Matlab CNN

数据格式

数据可以是ImageDatastore类型，也可以是4-D数组，四个维度分别是长度，宽度，通道数，第几个图片。

参数

输入层 imageInputLayer([imsize imsize 1])
卷积层1 convolution2dLayer(5,150)
激活函数层1 reluLayer()
标准化层 crossChannelNormalizationLayer(5,‘Alpha’,0.00005,‘Beta’,0.75,‘K’,1)
卷积层2 convolution2dLayer(3,300,‘Stride’,1,‘BiasLearnRateFactor’,2)
激活函数层2 reluLayer()
连接层 fullyConnectedLayer(1)
去最大值层 softmaxLayer()
分类层 classificationLayer()

回归网络最后一层必须是regressionLayer()，倒数第二层为全连接层。
对于回归网络，绘制根均方误差 (rmse) 而不是精度。

代码

%数据参数
[height,width,numChannels,~] = size(trainingImages);
imageSize = [height width numChannels];
numClasses = 10;

%结构
%输入层
inputLayer = imageInputLayer(imageSize);
%中间层
filterSize = [3 3];
numFilters = 32;
middleLayers = [ 
    convolution2dLayer(filterSize,numFilters,'Padding',1)
    reluLayer()
    crossChannelNormalizationLayer(5,'Alpha',0.00005,'Beta',0.75,'K',1)
    convolution2dLayer(filterSize,numFilters,'Padding',1)
    reluLayer()
    maxPooling2dLayer(3,'Stride',2)
    reluLayer()
    maxPooling2dLayer(3,'Stride',2)
    ];
%输出层
finalLayers = [
    fullyConnectedLayer(64)
    reluLayer()
    fullyConnectedLayer(numClasses)
    softmaxLayer()
    classificationLayer()
    ];
%各层合并
layers = [
    inputLayer
    middleLayers
    finalLayers
    ];

%参数
options = trainingOptions('sgdm', ...   
    'Momentum',0.8, ...
    'InitialLearnRate',0.001, ...    
    'LearnRateSchedule','piecewise', ...    
    'LearnRateDropFactor',0.2, ...    
    'LearnRateDropPeriod',8, ...    
    'L2Regularization',0.001, ...    
    'MaxEpochs',40, ...   
    'MiniBatchSize',128, ...   
    'Verbose',true, ...
    'Plots','training-progress');

%训练   
net = trainNetwork(trainingImages,trainingLabels,layers,options);

%测试
YTest = classify(net,testImages);

%准确率
accuracy = sum(YTest == testLabels)/numel(testLabels)