Numpy: efficient way to generate combinations from given ranges

I think what you’re looking for is np.mgrid. Unfortunately, this returns the array in a format that’s different from what you need, so you’ll need to do a little post-processing:

a = np.mgrid[0:4, 0:4, 0:11]     # All points in a 3D grid within the given ranges
a = np.rollaxis(a, 0, 4)         # Make the 0th axis into the last axis
a = a.reshape((4 * 4 * 11, 3))   # Now you can safely reshape while preserving order

Explanation

np.mgrid gives you a set of grid points in N-dimensional space. Let me try to show this with a smaller example, to make things clearer:

>>> a = np.mgrid[0:2, 0:2]
>>> a
array([[[0, 0],
        [1, 1]],

       [[0, 1],
        [0, 1]]])

Since I’ve given two sets of ranges, 0:2, 0:2, I get a 2D grid. What mgrid returns is the x-values and the y-values corresponding to the grid points (0, 0), (0, 1), (1, 0) and (1, 1) in 2D space. a[0] tells you what the x-values of the four points are, and a[1] tells you what the y-values are.

But what you really want is that list of actual grid points that I’ve written out, not the x- and y-values of those points separately. First instinct is to just reshape the array as desired:

>>> a.reshape((4, 2))
array([[0, 0],
       [1, 1],
       [0, 1],
       [0, 1]])

But clearly this doesn’t work, because it effectively reshapes the flattened array (the array obtained by just reading all elements in order), and that’s not what you want.

What you want to do is to look down the third dimension of a, and create an array:

[ [a[0][0, 0], a[1][0, 0]],
  [a[0][0, 1], a[1][0, 1]],
  [a[0][1, 0], a[1][1, 0]],
  [a[0][1, 1], a[1][1, 1]] ]

which reads “First tell me the first point (x1, y1), then the second point (x2, y2), …” and so on. Perhaps this is better explained with a figure, of sorts. This is what a looks like:

                you want to read
                in this direction
                 (0, 0)   (0, 1)
                   |        |
                   |        |
                   v        v

          /        0--------0            +----> axis0
 x-values "https://stackoverflow.com/"       /"https://stackoverflow.com/"
          "https://stackoverflow.com/"      / |    axis1 / |
          \     1--------1  |         L  |
                |  |     |  |            v
          /     |  0-----|--1           axis2
 y-values |     "https://stackoverflow.com/" /
          |     "https://stackoverflow.com/"/
          \     0--------1

                |        |
                |        |
                v        v
              (1, 0)   (1, 1)

np.rollaxis gives you a way to do this. np.rollaxis(a, 0, 3) in the above example says “take the 0th (or outermost) axis and make it into the last (or innermost) axis. (Note: only axes 0, 1 and 2 actually exist here. So saying “send the 0th axis to the 3rd position” is a way of telling python to put the 0th axis after the last axis). You might also want to read this.

>>> a = np.rollaxis(a, 0, 3)
>>> a
array([[[0, 0],
        [0, 1]],

       [[1, 0],
        [1, 1]]])

This is starting to look like what you want, except there’s an extra array dimension. We want to merge dimensions 0 and 1 to get just get a single array of grid points. But now that the flattened array reads in the manner that you expect, you can safely reshape it to give you the desired result.

>>> a = a.reshape((4, 2))
>>> a
array([[0, 0],
       [0, 1],
       [1, 0],
       [1, 1]])

The 3D version does just the same thing, except, I couldn’t make a figure for that, since it’d be in 4D.