|
wolffd@0
|
1 function [net, options, varargout] = netopt_weighted(net, options, x, t, eso_w, alg);
|
|
wolffd@0
|
2 %NETOPT Optimize the weights in a network model.
|
|
wolffd@0
|
3 %
|
|
wolffd@0
|
4 % Description
|
|
wolffd@0
|
5 %
|
|
wolffd@0
|
6 % NETOPT is a helper function which facilitates the training of
|
|
wolffd@0
|
7 % networks using the general purpose optimizers as well as sampling
|
|
wolffd@0
|
8 % from the posterior distribution of parameters using general purpose
|
|
wolffd@0
|
9 % Markov chain Monte Carlo sampling algorithms. It can be used with any
|
|
wolffd@0
|
10 % function that searches in parameter space using error and gradient
|
|
wolffd@0
|
11 % functions.
|
|
wolffd@0
|
12 %
|
|
wolffd@0
|
13 % [NET, OPTIONS] = NETOPT(NET, OPTIONS, X, T, ALG) takes a network
|
|
wolffd@0
|
14 % data structure NET, together with a vector OPTIONS of parameters
|
|
wolffd@0
|
15 % governing the behaviour of the optimization algorithm, a matrix X of
|
|
wolffd@0
|
16 % input vectors and a matrix T of target vectors, and returns the
|
|
wolffd@0
|
17 % trained network as well as an updated OPTIONS vector. The string ALG
|
|
wolffd@0
|
18 % determines which optimization algorithm (CONJGRAD, QUASINEW, SCG,
|
|
wolffd@0
|
19 % etc.) or Monte Carlo algorithm (such as HMC) will be used.
|
|
wolffd@0
|
20 %
|
|
wolffd@0
|
21 % [NET, OPTIONS, VARARGOUT] = NETOPT(NET, OPTIONS, X, T, ALG) also
|
|
wolffd@0
|
22 % returns any additional return values from the optimisation algorithm.
|
|
wolffd@0
|
23 %
|
|
wolffd@0
|
24 % See also
|
|
wolffd@0
|
25 % NETGRAD, BFGS, CONJGRAD, GRADDESC, HMC, SCG
|
|
wolffd@0
|
26 %
|
|
wolffd@0
|
27
|
|
wolffd@0
|
28 % Copyright (c) Ian T Nabney (1996-9)
|
|
wolffd@0
|
29
|
|
wolffd@0
|
30 optstring = [alg, '(''neterr_weighted'', w, options, ''netgrad_weighted'', net, x, t, eso_w)'];
|
|
wolffd@0
|
31
|
|
wolffd@0
|
32 % Extract weights from network as single vector
|
|
wolffd@0
|
33 w = netpak(net);
|
|
wolffd@0
|
34
|
|
wolffd@0
|
35 % Carry out optimisation
|
|
wolffd@0
|
36 [s{1:nargout}] = eval(optstring);
|
|
wolffd@0
|
37 w = s{1};
|
|
wolffd@0
|
38
|
|
wolffd@0
|
39 if nargout > 1
|
|
wolffd@0
|
40 options = s{2};
|
|
wolffd@0
|
41
|
|
wolffd@0
|
42 % If there are additional arguments, extract them
|
|
wolffd@0
|
43 nextra = nargout - 2;
|
|
wolffd@0
|
44 if nextra > 0
|
|
wolffd@0
|
45 for i = 1:nextra
|
|
wolffd@0
|
46 varargout{i} = s{i+2};
|
|
wolffd@0
|
47 end
|
|
wolffd@0
|
48 end
|
|
wolffd@0
|
49 end
|
|
wolffd@0
|
50
|
|
wolffd@0
|
51 % Pack the weights back into the network
|
|
wolffd@0
|
52 net = netunpak(net, w);
|
