[Numpy-tickets] [NumPy] #295: Double sort in numpy.median

[NumPy]

#295: Double sort in numpy.median
-----------------------+----------------------------------------------------
 Reporter:  faltet     |       Owner:  somebody   
     Type:  defect     |      Status:  new        
 Priority:  normal     |   Milestone:  1.0 Release
Component:  numpy.lib  |     Version:  devel      
 Severity:  normal     |    Keywords:             
-----------------------+----------------------------------------------------
 numpy.median is computing a sort twice for arrays with an even number of
 elements. The next patch is a cure for this:

 {{{
 --- ./numpy/lib/function_base.py        2006-09-26 10:05:27.000000000
 +0200
 +++ /usr/local/lib/python2.5/site-packages/numpy/lib/function_base.py
 2006-09-26 18:13:14.000000000 +0200
 @@ -1046,11 +1046,10 @@
      """median(m) returns a median of m along the first dimension of m.
      """
      sorted = msort(m)
 +    index = sorted.shape[0]/2
      if sorted.shape[0] % 2 == 1:
 -        return sorted[int(sorted.shape[0]/2)]
 +        return sorted[index]
      else:
 -        sorted = msort(m)
 -        index = sorted.shape[0]/2
          return (sorted[index-1]+sorted[index])/2.0

 }}}

 Also, and although median shouldn't have to be well defined in other
 values different from numbers, it can be forced to work decently for
 string, object and boolean arrays. Behind is a patch for this.

 {{{
 --- ./numpy/lib/function_base.py        2006-09-26 10:05:27.000000000
 +0200
 +++ /usr/local/lib/python2.5/site-packages/numpy/lib/function_base.py
 2006-09-26 18:13:27.000000000 +0200
 @@ -1046,11 +1046,10 @@
      """median(m) returns a median of m along the first dimension of m.
      """
      sorted = msort(m)
 -    if sorted.shape[0] % 2 == 1:
 -        return sorted[int(sorted.shape[0]/2)]
 +    index = sorted.shape[0]/2
 +    if sorted.shape[0] % 2 == 1 or sorted.dtype.kind in ["S", "O", "b"]:
 +        return sorted[index]
      else:
 -        sorted = msort(m)
 -        index = sorted.shape[0]/2
          return (sorted[index-1]+sorted[index])/2.0

 }}}

 Feel free to apply whatever patch you consider more appropriate.

-- 
Ticket URL: <http://projects.scipy.org/scipy/numpy/ticket/295>
NumPy <http://projects.scipy.org/scipy/numpy>
The fundamental package needed for scientific computing with Python.