manifoldsR 0.6.0
Breaking changes
Expect different coordinates than 0.5.0 at the same seed for most methods.
- Everything runs in
f32.use_high_precisionis gone,evoc()included. The backends do the numerically touchy bits inf64themselves. - New
knn_methoddefaults:"ivf"fortsne(),densne(),dreams(),pacmap()andphate(),"nndescent"for the rest.generate_knn_graph()stays on"kmknn". -
params_nn()dropsdiversify_prob,ef_budgetandextract_knn.ef_constructiondefaults to200L. -
pacmap()andphate()now follow their reference implementations.params_pacmap()dropsmn_candidate_startandmn_candidate_end,params_phate()gainsthreshandmds_iter. - Precomputed graphs:
pacmap()wants themn_near + 50wide,phate()6 * k + 5.umap(),densmap()andforceatlas2()cut wider graphs tokwith a warning. -
params_umap()randomiseddefaults toFALSE. Withn_epochs = NULL, only the SGD optimisers drop to 200 epochs above 10k samples. -
diffusion_maps()with spectral landmarks takes 100 singular vectors by default, was 10.t_maxmust be at least3. -
params_diffusion_maps()n_landmarksdefaults toNULL, the full operator. Was2048L. The full operator is faster on small to medium data sets. -
rs_approx_nearest_neighbours()renamesann_paramstonn_params.
Features
-
optimiser = "sgd_parallel"inparams_umap(): lock-free Hogwild! SGD forumap()anddensmap(). Not reproducible with more than one thread.
Performance
-
"sgd"for UMAP and densMAP is about 2.4x faster. - Spectral initialisation needs half the memory on connected graphs and a quarter on disconnected ones.
Bug fixes
- PCA initialisation no longer collapses large
f32layouts onto a few points. - PaCMAP no longer compresses the layout.
- Spectral landmark selection works at scale.
-
diffusion_maps()errors on points that lose all their edges instead of returningInfcoordinates.
manifoldsR 0.5.0
First CRAN release.
Breaking changes
- Typos fixed in exported names:
params_swiss_role(),params_swiss_role_biased(),rs_data_swiss_role(),rs_data_biased_swiss_role()andrs_intertia()are nowparams_swiss_roll(),params_swiss_roll_biased(),rs_data_swiss_roll(),rs_data_biased_swiss_roll()andrs_inertia(). The old names still work but throw a deprecation warning and will go in a future release.
Features
- Rayon’s thread pool is capped at two threads when
_R_CHECK_LIMIT_CORES_is set (i.e. underR CMD check --as-cran), unlessRAYON_NUM_THREADSis already set. -
use_high_precision = NAnow warns and falls back toTRUEinstead of silently doing so.
Internals
- All Rust crates are vendored into
src/rust/vendor.tar.xzand the build runs offline. The vendored set is filtered to the CRAN platforms. - Rust output goes through R’s console (
rprintln!) rather than stdout. - Vignettes are pkgdown-only articles now and no longer ship with the package.
- Examples added to all user-facing exports.
- Backend bumps:
manifolds-rs0.7.2 andbixverse-rs0.8.0.
manifoldsR 0.4.0
Features
- DREAMS from Kury, Kobak & Damrich: t-SNE with a pull towards a global reference embedding, so cluster positions keep their meaning. Use
dreams()withparams_dreams()(lambda, default0.15). PCA is the default reference; pass your own (e.g. a PHATE result) viareference. New vignette comparing it against PCA and t-SNE. - The FFT t-SNE variants (
approx_type = "fft"and"fft_3k") now work on Windows too.manifolds-rs0.7.0 swapped FFTW for a pure Rust FFT, so the Unix-only restriction intsne()anddensne()is gone. - Backend bumps:
manifolds-rs0.7.0,bixverse-rs0.7.0,ann-search-rs0.11.0,evoc-rs0.5.0 andfaer0.24.
manifoldsR 0.3.7
Features
- Quick-and-dirty Barnes-Hut t-SNE from
manifolds-rs0.6.0, after qdtsne. Useapprox_type = "bh_qd"intsne()anddensne(). The tree depth is capped via the newmax_depthparameter inparams_tsne()(default7L). Available on all platforms. -
evoc-rsbumped to 0.4.2 andann-search-rsto 0.9.3.
manifoldsR 0.3.3
Features
- Wired in ForceAtlas2 from manifolds-rs.
manifoldsR 0.3.2
Features
- Parameters all wrapped via devforge.
manifoldsR 0.3.0
Major release
Features
- A lot of the backend code in Rust changed. This gives in parts substantially faster kNN searches from ann-search-rs (
"v0.8.1").
manifoldsR 0.2.8
Features
- Implementations of dens-map and dens-sne, density-preserving versions of UMAP and tSNE, see Narayan et al.
manifoldsR 0.2.3
Features
- Updates to various Rust crates
- More control over verbosity over the functions.
- Improved tSNE on large scale:
late_exag_factoradded that can be used on large data sets to increase repulsion on the later epochs. Also, speed improvements for both versions of tSNE. - Faster PHATE (or rather fast again - k-means clustering iterations reduced here to avoid unncessary iterations).
manifoldsR 0.2.0
Scope of the package was a bit extended and now offers also some of the very fast clustering methods that power aspects of the approximate nearest neighbour searches (k-means) + EVõC clustering.
Features
- EVõC clustering implemented from the brilliant Leland McInnes.
- k-means clustering from
ann-search-rsin a full version and as a mini-batch version for memory constrained scenarios.
