class

YOLOV2ForObjectDetection

extendsPretrainedModel

YOLOV2ForObjectDetection(config: YOLOV2Config)

source edit

YOLOv2 / YOLO9000 object detector (Redmon & Farhadi, CVPR 2017).

The successor to YOLOv1 with several decisive upgrades: a deeper Darknet-19 backbone with batch normalisation, anchor boxes (5 k-means-clustered priors on COCO instead of YOLOv1's direct-regression "anchor-free" approach), and a passthrough layer that injects fine-grained stride-16 features into the stride-32 detection head via space-to-depth reorganisation. Anchors give per-cell predictions an inductive bias on shape that YOLOv1 lacks, and the passthrough recovers spatial detail lost in the final downsample — together they bump VOC 2007 mAP from 63.4% (YOLOv1) to 78.6% (paper Table 4).

The output tensor $(B, A \cdot (5 + C), fH, fW)$ encodes $A$ anchors per cell with $(t_x, t_y, t_w, t_h, \mathrm{conf})$ deltas plus $C$ class scores.

Parameters

configYOLOV2Config

Frozen architecture spec. Use yolo_v2 or yolo_v2_tiny for the standard variants.

Attributes

configYOLOV2Config

Stored copy of the config that built this model.

backbone_Darknet19 or _Darknet19Tiny

Darknet-19 trunk producing a 512-channel stride-16 route feature and a 1024-channel stride-32 trunk feature.

det1, det2, det3_conv_bn_lrelu

Stride-32 head convolutions; det3 consumes the concatenation of the 2048-channel space-to-depth passthrough and the 1024-channel det2 output.

prednn.Conv2d

Final 1x1 conv producing

A \cdot (5 + C)

channels (no BatchNorm, no activation).

Notes

See Redmon & Farhadi, "YOLO9000: Better, Faster, Stronger", CVPR 2017 (arXiv:1612.08242). Per-anchor box decoding uses the "direct location prediction" parameterisation:

\begin{aligned} b_x &= \sigma(t_x) + c_x, & b_w &= p_w \exp(t_w), \\ b_y &= \sigma(t_y) + c_y, & b_h &= p_h \exp(t_h), \end{aligned}

where $(c_x, c_y)$ is the cell top-left in feature coords and $(p_w, p_h)$ is the anchor's prior size. The sigmoid-on- $t_{x,y}$ bounds the centre to lie inside the responsible cell — a stability win over YOLOv1's unbounded regression.

Examples

>>> import lucid
>>> from lucid.models.vision.yolo._v2 import yolo_v2
>>> model = yolo_v2()
>>> 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: YOLOV2Config)

source edit

Instance methods

forward

→ObjectDetectionOutput

forward(x: Tensor, targets: list[dict[str, Tensor]] | None = None, image_size: tuple[int, int] = (416, 416))

source edit

Run YOLOv2 forward pass.

Parameters

xTensor

(B, C, H, W) input image tensor.

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

Optional list of B dicts for training loss. Each dict has: - "boxes": (M_i, 4) xyxy boxes normalised to [0,1]. - "labels": (M_i,) integer class indices.

image_sizetuple[int, int]= (416, 416)

(H, W) of the input image in pixels (default 416×416).

Returns

ObjectDetectionOutput

class:~lucid.models._output.ObjectDetectionOutput with:

postprocess

→list[dict[str, Tensor]]

postprocess(output: ObjectDetectionOutput, image_size: tuple[int, int] = (416, 416))

source edit

Per-class NMS on raw predictions.

Parameters

outputObjectDetectionOutput

Output from forward.

image_sizetuple[int, int]= (416, 416)

(H, W) of the input image.

Returns

list[dict[str, Tensor]]

List of per-image result dicts with "boxes", "scores",

class

YOLOV2ForObjectDetection

extendsPretrainedModel

YOLOV2ForObjectDetection(config: YOLOV2Config)

source edit

YOLOv2 / YOLO9000 object detector (Redmon & Farhadi, CVPR 2017).

The output tensor $(B, A \cdot (5 + C), fH, fW)$ encodes $A$ anchors per cell with $(t_x, t_y, t_w, t_h, \mathrm{conf})$ deltas plus $C$ class scores.

Parameters

configYOLOV2Config

Frozen architecture spec. Use yolo_v2 or yolo_v2_tiny for the standard variants.

Attributes

configYOLOV2Config

Stored copy of the config that built this model.

backbone_Darknet19 or _Darknet19Tiny

Darknet-19 trunk producing a 512-channel stride-16 route feature and a 1024-channel stride-32 trunk feature.

det1, det2, det3_conv_bn_lrelu

Stride-32 head convolutions; det3 consumes the concatenation of the 2048-channel space-to-depth passthrough and the 1024-channel det2 output.

prednn.Conv2d

Final 1x1 conv producing

A \cdot (5 + C)

channels (no BatchNorm, no activation).

Notes

See Redmon & Farhadi, "YOLO9000: Better, Faster, Stronger", CVPR 2017 (arXiv:1612.08242). Per-anchor box decoding uses the "direct location prediction" parameterisation:

\begin{aligned} b_x &= \sigma(t_x) + c_x, & b_w &= p_w \exp(t_w), \\ b_y &= \sigma(t_y) + c_y, & b_h &= p_h \exp(t_h), \end{aligned}

Examples

>>> import lucid
>>> from lucid.models.vision.yolo._v2 import yolo_v2
>>> model = yolo_v2()
>>> 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: YOLOV2Config)

source edit

Instance methods

forward

→ObjectDetectionOutput

forward(x: Tensor, targets: list[dict[str, Tensor]] | None = None, image_size: tuple[int, int] = (416, 416))

source edit

Run YOLOv2 forward pass.

Parameters

xTensor

(B, C, H, W) input image tensor.

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

Optional list of B dicts for training loss. Each dict has: - "boxes": (M_i, 4) xyxy boxes normalised to [0,1]. - "labels": (M_i,) integer class indices.

image_sizetuple[int, int]= (416, 416)

(H, W) of the input image in pixels (default 416×416).

Returns

ObjectDetectionOutput

class:~lucid.models._output.ObjectDetectionOutput with:

postprocess

→list[dict[str, Tensor]]

postprocess(output: ObjectDetectionOutput, image_size: tuple[int, int] = (416, 416))

source edit

Per-class NMS on raw predictions.

Parameters

outputObjectDetectionOutput

Output from forward.

image_sizetuple[int, int]= (416, 416)

(H, W) of the input image.

Returns

list[dict[str, Tensor]]

List of per-image result dicts with "boxes", "scores",