Proposing a means to embed learning rate finder (lr_find) before fitting multivariate models

[carusyte]

idea: #1262

[elephaint]

Thanks for your contribution;

Please follow our contributing guide; we use nbdev so direct changes to .py files are generally not allowed.

I like the idea as it seems simple and straightforward, can you provide a piece of code that showcases the functionality?

[marcopeix]

The changes needed to implement this is a bit more involved than just this piece of code. I also think it doesn't make sense to have it part of the trainer_kwargs since it is technically not a trainer kwarg. I have a working implementation, but it will add another parameter to BaseModel, so we also need to add it to all models in neuralforecast.

I'll create another PR using this as a starting point.

[marcopeix]

No need of this PR, we can do:

from pytorch_lightning.callbacks import LearningRateFinder
models = [NHITS(h=12, input_size=24, max_steps=10, learning_rate=1e-3, callbacks=[LearningRateFinder()])]