wolffd@0
|
1 function CPDpot = convert_dbn_CPDs_to_tables(bnet, evidence)
|
wolffd@0
|
2 % CONVERT_DBN_CPDS_TO_TABLES Convert CPDs of (possibly instantiated) DBN nodes to tables
|
wolffd@0
|
3 % CPDpot = convert_dbn_CPDs_to_tables(bnet, evidence)
|
wolffd@0
|
4 %
|
wolffd@0
|
5 % CPDpot{n,t} is a table containing P(n,t|pa(n,t), ev)
|
wolffd@0
|
6 % All hidden nodes are assumed to be discrete.
|
wolffd@0
|
7 % We assume the observed nodes are the same in every slice.
|
wolffd@0
|
8 %
|
wolffd@0
|
9 % Evaluating the conditional likelihood of long evidence sequences can be very slow,
|
wolffd@0
|
10 % so we take pains to vectorize where possible.
|
wolffd@0
|
11
|
wolffd@0
|
12 [ss T] = size(evidence);
|
wolffd@0
|
13 %obs_bitv = ~isemptycell(evidence(:));
|
wolffd@0
|
14 obs_bitv = zeros(1, 2*ss);
|
wolffd@0
|
15 obs_bitv(bnet.observed) = 1;
|
wolffd@0
|
16 obs_bitv(bnet.observed+ss) = 1;
|
wolffd@0
|
17
|
wolffd@0
|
18 ns = bnet.node_sizes(:);
|
wolffd@0
|
19 CPDpot = cell(ss,T);
|
wolffd@0
|
20
|
wolffd@0
|
21 for n=1:ss
|
wolffd@0
|
22 % slice 1
|
wolffd@0
|
23 t = 1;
|
wolffd@0
|
24 ps = parents(bnet.dag, n);
|
wolffd@0
|
25 e = bnet.equiv_class(n, 1);
|
wolffd@0
|
26 if ~any(obs_bitv(ps))
|
wolffd@0
|
27 CPDpot{n,t} = convert_CPD_to_table_hidden_ps(bnet.CPD{e}, evidence{n,t});
|
wolffd@0
|
28 else
|
wolffd@0
|
29 CPDpot{n,t} = convert_to_table(bnet.CPD{e}, [ps n], evidence(:,1));
|
wolffd@0
|
30 end
|
wolffd@0
|
31
|
wolffd@0
|
32 % special cases: c=child, p=parents, d=discrete, h=hidden, 1sl=1slice
|
wolffd@0
|
33 % if c=h=1 then c=d=1, since hidden nodes must be discrete
|
wolffd@0
|
34 % c=h c=d p=h p=d 1sl method
|
wolffd@0
|
35 % ---------------------------
|
wolffd@0
|
36 % 1 1 1 1 - replicate CPT
|
wolffd@0
|
37 % - 1 - 1 - evaluate CPT on evidence *
|
wolffd@0
|
38 % 0 1 1 1 1 dhmm
|
wolffd@0
|
39 % 0 0 1 1 1 ghmm
|
wolffd@0
|
40 % other loop
|
wolffd@0
|
41 %
|
wolffd@0
|
42 % * = any subset of the domain may be observed
|
wolffd@0
|
43
|
wolffd@0
|
44 % Example where all of the special cases occur - a hierarchical HMM
|
wolffd@0
|
45 % where the top layer (G) and leaves (Y) are observed and
|
wolffd@0
|
46 % all nodes are discrete except Y.
|
wolffd@0
|
47 % (O turns on if Y is an outlier)
|
wolffd@0
|
48
|
wolffd@0
|
49 % G ---------> G
|
wolffd@0
|
50 % | |
|
wolffd@0
|
51 % v v
|
wolffd@0
|
52 % S --------> S
|
wolffd@0
|
53 % | |
|
wolffd@0
|
54 % v v
|
wolffd@0
|
55 % Y Y
|
wolffd@0
|
56 % ^ ^
|
wolffd@0
|
57 % | |
|
wolffd@0
|
58 % O O
|
wolffd@0
|
59
|
wolffd@0
|
60 % Evaluating P(yt|St,Ot) is the ghmm case
|
wolffd@0
|
61 % Evaluating P(St|S(t-1),gt) is the eval CPT case
|
wolffd@0
|
62 % Evaluating P(gt|g(t-1) is the eval CPT case (hdom = [])
|
wolffd@0
|
63 % Evaluating P(Ot) is the replicated CPT case
|
wolffd@0
|
64
|
wolffd@0
|
65 % Cts parents (e.g., inputs) would require an additional special case for speed
|
wolffd@0
|
66
|
wolffd@0
|
67
|
wolffd@0
|
68 % slices 2..T
|
wolffd@0
|
69 [ss T] = size(evidence);
|
wolffd@0
|
70 self = n+ss;
|
wolffd@0
|
71 ps = parents(bnet.dag, self);
|
wolffd@0
|
72 e = bnet.equiv_class(n, 2);
|
wolffd@0
|
73
|
wolffd@0
|
74 if 1
|
wolffd@0
|
75 debug = 0;
|
wolffd@0
|
76 hidden_child = ~obs_bitv(n);
|
wolffd@0
|
77 discrete_child = myismember(n, bnet.dnodes);
|
wolffd@0
|
78 hidden_ps = all(~obs_bitv(ps));
|
wolffd@0
|
79 discrete_ps = mysubset(ps, bnet.dnodes);
|
wolffd@0
|
80 parents_in_same_slice = all(ps > ss);
|
wolffd@0
|
81
|
wolffd@0
|
82 if hidden_child & discrete_child & hidden_ps & discrete_ps
|
wolffd@0
|
83 CPDpot = helper_repl(bnet, evidence, n, CPDpot, obs_bitv, debug);
|
wolffd@0
|
84 elseif discrete_child & discrete_ps
|
wolffd@0
|
85 CPDpot = helper_eval(bnet, evidence, n, CPDpot, obs_bitv, debug);
|
wolffd@0
|
86 elseif discrete_child & hidden_ps & discrete_ps & parents_in_same_slice
|
wolffd@0
|
87 CPDpot = helper_dhmm(bnet, evidence, n, CPDpot, obs_bitv, debug);
|
wolffd@0
|
88 elseif ~discrete_child & hidden_ps & discrete_ps & parents_in_same_slice
|
wolffd@0
|
89 CPDpot = helper_ghmm(bnet, evidence, n, CPDpot, obs_bitv, debug);
|
wolffd@0
|
90 else
|
wolffd@0
|
91 if debug, fprintf('node %d, slow\n', n); end
|
wolffd@0
|
92 for t=2:T
|
wolffd@0
|
93 CPDpot{n,t} = convert_to_table(bnet.CPD{e}, [ps self], evidence(:,t-1:t));
|
wolffd@0
|
94 end
|
wolffd@0
|
95 end
|
wolffd@0
|
96 end
|
wolffd@0
|
97
|
wolffd@0
|
98 if 0
|
wolffd@0
|
99 for t=2:T
|
wolffd@0
|
100 CPDpot2{n,t} = convert_to_table(bnet.CPD{e}, [ps self], evidence(:,t-1:t));
|
wolffd@0
|
101 if ~approxeq(CPDpot{n,t}, CPDpot2{n,t})
|
wolffd@0
|
102 fprintf('CPDpot n=%d, t=%d\n',n,t);
|
wolffd@0
|
103 keyboard
|
wolffd@0
|
104 end
|
wolffd@0
|
105 end
|
wolffd@0
|
106 end
|
wolffd@0
|
107
|
wolffd@0
|
108
|
wolffd@0
|
109 end
|
wolffd@0
|
110
|
wolffd@0
|
111
|
wolffd@0
|
112
|
wolffd@0
|
113
|
wolffd@0
|
114 %%%%%%%
|
wolffd@0
|
115 function CPDpot = helper_repl(bnet, evidence, n, CPDpot, obs_bitv, debug)
|
wolffd@0
|
116
|
wolffd@0
|
117 [ss T] = size(evidence);
|
wolffd@0
|
118 if debug, fprintf('node %d, repl\n', n); end
|
wolffd@0
|
119 e = bnet.equiv_class(n, 2);
|
wolffd@0
|
120 CPT = convert_CPD_to_table_hidden_ps(bnet.CPD{e}, []);
|
wolffd@0
|
121 CPDpot(n,2:T) = num2cell(repmat(CPT, [1 1 T-1]), [1 2]);
|
wolffd@0
|
122
|
wolffd@0
|
123
|
wolffd@0
|
124
|
wolffd@0
|
125 %%%%%%%
|
wolffd@0
|
126 function CPDpot = helper_eval(bnet, evidence, n, CPDpot, obs_bitv, debug)
|
wolffd@0
|
127
|
wolffd@0
|
128 [ss T] = size(evidence);
|
wolffd@0
|
129 self = n+ss;
|
wolffd@0
|
130 ps = parents(bnet.dag, self);
|
wolffd@0
|
131 e = bnet.equiv_class(n, 2);
|
wolffd@0
|
132 ns = bnet.node_sizes(:);
|
wolffd@0
|
133 % Example: given CPT(p1, p2, p3, p4, c), where p1,p3 are observed
|
wolffd@0
|
134 % we create CPT([p2 p4 c], [p1 p3]).
|
wolffd@0
|
135 % We then convert all observed p1,p3 into indices ndx
|
wolffd@0
|
136 % and return CPT(:, ndx)
|
wolffd@0
|
137 CPT = CPD_to_CPT(bnet.CPD{e});
|
wolffd@0
|
138 domain = [ps self];
|
wolffd@0
|
139 % if dom is [3 7 8] and 3,8 are observed, odom_rel = [1 3], hdom_rel = 2,
|
wolffd@0
|
140 % odom = [3 8], hdom = 7
|
wolffd@0
|
141 odom_rel = find(obs_bitv(domain));
|
wolffd@0
|
142 hdom_rel = find(~obs_bitv(domain));
|
wolffd@0
|
143 odom = domain(odom_rel);
|
wolffd@0
|
144 hdom = domain(hdom_rel);
|
wolffd@0
|
145 if isempty(hdom)
|
wolffd@0
|
146 CPT = CPT(:);
|
wolffd@0
|
147 else
|
wolffd@0
|
148 CPT = permute(CPT, [hdom_rel odom_rel]);
|
wolffd@0
|
149 CPT = reshape(CPT, prod(ns(hdom)), prod(ns(odom)));
|
wolffd@0
|
150 end
|
wolffd@0
|
151 parents_in_same_slice = all(ps > ss);
|
wolffd@0
|
152 if parents_in_same_slice
|
wolffd@0
|
153 if debug, fprintf('node %d eval 1 slice\n', n); end
|
wolffd@0
|
154 data = cell2num(evidence(odom-ss,2:T)); %data(i,t) = val of i'th obs parent at t+1
|
wolffd@0
|
155 else
|
wolffd@0
|
156 if debug, fprintf('node %d eval 2 slice\n', n); end
|
wolffd@0
|
157 % there's probably a way of vectorizing this...
|
wolffd@0
|
158 data = zeros(length(odom), T-1);
|
wolffd@0
|
159 for t=2:T
|
wolffd@0
|
160 ev = evidence(:,t-1:t);
|
wolffd@0
|
161 ev = ev(:);
|
wolffd@0
|
162 ev2 = ev(odom);
|
wolffd@0
|
163 data(:,t-1) = cat(1, ev2{:});
|
wolffd@0
|
164 %data(:,t-1) = cell2num(ev2);
|
wolffd@0
|
165 end
|
wolffd@0
|
166 end
|
wolffd@0
|
167 ndx = subv2ind(ns(odom), data'); % ndx(t) encodes data(:,t)
|
wolffd@0
|
168 if isempty(hdom)
|
wolffd@0
|
169 CPDpot(n,2:T) = num2cell(CPT(ndx)); % a cell array of floats
|
wolffd@0
|
170 else
|
wolffd@0
|
171 CPDpot(n,2:T) = num2cell(CPT(:, ndx), 1); % a cell array of column vectors
|
wolffd@0
|
172 end
|
wolffd@0
|
173
|
wolffd@0
|
174 %%%%%%%
|
wolffd@0
|
175 function CPDpot = helper_dhmm(bnet, evidence, n, CPDpot, obs_bitv, debug)
|
wolffd@0
|
176
|
wolffd@0
|
177 if debug, fprintf('node %d, dhmm\n', n); end
|
wolffd@0
|
178 [ss T] = size(evidence);
|
wolffd@0
|
179 self = n+ss;
|
wolffd@0
|
180 ps = parents(bnet.dag, self);
|
wolffd@0
|
181 e = bnet.equiv_class(n, 2);
|
wolffd@0
|
182 ns = bnet.node_sizes(:);
|
wolffd@0
|
183 CPT = CPD_to_CPT(bnet.CPD{e});
|
wolffd@0
|
184 CPT = reshape(CPT, [prod(ns(ps)) ns(self)]); % what if no parents?
|
wolffd@0
|
185 %obslik = mk_dhmm_obs_lik(cell2num(evidence(n,2:T)), CPT);
|
wolffd@0
|
186 obslik = eval_pdf_cond_multinomial(cell2num(evidence(n,2:T)), CPT);
|
wolffd@0
|
187 CPDpot(n,2:T) = num2cell(obslik, 1);
|
wolffd@0
|
188
|
wolffd@0
|
189
|
wolffd@0
|
190 %%%%%%%
|
wolffd@0
|
191 function CPDpot = helper_ghmm(bnet, evidence, n, CPDpot, obs_bitv, debug)
|
wolffd@0
|
192
|
wolffd@0
|
193 if debug, fprintf('node %d, ghmm\n', n); end
|
wolffd@0
|
194 [ss T] = size(evidence);
|
wolffd@0
|
195 e = bnet.equiv_class(n, 2);
|
wolffd@0
|
196 S = struct(bnet.CPD{e});
|
wolffd@0
|
197 ev2 = cell2num(evidence(n,2:T));
|
wolffd@0
|
198 %obslik = mk_ghmm_obs_lik(ev2, S.mean, S.cov);
|
wolffd@0
|
199 obslik = eval_pdf_cond_gauss(ev2, S.mean, S.cov);
|
wolffd@0
|
200 CPDpot(n,2:T) = num2cell(obslik, 1);
|
wolffd@0
|
201
|