Crate kuji 

Compatibility crate for the kuji to drawset rename.

New code should depend on drawset.

Modules

gumbel

Gumbel-max sampling.

neighbor

Graph neighbor sampling.

qmc

Quasi-Monte Carlo sequences.

reservoir

Reservoir sampling.

thinning

Kernel thinning: coreset selection minimizing kernel discrepancy.

Structs

NeighborSampler

Sampler for graph neighborhoods.

ReservoirSampler

A reservoir sampler that maintains a uniform sample of size k from a stream.

ReservoirSamplerR

A reservoir sampler using Algorithm R (Vitter, 1985).

SobolGenerator

Incremental Sobol sequence generator using Gray-code enumeration.

WeightedReservoirSampler

A weighted reservoir sampler (Efraimidis-Spirakis, A-Res).

Functions

gumbel_max_sample

Sample an index using the Gumbel-max trick.

gumbel_noise

Generate Gumbel noise: G = -log(-log(U)) where U ~ Uniform(0, 1).

gumbel_softmax

Gumbel-Softmax: differentiable approximation to categorical sampling.

gumbel_topk_sample

Sample k indices without replacement using the Gumbel-top-k trick.

gumbel_topk_sample_with_rng

Gumbel-top-k with a caller-supplied RNG (for tests/benchmarks).

halton_point

Compute a single Halton point in [0, 1)^d.

halton_sequence

Generate n Halton points in [0, 1)^d, starting from index = 1.

kernel_herd

Kernel herding: deterministic sampling via greedy mean embedding matching.

kernel_thin

Greedy kernel thinning via MMD minimization.

mmd_sq_from_gram

Compute MMD^2 (biased) between a subset and the full set from a Gram matrix.

relaxed_topk_gumbel

Relaxed Top-K via Gumbel-Softmax.

sobol_scrambled

Generate n Owen-scrambled Sobol points in [0, 1)^d.

sobol_sequence

Generate n Sobol points in [0, 1)^d, starting from index 1 (the origin at index 0 is skipped).