MATLAB对Googlenet模型进行迁移学习

调用MATLAB中的Googlenet工具箱进行迁移学习。

%% 加载数据
clc;close all;clear;
Location = '';%这里输入自己的数据集
unzip('MerchData.zip');
imds = imageDatastore('MerchData',...  %若使用自己的数据集则改为Location（不加单引号）
                       'IncludeSubfolders',true,...
                       'LabelSource','foldernames');
[imdsTrain,imdsValidation] = splitEachLabel(imds,0.7,'randomized');%将数据集按7:3的比例分为训练集和测试集
%% 加载预训练网络
net = googlenet;
%% 从训练有素的网络中提取图层，并绘制图层图
lgraph = layerGraph(net);%从训练网络中提取layer graph

%绘制layer graph
% figure('Units','normalize','Position',[0.1 0.1 0.8 0.8]);
% plot(lgraph)
% net.Layers(1)

inputSize = net.Layers(1).InputSize;


%% 替换最终图层
% 为了训练Googlenet去分类新的图像，取代网络的最后三层。这三层为'loss3-classifier', 'prob', 和
% 'output'，包含如何将网络的提取的功能组合为类概率和标签的信息。在层次图中添加三层新层： a fully connected layer, a softmax layer, and a classification output layer
% 将全连接层设置为同新的数据集中类的数目相同的大小，为了使新层比传输层学习更快，增加全连接层的学习因子。
lgraph = removeLayers(lgraph,{'loss3-classifier','prob','output'});
numClasses = numel(categories(imdsTrain.Labels));
newLayers = [
              fullyConnectedLayer(numClasses,'Name','fc','weightLearnRateFactor',10,'BiasLearnRateFactor',10)
              softmaxLayer('Name','softmax')
              classificationLayer('Name','classoutput')];
lgraph = addLayers(lgraph,newLayers);

%将网络中最后一个传输层（pool5-drop_7x7_s1）连接到新层
lgraph = connectLayers(lgraph,'pool5-drop_7x7_s1','fc');

% 绘制新的图层
% figure('Units','normalized','Position',[0.3 0.3 0.4 0.4]);
% plot(lgraph)
% ylim([0,10])
 
 %% 冻结初始图层
 % 这个网络现在已经准备好训练新的图像集。或者你可以通过设置这些层的学习速率为0来“冻结”网络中早期层的权重
 %在训练过程中trainNetwork不会跟新冻结层的参数，因为冻结层的梯度不需要计算，冻结大多数初始层的权重对网络训练加速很重要。
 %如果新的数据集很小，冻结早期网络层也可以防止新的数据集过拟合。

 layers = lgraph.Layers;
 connections = lgraph.Connections;  
 layers(1:110) = freezeWeights(layers(1:110));%调用freezeWeights函数，设置开始的110层学习速率为0
 lgraph = createLgraphUsingConnections(layers,connections);%调用createLgraphUsingConnections函数，按原始顺序重新连接所有的层。


%% 训练网络
pixelRange = [-30 30];
imageAugmenter = imageDataAugmenter(...
                                    'RandXReflection',true,...
                                    'RandXTranslation',pixelRange,...
                                    'RandYTranslation',pixelRange);
%对输入数据进行数据加强
augimdsTrain = augmentedImageDatastore(inputSize(1:2),imdsTrain, ...
    'DataAugmentation',imageAugmenter);
 %  自动调整验证图像大小
augimdsValidation = augmentedImageDatastore(inputSize(1:2),imdsValidation);
 %设置训练参数
options = trainingOptions('sgdm', ...
    'MiniBatchSize',10, ...
    'MaxEpochs',6, ...
    'InitialLearnRate',1e-4, ...
    'ValidationData',augimdsValidation, ...
    'ValidationFrequency',3, ...  %设置验证频率
    'ValidationPatience',Inf, ...
    'Verbose',true ,...
    'Plots','training-progress');
 %开始训练网络
googlenetTrain = trainNetwork(augimdsTrain,lgraph,options);
 
 
%% 对验证图像进行分类
[YPred,probs] = classify(googlenetTrain,augimdsValidation);%使用训练好的网络进行分类
accuracy = mean(YPred == imdsValidation.Labels)%计算网络的精确度

%% 保存训练好的模型
 save googlenet_03 googlenetTrain;
% save  x  y;  保存训练好的模型y（注意：y为训练的模型，即y = trainNetwork()），取名为x

使用训练好的模型进行图像分类
我这里训练的模型是对细胞显微图像进行分类，包括BYST，GRAN，HYAL，MUCS，RBC，WBC，WBCC七种细胞。

%% 加载模型
clc;close all;clear;
load('-mat','E:\MATLAB_Code\googlenet_1');
%% 加载测试集
Location = 'E:\image_test\test_02';
imds = imageDatastore(Location,'includeSubfolders',true,'LabelSource','foldernames');
inputSize = googlenetTrain.Layers(1).InputSize; 
imdstest = augmentedImageDatastore(inputSize(1:2),imds);
tic;
YPred = classify(googlenetTrain,imdstest);
%使用训练好的模型对测试集进行分类
disp(['分类所用时间为：',num2str(toc),'秒']);
%% 显示分类结果，绘制混淆矩阵
byst = 'BYST';
BYST = numel(YPred,YPred == byst);
disp(['BYST = ',num2str(BYST)]);
gran = 'GRAN';
GRAN = numel(YPred,YPred == gran);
disp(['GRAN = ',num2str(GRAN)]);
hyal = 'HYAL';
HYAL = numel(YPred,YPred == hyal);
disp(['HYAL = ',num2str(HYAL)]);
mucs = 'MUCS';
MUCS = numel(YPred,YPred == mucs);
disp(['MUCS = ',num2str(MUCS)]);
rbc = 'RBC';
RBC = numel(YPred,YPred == rbc);
disp(['RBC = ',num2str(RBC)]);
wbc = 'WBC';
WBC = numel(YPred,YPred == wbc);
disp(['WBC = ',num2str(WBC)]);
wbcc = 'WBCC';
WBCC = numel(YPred,YPred == wbcc);
disp(['WBCC = ',num2str(WBCC)]);
sum = numel(YPred);
disp(['sum = ',num2str(sum)]);
%求出每个标签对应的分类数量
% numel(A)  返回数组A的数目
% numel(A,x) 返回数组A在x的条件下的数目
%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%计算精确度%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
 YTest = imds.Labels;
 accuracy = mean(YPred == YTest);
 disp(['accuracy = ',num2str(accuracy)]);
 % disp(x)   变量x的值
 % num2str(x)  将数值数值转换为表示数字的字符数组
 %%
 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%随机显示测试分类后的图片%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
idx = randperm(numel(imds.Files),16);
figure
for i = 1:16
    subplot(4,4,i);
    I = readimage(imds,idx(i));
    imshow(I);
    label = YPred(idx(i));
    title(string(label));
end

%% 绘制混淆矩阵
predictLabel = YPred;%通过训练好的模型分类后的标签
actualLabel = YTest;%原始的标签
plotconfusion(actualLabel,predictLabel,'Googlenet');%绘制混淆矩阵


%    plotconfusion(targets,outputs);绘制混淆矩阵，使用target（true）和output（predict）标签，将标签指定为分类向量或1到N的形式
%% 保存分类后的图片
x = numel(imds.Files);
% 图片保存位置
Location_BYST = 'E:\image_classification\Googlenet\BYST';
Location_GRAN = 'E:\image_classification\Googlenet\GRAN';
Location_HYAL = 'E:\image_classification\Googlenet\HYAL';
Location_MUCS = 'E:\image_classification\Googlenet\MUCS';
Location_RBC  = 'E:\image_classification\Googlenet\RBC';
Location_WBC  = 'E:\image_classification\Googlenet\WBC';
Location_WBCC = 'E:\image_classification\Googlenet\WBCC';
writePostfix = '.bmp';%图片保存后缀
for i = 1:x
    I = readimage(imds,i);
    Label = YPred(i);
    Name = YTest(i);
   switch Label
       case 'BYST'
           saveName = sprintf('%s%s%s_%d',Location_BYST,'\',Name,i,writePostfix);
           imwrite(I,saveName);
       case 'GRAN'
           saveName = sprintf('%s%s%s_%d',Location_GRAN,'\',Name,i,writePostfix);
           imwrite(I,saveName);
       case 'HYAL'
           saveName = sprintf('%s%s%s_%d',Location_HYAL,'\',Name,i,writePostfix);
           imwrite(I,saveName);
       case 'MUCS'
           saveName = sprintf('%s%s%s_%d',Location_MUCS,'\',Name,i,writePostfix);
           imwrite(I,saveName);
       case 'RBC'
           saveName = sprintf('%s%s%s_%d',Location_RBC,'\',Name,i,writePostfix);
           imwrite(I,saveName);
       case 'WBC'
           saveName = sprintf('%s%s%s_%d',Location_WBC,'\',Name,i,writePostfix);
           imwrite(I,saveName);
       case 'WBCC'
           saveName = sprintf('%s%s%s_%d',Location_WBCC,'\',Name,i,writePostfix);
           imwrite(I,saveName);
   end
    
    
end

结果：

附录
freezeWeights函数

function layers = freezeWeights(layers)

for ii = 1:size(layers,1)
    props = properties(layers(ii));
    for p = 1:numel(props)
        propName = props{p};
        if ~isempty(regexp(propName,'LearnRateFactor$','once'))
            layers(ii).(propName) = 0;
        end
    end
end
end

createLgraphUsingConnections函数

function lgraph = createLgraphUsingConnections(layers,connections)

lgraph = layerGraph();

for i = 1:numel(layers)
    lgraph = addLayers(lgraph,layers(i));
end
    

for c = 1:size(connections,1)
    lgraph = connectLayers(lgraph,connections.Source{c},connections.Destination{c});
end
end