Module yolo26 

YOLO26 training loss: CIoU box regression + BCE classification.

Pipeline (per training step)

1. Decode raw LTRB predictions to XYWH using the anchor grid.
2. Assign GT boxes to anchors (CPU-side TaskAlignedAssigner).
3. Box loss: CIoU forward (saved activations) then backward.
4. Cls loss: BCE cls backward on all anchors.
5. Decode backward: d_xywh → d_ltrb through the linear decode.
6. Return d_boxes [B, 4*A] and d_scores [B, nc*A] (host f32).

Re-exports

pub use super::anchor::AnchorGrid;

pub use super::assign::TaskAlignedAssigner;

Structs

Yolo26Loss

CUDA training loss for YOLO26: computes d_boxes and d_scores.

Functions

ciou_forward_cpu

CPU reference CIoU forward — returns saved activations (iou, v, alpha).