How to Define Custom Regression Output Layer with Robust (Huber) Loss

Question

Onur Kilic 2022-11-10

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1848048-how-to-define-custom-regression-output-layer-with-robust-huber-loss

编辑： Onur Kilic 2022-11-10

The following example provides a framework to define a custom regression layer for MAE loss:

https://www.mathworks.com/help/deeplearning/ug/define-custom-regression-output-layer.html

How can you define a similar layer for Huber loss to reduce the effects of outliers? I am specifically asking for when the feature input is a simple numeric array.

Thank you.

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Onur Kilic 2022-11-10

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1848048-how-to-define-custom-regression-output-layer-with-robust-huber-loss#answer_1096673

编辑：Onur Kilic 2022-11-10

在 MATLAB Online 中打开

I realized that I hadn't correctly formatted the predictions made by the network with dlarray. Below is the working regression layer based on Huber loss in case anyone needs it.

classdef huberRegressionLayer < nnet.layer.RegressionLayer ...
        & nnet.layer.Acceleratable % (Optional)
    
    methods
        function layer = huberRegressionLayer(name)           
            % layer = huberRegressionLayer(name) creates a
            % robust regression layer based on huber loss
            % and specifies the layer name.
            
            % Set layer name.
            layer.Name = name;
            % Set layer description.
            layer.Description = 'Huber loss';
        end
        function loss = forwardLoss(layer,Y,T)
            % loss = forwardLoss(layer, Y, T) returns the Huber loss between
            % the predictions Y and the training targets T.
            dlY = dlarray(Y,'CB');
            loss = huber(dlY,T,"TransitionPoint",1);
        end
    end
end