MATLAB中使用AlexNet、VGG、GoogLeNet进行迁移学习

直接贴代码，具体用法见注释：

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)),'%'])

MATLAB中训练神经网络一个非常大的优势就是训练过程中各项指标的可视化，并且最终也会生成一个混淆矩阵显示验证集的结果。