Release 2.0.4 #510

stefanradev93 · 2025-06-13T11:28:34Z

Summary

This PR introduces the next set of improvements and qualitative changes for the upcoming v2.0.4 release.

Major Changes

Added several diffusion model implementations
Rewritten and refactored all Approximators classes
Introduced new adapter transforms, including support for nested tensors (e.g., dictionaries of tensors) to better handle multimodal simulations
Added a new fusion network for multimodal amortized Bayesian inference (ABI)
Integrated standardization layers into each Approximation, eliminating the need for stateful adapters
- Note: this paves the way for fully stateless adapters in v2.0.5; all stateful functionality from the adapter will be removed in v2.0.5.
Enabled automatic standardization of point estimates
Enhanced support for tracking multiple loss terms
Enabled sampling of tensors with arbitrary shapes
Extended and improved tests
Support arbitrary data augmentation functions for data sets

Bug Fixes

Resolved serialization issues in edge cases and when using model comparison approximators
Fixed incorrect aggregation of validation losses
Corrected pairplot rendering issues

* made initial backend functions for adapter subsetting, need to still make the squeeze function and link it to the front end * added subsample functionality, to do would be adding them to testing procedures * made the take function and ran the linter * changed name of subsampling function * changed documentation, to be consistent with external notation, rather than internal shorthand * small formation change to documentation * changed subsample to have sample size and axis in the constructor * moved transforms in the adapter.py so they're in alphabetical order like the other transforms * changed random_subsample to maptransform rather than filter transform * updated documentation with new naming convention * added arguments of take to the constructor * added feature to specify a percentage of the data to subsample rather than only integer input * changed subsample in adapter.py to allow float as an input for the sample size * renamed subsample_array and associated classes/functions to RandomSubsample and random_subsample respectively * included TypeError to force users to only subsample one dataset at a time * ran linter * rerun formatter * clean up random subsample transform and docs * clean up take transform and docs * nitpick clean-up * skip shape check for subsampled adapter transform inverse * fix serialization of new transforms * skip randomly subsampled key in serialization consistency check --------- Co-authored-by: LarsKue <lars@kuehmichel.de>

Closes #466

…ositiveDefinite link (#469) * Refactor fill_triangular_matrix * stable positive definite link, fix for #468 * Minor changes to docstring * Remove self.built=True that prevented registering layer norm in build() * np -> keras.ops

* Remove old rounds data set, add documentation, and augmentation options to data sets * Enable augmentation to parts of the data or the whole data * Improve doc * Enable augmentations in workflow * Fix silly type check and improve readability of for loop * Bring back num_batches

y = (x - mu) / sigma log p(y) = log p(x) - log(sigma)

Fixed Jacobian computation of standardize transform

The two lines were switched, leading to performance degradation.

…eNetwork does)

This commit contains the following changes (see PR #408 for discussions) - DiffusionModel following the formalism in Kingma et. al (2023) [1] - Stochastic sampler to solve SDEs - Tests for the diffusion model [1] https://arxiv.org/abs/2303.00848 --------- Co-authored-by: arrjon <jonas.arruda@uni-bonn.de> Co-authored-by: Jonas Arruda <69197639+arrjon@users.noreply.github.com> Co-authored-by: LarsKue <lars@kuehmichel.de>

- From the table in the `bayesflow.networks` module overview, one cannot tell which network belongs to which group. This commit adds short labels to indicate inference networks (IN) and summary networks (SN)

…simulators (#452) Adds option to drop, fill or error when different keys are encountered in the outputs of different simulators. Fixes #441. --------- Co-authored-by: Valentin Pratz <git@valentinpratz.de>

* Add classes and transforms to simplify multimodal training - Add class `MultimodalSummaryNetwork` to combine multiple summary networks, each for one modality. - Add transforms `Group` and `Ungroup`, to gather the multimodal inputs in one variable (usually "summary_variables") - Add tests for new behavior * [no ci] add tutorial notebook for multimodal data * [no ci] add missing training argument * rename MultimodalSummaryNetwork to FusionNetwork * [no ci] clarify that the network implements late fusion

* add tests for find_network

Very basic transform, just the inverse of expand_dims

Allow Python version 3.12 after successful CI run: https://github.com/bayesflow-org/bayesflow/actions/runs/14988542031

…erialization in keras (#493) * add custom sequential to fix #491 * revert using Sequential in classifier_two_sample_test.py * Add docstring to custom Sequential Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * fix copilot docstring * remove mlp override methods --------- Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Make docs optional dependencies compatible with python 3.10

* Add NNPE adapter * Add NNPE adapter tests * Only apply NNPE during training * Integrate stage differentiation into tests * Improve test coverage * Fix inverse and add to tests * Adjust class name and add docstring to forward method * Enable compatibility with #486 by adjusting scales automatically * Add dimensionwise noise application * Update exception handling * Fix tests

vpratz · 2025-06-14T11:52:49Z

I will do some local experiments. The trained diffusion model is performing quite well. The strange thing is that the test only fails with a JAX backend....

@stefanradev93 I cannot reproduce the failure locally, can you? I think skipping this specific test for the diffusion models (e.g. via a marker) would be acceptable, as we do not expect cycle consistency to strictly hold for the number of steps we want to put in here, and the errors can accumulate in more or less arbitrary ways. We might want to use (at least somewhat) trained models in the future, where we can expect more consistent results.

LarsKue · 2025-06-14T17:14:55Z

@vpratz I added a review to #503. I think we should not include it here, yet.

stefanradev93 · 2025-06-14T18:09:29Z

@stefanradev93 I cannot reproduce the failure locally, can you? I think skipping this specific test for the diffusion models (e.g. via a marker) would be acceptable, as we do not expect cycle consistency to strictly hold for the number of steps we want to put in here, and the errors can accumulate in more or less arbitrary ways. We might want to use (at least somewhat) trained models in the future, where we can expect more consistent results.

The error is stochastic and I think you are right. We should not rely on this test for now. It is still a bit strange that it's only the case for JAX.

LarsKue

I left a detailed review. See individual comments.

bayesflow/adapters/transforms/random_subsample.py

README.md

bayesflow/adapters/transforms/standardize.py

bayesflow/datasets/__init__.py

bayesflow/links/positive_definite.py

bayesflow/metrics/functional/root_mean_squared_error.py

bayesflow/networks/coupling_flow/coupling_flow.py

bayesflow/networks/flow_matching/flow_matching.py

bayesflow/utils/dispatch/find_noise_schedule.py

The implementation is a simple wrapper leveraging the batching capabilities of `rejection_sample`.

bayesflow/adapters/transforms/nan_to_num.py

bayesflow/workflows/basic_workflow.py

- the test is unstable for untrained diffusion models, as the networks output is not sufficiently smooth for the step size we use - remove the diffusion_model marker

* Move docstring to comment * Always cast to _resolve_scale * Fix typo

LarsKue · 2025-06-17T13:30:18Z

Thanks everyone for the massive work in contributing these features and also in cleaning up the parts that I commented so quickly!

Rename approximator.summaries to summarize with deprecation

eodole and others added 30 commits May 6, 2025 13:30

[no ci] docs: start of user guide - draft intro, gen models

4140afe

[no ci] add draft for data processing section

728b130

[no ci] user guide: add stub on summary/inference networks

d461d60

[no ci] user guide: add stub on additional topics

21690fd

[no ci] add early stage disclaimer to user guide

6ecd258

pin dependencies in docs, fixes snowballstemmer error

6990d08

fix: correct check for "no accepted samples" in rejection_sample

a4d58c9

Closes #466

Fixed log det jac computation of standardize transform

b28ae2c

y = (x - mu) / sigma log p(y) = log p(x) - log(sigma)

Merge pull request #471 from han-ol/fix-standardze-log-det-jac

0ed0e88

Fixed Jacobian computation of standardize transform

Fix fill_triangular_matrix

265f667

The two lines were switched, leading to performance degradation.

Deal with inference_network.log_prob to return dict (as PointInferenc…

ccf9ca0

…eNetwork does)

[no ci] Merge remote-tracking branch 'upstream/dev' into docs-user-guide

fef8b90

[no ci] networks docstrings: summary/inference network indicator (#462)

e13f7c4

- From the table in the `bayesflow.networks` module overview, one cannot tell which network belongs to which group. This commit adds short labels to indicate inference networks (IN) and summary networks (SN)

ModelComparisonSimulator: handle different outputs from individual …

35cd671

…simulators (#452) Adds option to drop, fill or error when different keys are encountered in the outputs of different simulators. Fixes #441. --------- Co-authored-by: Valentin Pratz <git@valentinpratz.de>

allow dispatch of summary/inference network from type

92426d6

* add tests for find_network

Add squeeze transform

25b73d3

Very basic transform, just the inverse of expand_dims

[no ci] fix examples in ExpandDims docstring

56ddd99

squeeze: adapt example, add comment for changing batch dims

98a6bca

Merge remote-tracking branch 'upstream/dev' into feat-squeeze-transform

3c0bcc4

Permit Python version 3.12 (#474)

4781e2e

Allow Python version 3.12 after successful CI run: https://github.com/bayesflow-org/bayesflow/actions/runs/14988542031

Change order in readme and reference new book [skip ci]

cd2c212

make docs optional dependencies compatible with python 3.10

12c72bc

Merge pull request #495 from daniel-habermann/fix-doc-dependencies

e13c944

Make docs optional dependencies compatible with python 3.10

Merge branch 'main' into dev

7071f1a

LarsKue requested changes Jun 14, 2025

View reviewed changes

stefanradev93 and others added 4 commits June 14, 2025 15:46

Remove empty file [no ci]

d42fe51

Revert layer type for coupling flow [skip ci]

c44333d

remove failing import due to removed find_noise_schedule.py [no ci]

9018ce6

Add utility function for batched simulations (#511)

b43f1cc

The implementation is a simple wrapper leveraging the batching capabilities of `rejection_sample`.

elseml mentioned this pull request Jun 16, 2025

Implement changes to NNPE adapter for #510 #514

Merged

Restore PositiveDefinite link with deprecation warning

a8e9a72

elseml approved these changes Jun 16, 2025

View reviewed changes

bayesflow/adapters/transforms/nan_to_num.py Show resolved Hide resolved

bayesflow/workflows/basic_workflow.py Show resolved Hide resolved

vpratz and others added 10 commits June 16, 2025 13:11

skip cycle consistency test for diffusion models

990df1e

- the test is unstable for untrained diffusion models, as the networks output is not sufficiently smooth for the step size we use - remove the diffusion_model marker

Implement changes to NNPE adapter for #510 (#514)

1d4f646

* Move docstring to comment * Always cast to _resolve_scale * Fix typo

[no ci] remove unnecessary serializable decorator on rmse

664f00a

fix type hint in squeeze [no ci]

5659773

reintroduce comment in jax approximator [no ci]

9030803

remove unnecessary getattr calls [no ci]

1cd8ffb

Rename local variable transformation_type

3c9797d

fix error type in diffusion model [no ci]

f7fe860

remove non-functional per_training_step from plots.loss

212af40

Update doc [skip ci]

676b0cd

stefanradev93 requested a review from LarsKue June 17, 2025 07:19

rename approximator.summaries to summarize with deprecation

329ebe7

vpratz mentioned this pull request Jun 17, 2025

Rename approximator.summaries to summarize with deprecation #516

Merged

address remaining comments

00f0a89

LarsKue approved these changes Jun 17, 2025

View reviewed changes

Merge pull request #516 from bayesflow-org/rename-summaries

057f3fd

Rename approximator.summaries to summarize with deprecation

stefanradev93 merged commit 0cdd16a into main Jun 17, 2025
16 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Release 2.0.4 #510

Release 2.0.4 #510

Uh oh!

stefanradev93 commented Jun 13, 2025 •

edited

Loading

Uh oh!

vpratz commented Jun 14, 2025

Uh oh!

LarsKue commented Jun 14, 2025

Uh oh!

stefanradev93 commented Jun 14, 2025 •

edited

Loading

Uh oh!

LarsKue left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

LarsKue commented Jun 17, 2025

Uh oh!

Uh oh!

Uh oh!

Release 2.0.4 #510

Release 2.0.4 #510

Uh oh!

Conversation

stefanradev93 commented Jun 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Major Changes

Bug Fixes

Uh oh!

vpratz commented Jun 14, 2025

Uh oh!

LarsKue commented Jun 14, 2025

Uh oh!

stefanradev93 commented Jun 14, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

LarsKue left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

LarsKue commented Jun 17, 2025

Uh oh!

Uh oh!

Uh oh!

stefanradev93 commented Jun 13, 2025 •

edited

Loading

stefanradev93 commented Jun 14, 2025 •

edited

Loading