NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #1554 most downloaded on PyPI
Uniform Manifold Approximation and Projection
Latest stable
0.6.0
Last release 12 days ago
10 Nov 2026
Ships fairly regularly
a new release about every 7 months
Some releases are documented
notes for 21 of 42 stable releases
Nothing withdrawn
no release was ever pulled
9 years old
47 releases · first in 2017
Drag across the ruler to compare, or tap a release to open it.
Embed combined models without densMAP instead of failing with KeyError by @raashish1601 in https://github.com/lmcinnes/umap/pull/1304
Full Changelog: https://github.com/lmcinnes/umap/compare/release-0.6.0rc1...release-0.6.0
0.6 introduces a new nearest-neighbor, initialization and layout-optimization pipeline. This is a transition release: it keeps the established (0.5.x) behavior by default while making the new pipeline available for evaluation. A future release will make the new pipeline -- recursive initialization with the Adam optimizer -- the default.
compatibility_layout now defaults to None. When it is not set, UMAP uses the 0.5.x pipeline and emits a FutureWarning that the default will change to False. Set compatibility_layout=True to keep the current behavior (and silence the warning), or compatibility_layout=False to use the new pipeline now. Selecting optimizer="adam" or optimizer="momentum" without setting compatibility_layout also opts in to the new pipeline. Combining compatibility_layout=True with a modern optimizer raises a ValueError.
With a fixed random_state, compatibility_layout=True reproduces the final 0.5.x code line bit-for-bit -- embeddings, transform, inverse_transform, DensMAP radii and AlignedUMAP -- given the same platform and dependency versions. That reference includes the 0.5.x bug fixes listed below, so a few configurations differ from 0.5.12 itself; see reproducibility.
init now defaults to "auto", which resolves to "spectral" under the compatibility layout and "recursive" otherwise. Resolved values are stored privately, so get_params and sklearn.base.clone round-trip the parameters as given.
Models pickled with umap-learn 0.5.x can be loaded, and continue to use the compatibility pipeline for transform, inverse_transform and update.
The public functions nearest_neighbors, fuzzy_simplicial_set, simplicial_set_embedding and the spectral layout functions keep their 0.5.x behavior by default; pass compatibility_layout=False to use the new pipeline.
AlignedUMAP accepts compatibility_layout with the same default and warning, and now has a full parameter docstring.
New layout optimizers optimizer="adam" (the default for the new pipeline) and optimizer="momentum". The original immediate-update optimizer is available as optimizer="compatibility".
Adam and momentum give reproducible results with a fixed random_state while still optimizing in parallel, for fitting, transform and DensMAP.
Optimizer selection is independent of DensMAP: combine densmap=True with optimizer="adam" or optimizer="momentum". The development-only names densmap_adam, densmap_momentum and standard are rejected.
init="recursive" adds a hierarchical label-propagation initialization that recursively coarsens the fuzzy graph, lays out the small graph, and expands the result. recursive_coarsening_ratio (2, 3 or 4; 4 is fastest) controls how gradually this happens. coarse_layout_weights, coarsen_on_layout_weights and recursive_init_kwds provide advanced control.
init_rescale ("isotropic", "per_axis" or "graph_edge") and init_scale_factor control how the initialization is centered and scaled before optimization.
Hard-negative sampling controls for the modern kernels: negative_selection_range, negative_sample_scale, negative_sample_scale_adaptation_samples and exclude_graph_neighbors. See optimizers for where each applies.
Faster nearest-neighbor descent settings and a parallel sparse kernel for the fuzzy union/intersection.
Optional RINND backend for building the nearest-neighbor graph (pip install umap-learn[rinnd]), selected with nearest_neighbor_backend ("auto", "rinnd" or "pynndescent"). See performance for the requirements.
A new tutorial, large_datasets, walks through these choices on three million arXiv papers.
inverse_transform validates the embedding dimension. For queries outside the embedding's convex hull it warns and extrapolates from the nearest embedded point, instead of silently using an unrelated simplex.
Fitting with the default unique=False no longer copies or reindexes the data.
The input fingerprint used to recognise transform of the training data is computed with a parallel BLAKE2 hash.
The input dtype is preserved in transform and fit_transform output.
Lower peak memory in fit and transform, and faster component layout, breadth-first search, k-nearest-neighbor selection, DensMAP density computation and AlignedUMAP relation handling.
Type checkers can use the bundled py.typed marker.
UMAP now requires NumPy 2.0 or later (with SciPy 1.13, Numba 0.60 and PyNNDescent 0.7 or later) and supports Python 3.10 to 3.14. On Python 3.14 the optional TensorFlow (ParametricUMAP) and RINND dependencies do not yet provide wheels.
Layout optimization no longer uses fast-math settings that could produce NaN embeddings (#1228).
smooth_knn_dist no longer gives an infinite sigma for points with a disconnected neighbor.
Stable tie-breaking in fast_knn_indices is restored.
Spectral initialization works for graphs with at most n_components + 1 nodes.
transform and inverse_transform work when n_epochs is a list; transform matches the sparsity of the fitted data and accepts precomputed distances when fitting used nearest-neighbor descent.
update initialisation no longer divides by zero.
csr_unique handles sparse rows that all have the same number of non-zeros, and empty unions no longer mutate sparse inputs.
Gradients are fixed for the Gaussian energy and symmetric KL metrics; Hellinger distances no longer produce NaN for proportional inputs; symmetric_kl no longer mutates its inputs; named special metrics respect their keyword arguments.
AlignedUMAP supports per-model n_epochs, and update no longer mutates the caller's relations or parameters.
ParametricUMAP no longer produces NaN gradients under XLA.
Fixes to trustworthiness_vector, plot diagnostics for small datasets, and plotting with hex background colours.
Your coding agent can read these notes before it upgrades. Set up the MCP server →
Preserve input dtype in transform and fit_transform output by @sebastianbreguel in https://github.com/lmcinnes/umap/pull/1253
Full Changelog: https://github.com/lmcinnes/umap/compare/release-0.5.12...release-0.6.0rc1
Remove nomic atlas reference from plotting.rst by @cakiki in https://github.com/lmcinnes/umap/pull/1238
Full Changelog: https://github.com/lmcinnes/umap/compare/release-0.5.11...release-0.5.12
Deterministic sort for fast_knn_indices by @bohemian-miser in https://github.com/lmcinnes/umap/pull/1236
Full Changelog: https://github.com/lmcinnes/umap/compare/release-0.5.10.post2...release-0.5.11
Full Changelog: https://github.com/lmcinnes/umap/compare/release-0.5.9.post1...release-0.5.9.post2
Full Changelog: https://github.com/lmcinnes/umap/compare/release-0.5.9.post1...release-0.5.9.post2
Update basic_usage.rst to replace deprecated function. by @rattle99 in #1188
Full Changelog: release-0.5.7...release-0.5.8
Full Changelog: https://github.com/lmcinnes/umap/compare/release-0.5.7...release-0.5.8
Update README.rst by @Slyderek in https://github.com/lmcinnes/umap/pull/1111
Full Changelog: https://github.com/lmcinnes/umap/compare/release-0.5.6...release-0.5.7
Setting random state could still lead to stochastic results by @dannywhuang in https://github.com/lmcinnes/umap/pull/1081
Full Changelog: https://github.com/lmcinnes/umap/compare/release-0.5.5...release-0.5.6
Fix pkg_resources deprecation warning by @laclouis5 in https://github.com/lmcinnes/umap/pull/1072
tbb requirement optional by @mithaler in https://github.com/lmcinnes/umap/pull/1066pkg_resources deprecation warning by @laclouis5 in https://github.com/lmcinnes/umap/pull/1072Full Changelog: https://github.com/lmcinnes/umap/compare/0.5.4...release-0.5.5
Update deprecated matplotlib colormap registry by @yannikschaelte in https://github.com/lmcinnes/umap/pull/929
get_feature_names_out for UMAP + tests by @CarloLepelaars in https://github.com/lmcinnes/umap/pull/939Full Changelog: https://github.com/lmcinnes/umap/compare/0.5.3...0.5.4
Use labels in AlignedUMAP if provided by @gclen in https://github.com/lmcinnes/umap/pull/802
Full Changelog: https://github.com/lmcinnes/umap/compare/0.5.2...0.5.3
Various bug fixes for transform, update, etc. Support for passing pre-defined knn-graphs as input.
Various bug fixes for transform, update, etc. Support for passing pre-defined knn-graphs as input.
A few minor fixes, including a bug fix against the upcoming numba 0.53
A few minor fixes, including a bug fix against the upcoming numba 0.53
Adds support for ParametricUMAP, DensMAP, AlignedUMAP and more.
Adds support for ParametricUMAP, DensMAP, AlignedUMAP and more.
ParametricUMAP learns embeddings with neural networks.
AlignedUMAP can align multiple embeddings using relations between datasets.
DensMAP can preserve local density information in embeddings.
UMAP now depends on PyNNDescent, but has faster more parallel performance as a result.
UMAP now supports an update method to add new data and retrain.
Various performance improvements and bug fixes.
Additional plotting support, including text searching in interactive plots.
Support for "maximal distances" in neighbor graphs.
Nothing published for this version
Nothing published for this version
Fix some issues with the latest scipy sparse lil_matrix handling.
Fix some issues with the latest scipy sparse lil_matrix handling.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Inverse transform method. Generate points in the original space corresponding to points in embedded space. (Thanks to Joseph Courtney)
Inverse transform method. Generate points in the original space corresponding to points in embedded space. (Thanks to Joseph Courtney)
Different embedding spaces. Support for embedding to a variety of different spaces other than Euclidean. (Thanks to Joseph Courtney)
New metrics, including Hellinger distance for sparse count data.
New discrete/label metrics, including hierarchical categories, counts, ordinal data, and string edit distance.
Support for parallelism in neighbor search and layout optimization. (Thanks to Tom White)
Support for alternative methods to handling duplicated data samples. (Thanks to John Healy)
New plotting methods for fast and easy plots.
Initial support for dataframe embedding -- still experimental, but worth trying.
Support for transform methods with sparse data.
Multithreading support when no random seed is set.
Nothing published for this version
Nothing published for this version
The is a release candidate for the new 0.4 version of UMAP which introduces a number of new features and various improvements and bug fixes.
The is a release candidate for the new 0.4 version of UMAP which introduces a number of new features and various improvements and bug fixes.
Remove object mode jit decorators that now cause warnings in numba.
Remove object mode jit decorators that now cause warnings in numba.
Nothing published for this version
Nothing published for this version
There was a bug in the layout algorithm for small data sets. That has now been resolved. Pushing a patch release to keep everyone up to date with work
There was a bug in the layout algorithm for small data sets. That has now been resolved. Pushing a patch release to keep everyone up to date with working code.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Supervised and semi-supervised dimension reduction. Support for using labels or partial labels for dimension reduction.
Supervised and semi-supervised dimension reduction. Support for using labels or partial labels for dimension reduction.
Transform method. Support for adding new unseen points to an existing embedding.
Performance improvements.
Fix the sparse correlation metric which was incorrect.
Fix the sparse correlation metric which was incorrect.
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
A new layout algorithm that handles large datasets (more) correctly.
A new layout algorithm that handles large datasets (more) correctly.
Performance improvements.
Nothing published for this version
Some minor bug fixes, some new features:
Some minor bug fixes, some new features:
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
One column per quarter.