gradff_snn.m

来自「神经网络的工具箱, 神经网络的工具箱,」· M 代码 · 共 77 行

function de_dx = gradff_snn(X, net, data)
%GRADFF_SNN Derivative of cost function with respect to weights and
%           biases.
%
% Syntax
%
%  dE_dX = gradff_snn(net, data)
%  dE_dX = gradff_snn(X, net, data)
%
% Description
%
%  GRADFF_SNN takes
%    net    - a net_struct
%    data   - the data for net.costFcn.name
%    X      - a vector containing connection weights and biases (optional; 
%             default these will be taken from net)
%  and returns
%    dE_dX  - dE_dX (where E is cost function)
%
%
%  See also
%
%    GETX_SNN
%

%#function lintf_snn exptf_snn logsigtf_snn radbastf_snn tansigtf_snn 
%#function dlintf_snn dexptf_snn dlogsigtf_snn dradbastf_snn dtansigtf_snn 
%#function wcf_snn dwcf_snn 

if (nargin == 2)         % gradff_snn(net, data)
   de_dx = gradff_all(X, net);
elseif (nargin == 3)
   net = setx_snn(net, X);
   de_dx = gradff_all(net, data);
end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function de_dx = gradff_all(net, data)

M = net.numLayers;
N{1} = net.weights{1}*data.P + repmat(net.biases{1}, 1, size(data.P,2));
V{1} = feval(net.transferFcn{1}, N{1});

if (M > 1)
    for m = 2:M
       N{m} = net.weights{m}*V{m-1} + repmat(net.biases{m}, 1, size(data.P,2));
       V{m} = feval(net.transferFcn{m}, N{m});
    end
end

delta{M} = feval(feval(net.transferFcn{M}, 'deriv'), N{M}, V{M}) .* ...
           feval(feval(net.costFcn.name, 'deriv'), net, data, V{M});

for m = [(M-1):-1:1]
    delta{m} = feval(feval(net.transferFcn{m}, 'deriv'), N{m}, V{m}) .* ...
               (net.weights{m+1}'*delta{m+1});
end

de_db{1} = sum(delta{1},2);
de_dw{1} = delta{1}*data.P';
if (M > 1)
   for m = 2:M
      de_db{m} = sum(delta{m},2);
      de_dw{m} = delta{m}*(V{m-1}');
   end
end

ind_1 = 1;
for m = 1:M
   [rows, columns] = size(de_dw{m});
   ind_2 = ind_1 + rows*columns;
   de_dx([ind_1:(ind_2-1)],1) = reshape(de_dw{m}, rows*columns, 1);
   ind_1 = ind_2 + rows;
   de_dx([ind_2:(ind_1-1)],1) = de_db{m};
end