[Numpy-discussion] Adding .abs() method to the array object

[Sebastian Berg]

On Tue, 2013-02-26 at 11:16 +0100, Todd wrote:
> On Tue, Feb 26, 2013 at 10:58 AM, Sebastian Berg
> <sebastian@sipsolutions.net> wrote:
>         On Mon, 2013-02-25 at 22:04 -0500, josef.pktd@gmail.com wrote:
>         > On Mon, Feb 25, 2013 at 9:58 PM,  <josef.pktd@gmail.com>
>         wrote:
>         > > On Mon, Feb 25, 2013 at 9:20 PM, Sebastian Berg
>         > > <sebastian@sipsolutions.net> wrote:
>         > >> On Mon, 2013-02-25 at 10:50 -0500, Skipper Seabold wrote:
>         > >>> On Mon, Feb 25, 2013 at 10:43 AM, Till Stensitzki
>         <mail.till@gmx.de>
>         > >>> wrote:
>         > >>> >
>         > >>> > First, sorry that i didnt search for an old thread,
>         but because i
>         > >>> disagree with
>         > >>> > conclusion i would at least address my reason:
>         > >>> >
>         
>         <snip>
>         > >> Two small things (not sure if it matters much). But first
>         almost all of
>         > >> these methods are related to the container and not the
>         elements. Second
>         > >> actually using a method arr.abs() has a tiny pitfall,
>         since abs would
>         > >> work on numpy types, but not on python types. This means
>         that:
>         > >>
>         > >> np.array([1, 2, 3]).max().abs()
>         > >>
>         > >> works, but
>         > >>
>         > >> np.array([1, 2, 3], dtype=object).max().abs()
>         > >>
>         > >> breaks. Python has a safe name for abs already...
>         > >
>         > >>>> (np.array([1, 2, 3], dtype=object)).max()
>         > > 3
>         > >>>> (np.array([1, 2, 3], dtype=object)).__abs__().max()
>         > > 3
>         > >>>> (np.array([1, 2, '3'], dtype=object)).__abs__()
>         > > Traceback (most recent call last):
>         > >   File "<stdin>", line 1, in <module>
>         > > TypeError: bad operand type for abs(): 'str'
>         > >
>         > >>>> map(abs, [1, 2, 3])
>         > > [1, 2, 3]
>         > >>>> map(abs, [1, 2, '3'])
>         > > Traceback (most recent call last):
>         > >   File "<stdin>", line 1, in <module>
>         > > TypeError: bad operand type for abs(): 'str'
>         >
>         > or maybe more useful
>         >
>         > >>> from decimal import Decimal
>         > >>> d = [Decimal(str(k)) for k in np.linspace(-1, 1, 5)]
>         > >>> map(abs, d)
>         > [Decimal('1.0'), Decimal('0.5'), Decimal('0.0'),
>         Decimal('0.5'), Decimal('1.0')]
>         >
>         > >>> np.asarray(d).__abs__()
>         > array([1.0, 0.5, 0.0, 0.5, 1.0], dtype=object)
>         > >>> np.asarray(d).__abs__()[0]
>         > Decimal('1.0')
>         >
>         > Josef
>         >
>         > >
>         > > I don't see a difference.
>         > >
>         > > (I don't expect to use max abs on anything else than
>         numbers.)
>         > >
>         
>         
>         The difference is about scalars only. And of course __abs__ is
>         fine, but
>         if numpy adds an abs method, its scalars would logically have
>         it too.
>         But then you diverge from python scalars. That has exactly the
>         downside
>         that you may write code that suddenly stops working for python
>         scalars
>         without noticing.
>         
>         I turned around the abs and max order here, so that the abs
>         works on the
>         scalar, not useful but just as an example.
> 
> 
> But doesn't this also apply to many existing methods?
> 
I do not think that it does, or at a different quality. Almost all of
those methods either logically work on the container not the array. I.e.
reshape, etc. or are the most common reductions like mean/max/min (which
are also only sensible for containers). Now those also work on numpy
scalars but you should rarely use them on scalars. The only example I
can see is round (there is no special method for that before python 3,
so you could argue that python did not provide a canonical way for
numpy).

Now this is not a huge argument, obviously scalars can have a lot of
specialized methods (like decimals for example), but in the case of abs,
python provides a default way of doing it which always works, and it
makes me tend to say that it is better to use that (even if I sometimes
write .abs() too...).

> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion@scipy.org
> http://mail.scipy.org/mailman/listinfo/numpy-discussion