nn_softplus {torch}

Softplus module

Description

Applies the element-wise function:

\mbox{Softplus}(x) = \frac{1}{\beta} * \log(1 + \exp(\beta * x))

Usage

nn_softplus(beta = 1, threshold = 20)

Arguments

beta

the \beta value for the Softplus formulation. Default: 1

threshold

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

Details

SoftPlus is a smooth approximation to the ReLU function and can be used to constrain the output of a machine to always be positive. For numerical stability the implementation reverts to the linear function when input \times \beta > threshold.

Shape

• Input: (N, *) where * means, any number of additional dimensions

• Output: (N, *), same shape as the input

Examples

if (torch_is_installed()) {
m <- nn_softplus()
input <- torch_randn(2)
output <- m(input)
}