Variable changes unwantedly during repeated run

Hello,

I am currently trying to build a custom Neural Network using the Deep Network Designer and was getting an error message for one of my custom layers.

In said layer I want to multiply its input + bias with the weight matrix, but for some reason after a couple cycles the input vector changes its size so that the product can't be calculated.

The layer is the following

classdef weightedAdditionLayer < nnet.layer.Layer
    % Example custom weighted addition layer.
    properties
        inputsize
        outputsize
    end    
    
    properties (Learnable)
        Weights
    end
    
    methods
        function layer = weightedAdditionLayer(a,b,name) 
            % Set layer name.
            layer.Name = name;

            % Set layer description.
            layer.Description = "Weighted addition of inputs";
            
            layer.inputsize = a;
            layer.outputsize = b;
        
            % Initialize layer weights.
            layer.Weights = rand(layer.inputsize,layer.outputsize); 
        end
        
        function Z = predict(layer, X)
            Z = layer.Weights.' * [1;X(:)];
        end
    end
end

The vector that changes size is [1;X(:)], which goes from the wanted (201 x 1) to (12513 x 1) after approx. 6 cycles every time.

Could there be any reason in and/or outside of the code that causes this behaviour?

[DGM]

I don't see anything in this code that defines or constrains X, so it can't tell us why it's a certain size. That information is somewhere else.

[Gianluca Fuwa]

After doing a lot of troubleshooting I've found out that the problem is the Input of my Network.

Its size is supposed to be (98 x 1) but for some reason the input that causes the error is (98 x 1 x 1 x 128) even though I double checked that all the arrays in my training set are of correct size.

I really don't know where this could come from since it appears to be independent from my network.

[Tarunbir Gambhir]

Can you try training your network using a dlarray with correct data format. You can convert your training data to dlarray and specify the data format, then use this dlarray for network training.

[Gianluca Fuwa]

So, I figured out that the problem was the mini-batch size in my training options which caused an input to become size (98 x 1 x 1 x 128) because the mini batch size was 128.

After changing it to 1 everything worked fine.