[Numpy-discussion] use index array of len n to select columns of n x m array

[Keith Goodman]

On Fri, Aug 6, 2010 at 3:01 AM, Martin Spacek <numpy at mspacek.mm.st> wrote:
> Keith Goodman wrote:
>  > Here's one way:
>  >
>  >>> a.flat[i + a.shape[1] * np.arange(a.shape[0])]
>  >     array([0, 3, 5, 6, 9])
>
>
> I'm afraid I made my example a little too simple. In retrospect, what I really
> want is to be able to use a 2D index array "i", like this:
>
>  >>> a = np.array([[ 0,  1,  2,  3],
>                   [ 4,  5,  6,  7],
>                   [ 8,  9, 10, 11],
>                   [12, 13, 14, 15],
>                   [16, 17, 18, 19]])
>  >>> i = np.array([[2, 1],
>                   [3, 1],
>                   [1, 1],
>                   [0, 0],
>                   [3, 1]])
>  >>> foo(a, i)
> array([[ 2,  1],
>        [ 7,  5],
>        [ 9,  9],
>        [12, 12],
>        [19, 17]])
>
> I think the flat iterator indexing suggestion is about the only thing that'll
> work. Here's the function I've pretty much settled on:
>
> def rowtake(a, i):
>     """For each row in a, return values according to column indices in the
>     corresponding row in i. Returned shape == i.shape"""
>     assert a.ndim == 2
>     assert i.ndim <= 2
>     if i.ndim == 1:
>         return a.flat[i + a.shape[1] * np.arange(a.shape[0])]
>     else: # i.ndim == 2
>         return a.flat[i + a.shape[1] * np.vstack(np.arange(a.shape[0]))]
>
> This is about half as fast as my Cython function, but the Cython function is
> limited to fixed dtypes and ndim:

You can speed it up by getting rid of two copies:

idx = np.arange(a.shape[0])
idx *= a.shape[1]
idx += i