wolffd@0
|
1 function CPD = hhmmQ_CPD(bnet, self, Qnodes, d, D, varargin)
|
wolffd@0
|
2 % HHMMQ_CPD Make the CPD for a Q node at depth D of a D-level hierarchical HMM
|
wolffd@0
|
3 % CPD = hhmmQ_CPD(bnet, self, Qnodes, d, D, ...)
|
wolffd@0
|
4 %
|
wolffd@0
|
5 % Fd(t-1) \ Q1:d-1(t)
|
wolffd@0
|
6 % \ |
|
wolffd@0
|
7 % \ v
|
wolffd@0
|
8 % Qd(t-1) -> Qd(t)
|
wolffd@0
|
9 % /
|
wolffd@0
|
10 % /
|
wolffd@0
|
11 % Fd+1(t-1)
|
wolffd@0
|
12 %
|
wolffd@0
|
13 % We assume parents are ordered (numbered) as follows:
|
wolffd@0
|
14 % Qd(t-1), Fd+1(t-1), Fd(t-1), Q1(t), ..., Qd(t)
|
wolffd@0
|
15 %
|
wolffd@0
|
16 % The parents of Qd(t) can either be just Qd-1(t) or the whole stack Q1:d-1(t) (allQ)
|
wolffd@0
|
17 % In either case, we will call them Qps.
|
wolffd@0
|
18 % If d=1, Qps does not exist. Also, the F1(t-1) -> Q1(t) arc is optional.
|
wolffd@0
|
19 % If the arc is missing, startprob does not need to be specified,
|
wolffd@0
|
20 % since the toplevel is assumed to never reset (F1 does not exist).
|
wolffd@0
|
21 % If d=D, Fd+1(t-1) does not exist (there is no signal from below).
|
wolffd@0
|
22 %
|
wolffd@0
|
23 % optional args [defaults]
|
wolffd@0
|
24 %
|
wolffd@0
|
25 % transprob - transprob(i,k,j) = prob transition from i to j given Qps = k ['leftright']
|
wolffd@0
|
26 % selfprob - prob of a transition from i to i given Qps=k [0.1]
|
wolffd@0
|
27 % startprob - startprob(k,j) = prob start in j given Qps = k ['leftstart']
|
wolffd@0
|
28 % startargs - other args to be passed to the sub tabular_CPD for learning startprob
|
wolffd@0
|
29 % transargs - other args will be passed to the sub tabular_CPD for learning transprob
|
wolffd@0
|
30 % allQ - 1 means use all Q nodes above d as parents, 0 means just level d-1 [0]
|
wolffd@0
|
31 % F1toQ1 - 1 means add F1(t-1) -> Q1(t) arc, 0 means level 1 never resets [0]
|
wolffd@0
|
32 %
|
wolffd@0
|
33 % For d=1, startprob(1,j) is only needed if F1toQ1=1
|
wolffd@0
|
34 % Also, transprob(i,j) can be used instead of transprob(i,1,j).
|
wolffd@0
|
35 %
|
wolffd@0
|
36 % hhmmQ_CPD is a subclass of tabular_CPD so we inherit inference methods like CPD_to_pot, etc.
|
wolffd@0
|
37 %
|
wolffd@0
|
38 % We create isolated tabular_CPDs with no F parents to learn transprob/startprob
|
wolffd@0
|
39 % so we can avail of e.g., entropic or Dirichlet priors.
|
wolffd@0
|
40 % In the future, we will be able to represent the transprob using a tree_CPD.
|
wolffd@0
|
41 %
|
wolffd@0
|
42 % For details, see "Linear-time inference in hierarchical HMMs", Murphy and Paskin, NIPS'01.
|
wolffd@0
|
43
|
wolffd@0
|
44
|
wolffd@0
|
45 ss = bnet.nnodes_per_slice;
|
wolffd@0
|
46 %assert(self == Qnodes(d)+ss);
|
wolffd@0
|
47 ns = bnet.node_sizes(:);
|
wolffd@0
|
48 CPD.Qsizes = ns(Qnodes);
|
wolffd@0
|
49 CPD.d = d;
|
wolffd@0
|
50 CPD.D = D;
|
wolffd@0
|
51 allQ = 0;
|
wolffd@0
|
52
|
wolffd@0
|
53 % find out which parents to use, to get right size
|
wolffd@0
|
54 for i=1:2:length(varargin)
|
wolffd@0
|
55 switch varargin{i},
|
wolffd@0
|
56 case 'allQ', allQ = varargin{i+1};
|
wolffd@0
|
57 end
|
wolffd@0
|
58 end
|
wolffd@0
|
59
|
wolffd@0
|
60 if d==1
|
wolffd@0
|
61 CPD.Qps = [];
|
wolffd@0
|
62 else
|
wolffd@0
|
63 if allQ
|
wolffd@0
|
64 CPD.Qps = Qnodes(1:d-1);
|
wolffd@0
|
65 else
|
wolffd@0
|
66 CPD.Qps = Qnodes(d-1);
|
wolffd@0
|
67 end
|
wolffd@0
|
68 end
|
wolffd@0
|
69
|
wolffd@0
|
70 Qsz = ns(self);
|
wolffd@0
|
71 Qpsz = prod(ns(CPD.Qps));
|
wolffd@0
|
72
|
wolffd@0
|
73 % set default arguments
|
wolffd@0
|
74 startprob = 'leftstart';
|
wolffd@0
|
75 transprob = 'leftright';
|
wolffd@0
|
76 startargs = {};
|
wolffd@0
|
77 transargs = {};
|
wolffd@0
|
78 CPD.F1toQ1 = 0;
|
wolffd@0
|
79 selfprob = 0.1;
|
wolffd@0
|
80
|
wolffd@0
|
81 for i=1:2:length(varargin)
|
wolffd@0
|
82 switch varargin{i},
|
wolffd@0
|
83 case 'transprob', transprob = varargin{i+1};
|
wolffd@0
|
84 case 'selfprob', selfprob = varargin{i+1};
|
wolffd@0
|
85 case 'startprob', startprob = varargin{i+1};
|
wolffd@0
|
86 case 'startargs', startargs = varargin{i+1};
|
wolffd@0
|
87 case 'transargs', transargs = varargin{i+1};
|
wolffd@0
|
88 case 'F1toQ1', CPD.F1toQ1 = varargin{i+1};
|
wolffd@0
|
89 end
|
wolffd@0
|
90 end
|
wolffd@0
|
91
|
wolffd@0
|
92 Qps = CPD.Qps + ss;
|
wolffd@0
|
93 old_self = self-ss;
|
wolffd@0
|
94
|
wolffd@0
|
95 if strcmp(transprob, 'leftright')
|
wolffd@0
|
96 LR = mk_leftright_transmat(Qsz, selfprob);
|
wolffd@0
|
97 transprob = repmat(reshape(LR, [1 Qsz Qsz]), [Qpsz 1 1]); % transprob(k,i,j)
|
wolffd@0
|
98 transprob = permute(transprob, [2 1 3]); % now transprob(i,k,j)
|
wolffd@0
|
99 end
|
wolffd@0
|
100 transargs{end+1} = 'CPT';
|
wolffd@0
|
101 transargs{end+1} = transprob;
|
wolffd@0
|
102 CPD.sub_CPD_trans = mk_isolated_tabular_CPD([old_self Qps], ns([old_self Qps self]), transargs);
|
wolffd@0
|
103 S = struct(CPD.sub_CPD_trans);
|
wolffd@0
|
104 CPD.transprob = myreshape(S.CPT, [Qsz Qpsz Qsz]);
|
wolffd@0
|
105
|
wolffd@0
|
106
|
wolffd@0
|
107 if strcmp(startprob, 'leftstart')
|
wolffd@0
|
108 startprob = zeros(Qpsz, Qsz);
|
wolffd@0
|
109 startprob(:,1) = 1;
|
wolffd@0
|
110 end
|
wolffd@0
|
111
|
wolffd@0
|
112 if (d==1) & ~CPD.F1toQ1
|
wolffd@0
|
113 CPD.sub_CPD_start = [];
|
wolffd@0
|
114 CPD.startprob = [];
|
wolffd@0
|
115 else
|
wolffd@0
|
116 startargs{end+1} = 'CPT';
|
wolffd@0
|
117 startargs{end+1} = startprob;
|
wolffd@0
|
118 CPD.sub_CPD_start = mk_isolated_tabular_CPD(Qps, ns([Qps self]), startargs);
|
wolffd@0
|
119 S = struct(CPD.sub_CPD_start);
|
wolffd@0
|
120 CPD.startprob = myreshape(S.CPT, [Qpsz Qsz]);
|
wolffd@0
|
121 end
|
wolffd@0
|
122
|
wolffd@0
|
123 CPD = class(CPD, 'hhmmQ_CPD', tabular_CPD(bnet, self));
|
wolffd@0
|
124
|
wolffd@0
|
125 CPD = update_CPT(CPD);
|
wolffd@0
|
126
|