How to import input into the custom forwardLoss and backwardLoss function?

In my custom regression output layer, both of the forwardLoss and backwardLoss function depend on the input X. function loss = forwardLoss(layer,Y T) , function dLdY=backwardLoss(layer,Y，T）. How can I import input X into the forwardLoss and backwardLoss function.

[BartGoris]

I believe it is not possible. A solution to this is to use an additional layer just before the regressionOutputLayer. This layer can handle additional inputs and can output the actuall loss image that you would like. Next, you can create a custom regression layer which just accepts the already calculated loss and sums all the elements of it.

[Rebecca Plant]

Hi Hongbo Xu, I'm struggling with the same thing. Did you find a solution that worked for you?

[Henrik Wenzel]

Hi there, did anybody find a solution for this?

Answer 1

[Ayush Aniket]

In addition to @BartGoris's approach, another way to achieve this is to write your own training loop using dlnetwork and dlfeval. This way, you can compute the loss with access to X, Y, and T as needed, as shown below:

% Forward pass
Y = forward(net, X);

% Compute custom loss using X, Y, and T
loss = myCustomLoss(X, Y, T);

% Backward pass
gradients = dlgradient(loss, net.Learnables);

Refer to the following documentation link to read more about the process of writing custom training loops: https://www.mathworks.com/help/deeplearning/ug/define-model-gradients-function-for-custom-training-loop.html