test for nan

**David M. Cooke** · Jul 18 '05, 09:54 AM

Re: test for nan

At some point, John Hunter <jdhunter@ace.b sd.uchicago.edu > wrote:
[color=blue]
> I have a C extension module that is returning some doubles. When the
> doubles get out of range, the numbers print as 'nan'.
>
> Is there a better way to test for NaN than
>
> str(p1)=='nan'
>
> where p1 is a float?[/color]

The string representation of NaN is system-dependent. On Windows it's
something like #NaN. You'd be better off with Jeff Epler's suggestion
of wrapping isnan() (since you already have a C extension module, you
could through it in there).

If you're on something that uses IEEE floating-point representations ,
something like this in pure python should work:

import struct
def isnan(x):
s = struct.pack('d' , x)
if struct.pack('h' , 1) == '\x01\x00':
return s == '\x00\x00\x00\x 00\x00\x00\xf0\ x7f':
else:
return s == '\x7f\xf8\x00\x 00\x00\x00\x00\ x00':

The test for endianness is there since struct.unpack(' <d', x)
complains that frexp() is out of range.

--
|>|\/|<
/--------------------------------------------------------------------------\
|David M. Cooke
|cookedm(at)phy sics(dot)mcmast er(dot)ca

**Michael Hudson** · Jul 18 '05, 09:55 AM

Re: test for nan

John Hunter <jdhunter@ace.b sd.uchicago.edu > writes:
[color=blue]
> I have a C extension module that is returning some doubles. When the
> doubles get out of range, the numbers print as 'nan'.
>
> Is there a better way to test for NaN than
>
> str(p1)=='nan'
>
> where p1 is a float?[/color]

I'd hope so: that's exceeding unportable. Of course, all things to do
with nans are unportable, so you're going to have to tell us more of
your requirements...

Cheers,
mwh

--
I have *both* hands clapping, but I'm still not sure it's a sound.
When I tried deciding if it were a sound while clapping only one
hand, I fell off my chair.
-- Peter Hansen, Zen master, comp.lang.pytho n

**Peter Maas** · Jul 18 '05, 09:57 AM

Re: test for nan

John Hunter wrote:[color=blue]
> I have a C extension module that is returning some doubles. When the
> doubles get out of range, the numbers print as 'nan'.
>
> Is there a better way to test for NaN than
>
> str(p1)=='nan'
>
> where p1 is a float?
>
> python2.3
>
> Thanks,
> John Hunter
>[/color]

As far as I know NaN is th only value that yields false when compared
to itself. This leads to

def isNaN(x):
return (x == x) == False

Mit freundlichen Gruessen,

Peter Maas

--
-------------------------------------------------------------------
Peter Maas, M+R Infosysteme, D-52070 Aachen, Hubert-Wienen-Str. 24
Tel +49-241-93878-0 Fax +49-241-93878-20 eMail peter.maas@mplu sr.de
-------------------------------------------------------------------

**Peter Maas** · Jul 18 '05, 09:57 AM

Re: test for nan

Peter Maas wrote:[color=blue]
> John Hunter wrote:[color=green]
>> Is there a better way to test for NaN than
>>
>> str(p1)=='nan'
>>
>> where p1 is a float?[/color]
> [...]
> As far as I know NaN is th only value that yields false when compared
> to itself. This leads to
>
> def isNaN(x):
> return (x == x) == False[/color]

Shorter:

def isNaN(x):
return x != x

Mit freundlichen Gruessen,

Peter Maas

--
-------------------------------------------------------------------
Peter Maas, M+R Infosysteme, D-52070 Aachen, Hubert-Wienen-Str. 24
Tel +49-241-93878-0 Fax +49-241-93878-20 eMail peter.maas@mplu sr.de
-------------------------------------------------------------------

**Peter Otten** · Jul 18 '05, 09:57 AM

Re: test for nan

Peter Maas wrote:
[color=blue]
> As far as I know NaN is th only value that yields false when compared
> to itself. This leads to
>
> def isNaN(x):
> return (x == x) == False[/color]

That is not portable:

Python 2.3.3 (#1, Jan 3 2004, 13:57:08)
[GCC 3.2] on linux2
Type "help", "copyright" , "credits" or "license" for more information.[color=blue][color=green][color=darkred]
>>> nan = float("nan")
>>> nan != nan[/color][/color][/color]
False

Peter

**Peter Maas** · Jul 18 '05, 09:57 AM

Re: test for nan

Peter Otten wrote:[color=blue][color=green]
>>def isNaN(x):
>> return (x == x) == False[/color]
>
>
> That is not portable:
>
> Python 2.3.3 (#1, Jan 3 2004, 13:57:08)
> [GCC 3.2] on linux2
> Type "help", "copyright" , "credits" or "license" for more information.
>[color=green][color=darkred]
>>>>nan = float("nan")
>>>>nan != nan[/color][/color]
>
> False[/color]

Thanks for correction. I firmly believed that NaN != NaN -> True
was a property of IEEE 754 and of all software using this standard.
For gcc/Linux you could use

def isNaN(x):
return (x == 0) and (x == 1)

and utilize sys.platform. :)

Mit freundlichen Gruessen,

Peter Maas

--
-------------------------------------------------------------------
Peter Maas, M+R Infosysteme, D-52070 Aachen, Hubert-Wienen-Str. 24
Tel +49-241-93878-0 Fax +49-241-93878-20 eMail peter.maas@mplu sr.de
-------------------------------------------------------------------

**Michael Hudson** · Jul 18 '05, 09:57 AM

Re: test for nan

Peter Maas <fpetermaas@net scape.net> writes:
[color=blue]
> John Hunter wrote:[color=green]
> > I have a C extension module that is returning some doubles. When the
> > doubles get out of range, the numbers print as 'nan'.
> > Is there a better way to test for NaN than
> > str(p1)=='nan'
> > where p1 is a float?
> > python2.3
> > Thanks,
> > John Hunter
> >[/color]
>
> As far as I know NaN is th only value that yields false when
> compared to itself.[/color]

There can be more than one NaN, by the way.
[color=blue]
> This leads to
>
> def isNaN(x):
> return (x == x) == False[/color]

This will work with 2.3 on Windows (I believe), current CVS on Windows
(if compiled with VC7.1), current CVS on Linux (assuming an even
vaguely recent gcc), but not current CVS on Windows compiled with VC6,
nor Python 2.3 on Linux/gcc. Confused yet?

Other platforms I haven't the faintest idea about.

Cheers,
mwh

--
41. Some programming languages manage to absorb change, but
withstand progress.
-- Alan Perlis, http://www.cs.yale.edu/homes/perlis-alan/quotes.html

**Peter Otten** · Jul 18 '05, 09:57 AM

Re: test for nan

Peter Maas wrote:
[color=blue]
> def isNaN(x):
> return (x == 0) and (x == 1)[/color]

This works here. Do you have an idea what the rationale behind this
behaviour (i. e. any number == nan) is?

Peter

**John Hunter** · Jul 18 '05, 09:57 AM

Re: test for nan

>>>>> "Michael" == Michael Hudson <mwh@python.net > writes:

Michael> This will work with 2.3 on Windows (I believe), current
Michael> CVS on Windows (if compiled with VC7.1), current CVS on
Michael> Linux (assuming an even vaguely recent gcc), but not
Michael> current CVS on Windows compiled with VC6, nor Python 2.3
Michael> on Linux/gcc. Confused yet?

Fortunately I only need linux for this particular app, so I can use
one of the platform dependent solutions, but the bevy of proposed
solutions and gotchas have definitely been interesting.

Out of curiosity, are there any platforms where this is known to fail?

def is_nan(x):
return str(x).lower(). find('nan')>=0

JDH

**Roel Schroeven** · Jul 18 '05, 09:57 AM

Re: test for nan

John Hunter wrote:[color=blue]
> Out of curiosity, are there any platforms where this is known to fail?
>
> def is_nan(x):
> return str(x).lower(). find('nan')>=0[/color]

Yes...

Python 2.3.3 (#51, Dec 18 2003, 20:22:39) [MSC v.1200 32 bit (Intel)] on
win32

IDLE 1.0.2[color=blue][color=green][color=darkred]
>>> def is_nan(x):[/color][/color][/color]
return str(x).lower(). find('nan') >= 0
[color=blue][color=green][color=darkred]
>>> inf=float('1e99 99')
>>> inf[/color][/color][/color]
1.#INF[color=blue][color=green][color=darkred]
>>> nan=inf-inf
>>> is_nan(nan)[/color][/color][/color]
False[color=blue][color=green][color=darkred]
>>> nan[/color][/color][/color]
-1.#IND

--
"Codito ergo sum"
Roel Schroeven

**Peter Maas** · Jul 18 '05, 09:57 AM

Re: test for nan

Michael Hudson wrote:[color=blue]
> There can be more than one NaN, by the way.[/color]

I know that but if all NaNs have the same distinct algorithmic behaviour
that's not a problem.
[color=blue][color=green]
>>This leads to
>>
>>def isNaN(x):
>> return (x == x) == False[/color]
>
>
> This will work with 2.3 on Windows (I believe), current CVS on Windows
> (if compiled with VC7.1), current CVS on Linux (assuming an even
> vaguely recent gcc), but not current CVS on Windows compiled with VC6,
> nor Python 2.3 on Linux/gcc. Confused yet?[/color]

No:

def isNaN1(x):
return x != x

def isNaN2(x):
return (x == 0) and (x == 1)

if sys.platform == 'win32' and '2.3' in sys.version:
isNaN = isNaN1
elif sys.platform == 'win32' and 'good_CVS_VC_st uff' in sys.version:
isNaN = isNaN1
elif sys.platform == 'win32' and 'bad_CVS_VC_stu ff' in sys.version:
isNaN = isNaN2
elif sys.platform == 'linux2':
isNaN = isNaN2
else:
raise Exception, "Roll your own isNaN() for platform " \
"%s and version %s\n" % (sys.platform, sys.version)

bad/good_CVS_VC_stu ff has to be worked out, don't have the version strings
available.

Mit freundlichen Gruessen,

Peter Maas

--
-------------------------------------------------------------------
Peter Maas, M+R Infosysteme, D-52070 Aachen, Hubert-Wienen-Str. 24
Tel +49-241-93878-0 Fax +49-241-93878-20 eMail peter.maas@mplu sr.de
-------------------------------------------------------------------

test for nan

test for nan

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment

Comment