Numpy indexing reorders array

Question

I have just stumbled over a numpy index behavior which I do not quite understand. It seems like numpy is changing the order of my axes depending on the indexing schema. Unfortunately, I cannot find an explanation for the following in the documentation. Could somebody explain to me what is going on?

# This is expected: dimension 1 is reduced to length 1:
print np.ndarray(shape=(3,3,3,3))[:, [0], :, :].shape
>>> (3, 1, 3, 3)

# This is the unexpected behavior:
print np.ndarray(shape=(3,3,3,3))[:, [0], :, 0].shape
>>> (1, 3, 3)

I would expected the second command to yield (3, 1, 3). Why does does my shape of the first two dimensions change if I pick an element from the forth? Thanks a lot in advance!

Edit: I see this on numpy 1.11.0 with python 2.7.11

Though obscure, this behavior is documented.

user2379410
– user2379410

2016-06-21 09:37:26 +00:00
Commented Jun 21, 2016 at 9:37 — user2379410
– user2379410, Commented Jun 21, 2016 at 9:37

Jaime · Accepted Answer · 2016-06-21 09:35:15Z

It is expected behavior... NumPy treats axes indexed with slices and axes indexed with arrays differently. The default behavior is that the resulting shape will have the resulting shape of axes indexed with arrays before the resulting shape of axes indexed with slices. Which means that the odd one out is the first case, that should have been (1, 3, 3, 3). But someone thought that it would confuse users for some simple cases, like yours, so when all the axes indexed with arrays are grouped together, the resulting shape is placed in the position of the output shape where the cluster of array indexed axes was in the original array.

You can test it by yourself:

>>> print np.ndarray(shape=(3,3,3,3))[:, [0], [0], :].shape
(3, 1, 3)

As you have experienced, this cleverness is probably more confusing than a consistent behavior. The consensus among the NumPy developers is that it would be better to not treat clustered array indexed axes differently, and there is even a PR implementing indexer attributes that would allow this saner indexing.

For backwards compatibility the default indexing is unlikely to change any time soon, although it may start issuing deprecation warnings for some cases in the not so distant future.

Collectives™ on Stack Overflow

Numpy indexing reorders array

1 Answer 1

Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related