class

Relu6Backward

Relu6Backward

source edit

Python wrappers

pylucid.nn.functional.relu6function

Autograd node for ReLU6 $y = \mathrm{clamp}(x, 0, 6)$ (Krizhevsky 2010, MobileNet family).

Gradient is the indicator of the open interval $(0, 6)$ — built directly as an in-range storage mask and multiplied with the upstream gradient, avoiding a backend round-trip for this very simple piecewise rule.

Math

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

Attributes

schema_v1OpSchema

"relu6", AmpPolicy::KeepInput.

Static methods

static

dispatch

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

source edit

Forward — calls IBackend::relu6 to compute $y = \mathrm{clamp}(x, 0, 6)$ .

Methods

grad_formula

int grad_formula(const int & g)

source edit

Eager-mode backward: applies an in-range $(0, 6)$ mask to the upstream gradient.

class

Relu6Backward

Relu6Backward

source edit

Python wrappers

pylucid.nn.functional.relu6function

Autograd node for ReLU6 $y = \mathrm{clamp}(x, 0, 6)$ (Krizhevsky 2010, MobileNet family).

Math

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

Attributes

schema_v1OpSchema

"relu6", AmpPolicy::KeepInput.

Static methods

static

dispatch

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

source edit

Forward — calls IBackend::relu6 to compute $y = \mathrm{clamp}(x, 0, 6)$ .

Methods

grad_formula

int grad_formula(const int & g)

source edit

Eager-mode backward: applies an in-range $(0, 6)$ mask to the upstream gradient.