I have recently started building Neural Networks and I am currently working on a project to develop a NN for Van der Pol's equation. My problem is that I developed a sequential model with time and mu as inputs. But my loss function is custom (because I needed to compute derivatives etc) and it requires additional external custom parameters (2 arrays of input data and one scalar), which change based on input data.
My question is that, I have trained my model, but before predicting, I have to update these loss function parameters (external ones), so do I use model.compile before predicting? Because I have read alot and from what I assume, it degrades the performance of the trained model. Is there any other way to achieve this? please help.