Struct Yolo26Loss 

pub struct Yolo26Loss {
    pub grid: AnchorGrid,
    pub assigner: TaskAlignedAssigner,
    pub assigner_o2o: TaskAlignedAssigner,
    pub nc: usize,
    pub cap: Capability,
    /* private fields */
}

CUDA training loss for YOLO26: computes d_boxes and d_scores.

Fields

grid: AnchorGrid

Anchor grid (points + strides) for the model’s input resolution.

assigner: TaskAlignedAssigner

One2many assigner (top_k=10) — used during compute_grads and the o2m head of compute_grads_dual.

assigner_o2o: TaskAlignedAssigner

One2one assigner (top_k=1) — used for the o2o head of compute_grads_dual.

nc: usize

Number of detection classes.

cap: Capability

Target GPU compute capability for compiled kernels.

Implementations

impl Yolo26Loss

pub fn new(img_h: usize, img_w: usize, nc: usize, cap: Capability) -> Self

Builds the loss state for a model trained at img_h × img_w with nc classes.

pub fn compute_grads( &self, device: &CudaDevice<'_>, boxes: &[f32], scores: &[f32], gt_boxes_b: &[Vec<[f32; 4]>], gt_cls_b: &[Vec<usize>], ) -> Result<(Vec<f32>, Vec<f32>)>

Compute loss gradients for one batch (one2many head only).

boxes – raw LTRB predictions [B, 4*A] (host f32, channels-first) scores – class logits [B, nc*A] (host f32, channels-first)

Returns (d_boxes [B,4*A], d_scores [B,nc*A]) as host f32.

pub fn compute_grads_dual( &self, device: &CudaDevice<'_>, boxes_o2m: &[f32], scores_o2m: &[f32], boxes_o2o: &[f32], scores_o2o: &[f32], gt_boxes_b: &[Vec<[f32; 4]>], gt_cls_b: &[Vec<usize>], w_o2m: f32, w_o2o: f32, ) -> Result<(Vec<f32>, Vec<f32>, Vec<f32>, Vec<f32>)>

Compute dual-head loss gradients for consistent assignment training.

Runs TAL assignment independently for both heads using their respective assigners (top_k=10 for o2m, top_k=1 for o2o), then scales the resulting gradients by w_o2m and w_o2o respectively.

Loss weight schedule (ultralytics-style)

• w_o2m = 1.0 (constant throughout training)
• w_o2o = step / total_steps (ramps 0→1 linearly; caller controls the schedule)

Returns (d_boxes_o2m, d_scores_o2m, d_boxes_o2o, d_scores_o2o).