Run independent checks on an exported training dataset: the integrity
inventory (schema, safe paths, symbolic links, sizes and SHA-256 hashes,
missing and unlisted files), the manifest schema, split leakage (every
group and image in exactly one split, every tile in its image's split),
mask dimensions and orientation, integer label range against the class
legend, the recorded overlap counts, and a mask -> ROI -> mask round trip
on a sample of tiles. When the source x is supplied, each sampled tile is
also re-rasterised from the annotations and compared exactly. Tile files
outside the verified inventory are never opened. A check that cannot run
on a malformed manifest is reported as "fail" rather than raising an
error; only an unsupported major version aborts (PROTOCOL_MISMATCH).
Usage
at_training_check(
dataset,
x = NULL,
max_roundtrip_tiles = 20L,
call = rlang::caller_env()
)Arguments
- dataset
Path of a dataset written by
at_training_export().- x
Optional source annot_session or annot_project.
- max_roundtrip_tiles
Maximum number of tiles sampled for round trips.
- call
The calling environment, for error reporting.
See also
Other training:
at_training_export(),
at_training_import()