OmniCloudMask model weights

ocm_fetch_weights() downloads garry's mirror of the official OmniCloudMask v4 weights (two safetensors files, about 58 MB total, unmodified from upstream; MIT licensed by DPIRD-DMA, see the release's NOTICE.md for attribution and citation) into a per-user data directory, verifying each file's content hash. It is idempotent: files already present and intact are not re-downloaded. ocm_model() finds this directory automatically, so most users need nothing beyond a one-off ocm_fetch_weights().

Usage

ocm_fetch_weights(dir = NULL, quiet = FALSE)

ocm_load_weights(dir, models = c("regnety", "edgenext"))

Arguments

dir

For ocm_fetch_weights(): destination directory (default: tools::R_user_dir("garry", "data")/ocm-v4). For ocm_load_weights(): directory containing the OCM v4 .safetensors files.

quiet

Suppress progress output.

models

Which ensemble members to load ("regnety", "edgenext").

Value

ocm_fetch_weights() returns the weights directory, invisibly. ocm_load_weights() returns a list with weights (one entry per model), kernel_id, and paths.

Details

ocm_load_weights() is the lower-level loader ocm_model() uses: it reads the safetensors state dicts, folds batch norms into their convolutions, and returns the nested weight lists the native forward pass consumes. Results are cached as an .rds under tools::R_user_dir("garry", "cache"), keyed by the content hash of the weight files. Call it directly only to point at a non-standard weights directory, for example the Python package's download cache (~/.local/share/omnicloudmask/<version>/).

See also

ocm_model(), ocm_mask() and ocm_predict() for running the model; safetensors_read() for the underlying file format.