Decouple pypesto.ensemble.ensemble.Ensemble from pypesto.predict.amici_predictor.AmiciPredictor

[dweindl]

Currently, pypesto.ensemble.ensemble.Ensemble.predict is annotated to take a predictor: Callable. However, it assumes this to be an AmiciPredictor (or at least to have an amici_objective: AmiciObjective attribute), .e.g., here:

pyPESTO/pypesto/ensemble/ensemble.py

Line 855 in e389257

parameter_ids_objective = predictor.amici_objective.x_names

I don't think this is ideal. This makes it unnecessarily cumbersome to create a custom predictor, which should just be some callable that takes a parameter vector and parameter IDs and produces a PredictionResult, or an EnsemblePrediction.

I'd propose:

• Moving the functionality from pypesto.ensemble.ensemble.Ensemble._map_parameters_by_objective to the predictor
  • This gets rid of the AmiciPredictor dependency
  • However, AmiciPredictor.__call__ will then need the parameter IDs along with the parameter values. Breaking change.
  • Changing predictor: Callable to predictor: Callable[[Sequence], PredictionResult] (and respecting that)
• Consider removing pypesto.ensemble.ensemble.Ensemble.predict altogether, and rather offer that same functionality via
  WhateverPredictor.predict(ensemble: pypesto.ensemble.ensemble.Ensemble, ...). This seems more appropriate, since different predictors might want to offer different arguments.
• To facilitate parallelization without much reimplementation for every Predictor, it might make sense to introduce a Predictor base class that implements the respective functionality that's currently included in pypesto.ensemble.ensemble.Ensemble.predict .

[dweindl]

I think, before working on this, we'll need to discuss the desired ensemble functionality in pypesto.

The current ensemble prediction setup is 100% tailored to AMICI

pyPESTO/pypesto/result/predict.py

Lines 113 to 115 in 82b7d5d

	class PredictionResult:
	"""
	Light-weight wrapper around prediction from pyPESTO made by an AMICI model.

pyPESTO/pypesto/result/predict.py

Lines 28 to 34 in 82b7d5d

	class PredictionConditionResult:
	"""
	Light-weight wrapper for the prediction of one simulation condition.
	It should provide a common api how amici predictions should look like in
	pyPESTO.
	"""

Can we / do we want to generalize that? If not, I would suggest to rename those objects to something like AmiciPredictionResult/AmiciPredictionConditionResult to make it very clear that they are only intended to work with AmiciObjective, AmiciPredictor, ...