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

Good job Hans! Fyi, commit 955ac79 uncovered a bug in ModelComparisonApproximator's build_from_data method which d8d84c8 addresses.

Nice work :) While trying it out a bit, the additional dimension for each ensemble member in the data caught me by surprise, and it took me a while to figure out why a dimension was missing from my data during training.
I'm not sure what the best interface would be here, but I think it would be good to think about the design for a bit. Maybe the following questions will get us closer to what we want to have:

1. Should the ordinary dataset classes be supported? If yes, which mode should they operate in (does it need a warning?), if no, we might want to override fit to check for them and raise an error if they are passed.
2. Should there be multiple "modes" for the ...EnsembleDatasets, i.e., identical data vs. different data? Might be especially relevant for online training, as different data increases the required compute.
3. How do we handle custom dataset classes, what shapes do we expect from them?

As I did not take part in the discussions, maybe you already talked this through. In any way, I would be happy to hear your thoughts on this...

I have added serialization support, but deserialization fails when multiple approximators use the same weights, e.g. when they share a summary network. I'm not sure yet how this can be resolved, and if we want to enable serialization if we can not resolve it...

@han-ol could you perhaps provide a minimal code example for how the ensembles should work? On that basis, it might be then also easier to discuss interface questions, including those of @vpratz.

@paul-buerkner You can find one as part of the PR: https://github.com/bayesflow-org/bayesflow/blob/ensembles/examples/ApproximatorEnsemble%20example.ipynb

@han-ol and I had a discussion on the dataset question. One approach would be a general EnsembleDatasetWrapper with the following properties:

• it takes in an arbitrary dataset, e.g. an instance of OnlineDataset or OfflineDataset
• it determines the batch size either by reading dataset.batch_size or by sampling a batch from the dataset
• it has a parameter like unique_data_fraction to control if all ensemble members get the same data (0.0) or every member gets different data (1.0). For in between values, a bootstrap procedure can be used.
• the required number of simulations can be obtained by repeatedly sampling batches from dataset, or, for our own simulators, by changing the dataset.batch_size parameter. The latter would be a bit hacky, we would have to see if we want this.

In addition, in the approximator ensemble class, we can determine by the shape of inference_variables if a standard dataset (like OnlineDataset) was used. If so, we default to showing the same data to all ensemble members.

This gives us the possibility to use our existing datasets, and only requires this one additional class to add the capability to pass different data to different approximators.

I am strongly in favor of this idea. We also discussed in the past that this is a more elegant and catch-all solution.