Segmentation Fault when importing SimpleITK 2.0 before TensorFlow

[alkamid]

Likely related to #1192

OS: Ubuntu 20.04
Python 3.8.2 (tested with 3.7, too)

Make a virtualenv:

python -m venv sitk_venv
source sitk_venv/bin/activate
pip install tensorflow==2.3.1 SimpleITK==2.0.0

test.py:

import SimpleITK as sitk
import tensorflow as tf

$ python test.py
Segmentation fault (core dumped)

If we reverse the order of imports (TensorFlow first), it works fine. I'm not sure whether it's an issue with TF or SimpleITK, but it doesn't happen with SimpleITK==1.2.4.

[blowekamp]

Thank you for reporting this issue. We will look into it.

[blowekamp]

SimpleITK wheels are compiled per PEP 513 manylinux1, while tensor flow is manylinux2010. I believe these should be interoperable; I will need further investigation.

[blowekamp]

Here is an related piece with PyArrow: https://uwekorn.com/2019/09/15/how-we-build-apache-arrows-manylinux-wheels.html

Specifically the section "Annoying end users issue #2: 'Segmentation fault after import TensorFlow"' appear to be relevant.

[ltetrel]

I had the same issue.
Changing the import order as @alkamid suggested worked for me.

[blowekamp]

This issue has been addressed in PR #1210 and has been merged into master. It is available in the "latest" nightly packages. Please download and test the package to verify it resolves the issue. Then it can be merged into the release branch for 2.0.1.

[alkamid]

I can confirm that using this wheel (2.1.0.dev23) fixes the issue.

[blowekamp]

SimpleITK 2.0.1 have been uploaded to pypi, which addresses the import order issue with tensor flow.