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
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
|
function CPDpot = convert_dbn_CPDs_to_tables(bnet, evidence)
% CONVERT_DBN_CPDS_TO_TABLES Convert CPDs of (possibly instantiated) DBN nodes to tables
% CPDpot = convert_dbn_CPDs_to_tables(bnet, evidence)
%
% CPDpot{n,t} is a table containing P(n,t|pa(n,t), ev)
% All hidden nodes are assumed to be discrete.
% We assume the observed nodes are the same in every slice.
%
% Evaluating the conditional likelihood of long evidence sequences can be very slow,
% so we take pains to vectorize where possible.
[ss T] = size(evidence);
%obs_bitv = ~isemptycell(evidence(:));
obs_bitv = zeros(1, 2*ss);
obs_bitv(bnet.observed) = 1;
obs_bitv(bnet.observed+ss) = 1;
ns = bnet.node_sizes(:);
CPDpot = cell(ss,T);
for n=1:ss
% slice 1
t = 1;
ps = parents(bnet.dag, n);
e = bnet.equiv_class(n, 1);
if ~any(obs_bitv(ps))
CPDpot{n,t} = convert_CPD_to_table_hidden_ps(bnet.CPD{e}, evidence{n,t});
else
CPDpot{n,t} = convert_to_table(bnet.CPD{e}, [ps n], evidence(:,1));
end
% special cases: c=child, p=parents, d=discrete, h=hidden, 1sl=1slice
% if c=h=1 then c=d=1, since hidden nodes must be discrete
% c=h c=d p=h p=d 1sl method
% ---------------------------
% 1 1 1 1 - replicate CPT
% - 1 - 1 - evaluate CPT on evidence *
% 0 1 1 1 1 dhmm
% 0 0 1 1 1 ghmm
% other loop
%
% * = any subset of the domain may be observed
% Example where all of the special cases occur - a hierarchical HMM
% where the top layer (G) and leaves (Y) are observed and
% all nodes are discrete except Y.
% (O turns on if Y is an outlier)
% G ---------> G
% | |
% v v
% S --------> S
% | |
% v v
% Y Y
% ^ ^
% | |
% O O
% Evaluating P(yt|St,Ot) is the ghmm case
% Evaluating P(St|S(t-1),gt) is the eval CPT case
% Evaluating P(gt|g(t-1) is the eval CPT case (hdom = [])
% Evaluating P(Ot) is the replicated CPT case
% Cts parents (e.g., inputs) would require an additional special case for speed
% slices 2..T
[ss T] = size(evidence);
self = n+ss;
ps = parents(bnet.dag, self);
e = bnet.equiv_class(n, 2);
if 1
debug = 0;
hidden_child = ~obs_bitv(n);
discrete_child = myismember(n, bnet.dnodes);
hidden_ps = all(~obs_bitv(ps));
discrete_ps = mysubset(ps, bnet.dnodes);
parents_in_same_slice = all(ps > ss);
if hidden_child & discrete_child & hidden_ps & discrete_ps
CPDpot = helper_repl(bnet, evidence, n, CPDpot, obs_bitv, debug);
elseif discrete_child & discrete_ps
CPDpot = helper_eval(bnet, evidence, n, CPDpot, obs_bitv, debug);
elseif discrete_child & hidden_ps & discrete_ps & parents_in_same_slice
CPDpot = helper_dhmm(bnet, evidence, n, CPDpot, obs_bitv, debug);
elseif ~discrete_child & hidden_ps & discrete_ps & parents_in_same_slice
CPDpot = helper_ghmm(bnet, evidence, n, CPDpot, obs_bitv, debug);
else
if debug, fprintf('node %d, slow\n', n); end
for t=2:T
CPDpot{n,t} = convert_to_table(bnet.CPD{e}, [ps self], evidence(:,t-1:t));
end
end
end
if 0
for t=2:T
CPDpot2{n,t} = convert_to_table(bnet.CPD{e}, [ps self], evidence(:,t-1:t));
if ~approxeq(CPDpot{n,t}, CPDpot2{n,t})
fprintf('CPDpot n=%d, t=%d\n',n,t);
keyboard
end
end
end
end
%%%%%%%
function CPDpot = helper_repl(bnet, evidence, n, CPDpot, obs_bitv, debug)
[ss T] = size(evidence);
if debug, fprintf('node %d, repl\n', n); end
e = bnet.equiv_class(n, 2);
CPT = convert_CPD_to_table_hidden_ps(bnet.CPD{e}, []);
CPDpot(n,2:T) = num2cell(repmat(CPT, [1 1 T-1]), [1 2]);
%%%%%%%
function CPDpot = helper_eval(bnet, evidence, n, CPDpot, obs_bitv, debug)
[ss T] = size(evidence);
self = n+ss;
ps = parents(bnet.dag, self);
e = bnet.equiv_class(n, 2);
ns = bnet.node_sizes(:);
% Example: given CPT(p1, p2, p3, p4, c), where p1,p3 are observed
% we create CPT([p2 p4 c], [p1 p3]).
% We then convert all observed p1,p3 into indices ndx
% and return CPT(:, ndx)
CPT = CPD_to_CPT(bnet.CPD{e});
domain = [ps self];
% if dom is [3 7 8] and 3,8 are observed, odom_rel = [1 3], hdom_rel = 2,
% odom = [3 8], hdom = 7
odom_rel = find(obs_bitv(domain));
hdom_rel = find(~obs_bitv(domain));
odom = domain(odom_rel);
hdom = domain(hdom_rel);
if isempty(hdom)
CPT = CPT(:);
else
CPT = permute(CPT, [hdom_rel odom_rel]);
CPT = reshape(CPT, prod(ns(hdom)), prod(ns(odom)));
end
parents_in_same_slice = all(ps > ss);
if parents_in_same_slice
if debug, fprintf('node %d eval 1 slice\n', n); end
data = cell2num(evidence(odom-ss,2:T)); %data(i,t) = val of i'th obs parent at t+1
else
if debug, fprintf('node %d eval 2 slice\n', n); end
% there's probably a way of vectorizing this...
data = zeros(length(odom), T-1);
for t=2:T
ev = evidence(:,t-1:t);
ev = ev(:);
ev2 = ev(odom);
data(:,t-1) = cat(1, ev2{:});
%data(:,t-1) = cell2num(ev2);
end
end
ndx = subv2ind(ns(odom), data'); % ndx(t) encodes data(:,t)
if isempty(hdom)
CPDpot(n,2:T) = num2cell(CPT(ndx)); % a cell array of floats
else
CPDpot(n,2:T) = num2cell(CPT(:, ndx), 1); % a cell array of column vectors
end
%%%%%%%
function CPDpot = helper_dhmm(bnet, evidence, n, CPDpot, obs_bitv, debug)
if debug, fprintf('node %d, dhmm\n', n); end
[ss T] = size(evidence);
self = n+ss;
ps = parents(bnet.dag, self);
e = bnet.equiv_class(n, 2);
ns = bnet.node_sizes(:);
CPT = CPD_to_CPT(bnet.CPD{e});
CPT = reshape(CPT, [prod(ns(ps)) ns(self)]); % what if no parents?
%obslik = mk_dhmm_obs_lik(cell2num(evidence(n,2:T)), CPT);
obslik = eval_pdf_cond_multinomial(cell2num(evidence(n,2:T)), CPT);
CPDpot(n,2:T) = num2cell(obslik, 1);
%%%%%%%
function CPDpot = helper_ghmm(bnet, evidence, n, CPDpot, obs_bitv, debug)
if debug, fprintf('node %d, ghmm\n', n); end
[ss T] = size(evidence);
e = bnet.equiv_class(n, 2);
S = struct(bnet.CPD{e});
ev2 = cell2num(evidence(n,2:T));
%obslik = mk_ghmm_obs_lik(ev2, S.mean, S.cov);
obslik = eval_pdf_cond_gauss(ev2, S.mean, S.cov);
CPDpot(n,2:T) = num2cell(obslik, 1);
|
