Custom Layer- Incorrent number of outputs

Hi,

I'm trying to create a custom intermediate layer that can split up data. When I use checkLayer to validate the functionality it throws the error: "Incorrect number of output arguments for 'predict' in Layer splitDataLayer. Expected to have 1, but instead it has 4." although I've set the number of Outputs to 4 in the constructor.

classdef splitDataLayer < nnet.layer.Layer
    
    methods
        function obj = splitDataLayer(name)
            
            obj.Name = name;
            obj.numOutputs = 4;
            obj.OutputNames = {'out1','out2','out3','out4'};
        end
        
        function [Z1, Z2, Z3, Z4] = predict(~, X)
            
            Z1 = X(1, :, :, :);
            Z2 = X(2, :, :, :);
            Z3 = X(3, :, :, :);
            Z4 = X(4, :, :, :);  
     
        end        

        function [dLdX] = backward(~,~,~,~,~,~,dLdZ1,dLdZ2,dLdZ3,dLdZ4,~)
            
            dLdX = cat(1, dLdZ1,...
                          dLdZ2,...
                          dLdZ3,...
                          dLdZ4);
            
        end
    end
end    

As can be seen above, both the number of outputs as well as the output matrix in the predict function have been set correctly. So I don't know what could be wrong about the code and cause that error.

I would be happy for any help!

Answer 1

[Maria Duarte Rosa]

Hi Valentin,

Thank you for your question.

Multi-input/Multi-output custom layers are supported from R2019a. From your error message I suspect you are on a older release.

Your layer looks good though, apart from obj.numOutputs = 4; which should be obj.NumOutputs = 4;

When I correct for that all the checkLayer tests pass in R2019a:

layer = splitDataLayer('test');
validInputSize = [4 5 20 4]; % Some arbitrary dimensions
checkLayer(layer,validInputSize,'ObservationDimension',4)
Running nnet.checklayer.TestCase
.......... .......... ....
Done nnet.checklayer.TestCase
__________

Test Summary:
	 24 Passed, 0 Failed, 0 Incomplete, 0 Skipped.
	 Time elapsed: 1.3818 seconds.

I hope this helps.

[Valentin Steininger]

Hi Maria,

I also just realized that it's not available on 2018b as I tried the "custom weighted addition layer" example from the documenation and it threw the same error.

I'm just about to upgrade to 2019a.

Thanks for the answer!