On Thu, Aug 25, 2011 at 7:50 PM, Reinout van Rees email@example.com wrote:
On 26-08-11 03:06, Matthew Brett wrote:
'nibabel' needs to import numpy to install
With your latest reply I re-read your original. Hm, this looks like a real problem: if your setup.py already imports from numpy (which your statement sounds like), numpy really needs to be there already.
Yes, that is what made me think that a new recipe might be the only way, where the recipe would put numpy on the path during the develop step.
The only thing I can think of right now is to have numpy installed globally and to use buildout 1.4.4 + osc.recipe.sysegg to depend on numpy. If you call the sysegg part right at the beginning, you might get it to work.
Yes, if numpy is already on the path, it works, but we wanted to test against several different numpy versions, and it seemed ugly and fragile to make a separate virtualenv for each python / numpy version pair. We'd much rather use the buildout mechanism to specify these, if possible.
Alternative is to re-work your setup.py. Do you really need to import numpy?
Actually, no, not in this case, it is only used as a crude dependency check. But, for other packages we have, we do need numpy.distutils for the setup.py - for example to find the numpy include directory for compiling C modules. The only reason I was offering this package as an example is that it is rather small and easy to install compared to the others.
Thanks a lot,