Fix trainable parameters in distributions #520
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Codecov ReportAll modified and coverable lines are covered by tests ✅
|
stefanradev93
pushed a commit
that referenced
this pull request
Jul 2, 2025
* fix trainable parameters in distributions (#520) * Improve numerical precision in MVNScore.log_prob * add log_gamma diagnostic (#522) * add log_gamma diagnostic * add missing export for log_gamma * add missing export for gamma_null_distribution, gamma_discrepancy * fix broken unit tests * rename log_gamma module to sbc * add test_log_gamma unit test * add return information to log_gamma doc string * fix typo in docstring, use fixed-length np array to collect log_gammas instead of appending to an empty list * Breaking changes: Fix bugs regarding counts in standardization layer (#525) * standardization: add test for multi-input values (failing) This test reveals to bugs in the standarization layer: - count is updated multiple times - batch_count is too small, as the sizes from reduce_axes have to be multiplied * breaking: fix bugs regarding count in standardization layer Fixes #524 This fixes the two bugs described in c4cc133: - count was accidentally updated, leading to wrong values - count was calculated wrongly, as only the batch size was used. Correct is the product of all reduce dimensions. This lead to wrong standard deviations While the batch dimension is the same for all inputs, the size of the second dimension might vary. For this reason, we need to introduce an input-specific `count` variable. This breaks serialization. * fix assert statement in test * bump version to 2.0.5, adjust deprecation warnings * rename log_gamma to calibration_log_gamma (#527) --------- Co-authored-by: han-ol <[email protected]> Co-authored-by: Daniel Habermann <[email protected]>
stefanradev93
added a commit
that referenced
this pull request
Jul 19, 2025
…rovements (#539) * fix trainable parameters in distributions (#520) * Improve numerical precision in MVNScore.log_prob * add log_gamma diagnostic (#522) * add log_gamma diagnostic * add missing export for log_gamma * add missing export for gamma_null_distribution, gamma_discrepancy * fix broken unit tests * rename log_gamma module to sbc * add test_log_gamma unit test * add return information to log_gamma doc string * fix typo in docstring, use fixed-length np array to collect log_gammas instead of appending to an empty list * Breaking changes: Fix bugs regarding counts in standardization layer (#525) * standardization: add test for multi-input values (failing) This test reveals to bugs in the standarization layer: - count is updated multiple times - batch_count is too small, as the sizes from reduce_axes have to be multiplied * breaking: fix bugs regarding count in standardization layer Fixes #524 This fixes the two bugs described in c4cc133: - count was accidentally updated, leading to wrong values - count was calculated wrongly, as only the batch size was used. Correct is the product of all reduce dimensions. This lead to wrong standard deviations While the batch dimension is the same for all inputs, the size of the second dimension might vary. For this reason, we need to introduce an input-specific `count` variable. This breaks serialization. * fix assert statement in test * rename log_gamma to calibration_log_gamma (#527) * simple fix * Hotfix: numercial stability of non-log-stabilized sinkhorn plan (#531) * fix numerical stability issues in sinkhorn plan * improve test suite * fix ultra-strict convergence criterion in log_sinkhorn_plan * update dependencies * add comment about convergence check * update docsting to reflect fixes * sinkhorn_plan now returns a transport plan with uniform marginal distributions * add unit test for sinkhorn_plan * fix sinkhorn function by sampling from the logits of the transpose of the plan, instead of the plan directly * sinkhorn(x1, x2) now samples from log(plan) to receive assignments such that x2[assignments] matches x1 * re-enable test_assignment_is_optimal() for method='sinkhorn' * log_sinkhorn now correctly uses log_plan instead of keras.ops.exp(log_plan), log_sinkhorn_plan returns logits of the transport plan * add unit tests for log_sinkhorn_plan * fix faulty indexing with tensor for tensorflow backend * re-add numItermax for ot pot test --------- Co-authored-by: Daniel Habermann <[email protected]> * isinstance sequence * Pass correct training stage in compute_metrics (#534) * Pass correct training stage in CouplingFlow.compute_metrics * Pass correct training stage in CIF and PointInferenceNetwork * Custom test quantity support for calibration_ecdf (#528) * Custom test quantity support for calibration_ecdf * rename variable [no ci] * Consistent defaults for variable_keys/names in calibration_ecdf with test quantiles * Tests for calibration_ecdf with test_quantities * Remove redundant and simplify comments * Fix docstrings and typehints --------- Co-authored-by: stefanradev93 <[email protected]> * Log gamma test fix (#535) * fix test_calibration_log_gamma_end_to_end unit test failing too often than expected * set alpha to 0.1% in binom.ppf * fix typo in comment * Stateless adapters (#536) * Remove stateful adapter features * Fix tests * Fix typo * Remove nnpe from adapter * Bring back notes [skip ci] * Remove unncessary restriction to kwargs only [skip ci] * Remove old super call [skip ci] * Robustify type [skip ci] * remove standardize from multimodal sim notebook [no ci] * add draft module docstring to augmentations module [no ci] Feel free to modify. * adapt and run neurocognitive modeling notebook [no ci] * adapt cCM playground notebook [no ci] * adapt signature of Adapter.standardize * add parameters missed in previous commit * Minor NNPE polishing * remove stage in docstring from OnlineDataset --------- Co-authored-by: Lasse Elsemüller <[email protected]> Co-authored-by: Valentin Pratz <[email protected]> * Fix training strategies in BasicWorkflow * move multimodal data notebook to regular examples [no ci] * make pip install call on homepage more verbose [no ci] * remove deprecated summaries function The function was renamed to summarize in v2.0.4. * detail subsampling behavior docs for SIR simulator [no ci] fixes #518 * move DiffusionModel from experimental to networks Stabilizes the DiffusionModel class. A deprecation warning for the DiffusionModel class in the experimental module was added. * Add citation for resnet (#537) [no ci] * added citation for resnet * minor formatting --------- Co-authored-by: Valentin Pratz <[email protected]> * Bump up version [skip ci] --------- Co-authored-by: Valentin Pratz <[email protected]> Co-authored-by: han-ol <[email protected]> Co-authored-by: Daniel Habermann <[email protected]> Co-authored-by: Valentin Pratz <[email protected]> Co-authored-by: arrjon <[email protected]> Co-authored-by: Lars <[email protected]> Co-authored-by: Hans Olischläger <[email protected]> Co-authored-by: Lasse Elsemüller <[email protected]> Co-authored-by: Leona Odole <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR fixes two issues for the distributions when trainable parameters are used.
keras.initializers.get, as the arrays seem to be freed for some reason, leading to a weirdRuntimeError: Array has been deleted with shape=float32[4].when trying to access them.I encountered this when preparing #519, where I added a test for this. To reproduce it, check out the PR and run
nox -- save dev tests/test_compatibility/test_distributions.