matlab训练卷积神经网络预测手写数字倾斜角度

加载数据集

matlab提供了导入数据集的指令，数据集是内置的，可以自行查看digitTrain4DArrayData和digitTest4DArrayData的源码找到数据集的存储位置。

[XTrain,~,YTrain]=digitTrain4DArrayData;
[XValidation,~,YValidation]=digitTest4DArrayData;

可以随机查看几张数字

numTrainImages=numel(YTrain);
figure
idx=randperm(numTrainImages,20);
for ii=1:numel(idx)
    subplot(4,5,ii)
    imshow(XTrain(:,:,:,idx(ii)))
    drawnow
end

构建网络

构建网络的代码为，其中各种指令的含义见这篇文章

model=[
    imageInputLayer([28,28,1])
    convolution2dLayer([3,3],8,'Padding','same')
    batchNormalizationLayer
    reluLayer
    averagePooling2dLayer([2,2],'Stride',2)

    convolution2dLayer([3,3],16,'Padding','same')
    batchNormalizationLayer
    reluLayer
    averagePooling2dLayer([2,2],'Stride',2)

    convolution2dLayer([3,3],32,'Padding','same')
    batchNormalizationLayer
    reluLayer

    convolution2dLayer([3,3],64,'Padding','same')
    batchNormalizationLayer
    reluLayer

     dropoutLayer(0.2)
    fullyConnectedLayer(1)
    regressionLayer % 预测倾斜角度是回归操作，不是分类操作
];

训练网络

训练网络之前需要指定训练的参数，这里设置为，其中没有说明的参数见这篇文章

options=trainingOptions("sgdm", ...
    'MiniBatchSize',128, ...
    'MaxEpochs',30, ...
   'InitialLearnRate',0.001, ... % 初始学习率
   'LearnRateSchedule','piecewise', ...  % 学习率下降方式
   'LearnRateDropFactor',0.1, ... % 学习率下降因子，就是学习率每次更新的时候都乘以0.1
   'LearnRateDropPeriod',20, ... % 学习率下降周期，每迭代20次就下降一次
   'Shuffle','every-epoch', ...
   'ValidationData',{XValidation,YValidation}, ...
   'ValidationFrequency',floor(numel(YTrain)/128), ...
   'Plots','training-progress', ...
   'verbose',true);

随后便可以训练网络

net=trainNetwork(XTrain,YTrain,model,options)
YPredicted=predict(net,XValidation);
predictError=YValidation-YPredicted;
RMSE=sqrt(mean(predictError.^2))

完整代码

[XTrain,~,YTrain]=digitTrain4DArrayData;
[XValidation,~,YValidation]=digitTest4DArrayData;
numTrainImages=numel(YTrain);
figure
idx=randperm(numTrainImages,20);
for ii=1:numel(idx)
    subplot(4,5,ii)
    imshow(XTrain(:,:,:,idx(ii)))
    drawnow
end
model=[
    imageInputLayer([28,28,1])
    convolution2dLayer([3,3],8,'Padding','same')
    batchNormalizationLayer
    reluLayer
    averagePooling2dLayer([2,2],'Stride',2)

    convolution2dLayer([3,3],16,'Padding','same')
    batchNormalizationLayer
    reluLayer
    averagePooling2dLayer([2,2],'Stride',2)

    convolution2dLayer([3,3],32,'Padding','same')
    batchNormalizationLayer
    reluLayer

    convolution2dLayer([3,3],64,'Padding','same')
    batchNormalizationLayer
    reluLayer

     dropoutLayer(0.2)
    fullyConnectedLayer(1)
    regressionLayer % 预测倾斜角度是回归操作，不是分类操作

];
options=trainingOptions("sgdm", ...
    'MiniBatchSize',128, ...
    'MaxEpochs',30, ...
   'InitialLearnRate',0.001, ...
   'LearnRateSchedule','piecewise', ...
   'LearnRateDropFactor',0.1, ...
   'LearnRateDropPeriod',20, ...
   'Shuffle','every-epoch', ...
   'ValidationData',{XValidation,YValidation}, ...
   'ValidationFrequency',floor(numel(YTrain)/128), ...
   'Plots','training-progress', ...
   'verbose',true);

net=trainNetwork(XTrain,YTrain,model,options)
YPredicted=predict(net,XValidation);
predictError=YValidation-YPredicted;
RMSE=sqrt(mean(predictError.^2))