ini_proto.m

最详尽的神经网络源码

function [Wproto , yproto , lambda] = ini_proto(Xtrain , ytrain , Nproto_pclass , option)


% Initialize Prototypes Weights and class label
%
% Usage
% ------
%
% [Wproto , yproto , lambda] = ini_proto(Xtrain , ytrain , [Nproto_pclass] , [option])
%
% Inputs
% -------
%
% Xtrain            Train data (d x Ntrain)
% ytrain            Labels (1 x Ntrain), card(ytrain) = m
% Nproto_pclass     Number of prototype per class (1 x m)
%                   Nproto = sum(Nproto_pclass)
% option            =1 Wproto ~N(E[Xtrain|y=i] , Cov[Xtrain|y=i]); 2 = random vector from Xtrain 
%
%
% Outputs
% -------
%
% Wproto            Prototype Weights (d x Nproto)
% yproto            Prototype labels  (1 x Nproto)
% lambda            Lambda values
%
%  Author : S閎astien PARIS : [e-mail]
%  -------  Date : 04/09/2006


if (nargin < 4)
    
    option = 1;
    
end


[d , Ntrain] = size(Xtrain);

labels       = unique(ytrain);

m            = length(labels);

if (nargin < 3)

    Nproto_pclass = round(sqrt(Ntrain))*ones(1 , m);

end

if(size(Nproto_pclass , 2) ~= m)

    error('Nproto_pclass must be (1 x m) vector');

end

Nproto       = sum(Nproto_pclass);

Wproto       = zeros(d , Nproto);

yproto       = zeros(1 , Nproto);

co           = 1;

for i = 1 : m

    
    ind                             = (co:co + Nproto_pclass(i) - 1);
    
    indice                          = find(ytrain == labels(i));
    
    
    if (option == 1)
        
        Xi                              = Xtrain(: , indice);

        Ni                              = size(Xi , 2);

        EXi                             = sum(Xi , 2)/Ni;

        res                             = (Xi - EXi(: , ones(1 , Ni)));

        %    CovXi                           = res*res'/(Ni - 1);

        Sigmai                          = sqrt(sum(res.*res , 2)/(Ni - 1));

        ONproto                         = ones(1 , Nproto_pclass(i));


        %    Wproto(: , ind)                 = EXi(: , ONproto) + chol(CovXi)'*randn(d , Nproto_pclass(i));

        Wproto(: , ind)                 = EXi(: , ONproto) + Sigmai(: , ONproto).*randn(d , Nproto_pclass(i));

    else

        Ni                              = length(indice);
        
        tempindice                      = randperm(Ni);

        r                               = indice(tempindice(1:Nproto_pclass(i)));
        
        Wproto(: , ind)                 = Xtrain(: , r);


    end
    
    yproto(ind)                     = labels(i);

    co                              = co + Nproto_pclass(i);

end

lambda  = ones(d , 1)/d;