Deep Belief network笔记

参考https://blog.csdn.net/Rainbow0210/article/details/53010694?utm_medium=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2.channel_param&depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2.channel_param

Deep Belief network

概念

https://blog.csdn.net/Rainbow0210/article/details/53010694?utm_medium=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2.channel_param&depth_1-utm_source=distribute.pc_relevant.none-task-blog-BlogCommendFromMachineLearnPai2-2.channel_param

https://www.cnblogs.com/kemaswill/p/3203605.html

https://www.cnblogs.com/kemaswill/p/3266026.html

代码（matlab）

• 这里用了DBN（2个RBM）提取最后一层的隐含层，再加一层Dense完成MINIST分类
• DBN可用于无监督（相当于AE），有监督（类似再继续加入Dense层）

% 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
%对DBN的初始化
%除了输入层之外有两层，每层100个神经元，即为两个受限玻尔兹曼机
dbn.sizes = [100 100];
%训练次数
opts.numepochs =   2;
%每次随机的样本数量
opts.batchsize = 100;
%更新方向，目前不知道有什么用
opts.momentum  =   0;
%学习速率
opts.alpha     =   1;
%建立DBN
dbn = dbnsetup(dbn, train_x, opts);
%训练DBN
dbn = dbntrain(dbn, train_x, opts);
%至此，已完成了DBN的训练

%unfold dbn to nn
%将DBN训练得到的数据转化为NN的形式
nn = dbnunfoldtonn(dbn, 10);

%设置NN的阈值函数为Sigmoid函数
nn.activation_function = 'sigm';

%train nn
%训练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');

‘

• 构造RBM(w,vw,b,vb,c,vc)，

function dbn = dbnsetup(dbn, x, opts)
    %n是单个样本的特征维度，784
    n = size(x, 2);
    %dbn.sizes是rbm的维度，[784 100 100]
    dbn.sizes = [n, dbn.sizes];

    %numel(dbn.sizes)返回dbn.sizes中的元素个数，对于[784 100 100]，则为3
    %初始化每个rbm
    for u = 1 : numel(dbn.sizes) - 1
        %初始化rbm的学习速率
        dbn.rbm{u}.alpha    = opts.alpha;
        %学习方向
        dbn.rbm{u}.momentum = opts.momentum;
        %第一个rbm是784-100， 第二个rbm是100-100
        %对应的连接权重，初始值全为0
        dbn.rbm{u}.W  = zeros(dbn.sizes(u + 1), dbn.sizes(u));
        %用于更新的权重，下同，不再注释
        dbn.rbm{u}.vW = zeros(dbn.sizes(u + 1), dbn.sizes(u));
        %第一个rbm是784，第二个rbm是100
        %显层的偏置值，初始值全为0
        dbn.rbm{u}.b  = zeros(dbn.sizes(u), 1);
        dbn.rbm{u}.vb = zeros(dbn.sizes(u), 1);
        %第一个rbm是100，第二个rbm是100
        %隐层的偏置值，初始值全为0
        dbn.rbm{u}.c  = zeros(dbn.sizes(u + 1), 1);
        dbn.rbm{u}.vc = zeros(dbn.sizes(u + 1), 1);
    end
end

‘

• 构造两层RBM

function dbn = dbntrain(dbn, x, opts)
    % n = 1;
    % x = train_x，60000个样本，每个维度为784，即60000*784
    %n为dbn中有几个rbm，这里n=2
    n = numel(dbn.rbm);
    %充分训练第一个rbm
    dbn.rbm{1} = rbmtrain(dbn.rbm{1}, x, opts);
    %通过第一个rbm，依次训练后续的rbm
    for i = 2 : n
        %建立rbm
        x = rbmup(dbn.rbm{i - 1}, x);
        %训练rbm
        dbn.rbm{i} = rbmtrain(dbn.rbm{i}, x, opts);
    end

end

function x = rbmup(rbm, x)
    %sigm为sigmoid函数
    %通过隐层计算下一层
    x = sigm(repmat(rbm.c', size(x, 1), 1) + x * rbm.W');
end

‘

• 每运行一次rbmtrain，返回一个RBM(w,vw,b,vb,c,vc)，因此DBN由多层RBM组成，需要把v*w’作为下一层RBM的输入
• 吉布斯抽样sigmrnd见最后的代码片

function rbm = rbmtrain(rbm, x, opts)
    %矩阵x中的元素必须是浮点数，且取值为[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为样本数量，这里m = 60000
    m = size(x, 1);
    %训练批次，每一批是opts.batchsize个样本，注意这里opts.batchsize必须整除m
    numbatches = m / opts.batchsize;

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

    %opts.numepochs，训练次数
    for i = 1 : opts.numepochs

        %随机打乱1-m的数，也就是1-m的随机数，kk是1-m的随机数向量
        kk = randperm(m);

        %训练结果的eer
        err = 0;

        %对每一批数据进行训练
        for l = 1 : numbatches
            %取出opts.batchsize个待训练的样本
            %循环结束后所有样本都进行过训练，且仅训练了一次
            batch = x(kk((l - 1) * opts.batchsize + 1 : l * opts.batchsize), :);

            %赋值给v1
            %这里v1是100*784的矩阵
            v1 = batch;
            %通过v1计算h1的概率，吉布斯抽样
            h1 = sigmrnd(repmat(rbm.c', opts.batchsize, 1) + v1 * rbm.W');
            %通过h1计算v1的概率，吉布斯抽样
            v2 = sigmrnd(repmat(rbm.b', opts.batchsize, 1) + h1 * rbm.W);
            %通过v2计算h2的概率，吉布斯抽样
            h2 = sigm(repmat(rbm.c', opts.batchsize, 1) + v2 * rbm.W');

            %至此，h1,v1,h2,v2均已计算出来，即完成了对比散度算法的大半，只剩下相应权重的更新

            %权重更新的差值计算
            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;

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

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

    end
end

function X = sigmrnd(P)
%来自https://github.com/areslp/matlab/blob/master/DeepLearnToolbox-master/util/sigmrnd.m
%     X = double(1./(1+exp(-P)))+1*randn(size(P));
    X = double(1./(1+exp(-P)) > rand(size(P)));
end