Closed oooo111 closed 1 year ago
Hi, you can add a very small value (e.g. 0.00001) to the probabilities before calculating logP (i.e. _studentlogp and _teacherlogp) to avoid the occurrence of log0 and to accordingly solve the Inf issue. We have updated our code. Please check it.
About May I ask how to solve a large number of Inf problems in both student_logp and teacher_logp during model training?