How to give loss layers "learnable weight" instead of "constant loss_weight"

[Nick Chu]

I see that every layer can be given a constant loss_weight.
But I have multiple losses calculated from different level of side outputs.

And I want to get the weighted sum of those losses as the total loss, with the weights learned by the network itself.
Is it possible?
I have tried to add layers after the loss layers, but it seems that is not allowed......