[Numpy-discussion] ufunc oddities
Charles R Harris
charlesr.harris@gmail....
Sat May 24 22:32:39 CDT 2008
On Sat, May 24, 2008 at 9:18 PM, Robert Kern <robert.kern@gmail.com> wrote:
> On Sat, May 24, 2008 at 10:12 PM, Charles R Harris
> <charlesr.harris@gmail.com> wrote:
> >
> > On Sat, May 24, 2008 at 9:06 PM, Robert Kern <robert.kern@gmail.com>
> wrote:
> >>
> >> On Sat, May 24, 2008 at 10:02 PM, Charles R Harris
> >> <charlesr.harris@gmail.com> wrote:
> >> > So what about the rule that the array type takes precedence over the
> >> > scalar
> >> > type? That is broken for booleans.
> >>
> >> Yes, and if it wasn't an intentional special case (I don't recall
> >> discussing it on the list, but it might have been), then it's a bug
> >> and suitable for changing. The other behaviors are intentional and
> >> thus not suitable for changing.
> >
> > So how do we know which is which? How does one write a test for an
> > unspecified behavior? And it is clear that add.reduce breaks the rules
> for
> > booleans.
>
> It breaks it for every integer type, in fact.
>
> In [11]: add.reduce(array([1]*257, dtype=uint8))
> Out[11]: 257
>
> I thought we only did the accumulator-dtype changing for the .sum()
> method, but I was wrong.
>
I think the sum behavior has gone through changes over the last year and
that it is time to write down how things are supposed to work. So that we
can test that they actually work that way. This means specifying and
checking the type of all the outputs. And there are exceptions here also
In [33]: x = ones(2,dtype=int8)
In [34]: x.sum().dtype
Out[34]: dtype('int32')
In [35]: x = ones(2,dtype=int64)
In [36]: x.sum().dtype
Out[36]: dtype('int64')
So we have a default accumulator unless the precision is greater than the
default. Unless the numbers are floats
In [37]: x = ones(2,dtype=float32)
In [38]: x.sum().dtype
Out[38]: dtype('float32')
But the accumalator is of the same kind unless the kind is boolean, in which
case it is integer. Clear as a bell.
Chuck
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://projects.scipy.org/pipermail/numpy-discussion/attachments/20080524/e8659f31/attachment-0001.html
More information about the Numpy-discussion
mailing list