Skip to main content

smooth_l1_loss_forward

teeny_kernels::nn::loss::elementwise

Function smooth_l1_loss_forward 

Source 
pub fn smooth_l1_loss_forward<T: Triton, const BLOCK_SIZE: i32>(
    x_ptr: T::Pointer<f32>,
    y_ptr: T::Pointer<f32>,
    out_ptr: T::Pointer<f32>,
    n_elements: i32,
    beta: f32,
)
where
    T::I32Tensor: Tensor<i32, 1> + Comparison<i32, BoolTensor = T::BoolTensor>,
    T::Pointer<f32>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<f32>>>,
Expand description

Element-wise SmoothL1 (Huber variant) forward.

PyTorch convention with beta:

out = 0.5 * diff^2 / beta     if |diff| < beta
    = |diff| - 0.5 * beta     otherwise