Optimizing the GRU training process using Bayesian shows errors

Hi all, I'm having a problem with optimizing GRU parameters using Bayesian optimization, the code doesn't report an error, but some iterations of the Bayesian optimization process show ERROR. What should I do about it? Can you help me out, I would greatly appreciate it if you could help me out.

Answer 1

[Alan Weiss]

The error is coming from your code. Apparently, some points visited (that have, for example, NumOfUnits = 30, InitialLearnRate = 0.8 or 0.2, L2Regularization = 0.0048 or 7.5e-6) give NaN results to your objective function or nonlinear constraint functions.

You can test this outside of bayesopt to see where your code returns NaN.

If your code is running as expected, then there is nothing wrong with ignoring the iterations that lead to errors.

Alan Weiss

MATLAB mathematical toolbox documentation

[Yuanru Zou]

Thank you very much for your help！But how to test this outside of bayesopt?

[Alan Weiss]

Well, it depends how you do the Bayesian optimization. I suppose that you are using a fit function with the OptimizeHyperparameters argument, but I don't know offhand which fit function you are using. In any case, you can usually specify which parameter values the fit function should use instead of having bayesopt vary those parameters.

For more help, I'd need more detailed information, such as you exact function call and possibly some of your data.

Alan Weiss

MATLAB mathematical toolbox documentation

[Yuanru Zou]

Hi, I have used Bayesian optimization of GRU's hyperparameters: number of neurons in the hidden layer, InitialLearnRate and L2Regularization. In which I have written the objective function as follows:

function valError = BOFunction(optVars)
	inputn_train = evalin('base', 'inputn_train');
	outputn_train = evalin('base', 'outputn_train');

	inputSize = size(inputn_train,1);	
	outputSize = size(outputn_train,1);		

	
	opt.gru = [ ...
		sequenceInputLayer(inputSize)
		gruLayer(optVars.NumOfUnits,'outputmode','sequence','name','hidden') 
		fullyConnectedLayer(outputSize) 
		regressionLayer('name','out')];
	
	opt.opts = trainingOptions('adam', ...
		'MaxEpochs',50, ...
		'GradientThreshold',1,...
		'ExecutionEnvironment','cpu',...
		'InitialLearnRate',optVars.InitialLearnRate, ...
		'L2Regularization', optVars.L2Regularization, ...
		'LearnRateSchedule','piecewise', ...
		'LearnRateDropPeriod',40, ...                
		'LearnRateDropFactor',0.2, ...
		'Verbose',0, ...	
		'Plots','none'... 
		);
		

	net = trainNetwork(inputn_train, outputn_train, opt.gru, opt.opts);


	t_sim1 = predict(net, inputn_train); 
    error = t_sim1 - outputn_train;

	valError = sqrt(mean((error).^2));

end

This is the code that calls Bayesian optimization of GRU in my main program：

ObjFcn = @BOFunction;


	optimVars = [
    optimizableVariable('NumOfUnits', [2, 50], 'Type', 'integer')
    optimizableVariable('InitialLearnRate', [1e-3, 1], 'Transform', 'log')
    optimizableVariable('L2Regularization', [1e-10, 1e-2], 'Transform', 'log')
    ];

	BayesObject = bayesopt(ObjFcn, optimVars, ...   
        'MaxTime', Inf, ...                       
        'IsObjectiveDeterministic', false, ...
        'MaxObjectiveEvaluations', 30, ...      
        'Verbose', 1, ...                       
        'UseParallel', false);

	NumOfUnits       = BayesObject.XAtMinEstimatedObjective.NumOfUnits;       
	InitialLearnRate = BayesObject.XAtMinEstimatedObjective.InitialLearnRate; 
	L2Regularization = BayesObject.XAtMinEstimatedObjective.L2Regularization; 
	

	inputSize = size(inputn_train,1);	
	outputSize = size(outputn_train,1);		
	numhidden_units = NumOfUnits;


	gru = [ ...
		sequenceInputLayer(inputSize)
		gruLayer(numhidden_units,'outputmode','sequence','name','hidden') 
		fullyConnectedLayer(outputSize) 
		regressionLayer('name','out')];
	
	opts = trainingOptions('adam', ...
		'MaxEpochs',200, ...
		'GradientThreshold',1,...
		'ExecutionEnvironment','cpu',...
		'InitialLearnRate',InitialLearnRate, ...
		'L2Regularization', L2Regularization, ...
		'LearnRateSchedule','piecewise', ...
		'Verbose',true, ...
		'Plots','training-progress'... 
		);
		
	GRUnet = trainNetwork(inputn_train,outputn_train,gru,opts);	

[Alan Weiss]

I'm sorry, but I don't know much about deep learning, so I don't think that I can help you with your code. It looks like you are training a neural network and optimizing it to get a minimal mean squared error. I don't see anything obviously wrong, but then again I don't know what would cause the network training process or something else to throw an error. Usually in these systems, there is so much random going on (from the stochastic gradient descent to the data collection process) that things can get noisy or fail for a variety of reasons. In your case, I really don't know.

Sorry.

Alan Weiss

MATLAB mathematical toolbox documentation

[Yuanru Zou]

Okay, thanks for your patience and help, I hope you're doing well at work and in good health!