;
wolffd@0 74 Oprob(i,l,a)=1;
wolffd@0 75 end
wolffd@0 76 end
wolffd@0 77
wolffd@0 78
wolffd@0 79 % Define CPDs for slice
wolffd@0 80 bnet.CPD{eclass(W,1)} = tabular_CPD(bnet, W, 'CPT', Wstart);
wolffd@0 81 bnet.CPD{eclass(L,1)} = tabular_CPD(bnet, L, 'CPT', Lstart);
wolffd@0 82 bnet.CPD{eclass(F,1)} = tabular_CPD(bnet, F, 'CPT', Fprob);
wolffd@0 83 bnet.CPD{eclass(O,1)} = tabular_CPD(bnet, O, 'CPT', Oprob);
wolffd@0 84
wolffd@0 85 % Define CPDs for slice 2
wolffd@0 86 bnet.CPD{eclass(W,2)} = hhmmQ_CPD(bnet, W+ss, 'Fbelow', F, 'startprob', Wstart, 'transprob', Wtrans);
wolffd@0 87 bnet.CPD{eclass(L,2)} = hhmmQ_CPD(bnet, L+ss, 'Fself', F, 'Qps', W+ss, 'startprob', Lstart, 'transprob', Ltrans);
wolffd@0 88
wolffd@0 89 evidence = cell(ss,T);
wolffd@0 90 evidence{W,1}=1;
wolffd@0 91 sample = cell2num(sample_dbn(bnet, 'length', T, 'evidence', evidence));
wolffd@0 92 str = lower(sample(4,:))
wolffd@0 93
wolffd@0 94 engine = jtree_dbn_inf_engine(bnet);
wolffd@0 95 evidence = cell(ss,T);
wolffd@0 96 evidence(O,:) = num2cell(data);
wolffd@0 97 [engine, ll_dbn] = enter_evidence(engine, evidence);
wolffd@0 98
wolffd@0 99 gamma = zeros(nwords, T);
wolffd@0 100 for t=1:T
wolffd@0 101 m = marginal_nodes(engine, [W F], t);
wolffd@0 102 gamma(:,t) = m.T(:,2);
wolffd@0 103 end
wolffd@0 104 gamma
wolffd@0 105
wolffd@0 106 xidbn = zeros(nwords, nwords);
wolffd@0 107 for t=1:T-1
wolffd@0 108 m = marginal_nodes(engine, [W F W+ss], t);
wolffd@0 109 xidbn = xidbn + squeeze(m.T(:,2,:));
wolffd@0 110 end
wolffd@0 111
wolffd@0 112 % thee
wolffd@0 113 % xidbn(1,4) = 0.9412 the->e
wolffd@0 114 % (2,3)=0.0588 t->h
wolffd@0 115 % (3,4)=0.0588 h-e
wolffd@0 116 % (4,4)=0.0588 e-e