Why appear NAN in the Mini-batch-loss and Mini-batch-RMSE when Train a Convolutional Neural Network for Regression

Iam used same code steps in following link but modified with my work

https://www.mathworks.com/help/nnet/examples/train-a-convolutional-neural-network-for-regression.html

 traindata=rtrain_csiq;
 Y = rscore;
testdata=utest_csiq;

     layers = [ ...
     imageInputLayer([256 256 1])
    convolution2dLayer(12,25)
    reluLayer
    fullyConnectedLayer(1)
    regressionLayer];

 options = trainingOptions('sgdm','InitialLearnRate',0.001, ...     'MaxEpochs',15);

        net = trainNetwork(traindata,Y,layers,options)
        predictedTest = predict(net,testdata);

but the output as following

pls how can solve that..Thanks

Answer 1

[Amy]

Hi Ismail,

Sometimes this can happen if your data includes many regressors and/or large regression response values. This leads to larger losses that can become NaNs.

Two possible solutions:

1. Try a lower initial learning rate.
2. Normalize the responses (the variable Y in your example) so that the maximum value is 1. You can use the normc function to do this.

[Ismail T. Ahmed]

thanks Amy I applied your suggestion but still give me NaN I used 0.0001 for initial learning rate and used normc function of Y.

[AlexanderTUE]

Hi Amy, hi Ismail,

I has a similar problem in the past. It seems that the use of a single convolution connected layer is not enough for such big images sizes. I used three Conv layers with intial weigths. Please see the following QA https://de.mathworks.com/matlabcentral/answers/337587-how-to-avoid-nan-in-the-mini-batch-loss-from-traning-convolutional-neural-network

Alex