pickneglogsoftmax returning [inf]

> e_ = dy.pick(-dy.log(dy.softmax(e1)), k)

There is a chance for softmax to set 0 somewhere in the probability vector which will mess up the log. May be this should be handled internally? Otherwise, I can't think of any way of avoiding [inf] apart from if-condition, or a small additional noise value to avoid 0s. Is there any popular solution to this?

Start by reproducing the expression using dy.pick, dy.log, and dy.softmax, then trace those entry points to the relevant implementation. Determine how zero probabilities produce [inf] and review the existing numerical behavior around softmax and log. Done means the issue has a documented, tested resolution for this expression, if the project agrees it should be handled internally.

Written by the indexing model from the issue text.