Re: [easybuild] matplotlib installation fails with ImportError

[Maik Schmidt]

Hi,




Mystery solved, sorry for the confusion.

Kind regards,
Maik

Am 13.10.2016 um 12:25 schrieb Kenneth Hoste:

I noticed that in the site-packages directory of the Python installation, there was only "numpy" but no corresponding *egg-info file. Apparently, this caused setuptools to fetch the latest version again during the matplotlib setup. I have no idea why the file was gone, but after restoring it from one of our older file system snapshots, the matplotlib installation now works fine.

Hi Maik,








regards,

Kenneth

On 13/10/16 12:13, Maik Schmidt wrote:

I just tried to reinstall matplotlib-1.5.1-intel-2016a-Python-3.5.1.eb , and I'm not seeing that a newer numpy is being picked up at all... Can you share the full log file with us, preferably collected using eb --debug? Is there any chance you have numpy 1.11.2 installed system-wide, or somehow avaiable through $PYTHONPATH or in $HOME/.local/lib/python*/site-packages?

Kenneth,

I'm using the unmodified easyconfig files from the EasyBuild distribution (just with a different toolchain as I mentioned in my original mail). https://github.com/hpcugent/easybuild-easyconfigs/blob/master/easybuild/easyconfigs/p/Python/Python-3.5.1-intel-2016a.eb

and

https://github.com/hpcugent/easybuild-easyconfigs/blob/master/easybuild/easyconfigs/m/matplotlib/matplotlib-1.5.1-intel-2016a-Python-3.5.1.eb

Ole,


Regards,
Maik



Am 13.10.2016 um 11:44 schrieb Ole Holm Nielsen:

yes, the Python 2 installation works for me, too. It's just Python 3 that seems to have this problem, no idea why...

FWIW, we build matplotlib without problems using the foss2016b toolchain, see the attached easyconfig file, which is based upon the distribution file matplotlib-1.5.2-intel-2016b-Python-2.7.12.eb.

/Ole

On 10/13/2016 11:39 AM, Kenneth Hoste wrote:

Hi Maik,

On 13/10/16 11:37, Maik Schmidt wrote:

Hi Kenneth,

thanks for your input. I have tried with matplotlib 1.5.2 and 1.5.3,
but unfortunately, the error remains the same.

The numpy version in my Python 3.5.1 is the original one from the
respective easyconfig, namely 1.10.4. The matplotlib installer even
says so under

REQUIRED DEPENDENCIES AND EXTENSIONS
                 numpy: yes [version 1.10.4]

however, during the installation process, it somehow seems to have
fetched and referenced a newer version of numpy (1.11.2) in a temp
directory as can be seen in the log:




So I guess this is more a problem of the installer, always trying to
use the latest numpy version or so?

/scratch/hpcsupport/easybuild-tmp/eb-i1onxx/easy_install-v0f9q0bi/numpy-1.11.2/setup.py

when it found one? I've never seen that happening (but maybe I just
overlooked it).

Can you share the easyconfig files you're using?


regards,

Kenneth

That's surprising... Why would it try to download/install it's own numpy

Kind regards,
Maik


Am 13.10.2016 um 11:00 schrieb Kenneth Hoste:

Hi Maik,

On 13/10/16 10:43, Maik Schmidt wrote:

Hi all,

I've installed Python-3.5.1-intel-2016a.eb (with
--try-toolchain="intel,2016.03-GCC-5.3" but that shouldn't be
relevant to the problem) and am now trying to install

installing the extension matplotlib.

== 2016-10-13 10:31:49,338 build_log.py:163 ERROR cmd "

build " exited with exitcode 1 and output: ...

From the log I can see that an import exception is raised:

ImportError: No module named 'numpy._build_utils'

I have attached the full log.


Isn't this just some internal module of the numpy installer and
shouldn't even be used by other modules? Any advice on how to get
rid of this problem would be much appreciated. Thanks in advance.

matplotlib-1.5.1-intel-2016a-Python-3.5.1.eb, but it always fails on /sw/taurus/eb/Python/3.5.1-intel-2016.03-GCC-5.3/bin/python setup.py Is this normal? Why does it try to access numpy._build_utils anways?

This looks relevant: https://github.com/numpy/numpy/issues/6551 .

It seems to suggest that the numpy version you have included in your
Python 3.5.1 installation is too new for matplotlib 1.5.1?

Maybe try with matplotlib 1.5.2, if that's an option?


regards,

Kenneth

--
Maik Schmidt

Technische Universität Dresden
Zentrum für Informationsdienste und Hochleistungsrechnen (ZIH)
Willers-Bau A116
D-01062 Dresden
Telefon: +49 351 463-32836

smime.p7s
Description: S/MIME Cryptographic Signature