The export-targets epic (#784) ships two dialects: ultralytics (data.yaml with path and indexed names) and yolov5-yaml (data.yaml with nc and a names list). The original darknet trainer reads a different descriptor family — obj.data, obj.names, and train.txt index files — which is what YOLOv3 and YOLOv4 consume when trained outside the Ultralytics ecosystem.
Deferred from the epic:
- No shipped target requires it: YOLOv3 resolves to
ultralytics, and YOLOv4 is not a target in either shipped dialect — Ultralytics itself does not support v4, so a darknet dialect is the only home it could have.
- The writer is mechanical once wanted (the label line grammar is shared; only the descriptor differs), so it costs little to add later and nothing to defer.
Scope when picked up: a darknet dialect with obj.data / obj.names / train.txt (and valid.txt when a val fold exists), targets for the darknet-trained YOLOv3/YOLOv4 lineage, and a round-trip or golden-file proof consistent with the epic's acceptance style.
Not started; do not close this issue until a decision comment lands here.
The export-targets epic (#784) ships two dialects:
ultralytics(data.yamlwithpathand indexednames) andyolov5-yaml(data.yamlwithncand anameslist). The original darknet trainer reads a different descriptor family —obj.data,obj.names, andtrain.txtindex files — which is what YOLOv3 and YOLOv4 consume when trained outside the Ultralytics ecosystem.Deferred from the epic:
ultralytics, and YOLOv4 is not a target in either shipped dialect — Ultralytics itself does not support v4, so a darknet dialect is the only home it could have.Scope when picked up: a
darknetdialect withobj.data/obj.names/train.txt(andvalid.txtwhen a val fold exists), targets for the darknet-trained YOLOv3/YOLOv4 lineage, and a round-trip or golden-file proof consistent with the epic's acceptance style.Not started; do not close this issue until a decision comment lands here.