[Numpy-discussion] Proposal of new function: iteraxis()

[Andrew Giessel]

I like this, thank you Phil.

>From what I can see, the ordering of the returned slices when you use more
than one axis (ie: slices(a, [1,2]), increments the last axis fastest.
 Does this makes sense based on the default ordering of, say, nditer()?  I
know that C-order (row major) and Fortran order (column major) are two ways
of ordering the returned values- which does this default to? Is there a
default across numpy?

best,



On Fri, Apr 26, 2013 at 9:56 AM, Phil Elson <pelson.pub at gmail.com> wrote:

> I didn't find the rollaxis solution particularly obvious and also had to
> think about what rollaxis did before understanding its usefulness for
> iteration.
> Now that I've understood it, I'm +1 for the statement that, as it stands,
> the proposed iteraxis method doesn't add enough to warrant its inclusion.
>
> That said, I do think array iteration could be made simpler (or the
> function I've missed better documented!). I've put together an
> implementation of a "slices" function which can return subsets of an array
> based on the axes provided (a generalisation of iteraxis but implemented
> slightly differently):
>
> def slices(a, axes=-1):
>     indices = np.repeat(slice(None), a.ndim)
>     # turn axes into a 1d array of axes indices
>     axes = np.array(axes).flatten()
>
>     bad_indices = (axes < (-a.ndim + 1)) | axes > (a.ndim - 1)
>     if np.any(bad_indices):
>         raise ValueError('The axis index/indices were out of range.')
>
>     # Turn negative indices into real indices
>     axes[axes < 0] = a.ndim + axes[axes < 0]
>
>     if np.unique(axes).shape != axes.shape:
>         raise ValueError('Repeated axis indices were given.')
>
>     indexing_shape = np.array(a.shape)[axes]
>
>     for ind in np.ndindex(*indexing_shape):
>         indices[axes] = ind
>         yield a[tuple(indices)]
>
>
> This can be used simply with:
>
> >>> a = np.ones([2, 3, 4, 5])
> >>> for s in slices(a, 2):
> ...  print s.shape
> ...
> (2, 3, 5)
> (2, 3, 5)
> (2, 3, 5)
> (2, 3, 5)
>
>
> Or slightly with the slightly more complex:
>
> >>> len(list(slices(a, [2, -1])))
> 20
>
> Without focusing on my actual implementation, would this kind of interface
> be more desirable?
>
> Cheers,
>
>
>
>
> On 26 April 2013 12:33, Robert Kern <robert.kern at gmail.com> wrote:
>
>> On Fri, Apr 26, 2013 at 12:26 PM, Andrew Giessel
>> <andrew_giessel at hms.harvard.edu> wrote:
>> > I agree with Charles that rollaxis() isn't immediately intuitive.
>> >
>> > It seems to me that documentation like this doesn't belong in
>> rollaxis() but
>> > instead wherever people talk about indexing and/or iterating over an
>> array.
>> > Nothing about the iteration depends on rollaxis(),  rollaxis is just
>> giving
>> > you a different view of the array to call __getitem__() on, if I
>> understand
>> > correctly.
>>
>> Docstrings are perfect places to briefly describe and demonstrate
>> common use cases for a function. There is no problem with including
>> the example that I wrote in the rollaxis() docstring.
>>
>> In any case, whether you put the documentation in the rollaxis()
>> docstring or in one of the indexing/iteration sections, or
>> (preferably) both, I strongly encourage you to do that first and see
>> how it goes before adding a new alias.
>>
>> --
>> Robert Kern
>> _______________________________________________
>> NumPy-Discussion mailing list
>> NumPy-Discussion at scipy.org
>> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>>
>
>
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion at scipy.org
> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>
>


-- 
Andrew Giessel, PhD

Department of Neurobiology, Harvard Medical School
220 Longwood Ave Boston, MA 02115
ph: 617.432.7971 email: andrew_giessel at hms.harvard.edu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20130426/652b38b7/attachment.html>