GPU: fast Louvain clustering on the data — rs_fast_cluster

GPU equivalent of bixverse::rs_fast_cluster_sc. Runs k-means clustering on the WGPU backend, followed by a kNN detection on the centroids to then run Louvain clustering on the graph and propagate the membership back to the original data. Everything after the k-means stays on the CPU.

Usage

rs_fast_cluster_gpu(
  embd,
  resolutions,
  n_centroids,
  fc_params,
  snn,
  return_kmeans,
  seed,
  verbose
)

Arguments

embd: Numeric matrix. The original embedding.
resolutions: Numeric vector. The Louvain resolutions to iterate through.
n_centroids: Optional integer. The number of clusters to find. If not provided, defaults to sqrt(nrow(embd)).
fc_params: Named list. See params_sc_fast_cluster_gpu().
snn: Boolean. Shall the kNN graph be additionally transformed into an sNN graph.
return_kmeans: Boolean. Shall the k-means centroids and assignments be returned alongside the memberships.
seed: Integer. For reproducibility.
verbose: Integer. 0L - quiet; 1L - normal verbosity; 2L - detailed verbosity.

Value

A list with the following elements:

membership - The memberships across the different resolutions.
k_means_cluster - Optional integer vector of k-means assignments.
centroids - Optional numeric matrix of k-means centroids.