safeSoftMax {sdetorus}

Safe softmax function for computing weights

Description

Computes the weights w_i = \frac{e^{p_i}}{\sum_{j=1}^k e^{p_j}} from p_i, i=1,\ldots,k in a safe way to avoid overflows and to truncate automatically to zero low values of w_i.

Usage

safeSoftMax(logs, expTrc = 30)

Arguments

logs

matrix of logarithms where each row contains a set of p_1,\ldots,p_k to compute the weights from.

expTrc

truncation for exponential: exp(x) with x <= -expTrc is set to zero. Defaults to 30.

Details

The logs argument must be always a matrix.

Value

A matrix of the size as logs containing the weights for each row.

Examples

# A matrix
safeSoftMax(rbind(1:10, 20:11))
rbind(exp(1:10) / sum(exp(1:10)), exp(20:11) / sum(exp(20:11)))

# A row-matrix
safeSoftMax(rbind(-100:100), expTrc = 30)
exp(-100:100) / sum(exp(-100:100))