yolo_v2_tiny¶
ConvNet One-Stage Detector Object Detection
The yolo_v2_tiny function returns an instance of the YOLO_V2 model, configured with a lightweight version of the original YOLO-v2 architecture using a compact DarkNet-19 variant.
Total Parameters: 15,863,821
Function Signature¶
def yolo_v2_tiny(num_classes: int = 20, **kwargs) -> YOLO_V2
Parameters¶
num_classes (int, default=20): Number of object categories to detect. This sets the number of output class scores per anchor in the final detection head.
kwargs: Additional keyword arguments passed directly to the YOLO_V2 constructor.
Common kwargs include:
darknet (nn.Module, optional): custom lightweight backbone
num_anchors (int, default=5): number of anchor boxes (defaults to 5)
image_size (int, default=416): input image resolution
Returns¶
YOLO_V2: An initialized instance of the YOLO_V2 detection model using a tiny backbone.
Example Usage¶
>>> from lucid.models import yolo_v2_tiny
>>> model = yolo_v2_tiny(num_classes=20)
>>> print(model)
>>> x = lucid.rand(1, 3, 416, 416)
>>> out = model(x)
>>> print(out.shape) # shape: (1, 125, 13, 13) for Pascal-VOC (20 classes, 5 anchors)
Note
The tiny version uses a compressed version of the DarkNet-19 backbone, significantly reducing model size and computation while maintaining similar output format to the original YOLO-v2.
The model follows the YOLO-v2 head design with 5 anchor boxes and a grid output shape of \(S \times S \times (B \times (5 + C))\), where \(C\) is the number of classes, \(B=5\) is the number of anchors, and \(S=13\) when input image size is 416.