How to increase the training and testing accuracy in CNN training?

Question

Ravish Raj 2017-6-20

1
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/345435-how-to-increase-the-training-and-testing-accuracy-in-cnn-training

回答： Salma Hassan 2017-11-20

I am using MATLAB for CNN training. I have a data set of 27,000 images and angles corresponding to that images. My sample code is : %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

layers = [imageInputLayer([32 32 1])

convolution2dLayer(5,50)

reluLayer()

maxPooling2dLayer(2,'Stride',2)

fullyConnectedLayer(size(categories(trainAngle)))

softmaxLayer

classificationLayer];

options = trainingOptions('sgdm', 'MaxEpochs', 50,'InitialLearnRate', 0.0003);

convnet = trainNetwork(trainZ, trainAngle, layers,options);

% trainZ is my 4D matrix of images and trainAngle is 2D array of angles corresponding to images!

resultant_Train = classify(convnet,trainZ); %Training data

resultant_Valid = classify(convnet,validZ); %Validation data

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

My training accuracy is 70%

but test accuracy is only 2%;

I am completely blank what to do next. Do you have any suggestion? How can I improve my test accuracy?

Can someone also suggest how can i use adam in place of sgdm in optimizer?

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

MatlabUserN 2017-6-21

Well increase the number of layers. minimum number of network layers should be 7. Make the network denser as the name suggest deep CNN. increase the number of epochs.

请先登录，再进行评论。

请先登录，再回答此问题。