A Safely Imitating a Neural Policy