class

HuberLossBackward

HuberLossBackward

source edit

Python wrappers

Autograd node for Huber (smooth L1) loss.

Combines a quadratic regime for small residuals with a linear regime for large residuals, providing robustness to outliers without sacrificing differentiability at zero:

\ell_\delta(\hat{y}, y) = \begin{cases} \tfrac{1}{2} (\hat{y} - y)^2 & |\hat{y} - y| \le \delta \\ \delta \bigl(|\hat{y} - y| - \tfrac{1}{2} \delta\bigr) & |\hat{y} - y| > \delta \end{cases}

The backward emits $\nabla_{\hat y} = \operatorname{clamp}(\hat y - y, -\delta, \delta) \cdot g$ (scaled by $1/N$ for Mean).

Math

\ell_\delta(r) = \begin{cases} \tfrac{1}{2} r^2 & |r| \le \delta \\ \delta |r| - \tfrac{\delta^2}{2} & \text{otherwise} \end{cases}, \quad r = \hat y - y

Attributes

schema_v1OpSchema

"huber_loss", AmpPolicy::ForceFP32.

reduction_Reduction

Reduction mode.

delta_double

Transition threshold between quadratic and linear regimes; must be strictly positive. Default 1.0 (SmoothL1).

orig_shape_Shape

Element-wise shape before reduction.

Notes

Smaller $\delta$ behaves more like an absolute-error loss (robust to outliers); larger $\delta$ behaves more like MSE (sensitive to outliers).

Static methods

static

forward

→TensorImplPtr

int forward(const int & input, const int & target, double delta, Reduction reduction)

source edit

Compute the Huber loss with autograd wiring.

Parameters

inputTensorImplPtr

Predicted tensor.

targetTensorImplPtr

Ground-truth tensor of the same shape as input.

deltadouble

Transition threshold; must be > 0.

reductionReduction

Mean, Sum, or None.

Returns

TensorImplPtr

Reduced or per-element loss tensor.

Raises

:invalid_argument

If delta <= 0.

Methods

apply

int apply(int grad_out)

source edit

Backward pass: gradients w.r.t. input and target.

class

HuberLossBackward

HuberLossBackward

source edit

Python wrappers

Autograd node for Huber (smooth L1) loss.

Combines a quadratic regime for small residuals with a linear regime for large residuals, providing robustness to outliers without sacrificing differentiability at zero:

\ell_\delta(\hat{y}, y) = \begin{cases} \tfrac{1}{2} (\hat{y} - y)^2 & |\hat{y} - y| \le \delta \\ \delta \bigl(|\hat{y} - y| - \tfrac{1}{2} \delta\bigr) & |\hat{y} - y| > \delta \end{cases}

The backward emits $\nabla_{\hat y} = \operatorname{clamp}(\hat y - y, -\delta, \delta) \cdot g$ (scaled by $1/N$ for Mean).

Math

\ell_\delta(r) = \begin{cases} \tfrac{1}{2} r^2 & |r| \le \delta \\ \delta |r| - \tfrac{\delta^2}{2} & \text{otherwise} \end{cases}, \quad r = \hat y - y

Attributes

schema_v1OpSchema

"huber_loss", AmpPolicy::ForceFP32.

reduction_Reduction

Reduction mode.

delta_double

Transition threshold between quadratic and linear regimes; must be strictly positive. Default 1.0 (SmoothL1).

orig_shape_Shape

Element-wise shape before reduction.

Notes

Smaller $\delta$ behaves more like an absolute-error loss (robust to outliers); larger $\delta$ behaves more like MSE (sensitive to outliers).

Static methods

static

forward

→TensorImplPtr

int forward(const int & input, const int & target, double delta, Reduction reduction)

source edit

Compute the Huber loss with autograd wiring.

Parameters

inputTensorImplPtr

Predicted tensor.

targetTensorImplPtr

Ground-truth tensor of the same shape as input.

deltadouble

Transition threshold; must be > 0.

reductionReduction

Mean, Sum, or None.

Returns

TensorImplPtr

Reduced or per-element loss tensor.

Raises

:invalid_argument

If delta <= 0.

Methods

apply

int apply(int grad_out)

source edit

Backward pass: gradients w.r.t. input and target.