在安装需要 numpy 的包时，如何防止 pip 仅用 numpy 替换 numpy+mkl？

Question

When I pip install (or pip install --upgrade ) packages that require numpy , they have a tendency to uninstall my existing numpy+mkl (which has a high enough version to satisfy the numpy version requirement). Afterwards, they install numpy without +mkl , which causes problems for other packages that do require MKL. An example for which this happens is gym (which has 'numpy>=1.10.4' in its install_requires in setup.py ).

I understand that this is related to the +mkl suffix that probably somehow messes with the versions, and understand I can fix it afterwards by downloading and installing numpy+mkl from https://www.lfd.uci.edu/~gohlke/pythonlibs/ , but it gets annoying to manually do this every time over again when upgrading a package like gym to a new version. Is there any way to prevent numpy+mkl from getting uninstalled during the pip install --upgrade ?

---

For me, this is happening on Windows 10 , Python 3.6 . I did not yet check if the same happens on Linux , but would be interested in an answer for that too if it's different there.

My currently installed version of numpy+mkl (which often gets automatically uninstalled) is 1.13.3+mkl .

Answer 1

Using --upgrade-strategy , as suggested by cgohlke in a comment, addresses this problem. So, taking the example where we want to install gym from scratch without it replacing our existing numpy+mkl installation with regular numpy , the full command to run is:

pip install --upgrade-strategy only-if-needed gym

Or, if we just want to upgrade an existing installation , we also add --upgrade

pip install --upgrade --upgrade-strategy only-if-needed gym

Answer 2

I have found that some packages force pip to reinstall numpy. The safest (and only) way to ensure that numpy is installed with mkl (from conda) is to uninstall using conda and pip and then reinstall using conda:

conda uninstall numpy
pip uninstall nump
conda install numpy