Skip to main content

yolo_bce_cls_loss_backward

vision_rs::models::yolo::kernels::loss::cls

Function yolo_bce_cls_loss_backward 

Source 
pub fn yolo_bce_cls_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>,
    d_pred_ptr: T::Pointer<D>,
    N: i32,
    C: 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

BCE classification loss backward: upstream gradient + forward inputs → ∂L/∂pred.

Per element: d_pred[c,n] = dy[n] · (sigmoid(pred[c,n]) − target[c,n])

Uses the numerically stable sigmoid: σ(x) = 1 / (1 + exp(−x)). No saved activations are needed because the gradient depends only on the forward inputs (pred logits and targets).

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