Add keepdims argument to argmin and argmax #8710

Activity

1. eric-wieser

   added

   01 - Enhancement

   on Feb 27, 2017
2. eric-wieser commented on Feb 27, 2017

   

   eric-wieser

   on Feb 27, 2017

   MemberAuthor

   More actions

   This might actually be done best by making argmin a gufunc?
3. eric-wieser commented on Feb 28, 2017

   

   eric-wieser

   on Feb 28, 2017

   Last edited by eric-wieser

   MemberAuthor

   More actions

   Trivial implementation:

   def argmin(*args, keepdims=False, **kwargs):
       res = np.argmin(*args, **kwargs)
       axis = kwargs.get('axis', none)
       if axis is not None:
           res = np.expand_dims(res, axis=axis)
       return res
4. shoyer commented on Feb 28, 2017

   

   shoyer

   on Feb 28, 2017

   Member

   More actions

   This might actually be done best by making argmin a gufunc?

   The challenge here is that only ufunc.reduce has the logic for axis handling. That's not something that exists for general ufuncs, though it probably should (#5197), at least for simple cases like these where the ufunc signature looks like (n)->().
5. eric-wieser commented on Dec 1, 2018

   

   eric-wieser

   on Dec 1, 2018

   MemberAuthor

   More actions

   The next problem is that argmin supports axis=None, which ufuns do not. Perhaps argmin should just wrap a ufunc?
6. FabianSchuetze commented on Mar 9, 2019

   

   FabianSchuetze

   on Mar 9, 2019

   More actions

   I would be glad about this features. As I was thinking of trying to contribute to numpy for some time, would somebody mind if I simply tried to implement the feature according to the development workflow?
7. yarki commented on Jul 14, 2019

   

   yarki

   on Jul 14, 2019

   More actions

   Folks, below is the workaround I'm currently using. Please reply if you see a better way to preserve original dimensions. Thanks.

   def argmax_keepdims(x, axis):
       """
       Returns the indices of the maximum values along an axis.
   
       The axis which is reduced is left in the result as dimension with size one.
       The result will broadcast correctly against the input array.
   
       Original numpy.argmax() implementation does not currently support the keepdims parameter.
       See https://github.com/numpy/numpy/issues/8710 for further information.
       """
       output_shape = list(x.shape)
       output_shape[axis] = 1
       return np.argmax(x, axis=axis).reshape(output_shape)

   👍React with 👍4Reacted by David Refaeli, Yuantao Feng, Discreater and Jianyu Li
8. arvoelke commented on Sep 12, 2020

   

   arvoelke

   on Sep 12, 2020

   Contributor

   More actions

   In my case np.unravel_index(np.argmax(x), x.shape) did what I needed (flattens the array, takes the argmax, and then reinterprets the index with respect to the original shape). Your mileage may vary depending on what you are trying to do of course.
9. seberg

   closed this as completedin #19211on Jul 7, 2021