[Numpy-discussion] Speed up function on cross product of two sets?

[Arnd Baecker]

Hi,

On Sun, 2 Apr 2006, Zachary Pincus wrote:

> Hi folks,
>
> I have a inner loop that looks like this:
> out = []
> for elem1 in l1:
>    for elem2 in l2:
>      out.append(do_something(l1, l2))
> result = do_something_else(out)
>
> where do_something and do_something_else are implemented with only
> numpy ufuncs, and l1 and l2 are numpy arrays.
>
> As an example, I need to compute the median distance from any element
> in one set to any element in another set.
>
> What's the best way to speed this sort of thing up with numpy (e.g.
> push as much down into the underlying C as possible)? I could re-
> write do_something with the numexpr tools (which are very cool), but
> that doesn't address the fact that I've still got nested loops living
> in Python.

If do_something eats arrays, you could try:

  result = do_something(l1[:,NewAxis], l2)

E.g.:

  from numpy import *
  l1 = linspace(0.0, pi, 10)
  l2 = linspace(0.0, pi, 3)
  def f(y, x):
      return sin(y)*cos(x)

  print f(l1[:,NewAxis], l2)


((Note that I just learned in some other thread that with numpy there is
an alternative to NewAxis, but I haven't figured out which that is ...))

Best, Arnd