-
Notifications
You must be signed in to change notification settings - Fork 0
/
imgpPredict.m
78 lines (66 loc) · 2.12 KB
/
imgpPredict.m
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
function [ yp sig2 omega ypc ] = imgpPredict(model, vardist, Xtest, omega)
robotics = 0;
C = vardist.C;
D = model.D;
N = model.N;
Nstar = size(Xtest, 1);
if nargin < 4
omega = ones(C,1);
for c = 2:C
omega(c) = omega(c-1)*(1-vardist.beta1(c-1)/(vardist.beta1(c-1)+vardist.beta2(c-1)));
end
omega(1:C-1) = omega(1:C-1).*vardist.beta1(1:C-1)./(vardist.beta1(1:C-1)+vardist.beta2(1:C-1));
else
disp('omega is _given_ for predictions');
assert( length(omega) == C );
end
disp([ 'omega ' 'gamma ' 'sort(omega) ' 'sort(gamma) ' ]);
[ z1 t1 ] = sort(omega,'descend');
[ z2 t2 ] = sort(sum(vardist.gamma),'descend');
disp( [ omega sum(vardist.gamma)' t1 t2'; ] );
fprintf(1,'omega sum = %f\n', sum(omega));
yp = zeros(Nstar,1);
out1 = zeros(Nstar,1);
mix = gmm(D, C, 'full');
for c = 1:C
mix.priors(c) = omega(c);
mix.centres(c,:) = vardist.g(:,c)';
mix.covars(:,:,c) = inv(vardist.W(:,:,c))/vardist.nu(c);
end
post = gmmpost(mix, Xtest);
for c = 1:C
Kc = feval(model.GP{c}.covfunc, model.GP{c}.logtheta, model.X);
[Kss, Kstar] = feval(model.GP{c}.covfunc, model.GP{c}.logtheta, model.X, Xtest);
Bc = vardist.B(:,c);
V = solve_chol_zeros(Kc, Bc, Kstar, max(diag(Kc))+model.Likelihood.sigma2+1e10);
V = V';
% old implementation
% Lc = chol(Kc+diag(Bc),'lower');
% V = (Lc'\(Lc\(Kstar)))';
mustarc = model.GP{c}.mean + V*(model.Y(:)-model.GP{c}.mean);
% wo/ approximation
% mustarc = model.GP{c}.mean + V*(model.Y(:)-(Kc+diag(Bc))*inv(Kc+model.Likelihood.minSigma2*eye(N))*model.GP{c}.mean*ones(N,1));
sigma2starc = Kss - sum(V.*Kstar',2) + model.Likelihood.sigma2;
yp = yp + post(:,c).*mustarc;
if robotics
out1 = out1 + post(:,c).^2.*sigma2starc;
else
out1 = out1 + post(:,c).*(mustarc.^2+sigma2starc);
end
end
if robotics
sig2 = out1;
else
sig2 = out1 - yp.^2;
end
if nargout > 2
ypc = zeros(Nstar,C);
for c = 1:C
Kc = feval(model.GP{c}.covfunc, model.GP{c}.logtheta, model.X);
[Kss, Kstar] = feval(model.GP{c}.covfunc, model.GP{c}.logtheta, model.X, Xtest);
Bc = vardist.B(:,c);
Lc = chol(Kc+diag(Bc),'lower');
V = (Lc'\(Lc\(Kstar)))';
ypc(:,c) = model.GP{c}.mean + V*(model.Y(:)-model.GP{c}.mean);
end
end