|
wolffd@0
|
1 function [y, z, a] = mlpfwd(net, x)
|
|
wolffd@0
|
2 %MLPFWD Forward propagation through 2-layer network.
|
|
wolffd@0
|
3 %
|
|
wolffd@0
|
4 % Description
|
|
wolffd@0
|
5 % Y = MLPFWD(NET, X) takes a network data structure NET together with a
|
|
wolffd@0
|
6 % matrix X of input vectors, and forward propagates the inputs through
|
|
wolffd@0
|
7 % the network to generate a matrix Y of output vectors. Each row of X
|
|
wolffd@0
|
8 % corresponds to one input vector and each row of Y corresponds to one
|
|
wolffd@0
|
9 % output vector.
|
|
wolffd@0
|
10 %
|
|
wolffd@0
|
11 % [Y, Z] = MLPFWD(NET, X) also generates a matrix Z of the hidden unit
|
|
wolffd@0
|
12 % activations where each row corresponds to one pattern.
|
|
wolffd@0
|
13 %
|
|
wolffd@0
|
14 % [Y, Z, A] = MLPFWD(NET, X) also returns a matrix A giving the summed
|
|
wolffd@0
|
15 % inputs to each output unit, where each row corresponds to one
|
|
wolffd@0
|
16 % pattern.
|
|
wolffd@0
|
17 %
|
|
wolffd@0
|
18 % See also
|
|
wolffd@0
|
19 % MLP, MLPPAK, MLPUNPAK, MLPERR, MLPBKP, MLPGRAD
|
|
wolffd@0
|
20 %
|
|
wolffd@0
|
21
|
|
wolffd@0
|
22 % Copyright (c) Ian T Nabney (1996-2001)
|
|
wolffd@0
|
23
|
|
wolffd@0
|
24 % Check arguments for consistency
|
|
wolffd@0
|
25 errstring = consist(net, 'mlp', x);
|
|
wolffd@0
|
26 if ~isempty(errstring);
|
|
wolffd@0
|
27 error(errstring);
|
|
wolffd@0
|
28 end
|
|
wolffd@0
|
29
|
|
wolffd@0
|
30 ndata = size(x, 1);
|
|
wolffd@0
|
31
|
|
wolffd@0
|
32 z = tanh(x*net.w1 + ones(ndata, 1)*net.b1);
|
|
wolffd@0
|
33 a = z*net.w2 + ones(ndata, 1)*net.b2;
|
|
wolffd@0
|
34
|
|
wolffd@0
|
35 switch net.outfn
|
|
wolffd@0
|
36
|
|
wolffd@0
|
37 case 'linear' % Linear outputs
|
|
wolffd@0
|
38
|
|
wolffd@0
|
39 y = a;
|
|
wolffd@0
|
40
|
|
wolffd@0
|
41 case 'logistic' % Logistic outputs
|
|
wolffd@0
|
42 % Prevent overflow and underflow: use same bounds as mlperr
|
|
wolffd@0
|
43 % Ensure that log(1-y) is computable: need exp(a) > eps
|
|
wolffd@0
|
44 maxcut = -log(eps);
|
|
wolffd@0
|
45 % Ensure that log(y) is computable
|
|
wolffd@0
|
46 mincut = -log(1/realmin - 1);
|
|
wolffd@0
|
47 a = min(a, maxcut);
|
|
wolffd@0
|
48 a = max(a, mincut);
|
|
wolffd@0
|
49 y = 1./(1 + exp(-a));
|
|
wolffd@0
|
50
|
|
wolffd@0
|
51 case 'softmax' % Softmax outputs
|
|
wolffd@0
|
52
|
|
wolffd@0
|
53 % Prevent overflow and underflow: use same bounds as glmerr
|
|
wolffd@0
|
54 % Ensure that sum(exp(a), 2) does not overflow
|
|
wolffd@0
|
55 maxcut = log(realmax) - log(net.nout);
|
|
wolffd@0
|
56 % Ensure that exp(a) > 0
|
|
wolffd@0
|
57 mincut = log(realmin);
|
|
wolffd@0
|
58 a = min(a, maxcut);
|
|
wolffd@0
|
59 a = max(a, mincut);
|
|
wolffd@0
|
60 temp = exp(a);
|
|
wolffd@0
|
61 y = temp./(sum(temp, 2)*ones(1, net.nout));
|
|
wolffd@0
|
62
|
|
wolffd@0
|
63 otherwise
|
|
wolffd@0
|
64 error(['Unknown activation function ', net.outfn]);
|
|
wolffd@0
|
65 end
|
