Parameter groups¶

The escape hatch for every knob that is not a constructor argument. A field left at None is not sent, and the crate's default applies.

NeighbourParams `dataclass` ¶

NeighbourParams(
    n_tree: int | None = None,
    search_budget: int | None = None,
    m: int | None = None,
    ef_construction: int | None = None,
    ef_search: int | None = None,
    diversify_prob: float | None = None,
    delta: float | None = None,
    ef_budget: int | None = None,
    extract_knn: bool | None = None,
    bt_budget: float | None = None,
    n_list: int | None = None,
    n_probes: int | None = None,
)

Backend-specific knobs for the CPU neighbour search.

Only the fields belonging to the backend you chose have any effect; the rest are ignored by the core. metric is not here because it is a constructor argument on every estimator.

Attributes:

Name	Type	Description
`n_tree`	`int \| None`	Annoy. Trees in the forest. More means better recall and a slower build.
`search_budget`	`int \| None`	Annoy. Candidates inspected per query. `None` gives `k * n_tree * 20`.
`m`	`int \| None`	HNSW. Edges per node on the upper layers, `2 * m` on layer 0.
`ef_construction`	`int \| None`	HNSW. Candidate list width during the build.
`ef_search`	`int \| None`	HNSW. Beam width at query time, the recall knob.
`diversify_prob`	`float \| None`	NN-Descent. Diversification probability applied to the finished graph.
`delta`	`float \| None`	NN-Descent. Convergence threshold, as a fraction of neighbours updated in an iteration.
`ef_budget`	`int \| None`	NN-Descent. Beam budget when querying. `None` picks one. No effect when `extract_knn` is on, since no search runs.
`extract_knn`	`bool \| None`	NN-Descent. Return the graph the descent already built instead of searching it. On by default: a self-kNN query re-searches a graph that is already a kNN graph. Measured on 20k points in 50D, identical recall and about 25% faster at `k=15`; at `k=50` the graph is widened to cover the request, which is slower to build but reaches perfect recall where the beam search drops to 0.989.
`bt_budget`	`float \| None`	Ball tree. Fraction of the dataset to visit per query.
`n_list`	`int \| None`	IVF. Voronoi cells. `None` gives `sqrt(n)`.
`n_probes`	`int \| None`	IVF. Cells visited per query. `None` gives `sqrt(n_list)`.

NeighbourParamsGpu `dataclass` ¶

NeighbourParamsGpu(
    n_list: int | None = None,
    n_probes: int | None = None,
    k: int | None = None,
    k_build: int | None = None,
    n_tree: int | None = None,
    delta: float | None = None,
    rho: float | None = None,
    beam_width: int | None = None,
    max_beam_iters: int | None = None,
    n_entry_points: int | None = None,
    extract_knn: bool | None = None,
)

Backend-specific knobs for the GPU neighbour search.

A different set from NeighbourParams, not a subset: the device backends build a CAGRA graph and search it with a beam, neither of which has a CPU counterpart.

Attributes:

Name	Type	Description
`n_list`	`int \| None`	IVF-GPU. Voronoi cells. `None` gives `sqrt(n)`.
`n_probes`	`int \| None`	IVF-GPU. Cells visited per query.
`k`	`int \| None`	NN-Descent-GPU. Node degree after pruning. `None` gives 30, except under t-SNE where it is backfilled to `3 * perplexity`.
`k_build`	`int \| None`	NN-Descent-GPU. Node degree before pruning.
`n_tree`	`int \| None`	NN-Descent-GPU. Trees used to seed the graph.
`delta`	`float \| None`	NN-Descent-GPU. Convergence threshold.
`rho`	`float \| None`	NN-Descent-GPU. Sampling rate per iteration.
`beam_width`	`int \| None`	NN-Descent-GPU. Beam width when querying.
`max_beam_iters`	`int \| None`	NN-Descent-GPU. Beam iterations when querying.
`n_entry_points`	`int \| None`	NN-Descent-GPU. Entry points per query.
`extract_knn`	`bool \| None`	NN-Descent-GPU. Return the built CAGRA graph instead of searching it. On by default, as for the CPU backend, and worth about a third off a full GPU embedding: 0.74s to 0.48s at 20k points in 50D. It can cost bit-reproducibility on some inputs; see the reproducibility section of the guide. `beam_width`, `max_beam_iters` and `n_entry_points` do nothing when it is on.

UmapGraph `dataclass` ¶

UmapGraph(
    bandwidth: float | None = None,
    local_connectivity: float | None = None,
    mix_weight: float | None = None,
)

Fuzzy simplicial set construction.

Attributes:

Name	Type	Description
`bandwidth`	`float \| None`	Convergence tolerance for the smooth-kNN binary search that finds each point's sigma.
`local_connectivity`	`float \| None`	Neighbours assumed to sit at distance zero. Raising it makes the local neighbourhood denser and the embedding tighter.
`mix_weight`	`float \| None`	Balance between the fuzzy union and the directed graph during symmetrisation. `1.0` is the plain union.

UmapOptim `dataclass` ¶

UmapOptim(
    a: float | None = None,
    b: float | None = None,
    gamma: float | None = None,
    neg_sample_rate: int | None = None,
    beta1: float | None = None,
    beta2: float | None = None,
    eps: float | None = None,
)

UMAP optimiser knobs beyond the epochs and learning rate.

Attributes:

Name	Type	Description
`a`	`float \| None`	Repulsion curve numerator. Fitted from `min_dist` and `spread` unless you set it, and setting one of `a` / `b` without the other is rarely what you want.
`b`	`float \| None`	Repulsion curve exponent. See `a`.
`gamma`	`float \| None`	Weight on the repulsive term.
`neg_sample_rate`	`int \| None`	Negative samples drawn per positive edge.
`beta1`	`float \| None`	Adam first-moment decay. The crate uses 0.5 for UMAP rather than the usual 0.9.
`beta2`	`float \| None`	Adam second-moment decay.
`eps`	`float \| None`	Adam denominator epsilon.

TsneOptim `dataclass` ¶

TsneOptim(
    early_exag_iter: int | None = None,
    early_exag_factor: float | None = None,
    late_exag_factor: float | None = None,
    theta: float | None = None,
    n_interp_points: int | None = None,
)

t-SNE optimiser knobs beyond the epochs and learning rate.

Attributes:

Name	Type	Description
`early_exag_iter`	`int \| None`	Iterations of early exaggeration.
`early_exag_factor`	`float \| None`	Multiplier on the affinities during those iterations.
`late_exag_factor`	`float \| None`	Multiplier for the remaining iterations. `None` disables it. Above roughly 100k points a value near 4 keeps cluster structure from dispersing once early exaggeration ends.
`theta`	`float \| None`	Barnes-Hut opening angle. Larger is faster and coarser; `0` makes it exact and very slow.
`n_interp_points`	`int \| None`	Interpolation points per box on the FFT path. No effect under Barnes-Hut.

PacmapOptim `dataclass` ¶

PacmapOptim(
    beta1: float | None = None,
    beta2: float | None = None,
    eps: float | None = None,
    phase1_end: int | None = None,
    phase2_end: int | None = None,
)

PaCMAP optimiser knobs beyond the epochs and learning rate.

The three phases are what PaCMAP does instead of early exaggeration: the mid-near weight starts high, decays to zero across phase 2, and phase 3 is near pairs and repulsion alone.

Attributes:

Name	Type	Description
`beta1`	`float \| None`	Adam first-moment decay.
`beta2`	`float \| None`	Adam second-moment decay.
`eps`	`float \| None`	Adam denominator epsilon.
`phase1_end`	`int \| None`	Last epoch of the mid-near dominant phase.
`phase2_end`	`int \| None`	Last epoch of the decay phase.

DensParams `dataclass` ¶

DensParams(
    frac: float | None = None,
    var_shift: float | None = None,
)

Density-preservation knobs beyond the weight.

Attributes:

Name	Type	Description
`frac`	`float \| None`	Fraction of the run, at the end, over which the density term is active. It is switched on late so the embedding has settled first.
`var_shift`	`float \| None`	Additive shift on the variance of the embedding log-radii, which keeps the correlation defined when the spread is tiny.

PhateDiffusion `dataclass` ¶

PhateDiffusion(
    bandwidth_scale: float | None = None,
    thresh: float | None = None,
    graph_symmetry: str | None = None,
    n_landmarks: int | None = None,
    landmark_method: str | None = None,
    n_svd: int | None = None,
    t_max: int | None = None,
)

PHATE diffusion operator knobs beyond decay, gamma and t.

Attributes:

Name	Type	Description
`bandwidth_scale`	`float \| None`	Multiplier on the adaptive kernel bandwidth.
`thresh`	`float \| None`	Affinities below this are zeroed, which is what keeps the operator sparse.
`graph_symmetry`	`str \| None`	`"add"`, `"multiply"`, `"mnn"` or `"none"`.
`n_landmarks`	`int \| None`	Landmarks to diffuse on instead of the full graph. Worth setting above roughly 50k points.
`landmark_method`	`str \| None`	`"spectral"`, `"random"` or `"density"`.
`n_svd`	`int \| None`	Components for spectral landmark selection.
`t_max`	`int \| None`	Largest diffusion time the VNE knee search will consider. Ignored when `t` is pinned.

Parameter groups¶

NeighbourParams dataclass ¶

NeighbourParamsGpu dataclass ¶

UmapGraph dataclass ¶

UmapOptim dataclass ¶

TsneOptim dataclass ¶

PacmapOptim dataclass ¶

DensParams dataclass ¶

PhateDiffusion dataclass ¶

NeighbourParams `dataclass` ¶

NeighbourParamsGpu `dataclass` ¶

UmapGraph `dataclass` ¶

UmapOptim `dataclass` ¶

TsneOptim `dataclass` ¶

PacmapOptim `dataclass` ¶

DensParams `dataclass` ¶

PhateDiffusion `dataclass` ¶