wolffd@0
|
1 function T = convert_to_table(CPD, domain, evidence)
|
wolffd@0
|
2 % CONVERT_TO_TABLE Convert a mlp CPD to a table, incorporating any evidence
|
wolffd@0
|
3 % T = convert_to_table(CPD, domain, evidence)
|
wolffd@0
|
4
|
wolffd@0
|
5 self = domain(end);
|
wolffd@0
|
6 ps = domain(1:end-1); % self' parents
|
wolffd@0
|
7 %cps = myintersect(ps, cnodes); % self' continous parents
|
wolffd@0
|
8 cnodes = domain(CPD.cpndx);
|
wolffd@0
|
9 cps = myintersect(ps, cnodes);
|
wolffd@0
|
10 odom = domain(~isemptycell(evidence(domain))); % obs nodes in the net
|
wolffd@0
|
11 assert(myismember(cps, odom)); % !ALL the CTS parents must be observed!
|
wolffd@0
|
12 ns(cps)=1;
|
wolffd@0
|
13 dps = mysetdiff(ps, cps); % self' discrete parents
|
wolffd@0
|
14 dobs = myintersect(dps, odom); % discrete obs parents
|
wolffd@0
|
15
|
wolffd@0
|
16 % Extract the params compatible with the observations (if any) on the discrete parents (if any)
|
wolffd@0
|
17
|
wolffd@0
|
18 if ~isempty(dobs),
|
wolffd@0
|
19 dvals = cat(1, evidence{dobs});
|
wolffd@0
|
20 ns_eff= CPD.sizes; % effective node sizes
|
wolffd@0
|
21 ens=ns_eff;
|
wolffd@0
|
22 ens(dobs) = 1;
|
wolffd@0
|
23 S=prod(ens(dps));
|
wolffd@0
|
24 subs = ind2subv(ens(dps), 1:S);
|
wolffd@0
|
25 mask = find_equiv_posns(dobs, dps);
|
wolffd@0
|
26 for i=1:length(mask),
|
wolffd@0
|
27 subs(:,mask(i)) = dvals(i);
|
wolffd@0
|
28 end
|
wolffd@0
|
29 support = subv2ind(ns_eff(dps), subs)';
|
wolffd@0
|
30 else
|
wolffd@0
|
31 ns_eff= CPD.sizes;
|
wolffd@0
|
32 support=[1:prod(ns_eff(dps))];
|
wolffd@0
|
33 end
|
wolffd@0
|
34
|
wolffd@0
|
35 W1=[]; b1=[]; W2=[]; b2=[];
|
wolffd@0
|
36
|
wolffd@0
|
37 W1 = CPD.W1(:,:,support);
|
wolffd@0
|
38 b1= CPD.b1(support,:);
|
wolffd@0
|
39 W2 = CPD.W2(:,:,support);
|
wolffd@0
|
40 b2= CPD.b2(support,:);
|
wolffd@0
|
41 ns(odom) = 1;
|
wolffd@0
|
42 dpsize = prod(ns(dps)); % overall size of the self' discrete parents
|
wolffd@0
|
43
|
wolffd@0
|
44 x = cat(1, evidence{cps});
|
wolffd@0
|
45 ndata=size(x,2);
|
wolffd@0
|
46
|
wolffd@0
|
47 if ~isempty(evidence{self}) %
|
wolffd@0
|
48 app=struct(CPD); %
|
wolffd@0
|
49 ns(self)=app.mlp{1}.nout; % pump up self to the original dimension if observed
|
wolffd@0
|
50 clear app; %
|
wolffd@0
|
51 end %
|
wolffd@0
|
52
|
wolffd@0
|
53 T =zeros(dpsize, ns(self)); %
|
wolffd@0
|
54 for i=1:dpsize %
|
wolffd@0
|
55 W1app = W1(:,:,i); %
|
wolffd@0
|
56 b1app = b1(i,:); %
|
wolffd@0
|
57 W2app = W2(:,:,i); %
|
wolffd@0
|
58 b2app = b2(i,:); % for each of the dpsize combinations of self'parents values
|
wolffd@0
|
59 z = tanh(x(:)'*W1app + ones(ndata, 1)*b1app); % we tabulate the corrisponding glm model
|
wolffd@0
|
60 a = z*W2app + ones(ndata, 1)*b2app; % (element of the cell array CPD.glim)
|
wolffd@0
|
61 appoggio = normalise(exp(a)); %
|
wolffd@0
|
62 T(i,:)=appoggio; %
|
wolffd@0
|
63 W1app=[]; W2app=[]; b1app=[]; b2app=[]; %
|
wolffd@0
|
64 z=[]; a=[]; appoggio=[]; %
|
wolffd@0
|
65 end %
|
wolffd@0
|
66
|
wolffd@0
|
67 if ~isempty(evidence{self})
|
wolffd@0
|
68 appoggio=[]; %
|
wolffd@0
|
69 appoggio=zeros(1,ns(self)); %
|
wolffd@0
|
70 r = evidence{self}; %...if self is observed => in output there's only the probability of the 'true' class
|
wolffd@0
|
71 for i=1:dpsize %
|
wolffd@0
|
72 appoggio(i)=T(i,r); %
|
wolffd@0
|
73 end
|
wolffd@0
|
74 T=zeros(dpsize,1);
|
wolffd@0
|
75 for i=1:dpsize
|
wolffd@0
|
76 T(i,1)=appoggio(i);
|
wolffd@0
|
77 end
|
wolffd@0
|
78 clear appoggio;
|
wolffd@0
|
79 ns(self) = 1;
|
wolffd@0
|
80 end
|