Description
🚀 Feature
scikit's implementation of logsumexp accepts a supplemental array of scaling factors that allow for subtraction instead of just addition.
https://docs.scipy.org/doc/scipy-0.19.0/reference/generated/scipy.misc.logsumexp.html
Motivation
I'm doing work with log probabilities and need to subtract some likelihoods from each other. Currently I have to move to linear probabilities first.
Pitch
If logsumexp is operating over a dimension of size N, accept an additional tensor of size (N,) of scaling factors, to match the functionality in scikit
https://docs.scipy.org/doc/scipy-0.19.0/reference/generated/scipy.misc.logsumexp.html
Alternatives
subtraction of probabilities in linear space, but this has numerical stability issues obviously.