Ensembles of Approximators

[han-ol]

This draft-PR is the result of discussions with @elseml and @stefanradev93.

The goal is fast and convenient support of approximator ensembles and the first steps for this are taken already.

1. We envision ApproximatorEnsemble as the abstraction at the heart of future workflows using ensembles.

   • fundamentally, it is a wrapper of a dictionary of arbitrary Approximator objects.
   • it overwrites the central methods compute_metrics, build, sample and passes inputs on to the respective ensemble member's methods.

2. Since ensembles should cover the sensitivity wrt all randomness in approximators, which are not just initialization, but also the random order of training batches, we need slightly modified datasets.

   • For now only OfflineEnsembleDataset is implemented, which makes sure that training batches have an additional dimension at the second axis, containing multiple independent random slices of the available offline samples.

A few things are missing, among them are

• predict/estimate methods for ApproximatorEnsemble (currently sample exists)
• tests for ApproximatorEnsemble
• doc strings for ApproximatorEnsemble
• OnlineEnsembleDataset
• DiskEnsembleDataset
• tests for ensemble datasets
• some Workflow

[han-ol]

You can check out the example notebook here: https://github.com/bayesflow-org/bayesflow/blob/ensembles/examples/ApproximatorEnsemble%20example.ipynb

[codecov]

Codecov Report

Attention: Patch coverage is 28.12500% with 46 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
bayesflow/approximators/approximator_ensemble.py	24.48%	37 Missing ⚠️
bayesflow/datasets/offline_ensemble_dataset.py	30.76%	9 Missing ⚠️

Files with missing lines	Coverage Δ
bayesflow/approximators/__init__.py	100.00% <100.00%> (ø)
bayesflow/datasets/__init__.py	100.00% <100.00%> (ø)
bayesflow/datasets/offline_ensemble_dataset.py	30.76% <30.76%> (ø)
bayesflow/approximators/approximator_ensemble.py	24.48% <24.48%> (ø)