Inconsistent output shapes of SIR and Lotka-Volterra simulators

[vpratz]

The time-series benchmark simulators currently produce different output shapes, I think we should unify this to always produce shape (batch_size, n_timesteps, 1), which is what the summary networks expect. It looks like the problem occurs due to the subsampling functionality. Labeling this as a bug as I suppose it is not intended behavior.

Code to reproduce:

import bayesflow as bf
import keras

keras.tree.map_structure(keras.ops.shape, bf.simulators.SIR().sample(2))
# {'parameters': (2, 2), 'observables': (2, 160, 1)}
keras.tree.map_structure(keras.ops.shape, bf.simulators.SIR(subsample=10).sample(2))
# {'parameters': (2, 2), 'observables': (2, 10)}
keras.tree.map_structure(keras.ops.shape, bf.simulators.LotkaVolterra().sample(2))
# {'parameters': (2, 4), 'observables': (2, 20)}

[vpratz]

Or is the intention that subsample replaces the summary network, and therefore also flattens? If so, we should document it in the docstring...

[stefanradev93]

The intention is to have flexibility. subsample makes it consistent with the original "benchmark".

[vpratz]

Closed by ebbddce for now. We might want to align the signature of SIR and LotkaVolterra in the future, i.e., introduce a flatten argument for SIR as well...