[SciPy-user] scipy super pack installer for win32: please test

David Cournapeau david@ar.media.kyoto-u.ac...
Wed Oct 22 10:27:11 CDT 2008


joep wrote:
>
> installs without problems on windowsXP, sse2
> The only problem I had was that I had a trunk version of scipy (on a
> non site-packages directory) in easy-install.pth, which I needed to
> comment out. Initially it loaded this version instead of the one
> installed
> in site-packages by the superpack.

Unfortunately, there is not much I can do here, since this is a
setuptools problem. In my experience, the easy-install.pth is fragile,
and when you start using it, you should be ready to edit it by hand
occasionally.

>
> One weird thing I found, when I run scipy.test() in a command shell,
> or ipython shell, then I get a large output with Depreciation Warnings
> and test output. When I run scipy.test() in an Idle shell, I don't get
> any Depreciation Warnings or test output. I have no idea why the
> output should be different.

Maybe something linked to the Idle shell: warning and deprecation can be
controlled by the python interpreter. I find it a bit strange to ignore
warnings by default, though.

>
> Another observation which is not clear but is not very important is
> that, when I compiled the trunk version a few weeks ago (with MingW)
> and I run scipy.test(), then I don't get the 2 test failures in
> test_lapack.test_all_lapack. (my lapack and atlas files were
> downloaded a long time ago from the scipy.org install description
> page)

Different flags/compilers/atlas sources can cause some differences in
the results. The errors in the tests are small but not negligeable; I
don't know if they are significant. The superpack uses ATLAS 3.8.2, and
it does not produce those errors on linux; so this may be caused by
relatively old compilers on windows (mingw gcc is really old if you use
the released ones). To complicate the matter, a given ATLAS build is not
reproducible, making those issues even more difficult to track.

Thanks for testing,

David


More information about the SciPy-user mailing list