What is proper format of ValidationData of TrainingOptions in TrainNetwork?

Currently, I am building a model that has one output layer through two input layers using the Deep Learning Toolbox.

My model is learning with datastore type input value through the TrainNetwork function.

('cds_Train' is train data and has same size and format with 'cds_Vali')

I refer to https://kr.mathworks.com/help/deeplearning/ref/trainnetwork.html#d123e51195.

Traning with Datastore works well.

However, it is difficult to proceed with verification due to an issue regarding the input type of Validation Data in Training Options.

st=1; sf=1;
for ii=1:nv
    for i=1:t
        % XfTest{sf,1}=P(ii,:);
        XfVali(sf,:)=P(ii,:);
        XsVali(sf,:)=T(ii,i:t+i-1); YVali(sf,:)=T(ii,t+i); 
        sf=sf+1;
    end
    Vali_sdata(st,:)=T(ii,1:end);
    Vali_fdata(st,:)=P(ii,1:end);
    st=st+1;
end

ds_XfVali = arrayDatastore(XfVali',"IterationDimension",2);
ds_XsVali = arrayDatastore(XsVali);
ds_YVali = arrayDatastore(YVali);
cds_XVali = combine(ds_XfVali,ds_XsVali);
cds_Vali = combine(ds_XfVali,ds_XsVali,ds_YVali);

%.
%.
%.
%.

options = trainingOptions('adam', ...
    'ExecutionEnvironment','cpu', ...
    'MaxEpochs',100, ...   
    'GradientThreshold',1, ...
    'Shuffle','every-epoch', ...
    'MiniBatchSize',mb, ... 
    'VerboseFrequency',200, ...
    'InitialLearnRate',0.005, ...
    'L2Regularization',0.0000001, ...
    'ValidationData',cds_Vali , ...
    'ValidationFrequency',25, ...
    'LearnRateSchedule','none', ...
    'Verbose',0, ...
    'Plots','training-progress');

[net, info]=trainNetwork(cds_Train, lgraph, options);

Case 1) 'ValidationData',cds_Vali, ...

Case 2) 'ValidationData',{cds_XVali,ds_YVali}, ...

Case 3) 'ValidationData',{ds_XsVali, ds_XfVali,ds_YVali}, ...

If DataStore is used for that value, I have tried various cases as above, but there are continuous errors.

Please let me know some idea or give your knowledge.

Answer 1

[Avadhoot]

Hi Daemo Lee,

I understand that you are facing an error in passing the validation data to your model. The Validation data input argument can be a datastore or a cell array. In your case, the datastore can be used as validation data. You have combined "XsVali", "XfVali", and "YVali" to create a combined datastore, which is read as a 1x3 cell array by the "read" function. This causes an error with the "ValidationData" input argument, as the data must be in the form of "{predictors, targets}".

You can change the lines of code which use the “combine” function in the following manner:

cds_XVali = combine(ds_XfVali,ds_XsVali,ReadOrder='sequential'); 
cds_Vali = combine(cdsXVali,ds_YVali); 

• The first line creates a sequential datastore that contains "ds_XfVali" and "ds_XsVali" in a concatenated form.
• The second line creates a combined datastore that horizontally concatenates "cds_XVali" and "ds_YVali".
• The result will be a 1x2 cell array in the proper format for validation data.

For more information on combining datastore refer the following link: https://www.mathworks.com/help/matlab/ref/matlab.io.datastore.combine.html?s_tid=doc_ta#:~:text=imshow(imtile(dataOut))%3B-,Combine,-Datastores%20Sequentially

For more details on the proper format for Validation Data refer the following:

https://www.mathworks.com/help/deeplearning/ref/trainingoptions.html#d126e214993:~:text=the%20trainnet%20function.-,trainNetwork,-Function

I hope it helps,

Regards,

Avadhoot.