summarize_snn.m

神经网络的工具箱, 神经网络的工具箱,

function net = summarize_snn(nets, data, alpha)
%SUMMARIZE_SNN summarize an ensemble of nets in one net.
%
% Syntax
%
%  net = summarize_snn(nets, wcfdata, alpha)
%
% Description
%
%  SUMMARIZE_SNN summarizes an ensemble of networks by training 
%  one network on the weighted averaged output of the ensemble.
%
%  SUMMARIZE_SNN(nets, wcfdata, alpha) takes
%    nets    - [1 x M] matrix of net_structs containing ensemble of networks.
%    wcfdata - wcfdata_struct containing dataset which nets are
%              trained on.
%    alpha   - [1 x M] matrix of network weighting factors.
%  and returns 
%    net     - a net_struct with a network trained on the weighted
%              average output of the ensemble.
%
% See also
%
%   SIMFF_AVR_SNN
%

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

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

names = fieldnames(data);
for f = 1:size(names,1)
%   next two lines equal to this one: 
%   eval(['data.' names{f} '= [data.' names{f} ' data.' names{f} '];']);
    cont = subsref(data, substruct('.', names{f}));
    data = subsasgn(data, substruct('.', names{f}) , [cont cont]);
end

tr_ind = 1:MU;
vv_ind = (MU+1):(2*MU);

covie = cov(data.P(:, vv_ind)');
invcov = real(sqrtm(pinv(covie)));
data.P(:, vv_ind) = data.P(:, vv_ind) + invcov*randn(NI, MU)*0.08;

yall = zeros(N, 2*MU, M);
for m = 1:M
    yall(:,:,m) = simff_snn(nets(1,m), data);
end

data.T(:, tr_ind) = average_outputs_snn(yall(:,tr_ind,:), alpha, errf);
data.T(:, vv_ind) = average_outputs_snn(yall(:,vv_ind,:), alpha, errf);

for m = 1:M
    distance(m) = wef_snn(yall(:,:,m), data.T(:, tr_ind), g, errf); 
end
[dummy, index] = min(distance);

clear yall;

indices = {tr_ind; vv_ind};
dataLV = subset_wcfdata_snn(data, tr_ind);
dataVV = subset_wcfdata_snn(data, vv_ind);
net = train_snn(nets(1,index), dataLV, dataVV);