DOC: argmin, argmax, argort, etc. should link to np.take_along_axis in their documentation #6078

Activity

1. shoyer commented on Jul 17, 2015

   

   shoyer

   on Jul 17, 2015

   Member

   More actions

   Indeed, this problem has also flummoxed me.

   Here's the helper function I wrote, which may be slightly more interpretable:

   def argsort_indices(a, axis=-1):
       """Like argsort, but returns an index suitable for sorting the
       the original array even if that array is multidimensional
       """
       a = np.asarray(a)
       ind = list(np.ix_(*[np.arange(d) for d in a.shape]))
       ind[axis] = a.argsort(axis)
       return tuple(ind)
2. mhvk commented on Jul 17, 2015

   

   mhvk

   on Jul 17, 2015

   ContributorAuthor

   More actions

   @shoyer - yours is definitely easier to read, in that it leaves the proper shaping to np.ix_... also, it doesn't deal with keepdims, right?
3. shoyer commented on Jul 17, 2015

   

   shoyer

   on Jul 17, 2015

   Member

   More actions

   also, it doesn't deal with keepdims, right?

   I guess it's less general in that it only works for argsort -- I didn't think about min/max, so I didn't need anything for keepdims.
4. ElieGouzien commented on May 11, 2017

   

   ElieGouzien

   on May 11, 2017

   Contributor

   More actions

   I think it would be lot better than today's implementation if the axis=None option returned something equivalent to :

   np.unravel_index(np.argsort(a, axis=None), a.shape)

   👍React with 👍1Reacted by igurin-invn
5. shoyer commented on May 12, 2017

   

   shoyer

   on May 12, 2017

   Member

   More actions

   I think it would be lot better than today's implementation if the axis=None option returned something equivalent to :
   np.unravel_index(np.argsort(a, axis=None), a.shape)

   As a rule, NumPy doesn't change existing behavior. So this will need to be new API of some sort.
6. ElieGouzien commented on Jun 29, 2017

   

   ElieGouzien

   on Jun 29, 2017

   Contributor

   More actions

   @shoyer Oh that's true.
   I've just submitted a pull request to add my proposition to the doc so that the user can easily find how to retrieve indices.
7. seberg commented on May 3, 2019

   

   seberg

   on May 3, 2019

   Member

   More actions

   I think this issue is now irrelevant, since we have np.take_along_axis instead, which solves the issue. I will however transform it into a Documentation issue, since those functions should list it under the See Also section at least.
8. seberg

   added

   04 - Documentation

   difficulty: Easy

   on May 3, 2019
9. seberg

   changed the title [-]Optional output of full fancy index for argmin, argmax, argort?[/-] [+]DOC: argmin, argmax, argort, etc. should link to `np.take_along_axis` in their documentation[/+] on May 3, 2019
10. mattip

    closed this as completedin #14799on Nov 4, 2019