Skip to main content

yolo_ciou_loss_backward

vision_rs::models::yolo::kernels::loss::ciou

Function yolo_ciou_loss_backward 

Source 
pub fn yolo_ciou_loss_backward<T: Triton, D: Float, const BLOCK_N: i32>(
    dy_ptr: T::Pointer<D>,
    pred_ptr: T::Pointer<D>,
    target_ptr: T::Pointer<D>,
    iou_ptr: T::Pointer<D>,
    v_ptr: T::Pointer<D>,
    alpha_ptr: T::Pointer<D>,
    d_pred_ptr: T::Pointer<D>,
    N: i32,
)
where
    T::I32Tensor: Tensor<i32, 1> + Comparison<i32, BoolTensor = T::BoolTensor>,
    T::Pointer<D>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<D>>>,
Expand description

Fused CIoU loss backward: computes ∂L/∂pred given the upstream gradient and the saved activations from the forward pass.

Only produces gradients w.r.t. pred; target boxes are treated as constants.

Gradient decomposes into three independent parts:

(a) IoU term: ∂(−IoU)/∂pred (b) Center-distance term: ∂(d²/(c²+ε))/∂pred (c) α·v term: ∂(α·v)/∂(pw,ph) — zero for (px,py)

The min/max branching for intersection and enclosing-box corners is re-derived from (pred, target) rather than saved, since pred+target fully determine which branch was active.

Grid: cdiv(N, BLOCK_N) — one CTA per anchor tile.