One of the hardest Pytorch bug I had to debug is due to how the logsumexp behave with -inf masked inputs. Consider the following example. I build a vector of 3 logits, and each logit is the result of a logsumexp.
add a skeleton here at some point
5 days ago