I was trying to write a custom SoftMax layer, but I had some problems writing buildNet at the end.The error message is as follows:The Abstract class cannot be instantiated.

classdef mySoftmaxLayer < nnet.layer.Layer

% & nnet.layer.Acceleratable

properties (Learnable)

% Layer learnable parameters

% Scaling coefficients

Weights

end

methods

function layer = mySoftmaxLayer(numInputs,args)

% layer = preluLayer(numChannels) creates a PReLU layer

% with numChannels channels.

%

% layer = preluLayer(numChannels,Name=name) also specifies the

% layer name

%arguments

% numInputs

% args.Name = "";

%end

layer.NumInputs = numInputs;

layer.Name = name;

layer.Description = "My softmax layer";

layer.Weights =zeros;

...

end

%function Z = forward(layer,varargin)

%

%end

function varargout = predict(layer,varargin)

X = varargin;

%Z = varargout;

%X1 = X{1};

%sz = size(X1);

%for n =1:layer.NumInputs

% varargout{n} = zeros(sz,'like',X1);

%end

sz = size(X);

varargout = zeros(sz,'like',X);

for n=1:layer.NumInputs

layer.Weights = layer.Weight + exp(X{n})

end

for n =1:layer.NumInputs

varargout{n} = exp(varargin{n})/layer.Weights;

end

%varargout = Z;

end

end

end

Answer 1

[Abhaya]

Hi @he he,

The error ‘The Abstract class cannot be instantiated’ indicates that the 'mySoftmaxLayer' class, which is extending MATLAB nnnet.layer.Layer class, is abstract in nature and can not be instantiated.

In MATLAB R2018a, when creating a custom layer by extending the nnet.layer.Layer class, you are required to implement the two abstract methods: predict and backward. Without defining these methods, the class remains abstract and cannot be instantiated.

To resolve this issue, you need to define both the predict and backward functions within the 'mySoftmaxLayer' class.

Please refer to the example below, which demonstrates how to implement a custom layer.

classdef mySoftmaxLayer < nnet.layer.Layer
    % Custom softmax layer that extends nnet.layer.Layer
    methods
        function layer = mySoftmaxLayer(name)
            layer.Name = name;
            layer.Description = 'Softmax Layer';
        end
        
        function Z = predict(layer, X)
            Z = exp(X - max(X, [], 1)); 
            Z = Z ./ sum(Z, 1);         
        end
        
        function gradients = backward(layer, X, ~, state,)
            S = predict(layer, X);  
            gradients = S .* (1 - S);  
        end
    end
end

To create an instance of the 'mySoftmaxLayer' class, you can exceute following command:

layer = mySoftmaxLayer('softmax')

Feel free to refer to the following documentation for more information on creating a custom deep learning layer:

https://www.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-layer-with-formatted-inputs.html

Hope this resolves the query.