Rearranging rows and columns of a numpy array (in a single operation)

Question

First let's put that out of the way: I'm already aware of the solution discussed here: Create NumPy array from another array by specifying rows and columns

Let's say I have an array

test = np.arange(12).reshape((3,4))

array([[ 0,  1,  2,  3],
       [ 4,  5,  6,  7],
       [ 8,  9, 10, 11]])

and I want to rearrange its rows and columns into test2

array([[ 1,  0,  3,  2],
       [ 9,  8, 11, 10],
       [ 5,  4,  7,  6]])

The solution I'm using right now is the same as the one already answered in the link above:

test2 = test[[0,2,1]][:,[1,0,3,2]]

However this solution creates an unnecessary intermediate copy of the data: it first creates a copy with shuffled rows only test[[0,2,1]], then creates a second copy with shuffled columns as well. For such a small array, no one cares but if the array is gigantic this solution feels sub-optimal. Does numpy allow rows and columns shuffling simultaneously? Something looking like test[[0,2,1],[1,0,3,2]], except numpy doesn't interpret this operation that way...

score 5 · Accepted Answer · answered Sep 04 '17 at 23:35

5

Yes, that's the purpose of np.ix_:

>>> test[np.ix_([0,2,1], [1,0,3,2])]
array([[ 1,  0,  3,  2],
       [ 9,  8, 11, 10],
       [ 5,  4,  7,  6]])

It will still create an intermediate index array from the two input arrays but it won't index the test array twice.

Note that np.ix_ just adds dimensions to your index arrays, these will be "broadcasted" when you index the array:

>>> np.ix_([0,2,1], [1,0,3,2])
(array([[0],
        [2],
        [1]]), array([[1, 0, 3, 2]]))

answered Sep 04 '17 at 23:35

MSeifert

145,886
38
333
352

This solution seems twice slower than my current one... – Julien Sep 04 '17 at 23:39
Actually it's (only) marginally faster when tested on bigger arrays. – Julien Sep 04 '17 at 23:41
Yes, the break even seems to be high. However with 1kk items it's already significantly faster (2 times on my computer). – MSeifert Sep 04 '17 at 23:43
Well I was hoping for better, but apparently that's as good as numpy allows... – Julien Sep 05 '17 at 05:54
It made a **huuuuge** difference reordering a `70k x 70k` boolean array `a`: **209.20s** and **35.14s** for `a[idxs][:, idxs]` and `a[np.ix_(idxs, idxs)` respectively – Saravanabalagi Ramachandran Aug 19 '19 at 11:49

Rearranging rows and columns of a numpy array (in a single operation)

1 Answers1