umapers 1.1.0

Python bindings for rust_umap

umapers

umapers is a Python UMAP library for dimensionality reduction, clustering workflows, and embedding visualization.

umap-learn is the established Python implementation used as the compatibility and benchmark baseline for this project. For matching parameters and workflows, umapers keeps the user-visible semantics familiar while running its own fitting and transform code instead of importing umap-learn at runtime.

The public API is intentionally small: fit an embedding, transform new rows, check quality, and reproduce runs with fixed seeds.

Installation

pip install umapers

Optional extras for plotting and diagnostics:

pip install "umapers[plot,diagnostics]"

Quickstart

import numpy as np
from umapers import Umap

x = np.random.default_rng(42).normal(size=(1000, 32)).astype(np.float32)

embedding = Umap(
    n_neighbors=15,
    n_components=2,
    n_epochs=200,
    random_seed=42,
).fit_transform(x)

Fit once and transform new rows:

model = Umap(random_seed=42).fit(x)
query_embedding = model.transform(x[:10])

Use labels to run categorical supervised UMAP:

labels = np.random.default_rng(7).integers(0, 4, size=x.shape[0])

embedding = Umap(
    target_metric="categorical",
    target_weight=0.5,
    random_seed=42,
).fit_transform(x, y=labels)

Results

The 1.1.0 release reports compare umapers with umap-learn on feature coverage, clustering quality, and runtime. On the real datasets tested here, downstream clustering quality is similar. On the synthetic scaling benchmarks, umapers is strongest on medium and large matrices; for very small datasets, fixed overhead can make umap-learn faster.

• Clustering quality: current_clustering_analysis_report.md
• Runtime scaling: synthetic_runtime_scaling_report.md
• Large-data probe: synthetic_large_runtime_probe.md
• Real-data runtime: runtime_vs_dataset_size.md
• Feature parity: current_feature_parity_report.md

Supported Workflows

Need	API
Fit a 2D or nD embedding	Umap(...).fit_transform(data)
Fit once, then transform new rows	model.fit(data) then model.transform(query)
Approximate inverse mapping	model.inverse_transform(embedding)
Categorical supervised embedding	target_metric="categorical", pass y=
Dense densMAP radii	densmap=True, output_dens=True
Sparse input	pass CSR/CSC/COO input to fit or fit_transform
Plot an embedding	umapers.plot.points(...)
Compute trustworthiness	umapers.diagnostics.trustworthiness_report(...)
Reuse a precomputed kNN graph	Umap.fit_transform_with_knn(...)
Parametric convenience workflow	ParametricUmap
Aligned convenience workflow	AlignedUmap

Known Gaps

• sparse-trained inverse_transform is not supported yet
• ParametricUmap is a lightweight convenience workflow, not a Keras/TensorFlow replacement
• AlignedUmap covers only part of the relation modes available in umap-learn
• approximate-neighbor behavior is not identical to pynndescent
• arbitrary precomputed graph input is not exposed as a general public API

Source Build

pip install --upgrade pip maturin
maturin develop --release --manifest-path umap_rs/Cargo.toml
python -I -m pytest -q umap_rs/tests/test_binding.py