Training Tool

June 22, 2026 · View on GitHub

Config-based training for oriented object detection.

Running

uv pip install -e .
odet train --config configs/oriented_rcnn/dota_le90_1x.json --batch-size 4

From the repo root you can also use make train (same default CONFIG).

What happens

  1. Dataset — paths and augmentations from JSON (dataset, augmentation sections)
  2. Model — model_type selects Oriented R-CNN, Rotated Faster R-CNN, or Rotated RetinaNet
  3. Training loop — AMP, gradient accumulation, checkpointing, TensorBoard (see Training user guide)
  4. Validation — metrics during training per evaluation.* in config

Key concepts

  • Config inheritance — _base_ merges fragments from configs/_base_/
  • Checkpoints — written under runs/<model_type>/<timestamp>/
  • Overrides — CLI flags override JSON fields where supported (odet train --help)

See Configuration for the full schema.