class

SeluBackward

SeluBackward

source edit

Python wrappers

pylucid.nn.functional.selufunction

Autograd node for SELU (Scaled ELU, Klambauer et al. 2017).

Uses the self-normalising constants $\alpha \approx 1.6732632$ and $\lambda \approx 1.0507010$ baked into the backend; the Python and C++ layers do not expose them to the caller. The backward delegates to the backend because the piecewise scaled-ELU formula depends on those fixed constants.

Math

\alpha\,(e^x - 1), & x < 0 \end{cases}$$ **References** Klambauer et al., "Self-Normalizing Neural Networks", [arXiv:1706.02515](https://arxiv.org/abs/1706.02515).

Attributes

schema_v1OpSchema

"selu", AmpPolicy::ForceFP32.

Static methods

static

dispatch

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

source edit

Forward — calls IBackend::selu to compute the Scaled ELU $y = \lambda\,x$ for $x \ge 0$ else $\lambda\,\alpha\,(e^x - 1)$ .

Methods

grad_formula

int grad_formula(const int & g)

source edit

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

class

SeluBackward

SeluBackward

source edit

Python wrappers

pylucid.nn.functional.selufunction

Autograd node for SELU (Scaled ELU, Klambauer et al. 2017).

Math

\alpha\,(e^x - 1), & x < 0 \end{cases}$$ **References** Klambauer et al., "Self-Normalizing Neural Networks", [arXiv:1706.02515](https://arxiv.org/abs/1706.02515).

Attributes

schema_v1OpSchema

"selu", AmpPolicy::ForceFP32.

Static methods

static

dispatch

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

source edit

Forward — calls IBackend::selu to compute the Scaled ELU $y = \lambda\,x$ for $x \ge 0$ else $\lambda\,\alpha\,(e^x - 1)$ .

Methods

grad_formula

int grad_formula(const int & g)

source edit

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