Detection — YOLO (Ultralytics)¶
The full Ultralytics integration documentation is at Ultralytics Integration.
This page summarises what WLAwareTrainer handles automatically and where
the example lives.
Example: weightslab/examples/Ultralytics/wl-detection/main.py
What the example does¶
One YOLO.train(trainer=WLAwareTrainer, ...) call gives you:
Per-sample box / cls / dfl loss and live NMS overlay (train split).
Per-sample IoU and post-NMS overlay (val split).
Aggregate
train/{box,cls,dfl}andval/{precision,recall,mAP50,...}curves in the studio.Discard control: samples removed in the studio are silently excluded from future batches on both splits.
Integration in three lines¶
wl.watch_or_edit(cfg, flag="hyperparameters", defaults=cfg, poll_interval=1.0)
wl.serve()
YOLO("yolo11n.pt").train(trainer=WLAwareTrainer, data=..., workers=0, amp=False)
wl.keep_serving()
Required kwargs: workers=0 (UID counter lives in the parent process),
amp=False (autocast doesn’t see through the ModelInterface wrapper).
See Ultralytics Integration for config walkthrough, platform notes, and the full end-to-end sequence.