How can I validate CNN after training?

Question

Kwasi 2022-3-14

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1671244-how-can-i-validate-cnn-after-training

评论： Kwasi 2022-3-21

I have 4 samples, each sample contains about 51,000 images. I train the network but each training ends with a suden fall of the validation accuracy. I have tried to increase the minibatch size, it helped a little but now I cannot increase further because I'm running out of memory

After reading some few helps online I think I can train the network without validation data but how do I validate after training?

Thank you in advance.

% reset(gpuDevice(1));
cz1 = fullfile('v1');
    imds = imageDatastore(cz1,'LabelSource','none','IncludeSubfolders',true,'FileExtensions','.mat','ReadFcn',@(filename)customreader(filename));
tbl = countEachLabel(imds)
    [trainingSet,validationSet, testSet] = splitEachLabel(imds,0.8, 0.1, 0.1 ... 
        ...%399,49,49 ...
         ...%100,10,10 ...
        ,'randomized');
  
layers = [
    imageInputLayer([480 640 1],'Normalization', ...
    'none','Name','input')
    
   %layer 1
    convolution2dLayer([3 3],64,'Stride',[1 1],'Padding',1)
    batchNormalizationLayer('Name','BN1')
    reluLayer('Name','relu1')
    maxPooling2dLayer([3 3],'Stride',2,'Name','MP1')
    
    
    
    %layer 4
    convolution2dLayer([3 3],64,'Stride',[1 1],'Padding',1,'Name','conv4')
    batchNormalizationLayer('Name','BN4')
    reluLayer('Name','relu4')
   
    %layer 5
    convolution2dLayer([3 3],64,'Stride',[2 2],'Padding',1,'Name','conv5')
    additionLayer(2,'Name','add2')
    batchNormalizationLayer('Name','BN5')
    
    
    fullyConnectedLayer(512,'Name','fc1')
    batchNormalizationLayer('Name','BN7')
    reluLayer('Name','relu7')
   
    %warstwa 7
    fullyConnectedLayer(256,'Name','fc2')
    reluLayer('Name','relu8')
    
    %warstwa 8
    fullyConnectedLayer(4,'Name','fc3')
    softmaxLayer('Name','softmax')
    classificationLayer('Name','classif')
 ]
lgraph = layerGraph(layers);
skipConv1 = convolution2dLayer(1,64,'Stride',2,'Name','skipConv1');
lgraph = addLayers(lgraph,skipConv1);
lgraph = connectLayers(lgraph,'MP1','skipConv1');
lgraph = connectLayers(lgraph,'skipConv1','add1/in2')
skipConv2 = convolution2dLayer(1,64,'Stride',2,'Name','skipConv2');
lgraph = addLayers(lgraph,skipConv2);
lgraph = connectLayers(lgraph,'MP3','skipConv2');
lgraph = connectLayers(lgraph,'skipConv2','add2/in2')
options = trainingOptions('sgdm', ...
    'InitialLearnRate',0.001, ...
    'MaxEpochs',10, ...
    'Shuffle','every-epoch', ...
    'ValidationData',validationSet, ...
    'ValidationFrequency',4000, ...
    'ValidationPatience',1, ...
    'LearnRateSchedule','piecewise', ...
    'LearnRateDropPeriod',1,...
    'Verbose',false, ...
    'Plots','training-progress', ...
    'MiniBatchSize', 20, ...
    'CheckpointPath', 't7g');
    net = trainNetwork(trainingSet,lgraph,options);
[YPred,scores] = classify(net,testSet,'MiniBatchSize',20);
[S,I] = maxk(scores',5);
YValidation = testSet.Labels;
accuracy = sum(YPred == YValidation)/numel(YValidation)
top5 = sum(sum(tbl.Label(I)' == YValidation))/numel(YValidation)
function data = customreader(filename)
    load(filename,'frame');
end

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Joss Knight 2022-3-16

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1671244-how-can-i-validate-cnn-after-training#answer_919349

The easiest way to validate after training for classification is to do exactly what you do in your example code to check the accuracy of your test set, but with your validation set. To compute the cross-entropy loss rather than accuracy you might need to implement the crossentropy function yourself. You could just pass your validation data in instead of training data and train for a few iterations to get some numbers.

The simplest way to get rid of this unusual fall in accuracy at the final iteration is to use moving average statistics for your batch normalization layers. Set the BatchNormalizationStatistics training option to 'moving'. (See the documentation.)