wolffd@0
|
1 function jac = rbfjacob(net, x)
|
wolffd@0
|
2 %RBFJACOB Evaluate derivatives of RBF network outputs with respect to inputs.
|
wolffd@0
|
3 %
|
wolffd@0
|
4 % Description
|
wolffd@0
|
5 % G = RBFJACOB(NET, X) takes a network data structure NET and a matrix
|
wolffd@0
|
6 % of input vectors X and returns a three-index matrix G whose I, J, K
|
wolffd@0
|
7 % element contains the derivative of network output K with respect to
|
wolffd@0
|
8 % input parameter J for input pattern I.
|
wolffd@0
|
9 %
|
wolffd@0
|
10 % See also
|
wolffd@0
|
11 % RBF, RBFGRAD, RBFBKP
|
wolffd@0
|
12 %
|
wolffd@0
|
13
|
wolffd@0
|
14 % Copyright (c) Ian T Nabney (1996-2001)
|
wolffd@0
|
15
|
wolffd@0
|
16 % Check arguments for consistency
|
wolffd@0
|
17 errstring = consist(net, 'rbf', x);
|
wolffd@0
|
18 if ~isempty(errstring);
|
wolffd@0
|
19 error(errstring);
|
wolffd@0
|
20 end
|
wolffd@0
|
21
|
wolffd@0
|
22 if ~strcmp(net.outfn, 'linear')
|
wolffd@0
|
23 error('Function only implemented for linear outputs')
|
wolffd@0
|
24 end
|
wolffd@0
|
25
|
wolffd@0
|
26 [y, z, n2] = rbffwd(net, x);
|
wolffd@0
|
27
|
wolffd@0
|
28 ndata = size(x, 1);
|
wolffd@0
|
29 jac = zeros(ndata, net.nin, net.nout);
|
wolffd@0
|
30 Psi = zeros(net.nin, net.nhidden);
|
wolffd@0
|
31 % Calculate derivative of activations wrt n2
|
wolffd@0
|
32 switch net.actfn
|
wolffd@0
|
33 case 'gaussian'
|
wolffd@0
|
34 dz = -z./(ones(ndata, 1)*net.wi);
|
wolffd@0
|
35 case 'tps'
|
wolffd@0
|
36 dz = 2*(1 + log(n2+(n2==0)));
|
wolffd@0
|
37 case 'r4logr'
|
wolffd@0
|
38 dz = 2*(n2.*(1+2.*log(n2+(n2==0))));
|
wolffd@0
|
39 otherwise
|
wolffd@0
|
40 error(['Unknown activation function ', net.actfn]);
|
wolffd@0
|
41 end
|
wolffd@0
|
42
|
wolffd@0
|
43 % Ignore biases as they cannot affect Jacobian
|
wolffd@0
|
44 for n = 1:ndata
|
wolffd@0
|
45 Psi = (ones(net.nin, 1)*dz(n, :)).* ...
|
wolffd@0
|
46 (x(n, :)'*ones(1, net.nhidden) - net.c');
|
wolffd@0
|
47 % Now compute the Jacobian
|
wolffd@0
|
48 jac(n, :, :) = Psi * net.w2;
|
wolffd@0
|
49 end |