[pymvpa] pynifti ref-count bugs?

[Michael Hanke]

Hi,

On Wed, Oct 08, 2008 at 11:26:43AM +0200, Gael Varoquaux wrote:
> On Wed, Oct 08, 2008 at 11:14:14AM +0200, Michael Hanke wrote:
> > The file should be correct, but the way you access the array causes
> > trouble. getDataArray() simply returns a pointer to the data allocated
> > by the nifti library. As you do not store the nifti image itself, that
> > data is destroyed immediately, hence segfault when accessing it.
> 
> Ha, OK. It did seem that the data was being derefenced, but it wasn't
> clear to me that this was the expected behavior.
> 
> > Just replace getDataArray() (or simply .data) with asarray(), which has
> > a parameter 'copy' that is True by default.
> 
> Why is the copy necessary? It seems that this should simply be a refcount
> increment, rather than the copy. Travis Oliphant has a nice blog post on
> a cool way to do this:
> http://blog.enthought.com/?p=62
Good pointer! I guess the reason why it is not implemented like this is
that the blog post is dated "September 9th, 2008" ;-)

> If you do this, than you can encourage people not to use the data
> pointer, which is imho fairly dangerous.
> 
> I think getDataArray should be renamed to getDataPointer because this is
> really a trap :). Obviously that will break everybodys code, so you can
> force them to move to using asarray :). OK, this is not a way to treate
> users.
This is the main reason why I kept it like this. This piece stems from a
time where I had pynifti as my private puppet. It has still been like
this when I first released it and since then users require something
they call "backward-compatibility". ;-) This is also the reason for
the class representing the nifti header being the parent for the
NiftiImage class and not just a member of it ...


> > > I am running a git latest version of pynifti. As far as I can tell, I am
> > > running a 1.1.0 version of nifticlib (I can't find an obvious indication
> > > of the version number, but the last entry in the Updates.txt is 1.1.0),
> > > compiled from source. The system is a Mandriva (I don't have an option,
> > > there :( ).
> > I recently started compiling RPM packages with the OpenSUSE build
> > service. I am a total novice to RPM, but I also have them for Mandriva
> > 2007 and 2008 -- including packages for nifticlibs. I'd be glad, if you
> > could try them and tell me if they work
> 
> > http://download.opensuse.org/repositories/home:/hankem/
> 
> > or
> 
> > http://software.opensuse.org/search?baseproject=Mandriva%3A2008&p=1&q=python-nifti
> 
> Nice, very useful. You should point this out to Lev Givon who is doing a
> phD in neuro-sciences, uses scipy a lot, and is a Mandriva packager.
Please do! If someone who actually knows how RPMs are supposed to work
would take a look at them -- that would be great. If he has an OpenSUSE
account (or gets one) he could tweak the stuff on his own (if necessary)
-- it is all public.


> It unfortunately doesn't solve my instal problems, because we can't
> deploy packages on the machines here, we need to build them on an NFS
> drive (yes, we have some infrastructure problems, but we don't control
> them).
:(

Michael

-- 
GPG key:  1024D/3144BE0F Michael Hanke
http://apsy.gse.uni-magdeburg.de/hanke
ICQ: 48230050