errors_snn.m

神经网络的工具箱, 神经网络的工具箱,

function [errors, exp_errors] = errors_snn(nets, datasets, do_not_use)
%ERRORS_SNN compute cost and expected cost for nets.
%
%  Syntax
%
%   [errors, exp_errors] = errors_snn(nets, datasets, do_not_use)
%
%     nets     - net_structs
%     datasets - dataset_structs
%     do_not_use - indices for networks not to be used in computing
%                  exp_errrors
%
%     errors     - costs
%     exp_errors - expected costs 
%

if (nargin < 3)
   do_not_use = [];
end

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;

%#function se_snn
%#function relerr_snn
%#function loglikelihood_snn 
%#function crosslogistic_snn
%#function crossentropy_snn
E = zeros(N, MU, M);
for m = 1:M
    ym = simff_snn(nets(1,m), data);
    if isstr(fn)
       ii = find(~isnan(g));
       E(ii+(m-1)*N*MU) = feval(fn, ym(ii), data.T(ii)); 
    else
       for i = 1:N
           ii = find(~isnan(g(i,:)));
           E(i, ii, m) = feval(fn{i,1}, ym(i, ii), data.T(i, ii));
       end
    end
end

errors.validation = zeros(1,M);
errors.training = zeros(1,M);
for m = 1:M
    
    nu = datasets(m).val_ind;
    ii = find(i0(:,nu));
    g_tmp = g(:, nu);
    nf = 1/sum(g_tmp(find(~isnan(g_tmp))));
    tmp = nf * g(:, nu) .* E(:, nu, m);
    errors.validation(m) = sum(tmp(ii));

    nu = datasets(m).trg_ind;
    ii = find(i0(:,nu));
    g_tmp = g(:, nu);
    nf = 1/sum(g_tmp(find(~isnan(g_tmp))));
    tmp = nf * g(:, nu) .* E(:, nu, m);
    errors.training(m) = sum(tmp(ii)); 

end

if (nargin > 1)
    qval = zeros(N, MU, M);
    qtr = zeros(N, MU, M);
    for m = 1:M
        nu = datasets(m).val_ind;
	nu = setdiff(nu, do_not_use);
        qval(:, nu, m) = i0(:, nu);
        nu = datasets(m).trg_ind;
	nu = setdiff(nu, do_not_use);
        qtr(:, nu, m) = i0(:, nu);
    end
    nval = sum(qval,3); 
    ntr = sum(qtr,3); 

    E_m_avr_val = zeros(N, MU);
    ii = find(nval);
    tmp = sum(qval.*E, 3); 
    E_m_avr_val(ii) = tmp(ii)./nval(ii);
    E_m_avr_tr = zeros(N, MU);
    ii = find(ntr);
    tmp = sum(qtr.*E, 3); 
    E_m_avr_tr(ii) = tmp(ii)./ntr(ii);

    for m = 1:M

        nu = datasets(m).val_ind;
        ii = find(i0(:,nu));
        g_tmp = g(:, nu);
        nf = 1/sum(g_tmp(find(~isnan(g_tmp))));
        tmp = nf * g(:, nu) .* E_m_avr_val(:, nu);
        exp_errors.validation(m) = sum(tmp(ii));

        nu = datasets(m).trg_ind;
        ii = find(i0(:,nu));
        g_tmp = g(:, nu);
        nf = 1/sum(g_tmp(find(~isnan(g_tmp))));
        tmp = nf * g(:, nu) .* E_m_avr_tr(:, nu);
        exp_errors.training(m) = sum(tmp(ii));

    end

end