|
wolffd@0
|
1 function [e, edata, eprior, y, a] = glmerr_weighted(net, x, t, eso_w)
|
|
wolffd@0
|
2 %GLMERR Evaluate error function for generalized linear model.
|
|
wolffd@0
|
3 %
|
|
wolffd@0
|
4 % Description
|
|
wolffd@0
|
5 % E = GLMERR(NET, X, T) takes a generalized linear model data
|
|
wolffd@0
|
6 % structure NET together with a matrix X of input vectors and a matrix
|
|
wolffd@0
|
7 % T of target vectors, and evaluates the error function E. The choice
|
|
wolffd@0
|
8 % of error function corresponds to the output unit activation function.
|
|
wolffd@0
|
9 % Each row of X corresponds to one input vector and each row of T
|
|
wolffd@0
|
10 % corresponds to one target vector.
|
|
wolffd@0
|
11 %
|
|
wolffd@0
|
12 % [E, EDATA, EPRIOR, Y, A] = GLMERR(NET, X, T) also returns the data
|
|
wolffd@0
|
13 % and prior components of the total error.
|
|
wolffd@0
|
14 %
|
|
wolffd@0
|
15 % [E, EDATA, EPRIOR, Y, A] = GLMERR(NET, X) also returns a matrix Y
|
|
wolffd@0
|
16 % giving the outputs of the models and a matrix A giving the summed
|
|
wolffd@0
|
17 % inputs to each output unit, where each row corresponds to one
|
|
wolffd@0
|
18 % pattern.
|
|
wolffd@0
|
19 %
|
|
wolffd@0
|
20 % See also
|
|
wolffd@0
|
21 % GLM, GLMPAK, GLMUNPAK, GLMFWD, GLMGRAD, GLMTRAIN
|
|
wolffd@0
|
22 %
|
|
wolffd@0
|
23
|
|
wolffd@0
|
24 % Copyright (c) Ian T Nabney (1996-9)
|
|
wolffd@0
|
25
|
|
wolffd@0
|
26 % Check arguments for consistency
|
|
wolffd@0
|
27 errstring = consist(net, 'glm', x, t);
|
|
wolffd@0
|
28 if ~isempty(errstring);
|
|
wolffd@0
|
29 error(errstring);
|
|
wolffd@0
|
30 end
|
|
wolffd@0
|
31
|
|
wolffd@0
|
32 [y, a] = glmfwd(net, x);
|
|
wolffd@0
|
33
|
|
wolffd@0
|
34 %switch net.actfn
|
|
wolffd@0
|
35 switch net.outfn
|
|
wolffd@0
|
36
|
|
wolffd@0
|
37 case 'softmax' % Softmax outputs
|
|
wolffd@0
|
38
|
|
wolffd@0
|
39 nout = size(a,2);
|
|
wolffd@0
|
40 % Ensure that sum(exp(a), 2) does not overflow
|
|
wolffd@0
|
41 maxcut = log(realmax) - log(nout);
|
|
wolffd@0
|
42 % Ensure that exp(a) > 0
|
|
wolffd@0
|
43 mincut = log(realmin);
|
|
wolffd@0
|
44 a = min(a, maxcut);
|
|
wolffd@0
|
45 a = max(a, mincut);
|
|
wolffd@0
|
46 temp = exp(a);
|
|
wolffd@0
|
47 y = temp./(sum(temp, 2)*ones(1,nout));
|
|
wolffd@0
|
48 % Ensure that log(y) is computable
|
|
wolffd@0
|
49 y(y<realmin) = realmin;
|
|
wolffd@0
|
50 e_app=sum(t.*log(y),2);
|
|
wolffd@0
|
51 edata = - eso_w'*e_app;
|
|
wolffd@0
|
52
|
|
wolffd@0
|
53 otherwise
|
|
wolffd@0
|
54 error(['Unknown activation function ', net.actfn]);
|
|
wolffd@0
|
55 end
|
|
wolffd@0
|
56
|
|
wolffd@0
|
57 [e, edata, eprior] = errbayes(net, edata);
|
