How to resolve "Unable to determine if experiment is for classification or regression because setup function returned invalid outputs"?

Hello,

I wanted to use Alexnet for transfer learning and in the meantime, I was going to tune initial learning rate using Baysian hyper paramter tuning in experiment manager. I have defined the function as below, even though when I run experiment it shows and error saying " Unable to determine if experiment is for classification or regression because setup function returned invalid outputs. Layers argument must be an array of layers or a layer graph." I cannot find out what I did wrong and how should I resove this issue, so any help would be highly appreciated.

function [aug_traData,aug_valData,Laye,options] = BayesOptExperiment_setup2(params)
imds=imageDatastore('C:\Users\hamedm1\Documents\MATLAB\final',...
    'IncludeSubfolders',true,'LabelSource','foldernames');
[traData, valData] = splitEachLabel(imds, 0.8, 'randomized');
N=227;
M=227;
aug_traData = augmentedImageDatastore([N M], traData);
aug_valData = augmentedImageDatastore([N M], valData);
labelCount = countEachLabel(imds);  %newly added to count number of
Num_out_class=numel(labelCount(:,1));
net = alexnet;
LL=net.Layers(1:end-3);
Lay=[LL
    fullyConnectedLayer(Num_out_class,'WeightLearnRateFactor',20,'BiasLearnRateFactor',20,'name','bb')
    softmaxLayer('name','cc')
    classificationLayer('name','dd')];
lgraph = layerGraph(Lay);
Laye=lgraph.Layers;

miniBatchSize = 18;
validationFrequency = 10;
options = trainingOptions('sgdm', ...
    'InitialLearnRate',params.InitialLearnRate, ...
    'MiniBatchSize',miniBatchSize,...
    'Shuffle','every-epoch','ValidationData',aug_valData,...
    'LearnRateSchedule','piecewise','ValidationFrequency',validationFrequency,...
    'Plots','training-progress');

end

Answer 1

[Ben]

I think the issue is that the 2nd returned output of BayesOptExperiment_setup2 needs to be the layer array or layerGraph object - in your case you've returned the validation data. As far as I can tell that shouldn't be necessary, for example see the setup function in Appendix 1 of the ofllowing example

https://www.mathworks.com/help/deeplearning/ug/experiment-using-bayesian-optimization.html

[Hamed]

Hi thanks for your answer, though I dont think it can be the main source of issue, since I used single variable for data and still had the sam problem.

Answer 2

[Michelle Patrick-Krueger]

Hello Hamed,

I was having the same problem. This helped me solve it.

The function that MATLAB gives has outputs:

function [TrainingData, layers, options] = Experiment_setup(params)

Change it to this and see if it helps:

function [XTrain, YTrain, layers, options] = Experiment_setup(params)

[Hamed]

Hi Michelle thanks for your answer, even though I already done it if you see the output of my function.