深度学习中dropout策略的理解

如今有空整理一下关于深度学习中怎么加入dropout方法来防止测试过程的过拟合现象。

首先了解一下dropout的实现原理：

这些理论的解释在百度上有不少。。。。

这里重点记录一下怎么实现这一技术

参考别人的博客，主要http://www.cnblogs.com/dupuleng/articles/4340293.html

讲解一下用Matlab中的深度学习工具箱怎么实现dropout

首先要载入工具包。DeepLearn Toolbox是一个很是有用的matlab deep learning工具包，下载地址：https://github.com/rasmusbergpalm/DeepLearnToolbox

要使用它首先要将该工具包添加到matlab的搜索路径中，

一、将包复制到matlab 的toolbox中，做者的路径是D:\program Files\matlab\toolbox\

二、在matlab的命令行中输入：　　

cd D:\program Files\matlab\toolbox\deepLearnToolbox\
addpath(gepath('D:\program Files\matlab\toolbox\deepLearnToolbox-master\')
savepath   %保存，这样就不须要每次都添加一次

三、验证添加是否成功，在命令行中输入  

which saesetup

果成功就会出现，saesetup.m的路径D:\program Files\matlab\toolbox\deepLearnToolbox-master\SAE\saesetup.m 

四、使用deepLearnToolbox 工具包，作一个简单的demo，将autoencoder模型使用dropout先后的结果进行比较。

load mnist_uint8;
train_x = double(train_x(1:2000,:)) / 255;
test_x  = double(test_x(1:1000,:))  / 255;
train_y = double(train_y(1:2000,:));
test_y  = double(test_y(1:1000,:));

%% //实验一without dropout
rand('state',0)
sae = saesetup([784 100]);
sae.ae{1}.activation_function  = 'sigm';
sae.ae{1}.learningRate         =  1;
opts.numepochs = 10;
opts.batchsize = 100;
sae = saetrain(sae , train_x , opts );
visualize(sae.ae{1}.W{1}(:,2:end)');

nn = nnsetup([784 100 10]);% //初步构造了一个输入-隐含-输出层网络，其中包括了
                           % //权值的初始化，学习率，momentum，激发函数类型，
                           % //惩罚系数，dropout等

nn.W{1} = sae.ae{1}.W{1};
opts.numepochs =  10;   %  //Number of full sweeps through data
opts.batchsize = 100;  %  //Take a mean gradient step over this many samples
[nn, ~] = nntrain(nn, train_x, train_y, opts);
[er, ~] = nntest(nn, test_x, test_y);
str = sprintf('testing error rate is: %f',er);
fprintf(str);

%% //实验二：with dropout
rand('state',0)
sae = saesetup([784 100]);
sae.ae{1}.activation_function  = 'sigm';
sae.ae{1}.learningRate         =  1;

opts.numepochs = 10;
opts.bachsize = 100;
sae = saetrain(sae , train_x , opts );
figure;
visualize(sae.ae{1}.W{1}(:,2:end)');

nn = nnsetup([784 100 10]);% //初步构造了一个输入-隐含-输出层网络，其中包括了
                           % //权值的初始化，学习率，momentum，激发函数类型，
                           % //惩罚系数，dropout等
nn.dropoutFraction = 0.5;  
nn.W{1} = sae.ae{1}.W{1};
opts.numepochs =  10;   %  //Number of full sweeps through data
opts.batchsize = 100;  %  //Take a mean gradient step over this many samples
[nn, L] = nntrain(nn, train_x, train_y, opts);
[er, bad] = nntest(nn, test_x, test_y);
str = sprintf('testing error rate is: %f',er);
fprintf(str);