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 \(input * \beta > threshold\).