Gradient clipping with custom feed-forward net

Everytime I am training my custom feed-forward net with 2 inputs and one output( timeseries) with the train(net,....) function:

after ~10 training epochs the value of the gradient reaches the prestet value and the training stops.

Changing the networks architecture is not an option in my case.

Is there a way to implement "gradient clipping" with a feed-forward net?

Or is there any other workaround for the "exploding gradient"?

[Christoph Aistleitner]

*gradient reaches the preset maximum

Answer 1

[Vineet Joshi]

Hi

The following documentation link will provide you suitable details regarding dealing with exploding gradients in MATLAB.

Gradient Clipping: Opions for training deep learning neural networks.

Gradient Clipping: Algorithm

Hope this helps.

Thanks

[Artem Lensky]

The answer you provided is not for a custom loop. See this example https://au.mathworks.com/help/deeplearning/ug/train-network-using-custom-training-loop.html there is the following line

[loss,gradients,state] = dlfeval(@modelLoss,net,X,T);

The question is how to apply clipping to gradients. Is there are standard Matlab function can do this for me or should I implement it myself.

Answer 2

[Artem Lensky]

Please check this link that illustrates several examples on how to implement training options that you would usually define via trainingOptions() and use with trainNetwork() but for customs loops. Here is an L2 clipping example given in the link above

function gradients = thresholdL2Norm(gradients,gradientThreshold)

    gradientNorm = sqrt(sum(gradients(:).^2));
    if gradientNorm > gradientThreshold
        gradients = gradients * (gradientThreshold / gradientNorm);
    end

end

You might also find this link useful https://au.mathworks.com/help/deeplearning/ug/detect-vanishing-gradients-in-deep-neural-networks.html that discuss detection of vanishing gradients in deep neural networks.