Sorting should work for array scalars.

[charris]

Just return the array, don't error.

In [11]: a.ndim
Out[11]: 0

In [12]: a.sort()

ValueError                                Traceback (most recent call last)
<ipython-input-12-e7eb8b51a6fa> in <module>()
----> 1 a.sort()

ValueError: axis(=-1) out of bounds

[tushar-rishav]

@charris Checking a.ndim for zero before calling a.sort here avoids the ValueError while computing np.sort(a) for array scalars. I've this tried this and it worked

if not a.ndim:
    a = asanyarray([a])
a.sort(axis, kind, order)
return a

Probably, doing just that won't help because we'll still get ValueError in case of a.sort(). I noticed that behind the curtain a.sort() calls PyArray_Sort which takes PyArrayObject instance , axis and order . Unfortunately, I couldn't get much insight about how ndarray.sort() has been implemented here.

#define PyArray_Sort  (*(int (*)(PyArrayObject *, int, NPY_SORTKIND))  PyArray_API[129])

As I am a beginner I might be missing something here. Could you please let me know where PyArrayObject has been declared?
Thanks in advance!

[charris]

The base sort is an array method, so that is what we want fixed.

[seberg]

I agree that it is silly that this exists and does not work, but on the other hand the default for the sort is axis=-1, which is correctly an error for 0-d arrays unlike most similar methods which use axis=None (unsupported currently in sort) where this is no problem.
I would be +1 to allow arr.sort(None), which would also make sense for 0-d. But I don't feel just removing the error is more canonical, so I am slightly against it unless there is at least some use case.

[charris]

I ran into this because of an argument that could be either a list or a scalar and I wanted to preserve that distinction while also sorting the result of asarray. So in that case it would avoid special casing.

[jaimefrio]

Do we want to fix this for the Python API only, or for PyArray_Sort as well? Should array scalars suceed regardless of the value of axis, or should they only do for axis=-1?

On a related note, is this behavior correct, or should the return also be an array scalar?

>>> np.argsort(np.array(8))
array([0])

[njsmith]

I'm not sure that this does make sense... sorting needs a dimension to
operate on, doesn't it? You could flatten, but sort is inplace so that
doesn't make much sense either...
On Jan 19, 2016 4:34 PM, "Jaime" notifications@github.com wrote:

Do we want to fix this for the Python API only, or for PyArray_Sort as
well? Should array scalars suceed regardless of the value of axis, or
should they only do for axis=-1?

On a related note, is this behavior correct, or should the return also be
an array scalar?

np.argsort(np.array(8))
array([0])

—
Reply to this email directly or view it on GitHub
#7051 (comment).

[charris]

@njsmith Array scalars can be operated on inplace. I guess I'm looking at 0-D arrays as containers rather than scalars even if they can't be indexed with integers.

[charris]

Mind, i'd just like the operation to be a NOP rather than an error.

[njsmith]

I'm just saying that sorting is an intrinsically one-dimensional operation
(it feels like to me). So I can see an argument for zerod_array.sort()
being equivalent to zerod_array.resize((1,)) (with the sort implicitly
doing an in-place flatten to get something one-dimensional and sortable,
and then sorting it, which would be a no-op because there's only one
element). But zerod_array.sort() doesn't make sense to me as a no-op that
leaves you with a zero-d output.
On Jan 19, 2016 5:07 PM, "Charles Harris" notifications@github.com wrote:

Mind, i'd just like the operation to be a NOP rather than an error.

—
Reply to this email directly or view it on GitHub
#7051 (comment).