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
|
function CPD = hhmmQ_CPD(bnet, self, varargin)
% HHMMQ_CPD Make the CPD for a Q node in a hierarchical HMM
% CPD = hhmmQ_CPD(bnet, self, ...)
%
% Fself(t-1) Qps(t)
% \ |
% \ v
% Qold(t-1) -> Q(t)
% /
% /
% Fbelow(t-1)
%
% Let ss = slice size = num. nodes per slice.
% This node is Q(t), and has mandatory parents Qold(t-1) (assumed to be numbered Q(t)-ss)
% and optional parents Fbelow, Fself, Qps.
% We require parents to be ordered (numbered) as follows:
% Qold, Fbelow, Fself, Qps, Q.
%
% If Fself=2, we use the transition matrix, else we use the prior matrix.
% If Fself node is omitted (eg. top level), we always use the transition matrix.
% If Fbelow=2, we may change state, otherwise we must stay in the same state.
% If Fbelow node is omitted (eg., bottom level), we may change state at every step.
% If Qps (Q parents) are specified, all parameters are conditioned on their joint value.
% We may choose any subset of nodes to condition on, as long as they as numbered lower than self.
%
% optional args [defaults]
%
% Fself - node number <= ss
% Fbelow - node number <= ss
% Qps - node numbers (all <= 2*ss) - uses 2TBN indexing
% transprob - transprob(i,k,j) = prob transition from i to j given Qps = k ['leftright']
% selfprob - prob of a transition from i to i given Qps=k [0.1]
% startprob - startprob(k,j) = prob start in j given Qps = k ['leftstart']
% startargs - other args to be passed to the sub tabular_CPD for learning startprob
% transargs - other args will be passed to the sub tabular_CPD for learning transprob
% fullstartprob - 1 means startprob depends on Q(t-1) [0]
% hhmmQ_CPD is a subclass of tabular_CPD so we inherit inference methods like CPD_to_pot, etc.
%
% We create isolated tabular_CPDs with no F parents to learn transprob/startprob
% so we can avail of e.g., entropic or Dirichlet priors.
% In the future, we will be able to represent the transprob using a tree_CPD.
%
% For details, see "Linear-time inference in hierarchical HMMs", Murphy and Paskin, NIPS'01.
ss = bnet.nnodes_per_slice;
ns = bnet.node_sizes(:);
% set default arguments
Fself = [];
Fbelow = [];
Qps = [];
startprob = 'leftstart';
transprob = 'leftright';
startargs = {};
transargs = {};
selfprob = 0.1;
fullstartprob = 0;
for i=1:2:length(varargin)
switch varargin{i},
case 'Fself', Fself = varargin{i+1};
case 'Fbelow', Fbelow = varargin{i+1};
case 'Qps', Qps = varargin{i+1};
case 'transprob', transprob = varargin{i+1};
case 'selfprob', selfprob = varargin{i+1};
case 'startprob', startprob = varargin{i+1};
case 'startargs', startargs = varargin{i+1};
case 'transargs', transargs = varargin{i+1};
case 'fullstartprob', fullstartprob = varargin{i+1};
end
end
CPD.fullstartprob = fullstartprob;
ps = parents(bnet.dag, self);
ndsz = ns(:)';
CPD.dom_sz = [ndsz(ps) ns(self)];
CPD.Fself_ndx = find_equiv_posns(Fself, ps);
CPD.Fbelow_ndx = find_equiv_posns(Fbelow, ps);
%CPD.Qps_ndx = find_equiv_posns(Qps+ss, ps);
CPD.Qps_ndx = find_equiv_posns(Qps, ps);
old_self = self-ss;
CPD.old_self_ndx = find_equiv_posns(old_self, ps);
Qps = ps(CPD.Qps_ndx);
CPD.Qsz = ns(self);
CPD.Qpsz = prod(ns(Qps));
CPD.Qpsizes = ns(Qps);
Qsz = CPD.Qsz;
Qpsz = CPD.Qpsz;
if strcmp(transprob, 'leftright')
LR = mk_leftright_transmat(Qsz, selfprob);
transprob = repmat(reshape(LR, [1 Qsz Qsz]), [Qpsz 1 1]); % transprob(k,i,j)
transprob = permute(transprob, [2 1 3]); % now transprob(i,k,j)
end
transargs{end+1} = 'CPT';
transargs{end+1} = transprob;
CPD.sub_CPD_trans = mk_isolated_tabular_CPD(ns([old_self Qps self]), transargs);
S = struct(CPD.sub_CPD_trans);
%CPD.transprob = myreshape(S.CPT, [Qsz Qpsz Qsz]);
CPD.transprob = S.CPT;
if strcmp(startprob, 'leftstart')
startprob = zeros(Qpsz, Qsz);
startprob(:,1) = 1;
end
if isempty(CPD.Fself_ndx)
CPD.sub_CPD_start = [];
CPD.startprob = [];
else
startargs{end+1} = 'CPT';
startargs{end+1} = startprob;
if CPD.fullstartprob
CPD.sub_CPD_start = mk_isolated_tabular_CPD(ns([self Qps self]), startargs);
S = struct(CPD.sub_CPD_start);
%CPD.startprob = myreshape(S.CPT, [Qsz Qpsz Qsz]);
CPD.startprob = S.CPT;
else
CPD.sub_CPD_start = mk_isolated_tabular_CPD(ns([Qps self]), startargs);
S = struct(CPD.sub_CPD_start);
%CPD.startprob = myreshape(S.CPT, [CPD.Qpsizes Qsz]);
CPD.startprob = S.CPT;
end
end
CPD = class(CPD, 'hhmmQ_CPD', tabular_CPD(bnet, self));
CPD = update_CPT(CPD);
|
