Pretrained Hub (library)

July 11, 2026 · View on GitHub

Manifest-driven downloads from Hugging Face Hub.

  • Manifest: manifest.json — repo_id, revision, and registered assets (slug → hashed filename + metadata).
  • Default repo: dl4eo/oriented-det-pretrained (override with ORIENTED_DET_HF_REPO_ID).
  • Local cache directory: <oriented-det-install>/pretrained/ (framework root, not the product repo). Override with ORIENTED_DET_PRETRAINED_DIR.

Asset ids (slugs)

Use manifest slugs (not mAP numbers) with hf:// and the CLI:

  • oriented_rcnn_dota_le90_1x
  • oriented_rcnn_dota_le90_3x
  • rotated_faster_rcnn_dota_le90_1x
  • rotated_faster_rcnn_dota_le90_3x
  • rotated_retinanet_dota_le90_1x
  • rotated_retinanet_dota_le90_3x

On-disk / Hub filenames include a SHA-256[:8] suffix (see tools/publish_checkpoint.py). Each published weight can have sidecar artifacts beside it in pretrained/: <weight-stem>.json for the exact final run config and <weight-stem>.log for the training log.

Usage

from oriented_det.pretrained import ensure_checkpoint

path = ensure_checkpoint("hf://oriented_rcnn_dota_le90_1x")

odet pretrained download oriented_rcnn_dota_le90_1x
odet pretrained list

Training (tools/train.py) and inference call ensure_checkpoint when checkpoint.load_from_checkpoint points at a registered asset that is missing locally. Simple inference (odet image-demo, odet preds, deploy/export paths through oriented_det.runtime.checkpoint) prefers the sidecar JSON only when the provided config is that checkpoint's manifest source_recipe. A different user-provided config is kept as-is.

Publishing weights

  1. python tools/publish_checkpoint.py <run_ckpt> pretrained/<basename_without_hash>
  2. Copy <experiment>/config.json and train.log to pretrained/<weight-stem>.json and .log
  3. Update manifest.json (filename, sha256, source_recipe, eval_map50, …). eval_map50 is mAP50 from odet preds on val tiles (same protocol as make eval-val), not compute_map_final from training — see pretrained/README.md.
  4. Upload: make upload-pretrained (.pth plus sidecar .json / .log when present)

See also pretrained/README.md at the repository root.