CNN With unbalanced Data

Question

Mohanad Alkhodari 2020-6-21

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/552412-cnn-with-unbalanced-data

评论： Kenta 2020-7-11

Hello

I have a question regarding CNN in MATLAB

I have a dataset with imbalanced classes. 70 for A 20 for B and 10 for C. How can I reduce the effect of this imbalance during training. The input images are binary images

When performing a leave-one-out scheme, the prediction is mostly going to be A every time.

The network archetecture am using is the exact same as AlexNet network

image_size = 512;
layers = [
    imageInputLayer([image_size image_size 1],'Normalization','none')
    
    convolution2dLayer(11,96,'Stride',4,'Padding',0)
    reluLayer
    crossChannelNormalizationLayer(5)
    
    maxPooling2dLayer(3,'Stride',2)
   
    groupedConvolution2dLayer(5,128,2,'Stride',1,'Padding',2)
    reluLayer
    crossChannelNormalizationLayer(5)
    
    maxPooling2dLayer(3,'Stride',2)
    convolution2dLayer(3,384,'Stride',1,'Padding',1)
    reluLayer
    
    groupedConvolution2dLayer(3,192,2,'Stride',1,'Padding',1)
    reluLayer
    groupedConvolution2dLayer(3,128,2,'Stride',1,'Padding',1)
    reluLayer
    
    maxPooling2dLayer(3,'Stride',2)
    
    fullyConnectedLayer(4096)
    reluLayer
    dropoutLayer(0.5)
    
    fullyConnectedLayer(4096)
    reluLayer
    dropoutLayer(0.5)
    
    fullyConnectedLayer(3,'WeightLearnRateFactor',1,'BiasLearnRateFactor',1)
    softmaxLayer
    classificationLayer];
% analyzeNetwork(layers)
options = trainingOptions('sgdm',...
    'ExecutionEnvironment','gpu',...
    'Minibatchsize',10,...
    'MaxEpochs',64,...
    'InitialLearnRate',0.0001,...
    'Shuffle','every-epoch',...
    'Verbose',false,...
    'Plots','training-progress');

My main question is how to make the network predicting the other two classes as well instead of just shooting for A because it is the highest amount of data. Is there a command to use or something wrong with the network?

I do not prefer data augmentation for the dataset am using as the shape and size are important factors.

Thank you