DBN_rbm_DBNmatlab_dbn_dbn预训练

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