Release Notes for Version 0.2.2

This release fixes issues in the previous save/load implementation and introduces a stable, versioned serialization format.

Save/Load System

• Replaced pickle-based saving with a portable structure:
  • metadata.json: model configuration
  • params.npz: guide parameters
  • structure.npz: observations, priors, factor order/signs
  • factor.h5ad / cov.h5ad: optional cached AnnData
• Normalization is no longer applied during load.
• Pos-transform settings, factor names, order, and signs now restore correctly.
• Variational parameters are reconstructed as proper torch.nn.Parameter objects.

Cache Handling

• Cache is restored only if corresponding .h5ad files exist.
• If no cache was saved, _cache remains None.

Tests

• Added comprehensive roundtrip save/load tests.
• Tests now use tmp_path to avoid directory conflicts.

This version corrects all known issues related to model persistence and ensures deterministic reconstruction across sessions.

Release Notes for Version 0.1.5

Hotfix NaNs in prior masks.

Release Notes for Version 0.1.4

Summary

This release includes enhancements, bug fixes, and new features to improve functionality and usability. Key updates involve fixing renaming issues after reordering, updating plotting defaults, adding new functionalities, and addressing bugs.

New Features

• muvi.pl.missingness_overview: Added a function to provide an overview of missing data.

Changes and Improvements

• Allow Factor Sorting by R2 and Renaming by Enrichment Significance: Enhanced factor sorting and renaming capabilities.
• Set Default Renaming to False: Adjusted default to set renaming to false for better control.
• Redefine n_factors as Number of Uninformed Factors: Updated definition for accuracy.

Bug Fixes

• Fix Renaming After Reordering: Resolved issues with renaming operations post-reordering.
• Fix Variance Explained Grouped: Fixed unwanted reordering with grouped variance explanation calculations.

Version 0.1.3

Skip normalization for Bernoulli likelihood.

Version 0.1.2

Release Summary

New Features and Enhancements

• Data Handling:

  • Deduplicate Indices: Remove duplicate indices to ensure data integrity.
  • Data Merging and Normalization: Implement methods for merging (union/intersection) and normalizing data.
  • Flexibility in Masks and Samples: Improve handling of missing/extra features in prior masks and allow redundant samples in covariates.
  • Integration with adata and mdata: Update from_adata and from_mdata methods, including support for obs_key to load observations from obsm if required.

• Plotting and Visualization:

  • Group Plots: Introduce group plots to facilitate visual analysis of grouped data.

• Model Loading:

  • CPUUnpickler: Add CPUUnpickler to load models trained on GPU, enabling model compatibility across different hardware setups.

• Feature Set Enhancements:

  • Posterior Feature Sets: Allow computing and exporting posterior feature sets using kneed.

• Miscellaneous:

  • TensorDict: Add tensordict support for more efficient mini-batching.
  • ASCII Representation: Add ASCII representations of data for better readability and debugging.

First release

v0.1.1

First release