Changelog

Changelog#

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]#

Added#

  • Add drvi.internal.DRVI, a developmental (internal-use-only) subclass of scvi.external.DRVI that re-adds a few experimental features from earlier drvi-py: opt-in residual connections between the same-width hidden layers of the encoder/decoder (residual=True), streaming (online) training metrics logged per epoch (track_streaming_metrics, latent-dimension stats plus label/latent mutual information when a labels_key is set), a sparse latent representation (get_sparse_latent_representation), gene-subsampled reconstruction for scalable training on very wide panels (n_genes_to_reconstruct=N), and gradient scaling from the decoder heads into the decoder body/encoder (gradient_scale). Its API is unstable and may change or be removed without notice.

[0.3.0]#

Changed#

  • The DRVI PyTorch model is no longer shipped in this package. It has been contributed to scvi-tools and now lives there as scvi.external.DRVI (requires scvi-tools >= 1.5.0, now the minimum dependency). drvi.model.DRVI is kept as an alias for scvi.external.DRVI for backward compatibility and may be deprecated from 0.4.0. New code should import the model directly as scvi.external.DRVI. Utility, plotting, metrics, and interpretability tools (drvi.utils) continue to be maintained here and work on top of the scvi-tools model.

Removed#

  • Removed the in-package model implementation and its internals: drvi.scvi_tools_based (model, module, training plan, neural-network components, data fields), drvi.nn_modules, and drvi.model.DRVIModule. Model capabilities not supported by scvi.external.DRVI are dropped and will not be reinstated. To move a model trained with drvi-py < 0.3.0 to scvi-tools, use drvi.utils.port_to_scvi_tools.

  • Emptied the drvi.utils.tl tools namespace (drvi.utils.tools): removed set_latent_dimension_stats, traverse_latent, calculate_differential_vars, get_split_effects, and iterate_on_top_differential_vars. The latent-traversal interpretability pipeline is superseded by the scvi-tools DRVI model’s built-in set_latent_dimension_stats / calculate_interpretability_scores / get_interpretability_scores. The drvi.utils.tl namespace is kept (now empty) as a stable import location.

  • Removed the traversal-based plotting functions drvi.utils.pl.show_top_differential_vars, plot_relevant_genes_on_umap, show_differential_vars_scatter_plot, differential_vars_heatmap, and make_heatmap_groups (they consumed the removed traversal outputs). plot_interpretability_scores (which visualizes the model’s get_interpretability_scores) and the latent-dimension plots remain.

[0.2.7] - 2026-07-09#

Added#

  • Add drvi.utils.port_to_scvi_tools to migrate a trained/saved drvi-py model into a checkpoint loadable by scvi.external.DRVI (available in scvi-tools 1.5.0). Pure checkpoint surgery — no model is instantiated. Unsupported capabilities raise DRVIPortError with a link to open an issue.

[0.2.6] - 2026-07-02#

Added#

  • Add drvi.utils.pl.plot_interpretability_scores plotting utility, now reused by DRVI.plot_interpretability_scores

  • Add binary_maximum_mutual_information_score (BMMI) disentanglement metric

  • Add min_max_thresholds argument to plot_latent_dims_in_umap to control color-scale clamping

  • Add "gelu" option and callable factory support for the encoder mean_activation

Changed#

  • In DiscreteDisentanglementBenchmark, default metrics are now ("SMI", "SPN"), samples are shuffled before evaluation, and the benchmark version is bumped to v3_1

  • Renamed discrete_mutual_info_score to discrete_scaled_mutual_info_score.

  • Unknown encoder mean_activation values now raise NotImplementedError instead of failing an assertion

  • LatentStats now logs non_vanished* counts as Python scalars

Removed#

  • Remove MI metrics local_mutual_info_score, and global_dim_mutual_info_score, along with the SMI-cont/SMI-disc benchmark keys (superseded by SMI and BMMI)

Fixed#

  • Fix typo in metric name spearman_correlataion_score -> spearman_correlation_score

[0.2.5] - 2026-04-21#

Added#

  • latent stat (n_vanished) logs during training

  • Disentanglement metric logs during training

  • Relax upper bounds of all dependencies

[0.2.4] - 2026-03-02#

Added#

  • Allow sparse input as X

  • Allow generating memory efficient sparse tensors as latent

[0.2.3] - 2026-02-27#

Added#

  • Added within distribution interpretability

  • An efficient implementation of out-of-distribution interpretability is added

  • plotting and getting relevant genes is now possible in DRVI model interface.

  • Setting latent dimension stats is now done in model interface. Previous util functions still work, but will show a deprecation warning.

  • Add tutorial for query to reference mapping

Changed#

  • The default value for vanished threshold in new the interface of set_latent_dimension_stats is 0.5 (previously 0.1).

  • Interpretability scores for the other direction of a non-vanished dimension is not shown if that direction is meaningless.

  • Main tutorial updated with the new interpretability interface

[0.2.2] - 2026-02-11#

Added#

  • Added “tutorials” optional dependency that was accidentally removed.

  • Code to allow loading models from previous versions with no problem

  • Add dispersion parameter to allow modeling batch dependent dispersion

Removed#

  • Removed “x_loglib”, “div_lib_x_loglib”, “x_loglib_all” library normalization techniques

  • Removed Vamp prior and GMM prior

  • Removed Legacy likelihood functions. Will raise error if used.

Changed#

  • gene_likelihood parameter now accepts different values compared to before. gene_likelihood parameter from old models will be mapped properly, but new users should look into the docs.

[0.2.1] - 2025-12-29#

Added#

  • Add support for Python 3.13

  • Allow subset reconstruction

  • Allow gradient scaling in the last layer

  • Allow setting vector size after mapping in “split_map@k” and “power@k” splitting functions.

  • Add support for Python 3.14

Removed#

  • Remove restrict dependencies. To ensure compatibility with old packages run for example: uvx --exclude-newer 2024-01-01 hatch run pytest

Changed#

  • Minor code improvements

  • Update to scverse template version 0.7.0

[0.2.0] - 2025-11-24#

Changed#

  • Update upper bound of all dependencies

  • Align with the latest changes of scvi-tools

Added#

  • Use cruft for scverse template management

[0.1.11] - 2025-11-19#

Added#

  • Add RnaSeqMixin from scvi-tools for RNA-seq specific methods (get_normalized_expression, differential_expression, posterior_predictive_sample, get_likelihood_parameters)

Fixed#

  • Fix bug in decode space handling where library size was not considered (issue #46)

  • Update decode_latent_samples logic (decode in log space by default)

  • Code improvements and bug fixes

[0.1.10] - 2025-11-18#

Removed#

  • Remove merlin data support and all related code

  • Remove merlin-dataloader dependency

Added#

  • Add dependabot for dependency update notification (not used now, for next releases)

[0.1.9] - 2025-07-01#

Added#

  • Add DRVI-APnoEXP baseline

Fixed#

  • Imorove documnetation for all classes and functions in repository

[0.1.8] - 2025-06-22#

Changed#

  • Discretize latent dimension values in MI for benchmarking due to this bug.

Fixed#

  • Fix a minor issue with drvi.utils.tl.traverse_latent

[0.1.7] - 2025-05-30#

Fixed#

  • Fix categorical lookup for reconstruction

[0.1.6] - 2025-05-22#

Changed#

  • Allow kwargs to pass through in plot_relevant_genes_on_umap

  • Update project CI structure

  • Extract tutorial notebooks to another repo to keep this repo clean

[0.1.5] - 2025-05-09#

Changed#

  • Refactor benchmarking code for better reusability

  • Revert callable for mean and var activation

[0.1.4] - 2025-04-17#

Fixed#

  • Limit anndata version for compatibility with old scvi-tools

[0.1.3] - 2025-02-12#

Added#

  • Introduce mean activation to make non-negative latents possible (docs will come later)

Fixed#

  • Better communication when Merlin is not installed

  • Raise error when interpretability is called on model with continues covariates

[0.1.2] - 2024-11-11#

Fixed#

  • No change in DRVI code

  • Fix github workflow, tests, docs, and pypi publishing pipelines

[0.1.0] - 2024-08-21#

Added#

  • Moved all files from repo to scverse cookiecutter project template