Mercurial > hg > camir-aes2014

annotate toolboxes/FullBNT-1.0.7/bnt/examples/dynamic/HHMM/Mgram/mgram3.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 % like mgram2, except we unroll the DBN so we can use smaller
wolffd@0 2 % state spaces for the early duration nodes:
wolffd@0 3 % the state spaces are D1 in {1}, D2 in {1,2}
wolffd@0 4
wolffd@0 5 past = 1;
wolffd@0 6
wolffd@0 7 words = {'the', 't', 'h', 'e'};
wolffd@0 8 data = 'the';
wolffd@0 9 nwords = length(words);
wolffd@0 10 word_len = zeros(1, nwords);
wolffd@0 11 word_prob = normalise(ones(1,nwords));
wolffd@0 12 word_logprob = log(word_prob);
wolffd@0 13 for wi=1:nwords
wolffd@0 14 word_len(wi)=length(words{wi});
wolffd@0 15 end
wolffd@0 16 D = max(word_len);
wolffd@0 17
wolffd@0 18
wolffd@0 19 alphasize = 26*2;
wolffd@0 20 data = letter2num(data);
wolffd@0 21 T = length(data);
wolffd@0 22
wolffd@0 23 % node numbers
wolffd@0 24 W = 1; % top level state = word id
wolffd@0 25 L = 2; % bottom level state = letter position within word
wolffd@0 26 F = 3;
wolffd@0 27 O = 4;
wolffd@0 28
wolffd@0 29 ss = 4;
wolffd@0 30 intra = zeros(ss,ss);
wolffd@0 31 intra(W,[F L O])=1;
wolffd@0 32 intra(L,[O F])=1;
wolffd@0 33
wolffd@0 34 inter = zeros(ss,ss);
wolffd@0 35 inter(W,W)=1;
wolffd@0 36 inter(L,L)=1;
wolffd@0 37 inter(F,[W L O])=1;
wolffd@0 38
wolffd@0 39 T = 3;
wolffd@0 40 dag = unroll_dbn_topology(intra, inter, T);
wolffd@0 41
wolffd@0 42 % node sizes
wolffd@0 43 ns = zeros(1,ss);
wolffd@0 44 ns(W) = nwords;
wolffd@0 45 ns(L) = D;
wolffd@0 46 ns(F) = 2;
wolffd@0 47 ns(O) = alphasize;
wolffd@0 48 ns = repmat(ns(:), [1 T]);
wolffd@0 49 for d=1:D
wolffd@0 50 ns(d,L)=d; % max duration
wolffd@0 51 end
wolffd@0 52 ns = ns(:);
wolffd@0 53
wolffd@0 54 % Equiv class in brackets for D=3
wolffd@0 55 % The Lt's are not tied until t>=D, since they have different sizes.
wolffd@0 56 % W1 and W2 are not tied since they have different parent sets.
wolffd@0 57
wolffd@0 58 % W1 (1) W2 (5) W3 (5) W4 (5)
wolffd@0 59 % L1 (2) L2 (6) L3 (7) L4 (7)
wolffd@0 60 % F1 (3) F2 (3) F3 (4) F3 (4)
wolffd@0 61 % O1 (4) O2 (4) O2 (4) O4 (4)
wolffd@0 62
wolffd@0 63 % Since we are not learning, we can dispense with tying
wolffd@0 64
wolffd@0 65 % Make the bnet
wolffd@0 66 Wnodes = unroll_set(W, ss, T);
wolffd@0 67 Lnodes = unroll_set(L, ss, T);
wolffd@0 68 Fnodes = unroll_set(F, ss, T);
wolffd@0 69 Onodes = unroll_set(O, ss, T);
wolffd@0 70
wolffd@0 71 bnet = mk_bnet(dag, ns);
wolffd@0 72 eclass = bnet.equiv_class;
wolffd@0 73
wolffd@0 74 % uniform start distrib over words, uniform trans mat
wolffd@0 75 Wstart = normalise(ones(1,nwords));
wolffd@0 76 Wtrans = mk_stochastic(ones(nwords,nwords));
wolffd@0 77 bnet.CPD{eclass(Wnodes(1))} = tabular_CPD(bnet, Wnodes(1), 'CPT', Wstart);
wolffd@0 78 for t=2:T
wolffd@0 79 bnet.CPD{eclass(Wnodes(t))} = hhmmQ_CPD(bnet, Wnodes(t), 'Fbelow', Fnodes(t-1), ...
wolffd@0 80 'startprob', Wstart, 'transprob', Wtrans);
wolffd@0 81 end
wolffd@0 82
wolffd@0 83 % always start in state d = length(word) for each bottom level HMM
wolffd@0 84 % and then count down
wolffd@0 85 % make downcounters
wolffd@0 86 RLtrans = mk_rightleft_transmat(D, 0); % 0 self loop prob
wolffd@0 87 Ltrans = repmat(RLtrans, [1 1 nwords]);
wolffd@0 88
wolffd@0 89 for t=1:T
wolffd@0 90 Lstart = zeros(nwords, min(t,D));
wolffd@0 91 for i=1:nwords
wolffd@0 92 l = length(words{i});
wolffd@0 93 Lstart(i,l)=1;
wolffd@0 94 if d==1
wolffd@0 95 bnet.CPD{eclass(Lnodes(1))} = tabular_CPD(bnet, Lnodes(1), 'CPT', Lstart);
wolffd@0 96 else
wolffd@0 97 bnet.CPD{eclass(Lnodes(t))} = hhmmQ_CPD(bnet, Lnodes(t), 'Fself', Fnodes(t-1), 'Qps', Wnodes(t), ...
wolffd@0 98 'startprob', Lstart, 'transprob', Ltrans);
wolffd@0 99 end
wolffd@0 100 end
wolffd@0 101 end
wolffd@0 102
wolffd@0 103
wolffd@0 104 % Finish when downcoutner = 1
wolffd@0 105 Fprob = zeros(nwords, D, 2);
wolffd@0 106 Fprob(:,1,2)=1;
wolffd@0 107 Fprob(:,2:end,1)=1;
wolffd@0 108
wolffd@0 109
wolffd@0 110 % Define CPDs for slice
wolffd@0 111 bnet.CPD{eclass(W,1)} = tabular_CPD(bnet, W, 'CPT', Wstart);
wolffd@0 112 bnet.CPD{eclass(L,1)} = tabular_CPD(bnet, L, 'CPT', Lstart);
wolffd@0 113 bnet.CPD{eclass(F,1)} = tabular_CPD(bnet, F, 'CPT', Fprob);
wolffd@0 114
wolffd@0 115
wolffd@0 116 % Define CPDs for slice 2
wolffd@0 117 bnet.CPD{eclass(W,2)} = hhmmQ_CPD(bnet, W+ss, 'Fbelow', F, 'startprob', Wstart, 'transprob', Wtrans);
wolffd@0 118 bnet.CPD{eclass(L,2)} = hhmmQ_CPD(bnet, L+ss, 'Fself', F, 'Qps', W+ss, 'startprob', Lstart, 'transprob', Ltrans);
wolffd@0 119
wolffd@0 120
wolffd@0 121 if 0
wolffd@0 122 % To test it is generating correctly, we create an artificial
wolffd@0 123 % observation process that capitalizes at the start of a new segment
wolffd@0 124 % Oprob(Ft-1,Qt,Dt,Yt)
wolffd@0 125 Oprob = zeros(2,nwords,D,alphasize);
wolffd@0 126 Oprob(1,1,3,letter2num('t'),1)=1;
wolffd@0 127 Oprob(1,1,2,letter2num('h'),1)=1;
wolffd@0 128 Oprob(1,1,1,letter2num('e'),1)=1;
wolffd@0 129 Oprob(2,1,3,letter2num('T'),1)=1;
wolffd@0 130 Oprob(2,1,2,letter2num('H'),1)=1;
wolffd@0 131 Oprob(2,1,1,letter2num('E'),1)=1;
wolffd@0 132 Oprob(1,2,1,letter2num('a'),1)=1;
wolffd@0 133 Oprob(2,2,1,letter2num('A'),1)=1;
wolffd@0 134 Oprob(1,3,1,letter2num('b'),1)=1;
wolffd@0 135 Oprob(2,3,1,letter2num('B'),1)=1;
wolffd@0 136 Oprob(1,4,1,letter2num('c'),1)=1;
wolffd@0 137 Oprob(2,4,1,letter2num('C'),1)=1;
wolffd@0 138
wolffd@0 139 % Oprob1(Qt,Dt,Yt)
wolffd@0 140 Oprob1 = zeros(nwords,D,alphasize);
wolffd@0 141 Oprob1(1,3,letter2num('t'),1)=1;
wolffd@0 142 Oprob1(1,2,letter2num('h'),1)=1;
wolffd@0 143 Oprob1(1,1,letter2num('e'),1)=1;
wolffd@0 144 Oprob1(2,1,letter2num('a'),1)=1;
wolffd@0 145 Oprob1(3,1,letter2num('b'),1)=1;
wolffd@0 146 Oprob1(4,1,letter2num('c'),1)=1;
wolffd@0 147
wolffd@0 148 bnet.CPD{eclass(O,2)} = tabular_CPD(bnet, O+ss, 'CPT', Oprob);
wolffd@0 149 bnet.CPD{eclass(O,1)} = tabular_CPD(bnet, O, 'CPT', Oprob1);
wolffd@0 150
wolffd@0 151 evidence = cell(ss,T);
wolffd@0 152 %evidence{W,1}=1;
wolffd@0 153 sample = cell2num(sample_dbn(bnet, 'length', T, 'evidence', evidence));
wolffd@0 154 str = num2letter(sample(4,:))
wolffd@0 155 end
wolffd@0 156
wolffd@0 157
wolffd@0 158
wolffd@0 159
wolffd@0 160 [log_obslik, obslik, match] = mk_mgram_obslik(lower(data), words, word_len, word_prob);
wolffd@0 161 % obslik(j,t,d)
wolffd@0 162 softCPDpot = cell(ss,T);
wolffd@0 163 ens = ns;
wolffd@0 164 ens(O)=1;
wolffd@0 165 ens2 = [ens ens];
wolffd@0 166 for t=2:T
wolffd@0 167 dom = [F W+ss L+ss O+ss];
wolffd@0 168 % tab(Ft-1, Q2, Dt)
wolffd@0 169 tab = ones(2, nwords, D);
wolffd@0 170 if past
wolffd@0 171 tab(1,:,:)=1; % if haven't finished previous word, likelihood is 1
wolffd@0 172 %tab(2,:,:) = squeeze(obslik(:,t,:)); % otherwise likelihood of this segment
wolffd@0 173 for d=1:min(t,D)
wolffd@0 174 tab(2,:,d) = squeeze(obslik(:,t,d));
wolffd@0 175 end
wolffd@0 176 else
wolffd@0 177 for d=1:max(1,min(D,T+1-t))
wolffd@0 178 tab(2,:,d) = squeeze(obslik(:,t+d-1,d));
wolffd@0 179 end
wolffd@0 180 end
wolffd@0 181 softCPDpot{O,t} = dpot(dom, ens2(dom), tab);
wolffd@0 182 end
wolffd@0 183 t = 1;
wolffd@0 184 dom = [W L O];
wolffd@0 185 % tab(Q2, Dt)
wolffd@0 186 tab = ones(nwords, D);
wolffd@0 187 if past
wolffd@0 188 %tab = squeeze(obslik(:,t,:));
wolffd@0 189 tab(:,1) = squeeze(obslik(:,t,1));
wolffd@0 190 else
wolffd@0 191 for d=1:min(D,T-t)
wolffd@0 192 tab(:,d) = squeeze(obslik(:,t+d-1,d));
wolffd@0 193 end
wolffd@0 194 end
wolffd@0 195 softCPDpot{O,t} = dpot(dom, ens(dom), tab);
wolffd@0 196
wolffd@0 197
wolffd@0 198 %bnet.observed = [];
wolffd@0 199 % uniformative observations
wolffd@0 200 %bnet.CPD{eclass(O,2)} = tabular_CPD(bnet, O+ss, 'CPT', mk_stochastic(ones(2,nwords,D,alphasize)));
wolffd@0 201 %bnet.CPD{eclass(O,1)} = tabular_CPD(bnet, O, 'CPT', mk_stochastic(ones(nwords,D,alphasize)));
wolffd@0 202
wolffd@0 203 engine = jtree_dbn_inf_engine(bnet);
wolffd@0 204 evidence = cell(ss,T);
wolffd@0 205 % we add dummy data to O to force its effective size to be 1.
wolffd@0 206 % The actual values have already been incorporated into softCPDpot
wolffd@0 207 evidence(O,:) = num2cell(ones(1,T));
wolffd@0 208 [engine, ll_dbn] = enter_evidence(engine, evidence, 'softCPDpot', softCPDpot);
wolffd@0 209
wolffd@0 210
wolffd@0 211 %evidence(F,:) = num2cell(2*ones(1,T));
wolffd@0 212 %[engine, ll_dbn] = enter_evidence(engine, evidence);
wolffd@0 213
wolffd@0 214
wolffd@0 215 gamma = zeros(nwords, T);
wolffd@0 216 for t=1:T
wolffd@0 217 m = marginal_nodes(engine, [W F], t);
wolffd@0 218 gamma(:,t) = m.T(:,2);
wolffd@0 219 end
wolffd@0 220
wolffd@0 221 gamma
wolffd@0 222
wolffd@0 223 xidbn = zeros(nwords, nwords);
wolffd@0 224 for t=1:T-1
wolffd@0 225 m = marginal_nodes(engine, [W F W+ss], t);
wolffd@0 226 xidbn = xidbn + squeeze(m.T(:,2,:));
wolffd@0 227 end
wolffd@0 228
wolffd@0 229 % thee
wolffd@0 230 % xidbn(1,4) = 0.9412 the->e
wolffd@0 231 % (2,3)=0.0588 t->h
wolffd@0 232 % (3,4)=0.0588 h-e
wolffd@0 233 % (4,4)=0.0588 e-e
wolffd@0 234
wolffd@0 235