Mailman 3 May 2007 - NumPy-Discussion

APL2007: Arrays and Objects
by Steven H. Rogers 27 May '07

27 May '07

I'd like to see some NumPy/SciPy participation in this. I'm thinking of submitting a "Python as a Tool of Thought" paper, playing off of Ken Iverson's Turing Award Lecture "Notation as a Tool of Thought". A NumPy/SciPy tutorial would be valuable as well. # Steve =================================== Announcing APL2007, Montreal, October 21-23 The APL 2007 conference, sponsored by ACM SIGAPL, has as its principal theme "Arrays and Objects" and, appropriately, is co-located with OOPSLA 2007, in Montreal this October. APL 2007 starts with a tutorial day on Sunday, October 21, followed by a two-day program on Monday and Tuesday, October 22 and 23. APLers are welcome to attend OOPSLA program events on Monday and Tuesday (and OOPSLA attendees are welcome to come to APL program events). Registrants at APL 2007 can add full OOPSLA attendance at a favorable price. Dates: Sunday Oct 21 Tutorials Monday, Tuesday Oct 22,23 APL 2007 program Monday-Friday Oct 22-26 OOPSLA program APL 2007 keynote speaker: Guy Steele, Sun Microsystems Laboratories Tutorials Using objects within APL Array language practicum Intro to [language] for other-language users ( We expect that there will be at least one introductory tutorial on "classic" APL, and hope to have introductions to a variety of array languages ) We solicit papers and proposals for tutorials, panels and workshops on all aspects of array-oriented programming and languages; this year we have particular interest in the themes of integrating the use of arrays and objects languages that support the use of arrays as a central and thematic technique marketplace and education: making practitioners aware of array thinking and array languages Our interest is in the essential use of arrays in programming in any language (though our historical concern has been the APL family of languages: classic APL, J, K, NIAL, ....). Dates: Tutorial, panel, and workshop proposals, and notice of intent to submit papers, are due by Friday June 15, to the Program Chair. Contributed papers, not more than 10 pages in length, are due by Monday, July 23, to the Program Chair. Details of form of submission can be obtained from the program chair. Deadline for detailed tutorial/panel/workshop information TBA. Cost (to SIGAPL and ACM members, approximate $US, final cost TBA) APL2007 registration $375 Tutorial day $250 Single conference days $200 Social events: Opening reception Monday Others TBA Conference venue: Palais de Congres, Montreal, Quebec, CANADA Conference hotel: Hyatt Regency Montreal Committee General Chair Guy Laroque Guy.Laroque(a)nrcan.gc.ca Program Chair Lynne C. Shaw Shaw(a)ACM.org Treasurer Steven H. Rogers Steve(a)SHRogers.com Publicity Mike Kent MKent(a)ACM.org Links APL2007 http://www.sigapl.org/apl2007 OOPSLA 2007 http://www.oopsla.org/oopsla2007 Palais de Congres http://www.congresmtl.com/ Hyatt Regency Montreal http://montreal.hyatt.com Guy Steele http://research.sun.com/people/mybio.php?uid=25706 ACM SIGAPL http://www.sigapl.org

1 0

Median / mean functionality confusing?
by Matthew Brett 27 May '07

27 May '07

Hi, Does anyone else find this unexpected? In [93]: import numpy as N In [94]: a = N.arange(10).reshape(5,2) In [95]: N.mean(a) Out[95]: 4.5 In [96]: N.median(a) Out[96]: array([4, 5]) i.e. shouldn't median have the same axis, dtype, default axis=None behavior as mean? Best, Matthew

3 7

IEEE floating point arithmetics
by lorenzo bolla 26 May '07

26 May '07

Hi all. I have to work with floating point arithmetics and I found a module called "double module" (http://symptotic.com/mj/double/public/double-module.html) that does what I'd like. Basically, I would like to find the nearest smaller and bigger floating point numbers, given a "real" real number (the function doubleToLongBits<http://symptotic.com/mj/double/public/double-module.html#doubleToLongBits> can be used to do just that.). Is there something similar in numpy? In other words: is there a way, in numpy, to convert a floating point number to its binary representation and back? Thank you and regards, L.

4 4

shelve and object arrays
by Mark.Miller 26 May '07

26 May '07

Greetings: I recently experimented with changing from use of string arrays in some of my code to object arrays. This change speeds up my simulations and produces identical numerical results relative to use of string arrays. However, it now appears that my code is having issues at the ends of my simulations when all data (and simulation parameters) are archived using shelve. Here's the message: /var/lib/torque/mom_priv/jobs/539356.uint.SC: line 9: 18640 Segmentation fault I have been unable to create a small stand-alone example that reproduces the problem. Any thoughts on this? -Mark

2 1

Re: [Numpy-discussion] FW: RE: Linux numpy 1.0.1 install failed
by Gong, Shawn (Contractor) 25 May '07

25 May '07

Hi Robert, I used exactly your text in site.cfg. The out.txt seems to get the lib_dir, but "import numpy" in python stills gives error. Please see the attached files. Should I ask system guys to move the libs to /usr/lib so I don't have to fight with site.cfg ? Thanks, Shawn -----Original Message----- From: numpy-discussion-bounces(a)scipy.org [mailto:numpy-discussion-bounces@scipy.org] On Behalf Of Robert Kern Sent: Friday, May 18, 2007 5:56 PM To: Discussion of Numerical Python Subject: Re: [Numpy-discussion] FW: RE: Linux numpy 1.0.1 install failed Gong, Shawn (Contractor) wrote: > > Hi Robert, > > I added ref to site.cfg, but still getting the same error message: see > "out.txt" file. > > It seems that install can't find the libraries in /usr/local/lib/atlas/ > (see the "out.txt" file Line 14: libraries lapack,blas not found in > /usr/local/lib/atlas) Delete your [atlas] section entirely. It is wrong. Follow the instructions in the example. To make this clear, use exactly the following text in your site.cfg: [blas_opt] library_dirs = /usr/local/lib/atlas libraries = f77blas, cblas, atlas [lapack_opt] library_dirs = /usr/local/lib/atlas libraries = lapack, f77blas, cblas, atlas

3 13

Single-Precision FFT
by Benjamin M. Schwartz 24 May '07

24 May '07

-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 I am working on two programs using NumPy for the OLPC project. In both cases the performance is limited by the FFT. The OLPC machine uses a AMD Geode CPU, which is generally slow, but especially bad at double-precision floating point. It would be a major improvement if we could compile NumPy to use complex64 for FFTs instead of complex128, and might even increase the probability of NumPy being provided as a learning tool to millions of children. I know that FFTW can be compiled to run in single precision. What would it take to make NumPy use a single-precision FFT library? If absolutely necessary, it might be possible to ship a patched version of NumPy, but any other solution would be preferable. A compile-time configuration option in NumPy would be ideal. - --Ben Schwartz -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.6 (GNU/Linux) Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org iD8DBQFGVdIlUJT6e6HFtqQRAlCIAJ9o6DFMasm/ZABr8WMdRlmy/bTTMACZAXYn dCxeFPJtMbd/2YuYt9+4hDM= =aRQL -----END PGP SIGNATURE-----

3 3

f2py and C functions/subroutines/structs
by lorenzo bolla 24 May '07

24 May '07

Hi all, I hope this is the right mailing list to post my question to. I'm trying to make some easy C code working with Python by using f2py. For example, take the test.c file: ---------------------------------- typedef struct data { int a; double b; } DATA; /*function with struct*/ DATA incr(DATA x) { DATA y; y.a = x.a + 1; y.b = x.b + 1.; return(y); } /*function: return value*/ int incr0(int x) { return(x+1); } /*subroutine: no return value*/ void incr1(int *x) { *x += 1; } ---------------------------------- If I do: $ f2py -c test.c -m prova all seems fine, but inside ipython I get this error: In [1]: import prova --------------------------------------------------------------------------- <type 'exceptions.ImportError'> Traceback (most recent call last) /xlv1/labsoi_devices/bollalo001/work/test/python/<ipython console> in <module>() <type 'exceptions.ImportError'>: dynamic module does not define init function (initprova) I think I'm missing a correct .pyf file to do: $ f2py -c prova2.pyf test.c I tried writing prova2.pyf by my own, because doing: $ f2py test.c -h prova2.pyf -m prova gives me an empty prova2.pyf, but I wasn't able to do it!!! Can anyone of use kindly show me how to write it? Or giving me a good tutorial to read? I found only very few information on www.scipy.org. Thank you in advance, Lorenzo.

3 4

Another flags question
by Albert Strasheim 24 May '07

24 May '07

Hello all Me vs the flags again. I found another case where the flags aren't what I would expect: In [118]: x = N.array(N.arange(24.0).reshape(6,4), order='F') In [119]: x Out[119]: array([[ 0., 1., 2., 3.], [ 4., 5., 6., 7.], [ 8., 9., 10., 11.], [ 12., 13., 14., 15.], [ 16., 17., 18., 19.], [ 20., 21., 22., 23.]]) In [120]: x[:,0:1] Out[120]: array([[ 0.], [ 4.], [ 8.], [ 12.], [ 16.], [ 20.]]) # start=0, stop=1, step=2 In [121]: x[:,0:1:2] Out[121]: array([[ 0.], [ 4.], [ 8.], [ 12.], [ 16.], [ 20.]]) In [122]: x[:,0:1].flags Out[122]: C_CONTIGUOUS : False F_CONTIGUOUS : True OWNDATA : False WRITEABLE : True ALIGNED : True UPDATEIFCOPY : False In [123]: x[:,0:1:2].flags Out[123]: C_CONTIGUOUS : False F_CONTIGUOUS : False OWNDATA : False WRITEABLE : True ALIGNED : True UPDATEIFCOPY : False In [124]: x[:,0:1].strides Out[124]: (8, 48) In [125]: x[:,0:1:2].strides Out[125]: (8, 96) The views are slightly different (as can be seen from at least the strides), but I'd expect F_CONTIGUOUS to be true in both cases. I'm guessing that somewhere this special case isn't being checked for, which translates into a "missed opportunity" for marking the view as contiguous. Probably not a bug per se, but I thought I'd mention it here. Cheers, Albert

2 1

numpy testing
by Brian Hawthorne 24 May '07

24 May '07

Hello, I'm having a bit of trouble with numpy testing not finding my tests. It seems like a bug, but maybe it's a subtle feature. I constructed the simplest possible straw man to illustrate my problem (attached as foo.tgz). The directory structure looks like this: a/ __init__.py foo.py tests/ test_foo.py To witness the problem: % tar zxf foo.tgz % python -c 'import a; a.test()' % python -c 'import a; a.testall()' The test method (of NumpyTest class) does not pick up the test, but testall does. I think I've followed all the naming rules, even included test_, check_, and bench_ methods in the test case, in case they were treated differently. Any suggestions? I encountered this problem while attempting to isolate another (maybe related) problem with numpy testing. I'm working on the nipy project ( neuroimaging.scipy.org) and we are using the numpy test system. During a refactoring session, I noticed that when I moved some existing functionality (along with its tests) into a new identically named module under a different package (like a.b.foo and a.c.foo), these particular tests were no longer found. I thought it might have something to do with the modules having the same name, but when I tried to isolate the problem in a mockup, I ran into the problem described above. So, first things first. Maybe it's something really dumb that I can't see because I've been staring at it for too long. Thanks in advance for any suggestions, Brian

2 2

Six-legged feature in median function
by Mike Ressler 24 May '07

24 May '07

Bumped into the following in numpy-1.0.2 and 1.0.3 (of course :-) on both 32-bit and 64-bit linux boxes: >>> import numpy as nm >>> a=nm.zeros(1000000,dtype='Int32')-30000 >>> nm.median(a) -30000.0 >>> a=nm.zeros(1000000,dtype='Int16')-30000 >>> nm.median(a) Warning: overflow encountered in short_scalars 2768.0 Something in the computation of median overflows when there are a large number of values near the ends of an Int16 (like my image data). I can work around this, of course, by casting the data as something larger, but this one could definitely use a tweak. Mike -- mike.ressler(a)alum.mit.edu

1 0