[Numpy-discussion] Current copy

[Travis Oliphant]

Tom Loredo wrote:

>Hi Travis,
>
>  
>
>>I'm actually preparing the 1.0 release version, instead. 
>>
>>Here's the latest, though...
>>    
>>
>
>Thanks!
>
>I hate to be a nuisance about this, but what's the status
>of the latest releases of numpy/scipy?  Numpy 0.9.8 gives
>a segfault on my FC3 box.  
>
NumPy 0.9.8 should be fine except for one test.  That tests gives a 
segfault because of a problem with Python that was fixed a while ago. 

As long as you don't create the new complex array scalars (i.e. using 
cdouble(10), complex128(3), etc.) you should be fine with all code 
running NumPy 0.9.8.    Just delete the file 
site-packages/numpy/core/tests/test_scalarmath.py to get the tests to run.

>I waited till today to try the
>SVN version (per your scipy-dev post) and just installed
>rev 2669.  It passes the numpy tests--good!---but when I
>followed it with an install of scipy-0.4.9, importing
>scipy gives an error:
>
>import linsolve.umfpack -> failed: cannot import name ArrayType
>
>When you mentioned that the SVN numpy now worked with scipy,
>  
>
>was it only with SVN scipy?
>  
>
Yes.  You need to re-compile scipy to work with SVN NumPy.      Usually 
Ed Schofield has been helping release SciPy for each new NumPy release 
to make installation easier.

>I'm asking all this, partly for my own info, but also because
>last week at an astrostatistics conference I was given a long
>slot of time where I gave a pretty hard sell of numpy/scipy.
>I'm imagining all these people going home and installing the
>latest releases and cursing me under their breaths!
>
>Is it just my FC3 box having issues with the current releases?
>If not, I think something should be said on the download page
>(e.g., maybe encourage people to use SVN for certain platforms).
>  
>

It's just the one test that's a problem (my system was more forgiving 
and didn't segfault so I didn't catch the problem).    I doubt people 
are using the construct that is causing the problems much anway -- it's 
a subtle bug that was in Python when a C-type inherited from the Python 
complex type.

I'd probably recommend using SVN NumPy/SciPy if you are comfortable with 
compilation because it's the quickest way to get bug-fixes.   But, some 
like installed packages.  That's why we are pushing to get 1.0 done as 
quickly as is reasonable.