MatConvNet卷积神经网络（四）——用自己的数据训练

尝试过从Matconvnet官网上下载的已经训练好的神经网络之后，最近自己训练了能够识别果树上红苹果的神经网络。先上图。源代码放在https://github.com/YunpengZhai/MATCONVNET

10/21/2016 更新：把滑动窗的代码放到了github上（结尾为**slide）



下面分享一下经验。

以下内容看之前，希望已经阅读过Matconvnet的官方文档matconvnet-manual，或者对机器学习的一些概念、卷积神经网络的原理具备基本的了解。

现在进入正题。

构建自己的神经网络，需要完成以下三个部分：

1.准备数据。

2.设计神经网络的结构。

3.设置参数，用数据训练网络。

一、准备数据。

数据在磁盘中的存放如下图：



之后，将文件中的图片导入、格式化、划分训练集测试集交叉验证集、求均值，然后以.mat格式存储在磁盘上。

%cnn_setup_data.m

<span style="font-size:14px;">function imdb =cnn_setup_data(datadir)

inputSize =[64,64];
subdir=dir(datadir);
imdb.images.data=[];
imdb.images.labels=[];
imdb.images.set = [] ;
imdb.meta.sets = {'train', 'val', 'test'} ;
image_counter=0;
trainratio=0.8;
for i=3:length(subdir)
imdb.meta.classes(i-2) = {subdir(i).name};
imgfiles=dir(fullfile(datadir,subdir(i).name));
imgpercategory_count=length(imgfiles)-2;
disp([i-2 imgpercategory_count]);
image_counter=image_counter+imgpercategory_count;
for j=3:length(imgfiles)
img=imread(fullfile(datadir,subdir(i).name,imgfiles(j).name));
img=imresize(img, inputSize(1:2));
img=single(img);
imdb.images.data(:,:,:,end+1)=single(img);
imdb.images.labels(end+1)= i-2;
if j-2<imgpercategory_count*trainratio
imdb.images.set(end+1)=1;
else
imdb.images.set(end+1)=3;
end
end
end

dataMean=mean(imdb.images.data,4);
imdb.images.data = single(bsxfun(@minus,imdb.images.data, dataMean)) ;
imdb.images.data_mean = single(dataMean);%!!!!!!!!!!!
end</span>

二、初始化神经网络
这一部分包括了对神经网络各个层的设计（比如每一层的种类、维度、正则化，以及在训练中的一些参数等）。

%cnn_mnist_init.m

<span style="font-size:14px;">function net = cnn_mnist_init(varargin)
% CNN_MNIST_LENET Initialize a CNN similar for MNIST
opts.batchNormalization = true ;
opts.networkType = 'simplenn' ;
opts = vl_argparse(opts, varargin) ;

rng('default');
rng(0) ;

f=1/100 ;
net.layers = {} ;
net.layers{end+1} = struct('type', 'conv', ...
'weights', {{f*randn(5,5,3,20, 'single'), zeros(1, 20, 'single')}}, ...
'stride', 1, ...
'pad', 0) ;
net.layers{end+1} = struct('type', 'pool', ...
'method', 'max', ...
'pool', [2 2], ...
'stride', 2, ...
'pad', 0) ;
net.layers{end+1} = struct('type', 'conv', ...
'weights', {{f*randn(10,10,20,50, 'single'),zeros(1,50,'single')}}, ...
'stride', 1, ...
'pad', 0) ;
net.layers{end+1} = struct('type', 'pool', ...
'method', 'max', ...
'pool', [2 2], ...
'stride', 2, ...
'pad', 0) ;
net.layers{end+1} = struct('type', 'conv', ...
'weights', {{f*randn(10,10,50,500, 'single'),  zeros(1,500,'single')}}, ...
'stride', 1, ...
'pad', 0) ;
net.layers{end+1} = struct('type', 'relu') ;
net.layers{end+1} = struct('type', 'conv', ...
'weights', {{f*randn(1,1,500,4, 'single'), zeros(1,4,'single')}}, ...
'stride', 1, ...
'pad', 0) ;
net.layers{end+1} = struct('type', 'softmaxloss') ;

% optionally switch to batch normalization
if opts.batchNormalization
net = insertBnorm(net, 1) ;
net = insertBnorm(net, 4) ;
net = insertBnorm(net, 7) ;
end

% Meta parameters
net.meta.inputSize = [64 64] ;
net.meta.trainOpts.learningRate = 0.0005 ;
net.meta.trainOpts.numEpochs = 30 ;
net.meta.trainOpts.batchSize = 200 ;

% Fill in defaul values
net = vl_simplenn_tidy(net) ;

% Switch to DagNN if requested
switch lower(opts.networkType)
case 'simplenn'
% done
case 'dagnn'
net = dagnn.DagNN.fromSimpleNN(net, 'canonicalNames', true) ;
net.addLayer('top1err', dagnn.Loss('loss', 'classerror'), ...
{'prediction', 'label'}, 'error') ;
net.addLayer('top5err', dagnn.Loss('loss', 'topkerror', ...
'opts', {'topk', 5}), {'prediction', 'label'}, 'top5err') ;
otherwise
assert(false) ;
end

% --------------------------------------------------------------------
function net = insertBnorm(net, l)
% --------------------------------------------------------------------
assert(isfield(net.layers{l}, 'weights'));
ndim = size(net.layers{l}.weights{1}, 4);
layer = struct('type', 'bnorm', ...
'weights', {{ones(ndim, 1, 'single'), zeros(ndim, 1, 'single')}}, ...
'learningRate', [1 1 0.05], ...
'weightDecay', [0 0]) ;
net.layers{l}.biases = [] ;
net.layers = horzcat(net.layers(1:l), layer, net.layers(l+1:end)) ;</span><span style="font-size:18px;">
</span>

该网络结构：



三、训练网络

%cnn_mnist.m

<span style="font-size:14px;">function [net, info] = cnn_mnist(varargin)
%CNN_MNIST  Demonstrates MatConvNet on MNIST

run(fullfile(fileparts(mfilename('fullpath')),...
'..', '..', 'matlab', 'vl_setupnn.m')) ;

opts.batchNormalization = false ;
opts.networkType = 'simplenn'
4000
;
[opts, varargin] = vl_argparse(opts, varargin) ;

sfx = opts.networkType ;
if opts.batchNormalization, sfx = [sfx '-bnorm'] ; end
datadir='E:\学习\机器学习\matconvnet-1.0-beta20\photos\multi-label';
opts.expDir = fullfile(vl_rootnn, 'data', ['mnist-zyp-' sfx]) ;
[opts, varargin] = vl_argparse(opts, varargin) ;

opts.dataDir = fullfile(vl_rootnn, 'data', 'mnist') ;
opts.imdbPath = fullfile(opts.expDir, 'imdb.mat');
opts.train = struct() ;
opts = vl_argparse(opts, varargin) ;
if ~isfield(opts.train, 'gpus'), opts.train.gpus = []; end;

% --------------------------------------------------------------------
%                                                         Prepare data
% --------------------------------------------------------------------

net = cnn_mnist_init('batchNormalization', opts.batchNormalization, ...
'networkType', opts.networkType) ;

if exist(opts.imdbPath, 'file')
imdb = load(opts.imdbPath) ;
else
imdb=cnn_setup_data(datadir);
mkdir(opts.expDir) ;
save(opts.imdbPath, '-struct', 'imdb') ;
end

net.meta.classes.name = arrayfun(@(x)sprintf('%d',x),1:2,'UniformOutput',false) ;

% --------------------------------------------------------------------
%                                                                Train
% --------------------------------------------------------------------

switch opts.networkType
case 'simplenn', trainfn = @cnn_train ;
case 'dagnn', trainfn = @cnn_train_dag ;
end

[net, info] = trainfn(net, imdb, getBatch(opts), ...
'expDir', opts.expDir, ...
net.meta.trainOpts, ...
opts.train, ...
'val', find(imdb.images.set == 3)) ;
net.meta.data_mean = imdb.images.data_mean;
net.layers{end}.class = [1] ;

% --------------------------------------------------------------------
function fn = getBatch(opts)
% --------------------------------------------------------------------
switch lower(opts.networkType)
case 'simplenn'
fn = @(x,y) getSimpleNNBatch(x,y) ;
case 'dagnn'
bopts = struct('numGpus', numel(opts.train.gpus)) ;
fn = @(x,y) getDagNNBatch(bopts,x,y) ;
end

% --------------------------------------------------------------------
function [images, labels] = getSimpleNNBatch(imdb, batch)
% --------------------------------------------------------------------
images = imdb.images.data(:,:,:,batch) ;
labels = imdb.images.labels(1,batch) ;

% --------------------------------------------------------------------
function inputs = getDagNNBatch(opts, imdb, batch)
% --------------------------------------------------------------------
images = imdb.images.data(:,:,:,batch) ;
labels = imdb.images.labels(1,batch) ;
if opts.numGpus > 0
images = gpuArray(images) ;
end
inputs = {'input', images, 'label', labels} ;</span><span style="font-size:18px;">
</span>

四、应用——测试程序

<span style="font-size:14px;">%初次运行一次，之后不再运行
%[net_bn, info_bn] = cnn_mnist('batchNormalization', true);
load('E:\学习\机器学习\matconvnet-1.0-beta20\data\mnist-zyp-simplenn-bnorm\imdb.mat');
im=imread('E:\学习\机器学习\matconvnet-1.0-beta20\photos\QQ截图20160922172145.png');
im=imresize(im,[64 64 ]);
imshow(im);
im = single(im);
im = im - images.data_mean;
res = vl_simplenn(net_bn, im,[],[],...
'accumulate', 0, ...
'mode', 'test', ...
'backPropDepth', inf, ...
'sync', 0, ...
'cudnn', 1) ;
scores = res(11).x(1,1,:);
[bestScore, best] = max(scores);
switch best
case 1
title('判断结果：不是苹果');
case 2
title('判断结果：1个苹果');
case 3
title('判断结果：2个苹果');
case 4
title('判断结果：3个苹果');
end</span><span style="font-size:18px;">
</span>

测试一下：



PS：写着写着就懒得写注释了。

配合滑动窗的话，结果如下：