input_relevance_snn.m

来自「神经网络的工具箱, 神经网络的工具箱,」· M 代码 · 共 104 行

function [expl_var, y, ds, Xall] = input_relevance_snn(net, data, F, groups)
%INPUT_RELEVANCE_SNN Determine order of relevance of the inputs.
%
% Syntax
%
%   [expl_var, indices] = input_relevance_snn(net, data)
%   [expl_var, indices] = input_relevance_snn(net, data, F)
%   [expl_var, indices] = input_relevance_snn(net, data, F, groups)
%
% Description
%
%   INPUT_RELEVANCE_SNN takes
%       net    - the network for which the input relevance is computed
%       data   - the training data
%       F      - fisher matrix (optional)
%       groups - cell matrix with group indices (optinal)
%   and returns
%       expl_var  - the remaining explained variance
%       indices   - indices of inputs (or groups of inputs) in order
%                   of removal
%
% Algorithm
%
%   See: P. van de Laar and T. Heskes; Pruning Using Parameter and
%   Neuronal Metrics, Neural Computation, 1999.
%

if (nargin < 4)   % groups = no groups;
   N0 = size(net.weights{1}, 2);
%   groups = mat2cell_snn([1:N0], 1, ones(1,N0));
   groups = tocell_snn([1:N0]);
end

if (nargin < 3) % F = fisherff_snn
      F = fisherff_snn(net, data);
end

% invert F; if not invertible, add a small diagonal term. 
oldwarning = lastwarn;
lastwarn('');
s = warning;
warning off;
F_inv = inv(F);
warning(s);
if lastwarn 
   options.disp = 0;
   l_max = eigs(F, 1, 'LM', options);
   F_inv = inv(F + 1e-7 * l_max * eye(size(F,1)));
end
lastwarn(oldwarning)

X = getx_snn(net);
NGT = prod(size(groups));
Xall = zeros(size(X,1), NGT); 
Xall(:,1) = X;

y = zeros(NGT,1);
ds = NaN * ones(NGT, NGT);
todo = [1:NGT];
for k = 1:NGT
    for i = todo 
        dX = prune_input_snn(net, data, groups{i}, F_inv, X);
        ds(k, i) = 0.5 * dX' * F * dX;
    end
    [minima, ind] = min(ds(k,:));
    winner = ind(1);
    y(k) = winner;
    E_quad(k) = minima;

    [dX, dF_inv] = ...
           prune_input_snn(net, data, groups{winner}, F_inv, X);
    F_inv = F_inv + dF_inv;
    X = X + dX;
    Xall(:, k+1) = X;
    todo = setdiff(todo, winner);
end

%#function wcf_snn argmin_wcf_snn
y_min = feval(feval(net.costFcn.name, 'argmin'), net, data);
E_min = feval(net.costFcn.name, net, ...
            setfield(data, 'Y', repmat(y_min,1,size(data.T,2))));

%E = zeros(size(Xall,2), 1);
%for k = 1:size(Xall,2)
%    E(k) = feval(net.costFcn.name, Xall(:,k), net, data);
%end
%expl_var = 100 * (E_min - E)/E_min;

E_0 = feval(net.costFcn.name, Xall(:,1), net, data);
E = zeros(NGT, 1);
for k = 1:NGT
    E(k) = E_0 + (E_quad(k)/E_quad(NGT))*(E_min - E_0);
end
expl_var = 100 * (E_min - [E_0; E])/E_min;




function B = tocell_snn(A)
B = cell(size(A));
for i = 1:prod(size(A))
    B{i} = A(i);
end