NN training process?

why mini batch accuracy (value) graph of training is goes down during training process?

Answer 1

[Shishir Singhal]

Hi,

Mini batch accuracy should likely to increase with no. of epochs.

But for your case, there can be of multiple reasons behind this:

• Mini-batch size
• Learning rate
• cost function.
• Network Architechture
• Quality of data and lot more.

It would be better if you provide more information about the NN model you are using.

Moreover, You can also refer to this link : https://www.mathworks.com/matlabcentral/answers/484326-mini-batch-loss-and-accuracy-trends?s_tid=answers_rc1-2_p2_MLT

If your case is similar like that.

[Raza Ali]

Thank you for your reply.

plz see the detail below

Network = [

imageInputLayer([256 256 3],"Name","imageinput")

convolution2dLayer([3 3],32,"Name","conv_1","BiasLearnRateFactor",2,"Padding","same")

reluLayer("Name","relu_1")

batchNormalizationLayer("Name","batchnorm")

convolution2dLayer([3 3],64,"Name","conv_2","BiasLearnRateFactor",2,"Padding","same")

reluLayer("Name","relu_2")

transposedConv2dLayer([3 3],2,"Name","transposed-conv","Cropping","same")

softmaxLayer("Name","softmax")

dicePixelClassificationLayer("Name","dice-pixel-class")];

options = trainingOptions('sgdm', ...

'LearnRateSchedule','piecewise',...

'LearnRateDropPeriod',10,...

'LearnRateDropFactor',0.3,...

'Momentum',0.9, ...

'InitialLearnRate',1e-3, ...

'L2Regularization',0.005, ...

'MaxEpochs',30, ...

'MiniBatchSize',2, ...

'Shuffle','every-epoch', ...

'VerboseFrequency',2,...

'Plots','training-progress');