Skip to contents

Applies element-wise, the function \(Softplus(x) = 1/\beta * log(1 + exp(\beta * x))\).

Usage

nnf_softplus(input, beta = 1, threshold = 20)

Arguments

input

(N,*) tensor, where * means, any number of additional dimensions

beta

the beta value for the Softplus formulation. Default: 1

threshold

values above this revert to a linear function. Default: 20

Details

For numerical stability the implementation reverts to the linear function when \(input * \beta > threshold\).