Adding Regularization to the PINN Loss function

[mathlab]

Adding Regularization to the PINN Loss function

[dario-coscia]

👋🏻 @yorkiva Thank you for your comment. The Condition class is used to define the condition to apply (e.g. function) and where to apply it (e.g. location). The L2 regulariser on the weights can be accomplished by setting the regularizer parameter in the PINN class.

Eventually, I agree with you that a class Loss should be implemented, to enable the user to be more flexible. We will release soon a beta version of the software where we plan to introduce these features, but for mantainability we can not merge these features on the current version.

[dario-coscia]

Hello @yorkiva :)

Just wanted to let you know that in the beta version we will soon release the possibility to add a custom loss for the PINN (#105), and other very cool features such as gradient clipping, batch gradient accumulation, ... since we will use lightining Trainer module in backhand to train the PINN .

Thank you for the very useful feedbacks 😄

[danielskatz]

Has this been resolved?

[yorkiva]

The regularizer in the PINN class only implements the L2 regularization. It should also be extended to incorporate L1 regularization. I have had some experience with comparing these regularizations and I have found that for some problems L1 regularization can be quite useful when the training data (i.e. boundary/initial condition) is noisy.
Since the authors mention that a general functionality for regularization would be added in the upcoming release, this issue can be closed for now.