af

README

af

Activation functions for neural networks.

These activation functions are included:

• Swish (x / (1 + exp(-x)))
• Sigmoid (1 / (1 + exp(-x)))
• SoftPlus (log(1 + exp(x)))
• Gaussian01 (exp(-(x * x) / 2.0))
• Sin (math.Sin(math.Pi * x))
• Cos (math.Cos(math.Pi * x))
• Linear (x)
• Inv (-x)
• ReLU (x >= 0 ? x : 0)
• Squared (x * x)

These math functions are included just for convenience:

• Abs (math.Abs)
• Tanh (math.Tanh)

One functions that takes two arguments is also included:

• PReLU (x >= 0 ? x : x * a)

Requirements

• Go 1.11 or later.

General information

• License: MIT
• Version: 0.3.2
• Author: Alexander F. Rødseth <xyproto@archlinux.org>

Documentation

Overview

Package af provides several activation functions that can be used in neural networks

Index

• Variables
• func PReLU(x, a float64) float64
• func ReLU(x float64) float64
• func Step(x float64) float64

Variables

var (
	Sigmoid    = swish.Sigmoid
	Swish      = swish.Swish
	SoftPlus   = swish.SoftPlus
	Gaussian01 = swish.Gaussian01
	Linear     = func(x float64) float64 { return x }
	Inv        = func(x float64) float64 { return -x }
	Sin        = func(x float64) float64 { return math.Sin(math.Pi * x) }
	Cos        = func(x float64) float64 { return math.Cos(math.Pi * x) }
	Squared    = func(x float64) float64 { return x * x }
	Tanh       = math.Tanh
	Abs        = math.Abs
)

The swish package offers optimized Swish, Sigmoid SoftPlus and Gaussian01 activation functions

Functions

func PReLU

func PReLU(x, a float64) float64

PReLU is the parametric rectified linear unit. `x >= 0 ? x : a * x`

func ReLU

func ReLU(x float64) float64

ReLU is the "rectified linear unit" `x >= 0 ? x : 0`

func Step

func Step(x float64) float64

Step function