etestboot_snn.m

神经网络的工具箱, 神经网络的工具箱,

function Etestval = etestboot_snn(nets, datasets)
%ETESTBOOT_SNN  Compute a bootstrap estimate for E_test.
%
% Syntax
%
%    Etest = etestboot_snn(nets, datasets)
%
% Description
%
%    nets       - [1 x M]
%      array of networks trained with costFcn = wcf_snn, 
%      with the same costFcn parameters (fn and gn) and data.
%    datasets	- [1 x M] array of dataset_structs
%
%    Etest	- [1 x M]
%      generalization error bootstrap estimate.
%
% Algorithm
%
%    See: T. Heskes; Balancing between bagging and bumping; 
%         Advances in Neural Information Processsing Systems 9.
%
% See also
%
%    TRAIN_SNN, WCF_SNN, BALANCE_SNN
%
%

data = datasets(1).data;

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

g = getg_snn(nets(1), data);
fn = nets(1).costFcn.fn;

i0 = zeros(N, MU);
i0(find(~isnan(g))) = 1;

Eimum = zeros(N, MU, M);
q = zeros(N, MU, M);
for m = 1:M
    ym = simff_snn(nets(1,m), data);
    if isstr(fn)
       ii = find(~isnan(g));
       Eimum(ii+(m-1)*N*MU) = feval(fn, ym(ii), data.T(ii)); 
    else
       for i = 1:size(ym,1)
           ii = find(~isnan(g(i,:)));
           Eimum(i, ii, m) = feval(fn{i,1}, ym(i, ii), data.T(i, ii));
       end
    end
    nu = datasets(m).val_ind;
    q(:, nu, m) = i0(:, nu);
end

n = sum(q,3); 
Eimumean = zeros(N, MU);
ii = find(n);
tmp = sum(q.*Eimum, 3); 
Eimumean(ii) = tmp(ii)./n(ii);

for m = 1:M
    mu = datasets(m).val_ind;
    nu = mu(randboot_snn(size(mu,2)));
    gval = g(:,nu);
    Eval = Eimum(:,nu,m);
    Emeanval = Eimumean(:,nu);
    ii = find(i0(:,nu));
    relE(m) = (col(gval(ii))'*col(Eval(ii))) / ...
              (col(gval(ii))'*col(Emeanval(ii)));
end

ii = find(~isnan(g));
ii = ii(:);
Emean = col(g(ii))'*col(Eimumean(ii))/sum(g(ii));  

Etestval = relE*Emean; 

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function A = col(A)
A = reshape(A, prod(size(A)), 1);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%