[Numpy-discussion] bug ! arr.mean() outside arr.min() .. arr.max() range

[Charles R Harris]

On 8/11/06, Sebastian Haase <haase at msg.ucsf.edu> wrote:
>
> Travis Oliphant wrote:
> > Sebastian Haase wrote:
> >> Hi!
> >> b is a non-native byteorder array of type int16
> >> but see further down: same after converting to native ...
> >>
> >>>>> repr(b.dtype)
> >>>>>
> >> 'dtype('>i2')'
> >>
> >
> > The problem is no-doubt related to "wrapping" for integers.  Your total
> is
> > getting too large to fit into the reducing data-type.
> >
> > What does
> >
> > d.sum() give you?
> I can't check that particular array until Monday...
>
> >
> > You can add d.mean(dtype='d') to force reduction over doubles.
> This almost sound like what I reported is something like a feature !?
> Is there a sensible / generic way to avoid those "accident" ? Maybe it
> must be the default to reduce int8, uint8, int16, uint16 into doubles !?


Hard to say. I always bear the precision in mind when accumulating numbers
but even so it is possible to get unexpected results. Even doubles can give
problems if there are a few large numbers mixed with many small numbers.
That said, folks probably expect means to be accurate and don't want modular
arithmetic, so doubles would probably be a better default. It would be
slower though. I think there was a discussion of this problem previously in
regard to the reduce methods.

Chuck
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20060811/f223fd69/attachment-0001.html>