ImportError is misleading when importing an old multiarray in NumPy 1.16.1

[tbekolay]

I install NumPy from source. I recently checked out the 1.16.1 tag and when I imported the built and installed NumPy, I get

ImportError: Something is wrong with the numpy installation. While importing we detected an older version of numpy in ['/home/tbekolay/.virtualenvs/pytest/lib/python3.7/site-packages/numpy']. One method of fixing this is to repeatedly uninstall numpy until none is found, then reinstall this version.

The directory pointed there is the directory of the currently installed NumPy. When I looked in numpy/core/__init__.py, I saw the other error message, which includes the git clean -xdf command. When I ran this in the git repo, I could build and install successfully. So I believe it would be useful for both of those cases to include the instructions that normally occur when multiarray can't be imported. (line 19 in core/__init__.py).

Alternatively, we could detect if the path given by sys.modules['numpy'].__path__ is the same as the one we would get by looking at __file__ to determine if it is a weird version error or just an old multiarray sticking around.

Reproducing code example:

Check out an old release like 1.15.1 and build from source. Check out 1.16.1, build again and install. Then

import numpy as np

Error message:

ImportError: Something is wrong with the numpy installation. While importing we detected an older version of numpy in ['/home/tbekolay/.virtualenvs/pytest/lib/python3.7/site-packages/numpy']. One method of fixing this is to repeatedly uninstall numpy until none is found, then reinstall this version.

Numpy/Python version information:

1.16.1 3.7.2+ (default, Feb  2 2019, 14:31:48) 
[GCC 8.2.0]

[charris]

@mattip I think we should try using __import__ when importing ufunc.py and multiarray.py, it allows forcing relative imports only without exploring higher level directories. See https://docs.python.org/3/library/functions.html#__import__, I think level=1 would do it.

[mattip]

@charris what error scenario would that mitigate? It would not help in this case.

[eric-wieser]

Pretty sure that the description of __import__ is out of date, and it probably respects pep420 packages in the same way as any other import

[charris]

@eric-wieser The documentation is for Python 3.7.2. What we need is a way to replicate the problem so we can test things.

[rgommers]

What we need is a way to replicate the problem so we can test things.

The description of this issue provides one. Faster way: use your regular in-place dev build, and just copy in a multiarray.so (grab from a wheel or whatever).

[charris]

@rgommers So the problem is duplicate files with different extensions in the same directory?

[rgommers]

@rgommers So the problem is duplicate files with different extensions in the same directory?

Unless I'm missing something, yes. It's this: https://github.com/numpy/numpy/pull/12850/files

Reports on this stuff are daily, I think we need a better solution here. I didn't follow very closely though, there were multiple proposed solutions but no water tight ones yet.

[mattip]

The daily reports are a different issue altogether. Anaconda recently broke setting os.environ['PATH'].

[rgommers]

@mattip you're right, that one is worse. This is also common though, e.g. gh-12938, gh-12736, gh-12826, #12957 (comment)

[PorkPy]

I had this error. It was due to multiple instances of numpy. It seems python is looking for numpy packages not numpy installations. Following the advice from the error handler in bash and uninstalling numpy won't remove the error because python can still see the uninstalled packages.
If these versions of numpy are needed for some other application, for the time being, just move them onto your desktop where python can't find them. This should resolve the error.
Hope this helps.
Edit: You may need to use 'pip3 install numpy' when using python3.
Edit2: I also can't use my code with python3.6 or 3.7 even if I upgrade numpy. Only 3.5 for me. If I try to use python >3.5 I get "ImportError: cannot import name 'multiarray' ".

[rgommers]

It seems python is looking for numpy packages not numpy installations.

that isn't a well-defined distinction

uninstalling numpy won't remove the error because python can still see the uninstalled packages.

then the uninstall was incomplete. uninstall with pip should remove both site-packages/numpy/ and the corresponding .dist-info or .egg-info directories.

[OnlyBelter]

If you use conda, you can update numpy by
conda update -n env_name -c defaults numpy

[wasiuyahya]

Use pip uninstall numpy
It's due to the duplicate of numpy

[nguyenvulong]

Try
pip uninstall numpy and
sudo pip uninstall numpy

repeatedly until you purge all numpy versions.
Replace pip by pip3 if you use python3.

[ctkqiang]

For Ubuntu:

I solve it by sudo apt-get install python3-numpy && sudo apt-get install python-numpy.
Hope this help.