clc;clear;

net = alexnet; %加载在ImageNet上预训练的网络模型
imageInputSize = [227 227 3];
%加载图像
allImages = imageDatastore('.\data227Alexnet',...
 'IncludeSubfolders',true,...
 'LabelSource','foldernames');
 %划分训练集和验证集
 [training_set,validation_set] = splitEachLabel(allImages,0.7,'randomized');
 %由于原始网络全连接层1000个输出，显然不适用于我们的分类任务，因此在这里替换
layersTransfer = net.Layers(1:end-3);
categories(training_set.Labels)
numClasses = numel(categories(training_set.Labels));
%新的网络
layers = [
 layersTransfer
 fullyConnectedLayer(numClasses,'Name', 'fc','WeightLearnRateFactor',1,'BiasLearnRateFactor',1)
 softmaxLayer('Name', 'softmax')
 classificationLayer('Name', 'classOutput')];

lgraph = layerGraph(layers);
plot(lgraph)
%对数据集进行扩增
augmented_training_set = augmentedImageSource(imageInputSize,training_set);

opts = trainingOptions('adam', ...
 'MiniBatchSize', 32,... % mini batch size, limited by GPU RAM, default 100 on Titan, 500 on P6000
 'InitialLearnRate', 1e-4,... % fixed learning rate
 'LearnRateSchedule','piecewise',...
 'LearnRateDropFactor',0.25,...
 'LearnRateDropPeriod',10,...
 'L2Regularization', 1e-4,... constraint
 'MaxEpochs',20,..
 'ExecutionEnvironment', 'gpu',...
 'ValidationData', validation_set,...
 'ValidationFrequency',80,...
 'ValidationPatience',8,...
 'Plots', 'training-progress')

net = trainNetwork(augmented_training_set, lgraph, opts);

save Alex_Public_32.mat net

[predLabels,predScores] = classify(net, validation_set);
plotconfusion(validation_set.Labels, predLabels)
PerItemAccuracy = mean(predLabels == validation_set.Labels);
title(['overall per image accuracy ',num2str(round(100*PerItemAccuracy)),'%'])