Skip to contents

Load a pretrained encoder into an edaphos_foundation_moco wrapper

Source: R/foundation_weights.R
foundation_weights_load.Rd

Restores a pretrained MoCo v2 encoder from a .pt state-dict file (or by name from the published registry) into an in-memory edaphos_foundation_moco object that is shape-compatible with the output of foundation_moco_pretrain_tiles(). The restored object is ready to feed into foundation_moco_embed(), foundation_fit_classifier() or foundation_fit_regressor() without any additional plumbing.

Usage

foundation_weights_load(
  source,
  n_channels = NULL,
  feature_dim = NULL,
  proj_dim = NULL,
  cache_dir = NULL,
  overwrite = FALSE,
  verbose = FALSE
)

Arguments

source

Either a character name registered in foundation_weights_list() — in which case the encoder is downloaded (cached) via foundation_weights_download() — or a path to a local .pt file. For the local path case, pass the matching metadata as n_channels / feature_dim / etc.

n_channels, feature_dim, proj_dim

Optional integers; when source is a local path these must describe the shape the .pt file was saved from. When source is a registry name they are pulled from the registry.

cache_dir, overwrite, verbose

Forwarded to foundation_weights_download() when source is a registry name.

Value

An edaphos_foundation_moco object with a populated $encoder_q (the query encoder; the momentum encoder is not restored because it is only needed during training).

Examples

if (FALSE) { # \dontrun{
  moco <- foundation_weights_load("edaphos-cerrado-moco-v1")
  emb  <- foundation_moco_embed(moco, patches)
} # }