Error using deep learning custom recurrent layer: Illegal attribute 'State'.

Hi everyone,

I'm facing an error with a custom intermediate layer using the Deep Learning Toolbox. The layer gets sequence input data, has two states and no learnable parameters. Its task is simply to output the difference between the current input and the input at the previous time step. I tried designing the layer according to the peephole LSTM example (https://de.mathworks.com/help/deeplearning/ug/define-custom-recurrent-deep-learning-layer.html). Unfortunately, when I try to construct the layer using

forwardDifferenceLayer(100,'Name','forward difference')

I get the error message:

Error using forwardDifferenceLayer

Illegal attribute 'State'.

Here's my code for the layer. Can anyone see what I'm doing wrong? Thanks a lot in advance.

classdef forwardDifferenceLayer < nnet.layer.Layer & nnet.layer.Formattable

    properties
        % (Optional) Layer properties.

        NumHiddenUnits
    end

    properties (State)
        % (Optional) Layer state parameters.

        HiddenState
        CellState
    end

    methods
        function layer = forwardDifferenceLayer(numHiddenUnits, args)
            % (Optional) Create a myLayer.
            % This function must have the same name as the class.
            
            arguments
                
                numHiddenUnits
                args.Name = '';
                
            end
            
            layer.Name = args.Name;
            layer.NumHiddenUnits = numHiddenUnits;
        end

        function [Z,hiddenState,cellState] = predict(layer,X)
            
            if isempty(layer.CellState)
                
                hiddenState = zeros(size(X));
                cellState = X;
                
            else
                
                hiddenState = X - layer.CellState;
                cellState = X;
                
            end
            
            Z = dlarray(single(hiddenState));
                
        end

        function layer = resetState(layer)
            % (Optional) Reset layer state.

            layer.HiddenState = zeros(layer.NumHiddenUnits,1);
            layer.CellState = zeros(layer.NumHiddenUnits,1);
        end
        
    end
end

Answer 1

[Swatantra Mahato]

Hi Rebecca,

The feature to "Define stateful custom layers" used in the example "Define Custom Recurrent Deep Learning Layer" was added in MATLAB R2021b as mentioned in the release notes

https://www.mathworks.com/help/deeplearning/release-notes.html

Hence 'State' is not recognized as a legal attribute

Hope this helps

[Rebecca Plant]

Hi Swatantra, thanks a lot for your answer. That totally makes sense. I'll update to 21b then.