class

HardSigmoidBackward

HardSigmoidBackward

source edit

Autograd node for the Hard Sigmoid $y = \mathrm{clamp}((x + 3)/6, 0, 1)$ (Howard et al., MobileNetV3).

The derivative is the indicator of the active linear region $-3 < x < 3$ , scaled by $1/6$ . Backward goes through the backend to share the piecewise kernel with the other hard-* activations.

Math

\qquad \frac{\partial y}{\partial x} = \begin{cases} 1/6, & -3 < x < 3 \\ 0, & \text{otherwise} \end{cases}$$

Attributes

schema_v1OpSchema

"hard_sigmoid", AmpPolicy::KeepInput.

Static methods

static

dispatch

int dispatch(backend::IBackend & be, const int & a, const int & s, Dtype dt)

source edit

Forward — calls IBackend::hard_sigmoid to compute $y = \mathrm{clamp}((x + 3)/6, 0, 1)$ .

Methods

grad_formula

int grad_formula(const int & g)

source edit

Eager-mode backward: dispatches to IBackend::hard_sigmoid_backward.

class

HardSigmoidBackward

HardSigmoidBackward

source edit

Autograd node for the Hard Sigmoid $y = \mathrm{clamp}((x + 3)/6, 0, 1)$ (Howard et al., MobileNetV3).

The derivative is the indicator of the active linear region $-3 < x < 3$ , scaled by $1/6$ . Backward goes through the backend to share the piecewise kernel with the other hard-* activations.

Math

\qquad \frac{\partial y}{\partial x} = \begin{cases} 1/6, & -3 < x < 3 \\ 0, & \text{otherwise} \end{cases}$$

Attributes

schema_v1OpSchema

"hard_sigmoid", AmpPolicy::KeepInput.

Static methods

static

dispatch

int dispatch(backend::IBackend & be, const int & a, const int & s, Dtype dt)

source edit

Forward — calls IBackend::hard_sigmoid to compute $y = \mathrm{clamp}((x + 3)/6, 0, 1)$ .

Methods

grad_formula

int grad_formula(const int & g)

source edit

Eager-mode backward: dispatches to IBackend::hard_sigmoid_backward.