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