DBNコード解析

回転:http://blog.csdn.net/Rainbow0210/article/details/53010694?locationNum=1&fps=1
DBNの実現(DeepLeran ToolBox): 
ここではDBNを無監督学習の枠組みとして使用し、ANNに「学習成果」を付与して分類を完成させる.
トレーニングセットは6000枚で、28*28の手書きデジタル画像です.テストセットは10000枚で、28*28の手書きデジタル画像です.対応するシングル画像の特徴次元は28*28=784です.

% function test_example_DBN
load mnist_uint8;

train_x = double(train_x) / 255;
test_x  = double(test_x)  / 255;
train_y = double(train_y);
test_y  = double(test_y);

%%  ex2 train a 100-100 hidden unit DBN and use its weights to initialize a NN
rand('state',0)
%train dbn
% DBN    
%          ，  100    ，           
dbn.sizes = [100 100];
%    
opts.numepochs =   2;
%         
opts.batchsize = 100;
%    ，         
opts.momentum  =   0;
%    
opts.alpha     =   1;
%  DBN
dbn = dbnsetup(dbn, train_x, opts);
%  DBN
dbn = dbntrain(dbn, train_x, opts);
%  ，    DBN   

%unfold dbn to nn
% DBN          NN   
nn = dbnunfoldtonn(dbn, 10);

%  NN      Sigmoid  
nn.activation_function = 'sigm';

%train nn
%  NN
opts.numepochs =  3;
opts.batchsize = 100;
nn = nntrain(nn, train_x, train_y, opts);
[er, bad] = nntest(nn, test_x, test_y);

assert(er < 0.10, 'Too big error');

1

,

4

5

,

,

8

9,

10

11

15

,

,

,

,

,

23

,

,

30,

31

,

34,

,

,

,

38,

40,

,

,

,

1

,

4

5

,

,

8

9,

10

11

15

,

,

,

,

,

23

,

,

30,

31

,

34,

,

,

,

38,

40,

,

,

,

function dbn = dbnsetup(dbn, x, opts)
    %n          ，784
    n = size(x, 2);
    %dbn.sizes rbm   ，[784 100 100]
    dbn.sizes = [n, dbn.sizes];

    %numel(dbn.sizes)  dbn.sizes      ，  [784 100 100]，  3
    %     rbm
    for u = 1 : numel(dbn.sizes) - 1
        %   rbm     
        dbn.rbm{u}.alpha    = opts.alpha;
        %    
        dbn.rbm{u}.momentum = opts.momentum;
        %   rbm 784-100，    rbm 100-100
        %       ，     0
        dbn.rbm{u}.W  = zeros(dbn.sizes(u + 1), dbn.sizes(u));
        %       ，  ，    
        dbn.rbm{u}.vW = zeros(dbn.sizes(u + 1), dbn.sizes(u));
        %   rbm 784，   rbm 100
        %      ，     0
        dbn.rbm{u}.b  = zeros(dbn.sizes(u), 1);
        dbn.rbm{u}.vb = zeros(dbn.sizes(u), 1);
        %   rbm 100，   rbm 100
        %      ，     0
        dbn.rbm{u}.c  = zeros(dbn.sizes(u + 1), 1);
        dbn.rbm{u}.vc = zeros(dbn.sizes(u + 1), 1);
    end
end

1

,

4

5

,

,

8

9,

10

11

15

,

,

,

,

,

23

,

,

1

,

4

5

,

,

8

9,

10

11

15

,

,

,

,

,

23

,

,

function dbn = dbntrain(dbn, x, opts)
    % n = 1;
    % x = train_x，60000   ，     784， 60000*784
    %n dbn    rbm，  n=2
    n = numel(dbn.rbm);
    %       rbm
    dbn.rbm{1} = rbmtrain(dbn.rbm{1}, x, opts);
    %     rbm，       rbm
    for i = 2 : n
        %  rbm
        x = rbmup(dbn.rbm{i - 1}, x);
        %  rbm
        dbn.rbm{i} = rbmtrain(dbn.rbm{i}, x, opts);
    end

end

1

,

4

5

,

,

8

9,

10

11

15

,

,

1

,

4

5

,

,

8

9,

10

11

15

,

,

function x = rbmup(rbm, x)
    %sigm sigmoid  
    %         
    x = sigm(repmat(rbm.c', size(x, 1), 1) + x * rbm.W');
end

1

,

4

5

,

1

,

4

5

,

function rbm = rbmtrain(rbm, x, opts)
    %  x          ，    [0,1]
    assert(isfloat(x), 'x must be a float');
    assert(all(x(:)>=0) && all(x(:)<=1), 'all data in x must be in [0:1]');

    %m     ，  m = 60000
    m = size(x, 1);
    %    ，    opts.batchsize   ，    opts.batchsize    m
    numbatches = m / opts.batchsize;

    %opts.batchsize     m
    assert(rem(numbatches, 1) == 0, 'numbatches not integer');

    %opts.numepochs，    
    for i = 1 : opts.numepochs

        %    1-m  ，   1-m    ，kk 1-m      
        kk = randperm(m);

        %     eer
        err = 0;

        %          
        for l = 1 : numbatches
            %  opts.batchsize       
            %               ，       
            batch = x(kk((l - 1) * opts.batchsize + 1 : l * opts.batchsize), :);

            %   v1
            %  v1 100*784   
            v1 = batch;
            %  v1  h1   ，     
            h1 = sigmrnd(repmat(rbm.c', opts.batchsize, 1) + v1 * rbm.W');
            %  h1  v1   ，     
            v2 = sigmrnd(repmat(rbm.b', opts.batchsize, 1) + h1 * rbm.W);
            %  v2  h2   ，     
            h2 = sigm(repmat(rbm.c', opts.batchsize, 1) + v2 * rbm.W');

            %  ，h1,v1,h2,v2      ，             ，          

            %         
            c1 = h1' * v1;
            c2 = h2' * v2;

            rbm.vW = rbm.momentum * rbm.vW + rbm.alpha * (c1 - c2)     / opts.batchsize;
            rbm.vb = rbm.momentum * rbm.vb + rbm.alpha * sum(v1 - v2)' / opts.batchsize;
            rbm.vc = rbm.momentum * rbm.vc + rbm.alpha * sum(h1 - h2)' / opts.batchsize;

            %    
            rbm.W = rbm.W + rbm.vW;
            rbm.b = rbm.b + rbm.vb;
            rbm.c = rbm.c + rbm.vc;

            %  err
            err = err + sum(sum((v1 - v2) .^ 2)) / opts.batchsize;
        end
        %    
        disp(['epoch ' num2str(i) '/' num2str(opts.numepochs)  '. Average reconstruction error is: ' num2str(err / numbatches)]);

    end
end