self-adaptive weight for the loss

[kmache]

Dear Lu Lu, I hope this message finds you well. I am using deepxde for my project. I want to make the loss component's weights trainable. let's say my loss is L(theta) = w1L1 + w2L2 + etc.. I write it like L(theta) = L1(theta., m(w1)) + L2(theta., m(w2)) + etc.. where m() is a non-negative, increasing differentiable function The objective is then minmaxL(theta), the min is taken over theta = (W, b) the parameters of the model, and the max is taken over (w1, w2) the loss components weights.

How can I implement this in deepxde? please

[lululxvi]

There is some code related in https://github.com/lu-group/deeponet-extrapolation