yolo_v3(pretrained: bool | str = False, weights: YOLOV3Weights | None = None, overrides: object = {})YOLOv3 with Darknet-53 backbone, 3-scale detection (Redmon & Farhadi, 2018).
Builds the paper-cited YOLOv3 detector: 53-layer residual Darknet backbone + FPN-style top-down head producing predictions at strides 8 / 16 / 32 with 3 anchors per cell per scale (9 anchors total, COCO k-means clustered). Default 80 COCO classes; reaches COCO test-dev AP of 33.0% at 416x416 input (paper Table 3).
Model Size
Parameters
pretrainedbool or str= FalsePretrained-weight selector.
False → random init; True →
the DEFAULT tag (YOLOV3Weights.COCO_2014, converted
from darknet's yolov3.weights); a tag string → that specific
checkpoint. Mutually exclusive with weights (which wins if
both are given).Explicit weights enum member, e.g.
YOLOV3Weights.COCO_2014.
Takes precedence over pretrained.**overridesobject= {}Keyword overrides forwarded into
YOLOV3Config.Returns
YOLOV3ForObjectDetectionDetector with the standard YOLOv3 configuration applied (or with
overrides merged on top of it).
Notes
See Redmon & Farhadi, "YOLOv3: An Incremental Improvement", 2018 (arXiv:1804.02767). The multi-scale head is the principal upgrade over YOLOv2 and the source of its small-object accuracy gain.
Examples
>>> import lucid
>>> from lucid.models.vision.yolo._v3 import yolo_v3
>>> model = yolo_v3()
>>> x = lucid.randn(1, 3, 416, 416)
>>> out = model(x)
>>> out.logits.shape[0]
1