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
|
function bnet = mk_hhmm3(varargin)
% MK_HHMM3 Make a 3 level Hierarchical HMM
% bnet = mk_hhmm3(...)
%
% 3-layer hierarchical HMM where level 1 only connects to level 2, not 3 or obs.
% This enforces sub-models (which differ only in their Q1 index) to be shared.
% Also, we enforce the fact that each model always starts in its initial state
% and only finishes in its final state. However, the prob. of finishing (as opposed to
% self-transitioning to the final state) can be learned.
% The fact that we always finish from the same state means we do not need to condition
% F(i) on Q(i-1), since finishing prob is indep of calling context.
%
% The DBN is the same as Fig 10 in my tech report.
%
% Q1 ----------> Q1
% | / |
% | / |
% | F2 ------- |
% | ^ \ |
% | /| \ |
% v | v v
% Q2-| --------> Q2
% /| | ^
% / | | /|
% | | F3 ---------/ |
% | | ^ \ |
% | v / v
% | Q3 -----------> Q3
% | |
% \ |
% v v
% O
%
% Q1 (slice 1) is clamped to be uniform.
% Q2 (slice 1) is clamped to always start in state 1.
% Q3 (slice 1) is clamped to always start in state 1.
% F3 by default will only finish if Q3 is in its last state (F3 is a tabular_CPD)
% F2 by default gets the default hhmmF_CPD params.
% Q1:Q3 (slice 2) by default gets the default hhmmQ_CPD params.
% O by default gets the default tabular/Gaussian params.
%
% Optional arguments in name/value format [default]
%
% Qsizes - sizes at each level [ none ]
% Osize - size of O node [ none ]
% discrete_obs - 1 means O is tabular_CPD, 0 means O is gaussian_CPD [0]
% Oargs - cell array of args to pass to the O CPD [ {} ]
% Q1args - args to be passed to constructor for Q1 (slice 2) [ {} ]
% Q2args - args to be passed to constructor for Q2 (slice 2) [ {} ]
% Q3args - args to be passed to constructor for Q3 (slice 2) [ {} ]
% F2args - args to be passed to constructor for F2 [ {} ]
% F3args - args to be passed to constructor for F3 [ {'CPT', finish in last Q3 state} ]
%
ss = 6; D = 3;
Q1 = 1; Q2 = 2; Q3 = 3; F3 = 4; F2 = 5; obs = 6;
Qnodes = [Q1 Q2 Q3]; Fnodes = [F2 F3];
names = {'Q1', 'Q2', 'Q3', 'F3', 'F2', 'obs'};
intra = zeros(ss);
intra(Q1, Q2) = 1;
intra(Q2, [F2 Q3 obs]) = 1;
intra(Q3, [F3 obs]) = 1;
intra(F3, F2) = 1;
inter = zeros(ss);
inter(Q1,Q1) = 1;
inter(Q2,Q2) = 1;
inter(Q3,Q3) = 1;
inter(F2,[Q1 Q2]) = 1;
inter(F3,[Q2 Q3]) = 1;
% get sizes of nodes
args = varargin;
nargs = length(args);
Qsizes = [];
Osize = 0;
for i=1:2:nargs
switch args{i},
case 'Qsizes', Qsizes = args{i+1};
case 'Osize', Osize = args{i+1};
end
end
if isempty(Qsizes), error('must specify Qsizes'); end
if Osize==0, error('must specify Osize'); end
% set default params
discrete_obs = 0;
Oargs = {};
Q1args = {};
Q2args = {};
Q3args = {};
F2args = {};
% P(Q3, F3)
CPT = zeros(Qsizes(3), 2);
% Each model can only terminate in its final state.
% 0 params will remain 0 during EM, thus enforcing this constraint.
CPT(:, 1) = 1.0; % all states turn F off ...
p = 0.5;
CPT(Qsizes(3), 2) = p; % except the last one
CPT(Qsizes(3), 1) = 1-p;
F3args = {'CPT', CPT};
for i=1:2:nargs
switch args{i},
case 'discrete_obs', discrete_obs = args{i+1};
case 'Oargs', Oargs = args{i+1};
case 'Q1args', Q1args = args{i+1};
case 'Q2args', Q2args = args{i+1};
case 'Q3args', Q3args = args{i+1};
case 'F2args', F2args = args{i+1};
case 'F3args', F3args = args{i+1};
end
end
ns = zeros(1,ss);
ns(Qnodes) = Qsizes;
ns(obs) = Osize;
ns(Fnodes) = 2;
dnodes = [Qnodes Fnodes];
if discrete_obs
dnodes = [dnodes obs];
end
onodes = [obs];
bnet = mk_dbn(intra, inter, ns, 'observed', onodes, 'discrete', dnodes, 'names', names);
eclass = bnet.equiv_class;
% SLICE 1
% We clamp untied nodes in the first slice, since their params can't be estimated
% from just one sequence
% uniform prior on initial model
CPT = normalise(ones(1,ns(Q1)));
bnet.CPD{eclass(Q1,1)} = tabular_CPD(bnet, Q1, 'CPT', CPT, 'adjustable', 0);
% each model always starts in state 1
CPT = zeros(ns(Q1), ns(Q2));
CPT(:, 1) = 1.0;
bnet.CPD{eclass(Q2,1)} = tabular_CPD(bnet, Q2, 'CPT', CPT, 'adjustable', 0);
% each model always starts in state 1
CPT = zeros(ns(Q2), ns(Q3));
CPT(:, 1) = 1.0;
bnet.CPD{eclass(Q3,1)} = tabular_CPD(bnet, Q3, 'CPT', CPT, 'adjustable', 0);
bnet.CPD{eclass(F2,1)} = hhmmF_CPD(bnet, F2, Qnodes, 2, D, F2args{:});
bnet.CPD{eclass(F3,1)} = tabular_CPD(bnet, F3, F3args{:});
if discrete_obs
bnet.CPD{eclass(obs,1)} = tabular_CPD(bnet, obs, Oargs{:});
else
bnet.CPD{eclass(obs,1)} = gaussian_CPD(bnet, obs, Oargs{:});
end
% SLICE 2
bnet.CPD{eclass(Q1,2)} = hhmmQ_CPD(bnet, Q1+ss, Qnodes, 1, D, Q1args{:});
bnet.CPD{eclass(Q2,2)} = hhmmQ_CPD(bnet, Q2+ss, Qnodes, 2, D, Q2args{:});
bnet.CPD{eclass(Q3,2)} = hhmmQ_CPD(bnet, Q3+ss, Qnodes, 3, D, Q3args{:});
|
