When exactly is model.compile() necessary during training?

[zaier66]

Hi DeepXDE developers 👋

First of all, thank you for maintaining such a powerful and extensible framework.

I have a question regarding the correct usage of model.compile() across different training phases. Based on the docs and code behavior, my current understanding is:

Cases where model.compile(...) is required:

• Switching optimizers (e.g., from Adam to L-BFGS)
• Changing learning rates, even if using the same optimizer
• Updating loss_weights (e.g., when switching from warm-up to full loss training)
• Using regularization (e.g., applying L2 via weight_decay)
• Adding new training points (e.g., in RAR)
• Recompiling within custom callbacks, such as adaptive loss weighting or learning rate schedules

Is this understanding correct?

[LordPontiag]

Yes, mostly correct.