|
wolffd@0
|
1 function [y, a] = glmfwd(net, x)
|
|
wolffd@0
|
2 %GLMFWD Forward propagation through generalized linear model.
|
|
wolffd@0
|
3 %
|
|
wolffd@0
|
4 % Description
|
|
wolffd@0
|
5 % Y = GLMFWD(NET, X) takes a generalized linear model data structure
|
|
wolffd@0
|
6 % NET together with a matrix X of input vectors, and forward propagates
|
|
wolffd@0
|
7 % the inputs through the network to generate a matrix Y of output
|
|
wolffd@0
|
8 % vectors. Each row of X corresponds to one input vector and each row
|
|
wolffd@0
|
9 % of Y corresponds to one output vector.
|
|
wolffd@0
|
10 %
|
|
wolffd@0
|
11 % [Y, A] = GLMFWD(NET, X) also returns a matrix A giving the summed
|
|
wolffd@0
|
12 % inputs to each output unit, where each row corresponds to one
|
|
wolffd@0
|
13 % pattern.
|
|
wolffd@0
|
14 %
|
|
wolffd@0
|
15 % See also
|
|
wolffd@0
|
16 % GLM, GLMPAK, GLMUNPAK, GLMERR, GLMGRAD
|
|
wolffd@0
|
17 %
|
|
wolffd@0
|
18
|
|
wolffd@0
|
19 % Copyright (c) Ian T Nabney (1996-2001)
|
|
wolffd@0
|
20
|
|
wolffd@0
|
21 % Check arguments for consistency
|
|
wolffd@0
|
22 errstring = consist(net, 'glm', x);
|
|
wolffd@0
|
23 if ~isempty(errstring);
|
|
wolffd@0
|
24 error(errstring);
|
|
wolffd@0
|
25 end
|
|
wolffd@0
|
26
|
|
wolffd@0
|
27 ndata = size(x, 1);
|
|
wolffd@0
|
28
|
|
wolffd@0
|
29 a = x*net.w1 + ones(ndata, 1)*net.b1;
|
|
wolffd@0
|
30
|
|
wolffd@0
|
31 switch net.outfn
|
|
wolffd@0
|
32
|
|
wolffd@0
|
33 case 'linear' % Linear outputs
|
|
wolffd@0
|
34 y = a;
|
|
wolffd@0
|
35
|
|
wolffd@0
|
36 case 'logistic' % Logistic outputs
|
|
wolffd@0
|
37 % Prevent overflow and underflow: use same bounds as glmerr
|
|
wolffd@0
|
38 % Ensure that log(1-y) is computable: need exp(a) > eps
|
|
wolffd@0
|
39 maxcut = -log(eps);
|
|
wolffd@0
|
40 % Ensure that log(y) is computable
|
|
wolffd@0
|
41 mincut = -log(1/realmin - 1);
|
|
wolffd@0
|
42 a = min(a, maxcut);
|
|
wolffd@0
|
43 a = max(a, mincut);
|
|
wolffd@0
|
44 y = 1./(1 + exp(-a));
|
|
wolffd@0
|
45
|
|
wolffd@0
|
46 case 'softmax' % Softmax outputs
|
|
wolffd@0
|
47 nout = size(a,2);
|
|
wolffd@0
|
48 % Prevent overflow and underflow: use same bounds as glmerr
|
|
wolffd@0
|
49 % Ensure that sum(exp(a), 2) does not overflow
|
|
wolffd@0
|
50 maxcut = log(realmax) - log(nout);
|
|
wolffd@0
|
51 % Ensure that exp(a) > 0
|
|
wolffd@0
|
52 mincut = log(realmin);
|
|
wolffd@0
|
53 a = min(a, maxcut);
|
|
wolffd@0
|
54 a = max(a, mincut);
|
|
wolffd@0
|
55 temp = exp(a);
|
|
wolffd@0
|
56 y = temp./(sum(temp, 2)*ones(1,nout));
|
|
wolffd@0
|
57 % Ensure that log(y) is computable
|
|
wolffd@0
|
58 y(y<realmin) = realmin;
|
|
wolffd@0
|
59
|
|
wolffd@0
|
60 otherwise
|
|
wolffd@0
|
61 error(['Unknown activation function ', net.outfn]);
|
|
wolffd@0
|
62 end
|
