simff_avr_snn.m

神经网络的工具箱, 神经网络的工具箱,

function [y_av, err_estimate] = simff_avr_snn(nets, alpha, P)
%SIMFF_AVR_SNN Compute weighted ensemble average output.
%
%  Syntax
%
%   [y_av, err_estimate] = simff_avr_snn(nets, alpha, P)
%
%  Description
%
%   SIMFF_AVR_SNN takes
%     nets   - [1 x M] struct array of net_structs with an ensemble of
%              trained networks.  
%     alpha  - [1 x M] matrix of weighting factors.  
%     P      - [N0 x MU] matrix of input patterns.
%   and returns
%     y_av   - [NL x MU] matrix of averaged outputs.
%     err_estimate - [NL x MU] matrix of ensemble average error estimate. 
%
%   (M = #networks in ensemble; N0 = #inputs; MU = #patterns)
%
%  Algorithm
%
%      See: T. Heskes; Balancing between bagging and bumping.
%      Advancesin Neural Information Processing Systems 9, pages
%      466-472, Cambridge, 1997, MIT Press. 
%
%  See also
%
%   BALANCE_SNN
%

M = size(nets,2);
N = size(nets(1).biases{nets(1).numLayers},1);
[NI, MU] = size(P);

errf = nets(1).costFcn.fn;

ym = zeros(N, MU, M);
for m = 1:M
    ym(:,:,m) = simff_snn(nets(1,m), P);    
end

alpha = alpha/(sum(alpha,1));
y_av = average_outputs_snn(ym, alpha, errf);

%#function se_snn
%#function relerr_snn
%#function loglikelihood_snn
%#function crosslogistic_snn
%#function crossentropy_snn
if (nargout > 1)
   if isstr(errf)
      err_estimate = (1/(1-sumsqr_snn(alpha))) * ...
                  reshape(...
                     reshape(...
                       feval(errf, ym, reshape(repmat(y_av,1,M), [N MU M])), ...
		       [N*MU M]) * alpha, ...
		     [N MU]);
   elseif iscell(errf)
      for i = 1:N
          err_estimate(i,:) = (1/(1-sumsqr_snn(alpha))) * ...
	          reshape(...
		     reshape(...
		       feval(errf{i}, reshape(ym(i,:,:), [MU M]), ...
		         reshape(repmat(y_av(i,:),1,M), [MU M])), ...
		       [MU M]) * alpha, ...
		     [1 MU]);
      end
   else
     error('Incorrect format for output errorfunction');
   end
end