wolffd@0
|
1 function CPD = update_ess2(CPD, fmarginal, evidence, ns, cnodes, hidden_bitv)
|
wolffd@0
|
2 % UPDATE_ESS Update the Expected Sufficient Statistics of a hhmm Q node.
|
wolffd@0
|
3 % function CPD = update_ess(CPD, fmarginal, evidence, ns, cnodes, idden_bitv)
|
wolffd@0
|
4
|
wolffd@0
|
5 % Figure out the node numbers associated with each parent
|
wolffd@0
|
6 dom = fmarginal.domain;
|
wolffd@0
|
7 self = dom(end); % by assumption
|
wolffd@0
|
8 old_self = dom(CPD.old_self_ndx);
|
wolffd@0
|
9 Fself = dom(CPD.Fself_ndx);
|
wolffd@0
|
10 Fbelow = dom(CPD.Fbelow_ndx);
|
wolffd@0
|
11 Qps = dom(CPD.Qps_ndx);
|
wolffd@0
|
12
|
wolffd@0
|
13 Qsz = CPD.Qsz;
|
wolffd@0
|
14 Qpsz = CPD.Qpsz;
|
wolffd@0
|
15
|
wolffd@0
|
16
|
wolffd@0
|
17 fmarg = add_ev_to_dmarginal(fmarginal, evidence, ns);
|
wolffd@0
|
18
|
wolffd@0
|
19
|
wolffd@0
|
20
|
wolffd@0
|
21 % hor_counts(old_self, Qps, self),
|
wolffd@0
|
22 % fmarginal(old_self, Fbelow, Fself, Qps, self)
|
wolffd@0
|
23 % hor_counts(i,k,j) = fmarginal(i,2,1,k,j) % below has finished, self has not
|
wolffd@0
|
24 % ver_counts(i,k,j) = fmarginal(i,2,2,k,j) % below has finished, and so has self (reset)
|
wolffd@0
|
25 % Since any of i,j,k may be observed, we write
|
wolffd@0
|
26 % hor_counts(counts_ndx{:}) = fmarginal(fmarg_ndx{:})
|
wolffd@0
|
27 % where e.g., counts_ndx = {1, ':', 2} if Qps is hidden but we observe old_self=1, self=2.
|
wolffd@0
|
28 % To create this counts_ndx, we write counts_ndx = mk_multi_ndx(3, obs_dim, obs_val)
|
wolffd@0
|
29 % where counts_obs_dim = [1 3], counts_obs_val = [1 2] specifies the values of dimensions 1 and 3.
|
wolffd@0
|
30
|
wolffd@0
|
31 counts_obs_dim = [];
|
wolffd@0
|
32 fmarg_obs_dim = [];
|
wolffd@0
|
33 obs_val = [];
|
wolffd@0
|
34 if hidden_bitv(self)
|
wolffd@0
|
35 effQsz = Qsz;
|
wolffd@0
|
36 else
|
wolffd@0
|
37 effQsz = 1;
|
wolffd@0
|
38 counts_obs_dim = [counts_obs_dim 3];
|
wolffd@0
|
39 fmarg_obs_dim = [fmarg_obs_dim 5];
|
wolffd@0
|
40 obs_val = [obs_val evidence{self}];
|
wolffd@0
|
41 end
|
wolffd@0
|
42
|
wolffd@0
|
43 % e.g., D=4, d=3, Qps = all Qs above, so dom = [Q3(t-1) F4(t-1) F3(t-1) Q1(t) Q2(t) Q3(t)].
|
wolffd@0
|
44 % so self = Q3(t), old_self = Q3(t-1), CPD.Qps = [1 2], Qps = [Q1(t) Q2(t)]
|
wolffd@0
|
45 dom = fmarginal.domain;
|
wolffd@0
|
46 self = dom(end);
|
wolffd@0
|
47 old_self = dom(1);
|
wolffd@0
|
48 Qps = dom(length(dom)-length(CPD.Qps):end-1);
|
wolffd@0
|
49
|
wolffd@0
|
50 Qsz = CPD.Qsizes(CPD.d);
|
wolffd@0
|
51 Qpsz = prod(CPD.Qsizes(CPD.Qps));
|
wolffd@0
|
52
|
wolffd@0
|
53 % If some of the Q nodes are observed (which happens during supervised training)
|
wolffd@0
|
54 % the counts will only be non-zero in positions
|
wolffd@0
|
55 % consistent with the evidence. We put the computed marginal responsibilities
|
wolffd@0
|
56 % into the appropriate slots of the big counts array.
|
wolffd@0
|
57 % (Recall that observed discrete nodes only have a single effective value.)
|
wolffd@0
|
58 % (A more general, but much slower, way is to call add_evidence_to_dmarginal.)
|
wolffd@0
|
59 % We assume the F nodes are never observed.
|
wolffd@0
|
60
|
wolffd@0
|
61 obs_self = ~hidden_bitv(self);
|
wolffd@0
|
62 obs_Qps = (~isempty(Qps)) & (~any(hidden_bitv(Qps))); % we assume that all or none of the Q parents are observed
|
wolffd@0
|
63
|
wolffd@0
|
64 if obs_self
|
wolffd@0
|
65 self_val = evidence{self};
|
wolffd@0
|
66 oldself_val = evidence{old_self};
|
wolffd@0
|
67 end
|
wolffd@0
|
68
|
wolffd@0
|
69 if obs_Qps
|
wolffd@0
|
70 Qps_val = subv2ind(Qpsz, cat(1, evidence{Qps}));
|
wolffd@0
|
71 if Qps_val == 0
|
wolffd@0
|
72 keyboard
|
wolffd@0
|
73 end
|
wolffd@0
|
74 end
|
wolffd@0
|
75
|
wolffd@0
|
76 if CPD.d==1 % no Qps from above
|
wolffd@0
|
77 if ~CPD.F1toQ1 % no F from self
|
wolffd@0
|
78 % marg(Q1(t-1), F2(t-1), Q1(t))
|
wolffd@0
|
79 % F2(t-1) P(Q1(t)=j | Q1(t-1)=i)
|
wolffd@0
|
80 % 1 delta(i,j)
|
wolffd@0
|
81 % 2 transprob(i,j)
|
wolffd@0
|
82 if obs_self
|
wolffd@0
|
83 hor_counts = zeros(Qsz, Qsz);
|
wolffd@0
|
84 hor_counts(oldself_val, self_val) = fmarginal.T(2);
|
wolffd@0
|
85 else
|
wolffd@0
|
86 marg = reshape(fmarginal.T, [Qsz 2 Qsz]);
|
wolffd@0
|
87 hor_counts = squeeze(marg(:,2,:));
|
wolffd@0
|
88 end
|
wolffd@0
|
89 else
|
wolffd@0
|
90 % marg(Q1(t-1), F2(t-1), F1(t-1), Q1(t))
|
wolffd@0
|
91 % F2(t-1) F1(t-1) P(Qd(t)=j| Qd(t-1)=i)
|
wolffd@0
|
92 % ------------------------------------------------------
|
wolffd@0
|
93 % 1 1 delta(i,j)
|
wolffd@0
|
94 % 2 1 transprob(i,j)
|
wolffd@0
|
95 % 1 2 impossible
|
wolffd@0
|
96 % 2 2 startprob(j)
|
wolffd@0
|
97 if obs_self
|
wolffd@0
|
98 marg = myreshape(fmarginal.T, [1 2 2 1]);
|
wolffd@0
|
99 hor_counts = zeros(Qsz, Qsz);
|
wolffd@0
|
100 hor_counts(oldself_val, self_val) = marg(1,2,1,1);
|
wolffd@0
|
101 ver_counts = zeros(Qsz, 1);
|
wolffd@0
|
102 %ver_counts(self_val) = marg(1,2,2,1);
|
wolffd@0
|
103 ver_counts(self_val) = marg(1,2,2,1) + marg(1,1,2,1);
|
wolffd@0
|
104 else
|
wolffd@0
|
105 marg = reshape(fmarginal.T, [Qsz 2 2 Qsz]);
|
wolffd@0
|
106 hor_counts = squeeze(marg(:,2,1,:));
|
wolffd@0
|
107 %ver_counts = squeeze(sum(marg(:,2,2,:),1)); % sum over i
|
wolffd@0
|
108 ver_counts = squeeze(sum(marg(:,2,2,:),1)) + squeeze(sum(marg(:,1,2,:),1)); % sum i,b
|
wolffd@0
|
109 end
|
wolffd@0
|
110 end % F1toQ1
|
wolffd@0
|
111 else % d ~= 1
|
wolffd@0
|
112 if CPD.d < CPD.D % general case
|
wolffd@0
|
113 % marg(Qd(t-1), Fd+1(t-1), Fd(t-1), Qps(t), Qd(t))
|
wolffd@0
|
114 % Fd+1(t-1) Fd(t-1) P(Qd(t)=j| Qd(t-1)=i, Qps(t)=k)
|
wolffd@0
|
115 % ------------------------------------------------------
|
wolffd@0
|
116 % 1 1 delta(i,j)
|
wolffd@0
|
117 % 2 1 transprob(i,k,j)
|
wolffd@0
|
118 % 1 2 impossible
|
wolffd@0
|
119 % 2 2 startprob(k,j)
|
wolffd@0
|
120 if obs_Qps & obs_self
|
wolffd@0
|
121 marg = myreshape(fmarginal.T, [1 2 2 1 1]);
|
wolffd@0
|
122 k = 1;
|
wolffd@0
|
123 hor_counts = zeros(Qsz, Qpsz, Qsz);
|
wolffd@0
|
124 hor_counts(oldself_val, Qps_val, self_val) = marg(1, 2,1, k,1);
|
wolffd@0
|
125 ver_counts = zeros(Qpsz, Qsz);
|
wolffd@0
|
126 %ver_counts(Qps_val, self_val) = marg(1, 2,2, k,1);
|
wolffd@0
|
127 ver_counts(Qps_val, self_val) = marg(1, 2,2, k,1) + marg(1, 1,2, k,1);
|
wolffd@0
|
128 elseif obs_Qps & ~obs_self
|
wolffd@0
|
129 marg = myreshape(fmarginal.T, [Qsz 2 2 1 Qsz]);
|
wolffd@0
|
130 k = 1;
|
wolffd@0
|
131 hor_counts = zeros(Qsz, Qpsz, Qsz);
|
wolffd@0
|
132 hor_counts(:, Qps_val, :) = marg(:, 2,1, k,:);
|
wolffd@0
|
133 ver_counts = zeros(Qpsz, Qsz);
|
wolffd@0
|
134 %ver_counts(Qps_val, :) = sum(marg(:, 2,2, k,:), 1);
|
wolffd@0
|
135 ver_counts(Qps_val, :) = sum(marg(:, 2,2, k,:), 1) + sum(marg(:, 1,2, k,:), 1);
|
wolffd@0
|
136 elseif ~obs_Qps & obs_self
|
wolffd@0
|
137 error('not yet implemented')
|
wolffd@0
|
138 else % everything is hidden
|
wolffd@0
|
139 marg = reshape(fmarginal.T, [Qsz 2 2 Qpsz Qsz]);
|
wolffd@0
|
140 hor_counts = squeeze(marg(:,2,1,:,:)); % i,k,j
|
wolffd@0
|
141 %ver_counts = squeeze(sum(marg(:,2,2,:,:),1)); % sum over i
|
wolffd@0
|
142 ver_counts = squeeze(sum(marg(:,2,2,:,:),1)) + squeeze(sum(marg(:,1,2,:,:),1)); % sum over i,b
|
wolffd@0
|
143 end
|
wolffd@0
|
144 else % d == D, so no F from below
|
wolffd@0
|
145 % marg(QD(t-1), FD(t-1), Qps(t), QD(t))
|
wolffd@0
|
146 % FD(t-1) P(QD(t)=j | QD(t-1)=i, Qps(t)=k)
|
wolffd@0
|
147 % 1 transprob(i,k,j)
|
wolffd@0
|
148 % 2 startprob(k,j)
|
wolffd@0
|
149 if obs_Qps & obs_self
|
wolffd@0
|
150 marg = myreshape(fmarginal.T, [1 2 1 1]);
|
wolffd@0
|
151 k = 1;
|
wolffd@0
|
152 hor_counts = zeros(Qsz, Qpsz, Qsz);
|
wolffd@0
|
153 hor_counts(oldself_val, Qps_val, self_val) = marg(1, 1, k,1);
|
wolffd@0
|
154 ver_counts = zeros(Qpsz, Qsz);
|
wolffd@0
|
155 ver_counts(Qps_val, self_val) = marg(1, 2, k,1);
|
wolffd@0
|
156 elseif obs_Qps & ~obs_self
|
wolffd@0
|
157 marg = myreshape(fmarginal.T, [Qsz 2 1 Qsz]);
|
wolffd@0
|
158 k = 1;
|
wolffd@0
|
159 hor_counts = zeros(Qsz, Qpsz, Qsz);
|
wolffd@0
|
160 hor_counts(:, Qps_val, :) = marg(:, 1, k,:);
|
wolffd@0
|
161 ver_counts = zeros(Qpsz, Qsz);
|
wolffd@0
|
162 ver_counts(Qps_val, :) = sum(marg(:, 2, k, :), 1);
|
wolffd@0
|
163 elseif ~obs_Qps & obs_self
|
wolffd@0
|
164 error('not yet implemented')
|
wolffd@0
|
165 else % everything is hidden
|
wolffd@0
|
166 marg = reshape(fmarginal.T, [Qsz 2 Qpsz Qsz]);
|
wolffd@0
|
167 hor_counts = squeeze(marg(:,1,:,:));
|
wolffd@0
|
168 ver_counts = squeeze(sum(marg(:,2,:,:),1)); % sum over i
|
wolffd@0
|
169 end
|
wolffd@0
|
170 end
|
wolffd@0
|
171 end
|
wolffd@0
|
172
|
wolffd@0
|
173 CPD.sub_CPD_trans = update_ess_simple(CPD.sub_CPD_trans, hor_counts);
|
wolffd@0
|
174
|
wolffd@0
|
175 if ~isempty(CPD.sub_CPD_start)
|
wolffd@0
|
176 CPD.sub_CPD_start = update_ess_simple(CPD.sub_CPD_start, ver_counts);
|
wolffd@0
|
177 end
|
wolffd@0
|
178
|