How can I efficiently sort an array of objects on two or more attributes in Numpy?
class Obj():
def __init__(self,a,b):
self.a = a
self.b = b
arr = np.array([],dtype=Obj)
for i in range(10):
arr = np.append(arr,Obj(i, 10-i))
arr_sort = np.sort(arr, order=a,b) ???
Thx, Willem-Jan
The order parameter only applies to structured arrays:
In [383]: arr=np.zeros((10,),dtype='i,i')
In [385]: for i in range(10):
...: arr[i] = (i,10-i)
In [386]: arr
Out[386]:
array([(0, 10), (1, 9), (2, 8), (3, 7), (4, 6), (5, 5), (6, 4), (7, 3), (8, 2), (9, 1)],
dtype=[('f0', '<i4'), ('f1', '<i4')])
In [387]: np.sort(arr, order=['f0','f1'])
Out[387]:
array([(0, 10), (1, 9), (2, 8), (3, 7), (4, 6), (5, 5), (6, 4), (7, 3), (8, 2), (9, 1)],
dtype=[('f0', '<i4'), ('f1', '<i4')])
In [388]: np.sort(arr, order=['f1','f0'])
Out[388]:
array([(9, 1), (8, 2), (7, 3), (6, 4), (5, 5), (4, 6), (3, 7), (2, 8),
(1, 9), (0, 10)],
dtype=[('f0', '<i4'), ('f1', '<i4')])
With a 2d array, lexsort provides a similar 'ordered' sort
In [402]: arr=np.column_stack((np.arange(10),10-np.arange(10)))
In [403]: np.lexsort((arr[:,1],arr[:,0]))
Out[403]: array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], dtype=int32)
In [404]: np.lexsort((arr[:,0],arr[:,1]))
Out[404]: array([9, 8, 7, 6, 5, 4, 3, 2, 1, 0], dtype=int32)
With your object array, I could extract the attributes into either of these structures:
In [407]: np.array([(a.a, a.b) for a in arr])
Out[407]:
array([[ 0, 10],
[ 1, 9],
[ 2, 8],
....
[ 7, 3],
[ 8, 2],
[ 9, 1]])
In [408]: np.array([(a.a, a.b) for a in arr],dtype='i,i')
Out[408]:
array([(0, 10), (1, 9), (2, 8), (3, 7), (4, 6), (5, 5), (6, 4), (7, 3),
(8, 2), (9, 1)],
dtype=[('f0', '<i4'), ('f1', '<i4')])
The Python sorted function will work on arr (or its list equivalent)
In [421]: arr
Out[421]:
array([<__main__.Obj object at 0xb0f2d24c>,
<__main__.Obj object at 0xb0f2dc0c>,
....
<__main__.Obj object at 0xb0f35ecc>], dtype=object)
In [422]: sorted(arr, key=lambda a: (a.b,a.a))
Out[422]:
[<__main__.Obj at 0xb0f35ecc>,
<__main__.Obj at 0xb0f3570c>,
...
<__main__.Obj at 0xb0f2dc0c>,
<__main__.Obj at 0xb0f2d24c>]
Your Obj class is missing a nice __str__ method. I have to use something like [(i.a, i.b) for i in arr] to see the values of the arr elements.
As I stated in the comment, for this example, a list is much nice than an object array.
In [423]: alist=[]
In [424]: for i in range(10):
...: alist.append(Obj(i,10-i))
list append is faster than the repeated array append. And object arrays don't add much functionality compared to a list, especially when 1d, and the objects are custom classes like this. You can't do any math on arr, and as you can see, sorting isn't any easier.
np.array([],dtype=Obj)?dtype=Objis treated asdtype=object, the genericobjectdtype. Elements of such an array can be anything, includingNone.