move_median off by 1?

[brentp]

Hi, with this code:


import numpy as np
import bottleneck as bn

a = np.arange(10)
print "\n".join(map(str, zip(a, bn.move_median(a, 3))))

I get this output:
(0, nan)
(1, nan)
(2, 1.0)
(3, 2.0)
(4, 3.0)
(5, 4.0)
(6, 5.0)
(7, 6.0)
(8, 7.0)
(9, 8.0)

It seems to me that the RHS should be [nan, 1 .. 8, nan]
instead of [nan, nan, 1..8]

am I missing something?
thanks,
-Brent

[Brent Pedersen]

nevermind, the current behavior makes complete sense.
(as usual, it takes posting to a list for me to understand my mistake).

[Keith Goodman]

On Wed, Jun 29, 2011 at 1:32 PM, brentp <bped...@gmail.com> wrote:
> Hi, with this code:
>
>
> import numpy as np
> import bottleneck as bn
>
> a = np.arange(10)
> print "\n".join(map(str, zip(a, bn.move_median(a, 3))))
>
> I get this output:
> (0, nan)
> (1, nan)
> (2, 1.0)
> (3, 2.0)
> (4, 3.0)
> (5, 4.0)
> (6, 5.0)
> (7, 6.0)
> (8, 7.0)
> (9, 8.0)
>
> It seems to me that the RHS should be [nan, 1 .. 8, nan]
> instead of [nan, nan, 1..8]

Looks like you'd like a centered moving median. The moving median in
bottleneck, however, puts the moving median in the position of the
last element in the window:

>> bn.move_median([0,1,2], 3)
array([ nan, nan, 1.])

BTW, how would a centered moving median work for an even window?

[Keith Goodman]

I do that all the time. It works even better, BTW, when you post to a
large list like NumPy.

[Brent Pedersen]

yes, i was looking for centered, came up with this:

def centered_median(a, window):
"""
centered moving median
"""
assert window % 2 != 0
w = (window - 1) / 2
t = bn.move_median(a, window)
t = np.roll(t, -w)
t[:w] = a[:w]
t[-w:] = a[-w:]
return t