"Example: Define Custom Deep Learning Layer with Learnable Parameters" doesn't work

Hellow, I wrote this question due to getting error for abstract method inheritance.

The example is here: https://kr.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-layer.html

In this example, The written class file in the part named 'Completed Layer' is just used to test for the below 'check validity of Layer Using checkLayer' part:

layer = preluLayer(20,'prelu');

validInputSize = [24 24 20];

checkLayer(layer,validInputSize,'ObservationDimension',4)

But, in 'layer = preluLayer(20, 'prelu'); part, the error message was created: "Creating an instance of the Abstract class is not allowd"

Actually, I know the nnet.layer.Layer object has 2 abstract method named as 'predict' and 'backward', and this example just define 'predict' so the error will be from undefining 'backward'. Even though, I just had followed the example but I got error.

Could you please some tips to resolve this problem?

Answer 1

[Asvin Kumar]

Thanks for pointing this out. Check out the complete file by running the following command:

edit(fullfile(matlabroot,'examples','nnet','main','preluLayer.m')) 

This file contains the implementation for the backward function too.

As mentioned in the example, it is not required to implement a backward function for layers whose predict/forward functions use only dlarray supported functions. In this example, the predict uses layer.Alpha which is created using randn. The randn function has a limitation for dlarray support. It works when used along with the ‘like’ argument.

Check that out at: https://www.mathworks.com/help/deeplearning/ug/list-of-functions-with-dlarray-support.html#mw_9ef507fb-d77e-4acc-a989-ab8a37483a2c

[Asvin Kumar]

Edit: This should not be an issue in R2019b and onwards.