Added
- Add {class}`scvi.criticism.PosteriorPredictiveCheck` for model evaluation {pr}`2058`.
- Add {func}`scvi.data.reads_to_fragments` for scATAC data {pr}`1946`
- Add default `stacklevel` for `warnings` in `scvi.settings` {pr}`1971`.
- Add scBasset motif injection procedure {pr}`2010`.
- Add importance sampling based differential expression procedure {pr}`1872`.
- Raise clearer error when initializing {class}`scvi.external.SOLO` from {class}`scvi.model.SCVI`
with extra categorical or continuous covariates {pr}`2027`.
- Add option to generate {class}`mudata.MuData` in {meth}`scvi.data.synthetic_iid` {pr}`2028`.
- Add option for disabling shuffling prior to splitting data in
{class}`scvi.dataloaders.DataSplitter` {pr}`2037`.
- Add {meth}`scvi.data.AnnDataManager.create_torch_dataset` and expose custom sampler ability
{pr}`2036`.
- Log training loss through Lightning's progress bar {pr}`2043`.
- Filter Jax undetected GPU warnings {pr}`2044`.
- Raise warning if MPS backend is selected for PyTorch,
see <https://github.com/pytorch/pytorch/issues/77764> {pr}`2045`.
- Add `deregister_manager` function to {class}`scvi.model.base.BaseModelClass`, allowing to clear
{class}`scvi.data.AnnDataManager` instances from memory {pr}`2060`.
- Add option to use a linear classifier in {class}`scvi.model.SCANVI` {pr}`2063`.
- Add lower bound 0.12.1 for Numpyro dependency {pr}`2078`.
- Add new section in scBasset tutorial for motif scoring {pr}`2079`.
Fixed
- Fix creation of minified adata by copying original uns dict {pr}`2000`. This issue arises with
anndata>=0.9.0.
- Fix {class}`scvi.model.TOTALVI` {class}`scvi.model.MULTIVI` handling of missing protein values
{pr}`2009`.
- Fix bug in {meth}`scvi.distributions.NegativeBinomialMixture.sample` where `theta` and `mu`
arguments were switched around {pr}`2024`.
- Fix bug in {meth}`scvi.dataloaders.SemiSupervisedDataLoader.resample_labels` where the labeled
dataloader was not being reinitialized on subsample {pr}`2032`.
- Fix typo in {class}`scvi.model.JaxSCVI` example snippet {pr}`2075`.
Changed
- Use sphinx book theme for documentation {pr}`1673`.
- {meth}`scvi.model.base.RNASeqMixin.posterior_predictive_sample` now outputs 3-d
{class}`sparse.GCXS` matrices {pr}`1902`.
- Add an option to specify `dropout_ratio` in {meth}`scvi.data.synthetic_iid` {pr}`1920`.
- Update to lightning 2.0 {pr}`1961`
- Hyperopt is new default searcher for tuner {pr}`1961`
- {class}`scvi.train.AdversarialTrainingPlan` no longer encodes data twice during a training step,
instead uses same latent for both optimizers {pr}`1961`, {pr}`1980`
- Switch back to using sphinx autodoc typehints {pr}`1970`.
- Disable default seed, run `scvi.settings.seed` after import for reproducibility {pr}`1976`.
- Deprecate `use_gpu` in favor of PyTorch Lightning arguments `accelerator` and `devices`, to be
removed in v1.1 {pr}`1978`.
- Docs organization {pr}`1983`.
- Validate training data and code URLs for {class}`scvi.hub.HubMetadata` and
{class}`scvi.hub.HubModelCardHelper` {pr}`1985`.
- Keyword arguments for encoders and decoders can now be passed in from the model level {pr}`1986`.
- Expose `local_dir` as a public property in {class}`scvi.hub.HubModel` {pr}`1994`.
- Use {func}`anndata.concat` internally inside {meth}`scvi.external.SOLO.from_scvi_model` {pr}`2013`.
- {class}`scvi.train.SemiSupervisedTrainingPlan` and {class}`scvi.train.ClassifierTrainingPlan`
now log accuracy, F1 score, and AUROC metrics {pr}`2023`.
- Switch to cellxgene census for backend for cellxgene data function {pr}`2030`.
- Change default `max_cells` and `truncation` in
{meth}`scvi.model.base.RNASeqMixin._get_importance_weights` {pr}`2064`.
- Refactor heuristic for default `max_epochs` as a separate function
{meth}`scvi.model._utils.get_max_epochs_heuristic` {pr}`2083`.
Removed
- Remove ability to set up ST data in {class}`~scvi.external.SpatialStereoscope.from_rna_model`,
which was deprecated. ST data should be set up using
{class}`~scvi.external.SpatialStereoscope.setup_anndata` {pr}`1949`.
- Remove custom reusable doc decorator which was used for de docs {pr}`1970`.
- Remove `drop_last` as an integer from {class}`~scvi.dataloaders.AnnDataLoader`, add typing and
code cleanup {pr}`1975`.
- Remove seqfish and seqfish plus datasets {pr}`2017`.
- Remove support for Python 3.8 (NEP 29) {pr}`2021`.