correct overlapping checkpoint evaluation
This commit is contained in:
@@ -35,6 +35,14 @@ excludes rather than relabels any positive tile that would become empty, and
|
||||
writes complete source/output hashes plus a row-level removal manifest. It
|
||||
never creates a training-release manifest or human-review claim.
|
||||
|
||||
`derive_yolo_nonoverlap_evaluation_view.py` selects the complete non-overlap
|
||||
grid from an existing overlapping validation corpus, records every excluded
|
||||
low-variance/no-data tile, leaves train directories empty and writes a
|
||||
checksum-bound `NO_TRAINING.json`. The operator training wrapper rejects that
|
||||
marker before release verification or CUDA allocation. Non-overlap proves only
|
||||
that tiles do not share pixels; checkpoint reports explicitly avoid claiming
|
||||
statistical independence for adjacent tiles from the same AOI.
|
||||
|
||||
Setup-, import-, demo- en maintenance-scripts voor GeoIntel.
|
||||
|
||||
## WALOUS source provisioning
|
||||
|
||||
Reference in New Issue
Block a user