Diffusion-map-style embedding — cuda_diffusion

Pairwise distances can use the cudaverse CUDA path. Kernel construction and eigendecomposition currently run on the CPU.

Usage

cuda_diffusion_map(
  x,
  n_components = 2L,
  sigma = NULL,
  diffusion_time = 1,
  metric = c("euclidean", "cosine"),
  device = c("auto", "cuda", "cpu"),
  reduced_dim = NULL
)

Arguments

x: Numeric observation-by-feature matrix, compatible cudaverse result, or a SingleCellExperiment with a reduced dimension.
n_components: Output dimensions.
sigma: Gaussian kernel bandwidth. Defaults to the median positive pairwise distance.
diffusion_time: Non-negative diffusion time exponent.
metric: Euclidean or cosine distance.
device: Device passed to cuda_distance().
reduced_dim: For a SingleCellExperiment, the reduced-dimension name to embed. See cuda_umap() for automatic selection.

Value

A cuda_embedding with the stable fields documented by cuda_umap(), stage-level distance/kernel/eigendecomposition provenance, an optional distance_input stage when resident native storage is reused, and an additional eigenvalues element.

Examples

cuda_diffusion_map(
  matrix(rnorm(120), 40, 3),
  n_components = 2,
  device = "cpu"
)
#> <cuda_embedding method=diffusion observations=40 dimensions=2 backend=base-eigen compute_device=cpu>