Stateless adapters

[stefanradev93]

This PR implements the transition towards stateless adapters, as announced in v2.0.4. In particular, it:

• Removes the stage argument from data sets and all data set builders.
• Remove the stage argument from the Adapter and all callers.
• Removes adaptive mean computation from the standardize transform, as the functionality is now superseded by the Standardize layers in approximators. The remaining standardize transform should now be used only with fixed means / stds.
• Removes the stage argument from NNPE transform. @elseml / @vpratz / @paul-buerkner This is actually a problem for the 1% of the use cases, as we don't want the noise during inference. As I thought about it, NNPE actually fits the philosophy of an augmentation (introduced in 2.0.3) rather than a transform (i.e., something strictly needed only during training). Thus, my proposed solution is to introduce an augmentations module and start collecting these transforms there (more to come). Let me know what you think.

[codecov]

Codecov Report

Attention: Patch coverage is 92.85714% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
bayesflow/augmentations/nnpe.py	91.66%	1 Missing ⚠️
bayesflow/datasets/disk_dataset.py	0.00%	1 Missing ⚠️

Files with missing lines	Coverage Δ
bayesflow/__init__.py	100.00% <ø> (ø)
bayesflow/adapters/adapter.py	85.89% <100.00%> (-0.35%)	⬇️
bayesflow/adapters/transforms/__init__.py	100.00% <ø> (ø)
bayesflow/adapters/transforms/standardize.py	100.00% <100.00%> (+6.52%)	⬆️
bayesflow/approximators/continuous_approximator.py	90.40% <100.00%> (ø)
...low/approximators/model_comparison_approximator.py	84.39% <100.00%> (ø)
bayesflow/augmentations/__init__.py	100.00% <100.00%> (ø)
bayesflow/datasets/offline_dataset.py	82.14% <100.00%> (-0.32%)	⬇️
bayesflow/datasets/online_dataset.py	78.12% <100.00%> (-0.67%)	⬇️
bayesflow/augmentations/nnpe.py	87.50% <91.66%> (ø)
... and 1 more

[vpratz]

I agree that NNPE is a better fit for an augmentation than for a transform. Moving it as part of this PR (with a deprecation for the transform) sounds sensible to me.

[elseml]

Agree, that should fit the use case perfectly well. Could you provide a first outline of the augmentation module structure for NNPE that you have in mind (that we can then iterate on if needed)?

[elseml]

Everything else besides the open NNPE stuff looks uncritical to me. Leaving a reminder for us to shift the NNPE tests once the augmentation is implemented.

[vpratz]

I like the changes overall, I'm currently removing the .standardize() calls from the experimental notebooks and will push the changes when I'm done.

The signature of Adapter.standardize() is now a bit awkward in my opinion, as mean and std are required, but are not listed in neither signature nor docstring. They are only passed via kwargs. Do we want to change this, @stefanradev93?

[stefanradev93]

Good spot! Let's introduce them as they are described in the transform.

…

On Sun, Jul 13, 2025, 11:04 AM Valentin Pratz ***@***.***> wrote: *vpratz* left a comment (bayesflow-org/bayesflow#536) <#536 (comment)> I like the changes overall, I'm currently removing the .standardize() calls from the experimental notebooks and will push the changes when I'm done. The signature of Adapter.standardize() is now a bit awkward in my opinion, as mean and std are required, but are not listed in neither signature nor docstring. They are only passed via kwargs. Do we want to change this, @stefanradev93 <https://github.com/stefanradev93>? — Reply to this email directly, view it on GitHub <#536 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFKWAGIIHWKO72L2CKVEFCT3IIOJJAVCNFSM6AAAAACBJ22KDSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTANRWG4ZTSNJTG4> . You are receiving this because you were mentioned.Message ID: ***@***.***>