YOLO_V4Config

class lucid.models.YOLO_V4Config(num_classes: int, anchors: list[list[tuple[int, int]]] = <factory>, strides: list[int] = <factory>, backbone: lucid.nn.module.Module | None = None, backbone_out_channels: tuple[int, int, int] | None = None, in_channels: tuple[int, int, int] = (256, 512, 1024), pos_iou_thr: float = 0.25, ignore_iou_thr: float = 0.5, obj_balance: tuple[float, float, float] = (4.0, 1.0, 0.4), cls_label_smoothing: float = 0.0, iou_aware_alpha: float = 0.5, iou_branch_weight: float = 1.0)

YOLO_V4Config stores the 3-scale anchor setup, strides, backbone selection, neck channel widths, assignment thresholds, and IoU-aware options used by lucid.models.YOLO_V4.

Class Signature

@dataclass
class YOLO_V4Config:
    num_classes: int
    anchors: list[list[tuple[int, int]]] = field(default_factory=...)
    strides: list[int] = field(default_factory=...)
    backbone: nn.Module | None = None
    backbone_out_channels: tuple[int, int, int] | None = None
    in_channels: tuple[int, int, int] = (256, 512, 1024)
    pos_iou_thr: float = 0.25
    ignore_iou_thr: float = 0.5
    obj_balance: tuple[float, float, float] = (4.0, 1.0, 0.4)
    cls_label_smoothing: float = 0.0
    iou_aware_alpha: float = 0.5
    iou_branch_weight: float = 1.0

Validation

• num_classes must be greater than 0.

• anchors must contain three scales with three positive integer anchors each.

• strides must contain exactly three positive values.

• backbone must be an nn.Module or None.

• backbone_out_channels must be omitted for the default backbone and required for a custom backbone.

• in_channels and backbone_out_channels must contain exactly three positive integers.

• IoU thresholds and iou_aware_alpha must be in [0, 1].

• obj_balance must contain exactly three positive values.

• cls_label_smoothing must be in [0, 1).

• iou_branch_weight must be non-negative.

Usage

import lucid.models as models
import lucid.nn as nn

class ToyBackbone(nn.Module):
    def forward(self, x):
        return x, x, x

config = models.YOLO_V4Config(
    num_classes=3,
    backbone=ToyBackbone(),
    backbone_out_channels=(3, 3, 3),
    in_channels=(3, 3, 3),
    iou_aware_alpha=0.0,
    iou_branch_weight=0.0,
)
model = models.YOLO_V4(config)