feat: EnsembleModel accepts pretrained global models

[madtoinou]

Fixes #1775 and fixes #1785.

Summary

• EnsembleModel accept pre-trained models if they are all global and all already fitted.
• RegressionEnsembleModel.fit() accepts an additional argument retrain_forecasting_model which is False by default to avoid catastrophic forgetting. retrain_forecasting_model=True is accepted only if all the forecasting models are global and pre-trained.

Other Information

• If a NaiveEnsembleModel was instantiated with pre-trained global models, the constructor sets self._fit_called=True so that predict() can be called directly without having to call fit(). I don't know if it's the most intuitive behavior but in case of single ts training/prediction, it makes the workflow more natural.
• Added some very basics tests

[codecov-commenter]

Codecov Report

Patch coverage is 93.02% of modified lines.

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Files Changed	Coverage
darts/models/forecasting/regression_model.py	ø
darts/timeseries.py	ø
darts/utils/historical_forecasts/utils.py	50.00%
darts/models/forecasting/ensemble_model.py	90.90%
...ts/models/forecasting/regression_ensemble_model.py	93.24%
darts/models/forecasting/baselines.py	100.00%
darts/models/forecasting/forecasting_model.py	100.00%
...orical_forecasts/optimized_historical_forecasts.py	100.00%

📢 Thoughts on this report? Let us know!.

[dennisbader]

Thanks @madtoinou , this is a great start 🚀

I have a couple of points:

• let's move retrain_forecasting_models to the constructor __init__() of the base EnsembleModel
• retrain_forecasting_models should be True by default to maintain the same behavior as we had previously.
  • if retrain_forecasting_models=True, we use the forecasting_models = [model.untrained_model() for model in forecasting_models]. Like this we will not mess with any pre-trained model.
    • we can log an info if there were any pre-trained models in the original forecasting_models
• I think it would be a good idea at this point to rename models to forecasting_models in all ensemble models (base, naive) to have a common naming. This is a breaking change, but I think it is acceptable along with all the other changes of the next release.
• In theory NaiveEnsembleModel could be used without fit() if all the models were pretrained. Should we allow this?

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[dennisbader]

Really nice, thanks @madtoinou ! 🚀

Had some last suggestions mainly regarding documentation and some typos.

One thing I thought could be nice to add:

• How about we allow -1 for regression_train_n_points when retrain_forecasting_models=False? Like this whenever the user calls fit(), the regression model would be trained on the entire input?

[dennisbader]

we need to return here before when retrain=False

if not self.retrain_forecasting_models:
    return self

[madtoinou]

I decided to update the training_series attribute of the models to make the behavior of predict() more intuitive (otherwise, there might be discrepancies between the forecasting models and the regressor, especially in the prediction time indexes)

[dennisbader]

Very nice, we're almost there 🚀

Had a couple of suggestions, the main ones revolve around the regression_train_n_points=-1 where I think we can just use the series as is for the input to generate the regression model training data.

[dennisbader]

Here the same comment as I added for regression ensemble model.

[dennisbader]

Looks great now, thank for this PR @madtoinou 🚀