Save a trained UMAP model plus the scaling needed to reproduce its input

The scaling parameters travel WITH the model. Re-deriving median/MAD from a new batch would scale it against itself, so the projection would land in a space subtly different from the one the model was trained in – a silent error that looks like a batch effect.

Usage

save_umap_model(
  model,
  path,
  scale_params,
  features,
  meta = list(),
  transform = NULL
)

Arguments

model

A trained uwot model.

path

File path.

scale_params

Scaling constants that must accompany the model for a projection to land in the same space.

features

Character vector of feature names the model was trained on.

meta

Named list of metadata stored alongside the model. Default list().

transform

the run's transform from make_transform(), whose method and parameters are stored so a projection can carry a later run onto it.