How to implement learning rate anealing?

[AJAXJR24]

Dear @lululxvi
Hi and thanks for your help.

1. I want to know How can i implement the underlying LRA algorithm in deepxde.
2. How to define the equation number 15?

[haison19952013]

• While your question is about loss weights, not learning rate, dynamically updating them in TensorFlow is achievable.

• The key is to create custom callbacks that adjust the weights during training. This requires a deep understanding of TensorFlow's internal workings, particularly the autograph functionality.

• If diving into code complexity isn't your preference, consider alternative PINN libraries like sciann or modulus that offer built-in dynamic loss weight functionalities.

• Or you can refer to (this blog) to learn about constructing a PINN and the adaptive loss weights from the scratch

[praksharma]

No this is not implemented in DeepXDE and to be honest, LRA (learning rate annealing) isn't very effectively on many problems. Anyways it is implemented in NVIDIA Modulus. Personally, I would say that just stick to constant coefficients and use deeper networks.
Here is one of my paper where solved PDEs with discontinuous solutions without any adaptive cofficients: paper