class

YOLOV3TinyForObjectDetection

extendsObjectDetectionModel

YOLOV3TinyForObjectDetection(config: YOLOV3Config)

source edit

YOLOv3-Tiny — the 2-scale lightweight YOLOv3 (Redmon & Farhadi, 2018).

A shallow Darknet backbone (six conv + max-pool stages, no residual connections) feeding two detection scales — stride 32 and stride 16 — each through yolov3-tiny.cfg's two-convolution head rather than the full model's five. Boxes, objectness and per-class sigmoids are decoded exactly as in YOLOV3ForObjectDetection, so the two share their loss and postprocess.

Parameters

configYOLOV3Config

Frozen architecture spec describing exactly two scales. yolo_v3_tiny supplies yolov3-tiny.cfg's table: six anchors, strides (16, 32), masks ((0, 1, 2), (3, 4, 5)). The default YOLOV3Config describes the full model's three scales and is refused with a ValueError.

Attributes

configYOLOV3Config

Stored copy of the config that built this model.

backbone_Darknet53Tiny

Plain conv/max-pool backbone producing the stride-16 (256ch) and stride-32 (256ch bottleneck) features.

p5_compressnn.Sequential

Stride-32 conv 512 3x3 expansion.

p5_predictnn.Conv2d

Stride-32 1x1 predictor emitting

3(5 + C)

channels.

p5_to_p4_conv_ConvBnLeaky

conv 128 1x1 on the stride-32 bottleneck before upsampling.

p4_compressnn.Sequential

Stride-16 conv 256 3x3 over the upsampled-and-concatenated 384 channels.

p4_predictnn.Conv2d

Stride-16 1x1 predictor emitting

3(5 + C)

channels.

Notes

See Redmon & Farhadi, "YOLOv3: An Incremental Improvement", 2018 (arXiv:1804.02767). Tiny is the variant released alongside the full model for real-time and embedded inference.

Examples

>>> import lucid
>>> from lucid.models.vision.yolo import yolo_v3_tiny
>>> model = yolo_v3_tiny().eval()
>>> x = lucid.randn(1, 3, 416, 416)
>>> out = model(x)
>>> dets = model.postprocess(out, image_sizes=[(416, 416)])
>>> sorted(dets[0])
['boxes', 'labels', 'scores']

Used by 1

lucid.models.vision.yolo

Constructors

dunder

init

→None

__init__(config: YOLOV3Config)

source edit

Instance methods

forward

→ObjectDetectionOutput

forward(x: Tensor, targets: list[dict[str, Tensor]] | None = None)

source edit

Run YOLOv3.

Parameters

x

(B, C, H, W) image batch.

targets

Optional list of target dicts per image.

Returns

ObjectDetectionOutput

ObjectDetectionOutput: logits : (B, total_anchors, C) raw class logits. pred_boxes: (B, total_anchors, 4) xyxy decoded boxes. loss : loss scalar when targets provided.

postprocess

→list[dict[str, Tensor]]

postprocess(output: ObjectDetectionOutput, image_sizes: list[tuple[int, int]])

source edit

Filter by score, clip boxes, apply per-class NMS.

Parameters

outputObjectDetectionOutput

Forward pass output.

image_sizeslist[tuple[int, int]]

List of (H, W) per image.

Returns

list[dict[str, Tensor]]

Per-image list of dicts with "boxes", "scores", "labels", on

>>> import lucid >>> from lucid.models.vision.yolo import yolo_v3_tiny >>> model = yolo_v3_tiny().eval() >>> x = lucid.randn(1, 3, 416, 416) >>> out = model(x) >>> dets = model.postprocess(out, image_sizes=[(416, 416)]) >>> sorted(dets[0]) ['boxes', 'labels', 'scores']