yolo_v3_tiny¶
ConvNet One-Stage Detector Object Detection
The yolo_v3_tiny function returns an instance of the YOLO_V3 model configured with a lightweight backbone similar to the official YOLOv3-Tiny design.
Total Parameters: 23,450,997 (MS-COCO)
Function Signature¶
def yolo_v3_tiny(num_classes: int = 80, **kwargs) -> YOLO_V3
Parameters¶
num_classes (int, default=80): Number of object categories to detect. This controls the number of class predictions per anchor.
kwargs: Additional keyword arguments forwarded to the YOLO_V3 constructor.
Preconfigured values:
darknet is set to a lightweight custom _DarkNet_53_Tiny backbone
darknet_out_channels_arr=[128, 256, 512] (used to match the 3 feature map channels with head input)
Returns¶
YOLO_V3: A YOLOv3 model instance with reduced complexity and parameter count, optimized for real-time applications and resource-limited environments.
Example Usage¶
>>> from lucid.models import yolo_v3_tiny
>>> model = yolo_v3_tiny(num_classes=80)
>>> print(model)
>>> x = lucid.rand(1, 3, 416, 416)
>>> out = model(x)
>>> for o in out:
... print(o.shape)
Note
This Tiny variant uses the same 3-scale detection architecture as the full YOLOv3 model, but with a more efficient backbone. Each detection head processes feature maps of sizes 13x13, 26x26, and 52x52 (for 416x416 input), using 3 anchors per scale.