In my experience, you'll have some parts of your code in which you have to do a for loop or something else that cannot be (cleanly) expressed using numpy and th

In my experience, you'll have some parts of your code in which you have to do a for loop or something else that cannot be (cleanly) expressed using numpy and that's where performance starts to suffer. The current solution is to use something like cython, but if pypy simplifies that, I think that's great.