MATLAB进行深度学习网络训练

---

前言

在 MATLAB 中进行深度学习网络训练主要分为数据准备、网络构建、训练配置和模型评估四个核心步骤。以下是详细教程：

---

二、数据准备

1. 加载与组织数据

% 图像数据（以MNIST为例）
digitDatasetPath = fullfile(matlabroot, 'toolbox', 'nnet', ...
    'nndemos', 'nndatasets', 'DigitDataset');
imds = imageDatastore(digitDatasetPath, ...
    'IncludeSubfolders', true, ...
    'LabelSource', 'foldernames');

% 文本数据（以IMDB影评为例）
tbl = readtable('imdb_reviews.csv');
documents = tokenizedDocument(tbl.Review);

2. 数据增强与划分

% 图像数据增强
augmenter = imageDataAugmenter(...
    'RandRotation', [-15, 15], ...
    'RandXReflection', true);
augimds = augmentedImageDatastore([224 224 3], imds, ...
    'DataAugmentation', augmenter);

% 划分训练/验证/测试集
[imdsTrain, imdsVal, imdsTest] = partitionCamVidData(imds, ...
    'TrainSet', 0.7, 'ValSet', 0.15);
    ```
# 二、网络构建
1. 自定义网络（以 CNN 为例）
```c
layers = [
    imageInputLayer([224 224 3])
    
    convolution2dLayer(3, 16, 'Padding', 'same')
    batchNormalizationLayer
    reluLayer
    
    maxPooling2dLayer(2, 'Stride', 2)
    
    convolution2dLayer(3, 32, 'Padding', 'same')
    batchNormalizationLayer
    reluLayer
    
    maxPooling2dLayer(2, 'Stride', 2)
    
    fullyConnectedLayer(128)
    reluLayer
    dropoutLayer(0.5)
    
    fullyConnectedLayer(numClasses)
    softmaxLayer
    classificationLayer
];

2. 使用预训练模型（迁移学习）

% 加载预训练ResNet-50
net = resnet50;

% 修改网络结构（替换最后几层）
lgraph = layerGraph(net);
lgraph = removeLayers(lgraph, {'fc1000', 'prob', 'ClassificationLayer_Predictions'});
lgraph = addLayers(lgraph, [
    fullyConnectedLayer(numClasses, 'Name', 'fc')
    softmaxLayer('Name', 'softmax')
    classificationLayer('Name', 'classoutput')
]);

% 连接层
lgraph = connectLayers(lgraph, 'avg_pool', 'fc');

三、训练配置与执行

1. 设置训练选项

options = trainingOptions('sgdm', ...
    'InitialLearnRate', 0.001, ...
    'MaxEpochs', 10, ...
    'MiniBatchSize', 64, ...
    'Shuffle', 'every-epoch', ...
    'ValidationData', imdsVal, ...
    'ValidationFrequency', 30, ...
    'Verbose', false, ...
    'Plots', 'training-progress', ...
    'LearnRateSchedule', 'piecewise', ...
    'LearnRateDropFactor', 0.1, ...
    'LearnRateDropPeriod', 5);
    ```
2. 执行训练
```c
% 训练网络
net = trainNetwork(imdsTrain, layers, options);

% 迁移学习训练（冻结部分层）
options = trainingOptions('sgdm', ...
    'InitialLearnRate', 0.0001, ...
    'LayersToFreeze', 'all-10');  % 冻结前10层
net = trainNetwork(imdsTrain, lgraph, options);

四、模型评估与优化

1. 性能评估

% 在测试集上评估
YPred = classify(net, imdsTest);
YTest = imdsTest.Labels;
accuracy = mean(YPred == YTest);
fprintf('测试集准确率: %.2f%%\n', accuracy*100);

% 混淆矩阵
cm = confusionmat(YTest, YPred);
figure
confusionchart(cm, categories(YTest));

2. 模型优化

% 超参数优化
hyperparams = struct(...
    'LearnRate', optimizableVariable('log', [1e-4, 1e-2]), ...
    'BatchSize', optimizableVariable('discrete', [32, 64, 128]));

results = hyperparameterOptimization(@myTrainingFcn, hyperparams, ...
    'MaxObjectiveEvaluations', 20);
    

五、高级技巧

1. GPU 加速

options = trainingOptions('sgdm', ...
    'ExecutionEnvironment', 'gpu');

2. 模型解释

% 类激活映射（CAM）
I = imread('test_image.jpg');
[YPred, scores] = classify(net, I, 'OutputAs', 'probabilities');
cam = activation(net, I, 'fc7', 'OutputAs', 'image');
imshow(cam, 'AlphaData', cam)
title(string(YPred));

3. 模型部署

% 导出为ONNX格式
exportONNXNetwork(net, 'model.onnx');

% 生成C/C++代码
codegen -config:deeplearning net -args {zeros([224 224 3 'single'])};

六、实战案例：COVID-19 肺部 CT 图像分类

% 加载数据
imds = imageDatastore('covid_ct_scans', ...
    'IncludeSubfolders', true, ...
    'LabelSource', 'foldernames');

% 数据增强
augmenter = imageDataAugmenter(...
    'RandRotation', [-10, 10], ...
    'RandXScale', [0.9, 1.1], ...
    'RandYScale', [0.9, 1.1]);
augimds = augmentedImageDatastore([224 224 3], imds, ...
    'DataAugmentation', augmenter);

% 加载预训练ResNet-18
net = resnet18;
lgraph = layerGraph(net);

% 修改网络
lgraph = removeLayers(lgraph, {'fc1000', 'prob', 'ClassificationLayer_Predictions'});
lgraph = addLayers(lgraph, [
    globalAveragePooling2dLayer('Name', 'globalpool')
    fullyConnectedLayer(2, 'Name', 'fc')
    softmaxLayer('Name', 'softmax')
    classificationLayer('Name', 'classoutput')
]);
lgraph = connectLayers(lgraph, 'avg_pool', 'globalpool');

% 训练选项
options = trainingOptions('sgdm', ...
    'InitialLearnRate', 0.0001, ...
    'MaxEpochs', 15, ...
    'MiniBatchSize', 16, ...
    'ValidationData', augimds, ...
    'ValidationFrequency', 30, ...
    'Verbose', false, ...
    'Plots', 'training-progress');

% 训练网络
net = trainNetwork(augimds, lgraph, options);

% 评估模型
YPred = classify(net, augimds);
YTest = augimds.Labels;
accuracy = mean(YPred == YTest);
fprintf('CT图像分类准确率: %.2f%%\n', accuracy*100);