comparison toolboxes/FullBNT-1.0.7/netlab3.3/glmhess.m @ 0:e9a9cd732c1e tip

first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
comparison
equal deleted inserted replaced
-1:000000000000 0:e9a9cd732c1e
1 function [h, hdata] = glmhess(net, x, t, hdata)
2 %GLMHESS Evaluate the Hessian matrix for a generalised linear model.
3 %
4 % Description
5 % H = GLMHESS(NET, X, T) takes a GLM network data structure NET, a
6 % matrix X of input values, and a matrix T of target values and returns
7 % the full Hessian matrix H corresponding to the second derivatives of
8 % the negative log posterior distribution, evaluated for the current
9 % weight and bias values as defined by NET. Note that the target data
10 % is not required in the calculation, but is included to make the
11 % interface uniform with NETHESS. For linear and logistic outputs, the
12 % computation is very simple and is done (in effect) in one line in
13 % GLMTRAIN.
14 %
15 % [H, HDATA] = GLMHESS(NET, X, T) returns both the Hessian matrix H and
16 % the contribution HDATA arising from the data dependent term in the
17 % Hessian.
18 %
19 % H = GLMHESS(NET, X, T, HDATA) takes a network data structure NET, a
20 % matrix X of input values, and a matrix T of target values, together
21 % with the contribution HDATA arising from the data dependent term in
22 % the Hessian, and returns the full Hessian matrix H corresponding to
23 % the second derivatives of the negative log posterior distribution.
24 % This version saves computation time if HDATA has already been
25 % evaluated for the current weight and bias values.
26 %
27 % See also
28 % GLM, GLMTRAIN, HESSCHEK, NETHESS
29 %
30
31 % Copyright (c) Ian T Nabney (1996-2001)
32
33 % Check arguments for consistency
34 errstring = consist(net, 'glm', x, t);
35 if ~isempty(errstring);
36 error(errstring);
37 end
38
39 ndata = size(x, 1);
40 nparams = net.nwts;
41 nout = net.nout;
42 p = glmfwd(net, x);
43 inputs = [x ones(ndata, 1)];
44
45 if nargin == 3
46 hdata = zeros(nparams); % Full Hessian matrix
47 % Calculate data component of Hessian
48 switch net.outfn
49
50 case 'linear'
51 % No weighting function here
52 out_hess = [x ones(ndata, 1)]'*[x ones(ndata, 1)];
53 for j = 1:nout
54 hdata = rearrange_hess(net, j, out_hess, hdata);
55 end
56 case 'logistic'
57 % Each output is independent
58 e = ones(1, net.nin+1);
59 link_deriv = p.*(1-p);
60 out_hess = zeros(net.nin+1);
61 for j = 1:nout
62 inputs = [x ones(ndata, 1)].*(sqrt(link_deriv(:,j))*e);
63 out_hess = inputs'*inputs; % Hessian for this output
64 hdata = rearrange_hess(net, j, out_hess, hdata);
65 end
66
67 case 'softmax'
68 bb_start = nparams - nout + 1; % Start of bias weights block
69 ex_hess = zeros(nparams); % Contribution to Hessian from single example
70 for m = 1:ndata
71 X = x(m,:)'*x(m,:);
72 a = diag(p(m,:))-((p(m,:)')*p(m,:));
73 ex_hess(1:nparams-nout,1:nparams-nout) = kron(a, X);
74 ex_hess(bb_start:nparams, bb_start:nparams) = a.*ones(net.nout, net.nout);
75 temp = kron(a, x(m,:));
76 ex_hess(bb_start:nparams, 1:nparams-nout) = temp;
77 ex_hess(1:nparams-nout, bb_start:nparams) = temp';
78 hdata = hdata + ex_hess;
79 end
80 otherwise
81 error(['Unknown activation function ', net.outfn]);
82 end
83 end
84
85 [h, hdata] = hbayes(net, hdata);
86
87 function hdata = rearrange_hess(net, j, out_hess, hdata)
88
89 % Because all the biases come after all the input weights,
90 % we have to rearrange the blocks that make up the network Hessian.
91 % This function assumes that we are on the jth output and that all outputs
92 % are independent.
93
94 bb_start = net.nwts - net.nout + 1; % Start of bias weights block
95 ob_start = 1+(j-1)*net.nin; % Start of weight block for jth output
96 ob_end = j*net.nin; % End of weight block for jth output
97 b_index = bb_start+(j-1); % Index of bias weight
98 % Put input weight block in right place
99 hdata(ob_start:ob_end, ob_start:ob_end) = out_hess(1:net.nin, 1:net.nin);
100 % Put second derivative of bias weight in right place
101 hdata(b_index, b_index) = out_hess(net.nin+1, net.nin+1);
102 % Put cross terms (input weight v bias weight) in right place
103 hdata(b_index, ob_start:ob_end) = out_hess(net.nin+1,1:net.nin);
104 hdata(ob_start:ob_end, b_index) = out_hess(1:net.nin, net.nin+1);
105
106 return