class

YOLOV3ForObjectDetection

extendsPretrainedModel

YOLOV3ForObjectDetection(config: YOLOV3Config)

source edit

YOLOv3 multi-scale object detector (Redmon & Farhadi, 2018).

Adds a 53-layer Darknet-53 backbone (residual connections, more capacity than Darknet-19) and an FPN-style multi-scale detection head that predicts on three feature levels — $P_3$ (stride 8), $P_4$ (stride 16), $P_5$ (stride 32) — with three anchors per cell per level (9 anchors total). This multi-scale detection dramatically improves small-object recall over YOLOv2, bringing COCO AP up to 33.0% (paper Table 3). Per-class predictions use independent sigmoids rather than softmax — supporting multi-label classification.

Parameters

configYOLOV3Config

Frozen architecture spec. Use yolo_v3 for the full Darknet-53 model or yolo_v3_tiny for the lightweight 2-scale variant.

Attributes

configYOLOV3Config

Stored copy of the config that built this model.

backbone_Darknet53

Residual backbone producing P3 / P4 / P5 features at strides 8 / 16 / 32 (128 / 256 / 1024 channels).

p5_compress, p5_predictnn.Sequential

Stride-32 detection head: five-conv compress + 3x3 + 1x1 prediction conv producing

3(5 + C)

channels.

p4_compress, p4_predict, p4_to_p3_convnn.Sequential

Stride-16 head with FPN-style top-down upsample + concatenation with the P4 backbone feature.

p3_compress, p3_predict, p5_to_p4_convnn.Sequential

Stride-8 head with another top-down upsample + concatenation with the P3 backbone feature.

Notes

See Redmon & Farhadi, "YOLOv3: An Incremental Improvement", 2018 (arXiv:1804.02767). Per-anchor box decoding reuses the YOLOv2 direct-location parameterisation; the multi-class objective swaps YOLOv2's softmax for per-class binary cross-entropy:

\mathcal{L}_\mathrm{cls} = -\sum_{c = 1}^{C} \bigl[\,y_c \log \sigma(s_c) + (1 - y_c) \log(1 - \sigma(s_c))\bigr],

enabling multi-label predictions where a single object can belong to multiple categories (e.g. "Woman" + "Person").

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

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

xTensor

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

targetslist[dict[str, Tensor]] | None= None

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".

class

YOLOV3ForObjectDetection

extendsPretrainedModel

YOLOV3ForObjectDetection(config: YOLOV3Config)

source edit

YOLOv3 multi-scale object detector (Redmon & Farhadi, 2018).

Parameters

configYOLOV3Config

Frozen architecture spec. Use yolo_v3 for the full Darknet-53 model or yolo_v3_tiny for the lightweight 2-scale variant.

Attributes

configYOLOV3Config

Stored copy of the config that built this model.

backbone_Darknet53

Residual backbone producing P3 / P4 / P5 features at strides 8 / 16 / 32 (128 / 256 / 1024 channels).

p5_compress, p5_predictnn.Sequential

Stride-32 detection head: five-conv compress + 3x3 + 1x1 prediction conv producing

3(5 + C)

channels.

p4_compress, p4_predict, p4_to_p3_convnn.Sequential

Stride-16 head with FPN-style top-down upsample + concatenation with the P4 backbone feature.

p3_compress, p3_predict, p5_to_p4_convnn.Sequential

Stride-8 head with another top-down upsample + concatenation with the P3 backbone feature.

Notes

\mathcal{L}_\mathrm{cls} = -\sum_{c = 1}^{C} \bigl[\,y_c \log \sigma(s_c) + (1 - y_c) \log(1 - \sigma(s_c))\bigr],

enabling multi-label predictions where a single object can belong to multiple categories (e.g. "Woman" + "Person").

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

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

xTensor

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

targetslist[dict[str, Tensor]] | None= None

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".