We should still keep the RNG as standalone functions, but then have torch-distributions call them. We want to avoid what Boost does, i.e. having to instantiate a class every time you want to compute a PDF or a CDF or sample.
Note that a proper torch-distributions package is out of the scope of a single week, of course :-) we can start it, and then build upon it as time goes. Let's discuss it when you're back!
jucor
commented
11 years ago
That makes it more easily extensible, and is in keeping with scipy. A structure like https://github.com/scipy/scipy/blob/master/scipy/stats/distributions.py might be our best guess for it -- very much in line with what you suggested at some point, @d11 .
We should still keep the RNG as standalone functions, but then have torch-distributions call them. We want to avoid what Boost does, i.e. having to instantiate a class every time you want to compute a PDF or a CDF or sample.
Note that a proper torch-distributions package is out of the scope of a single week, of course :-) we can start it, and then build upon it as time goes. Let's discuss it when you're back!