How to Realize 'Gradient Reversal Layer' ?

How can i complete a 'Gradient Reversal Layer' in matlab like in pytorch or tensorflow?

It is normally used in transfer learning network when a GAN-like loss is adopted.

Could i realize it by define a custom layer?

It is very grateful if you can offer an example of some detailed advice. Thank you for your help.

[Percy Hu]

[Percy Hu]

The gradient reversal layer has no parameters associated with it. During the forward propagation, the GRL acts as an identity transformation. During the backpropagation however, the GRL takes the gradient from the subsequent level and changes its sign, i.e., multiplies it by -1, before passing it to the preceding layer. Implementing such a layer using existing object-oriented packages for deep learning is simple, requiring only to dene procedures for the forward propagation (identity transformation), and backpropagation (multiplying by -1). The layer requires no parameter update.

Answer 1

[Philip Brown]

It looks like you should be able to do this by writing your own custom layer. See the "Intermediate Layer Template" for some code to get started.

There's a custom layer used in a visualization example which does something a little bit similar (to modify the behavior of a ReLU gradient), here.

I think the custom layer code you need looks something like this:

classdef GradientReversalLayer < nnet.layer.Layer

    methods
        function Z = predict(layer, X)
            Z = X;  % Identity
        end
        
        function dLdX = backward(layer, X, dLdZ)
            dLdX = -dLdZ;  % Reverse gradient
        end
    end
    
end

If you want to define the constant you multiple the gradient by, you could make it a property of the custom layer and include that in your backward function.

[Percy Hu]

thank you very much for your help. It is really kind of you for your detailed explanation. and i hope your answer will help more people.