onnx: Import error on Raspberry Pi 3

I installed onnx on Raspberry Pi 3 in a non-Anaconda environment(Python 3.5.3).

sudo apt-get install protobuf-compiler libprotoc-dev
pip install onnx

I seemed to succeed in installing.

Installing collected packages: numpy, six, setuptools, protobuf, onnx
Successfully installed numpy-1.13.3 onnx-1.0.0 protobuf-3.5.1 setuptools-38.2.5 six-1.11.0

However, I couldn’t import onnx.

>>> import onnx
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/pi/.local/lib/python3.5/site-packages/onnx/__init__.py", line 10, in <module>
    import onnx.helper  # noqa
  File "/home/pi/.local/lib/python3.5/site-packages/onnx/helper.py", line 15, in <module>
    import onnx.defs as defs
  File "/home/pi/.local/lib/python3.5/site-packages/onnx/defs/__init__.py", line 6, in <module>
    import onnx.onnx_cpp2py_export.defs as C
ImportError: /home/pi/.local/lib/python3.5/site-packages/onnx/onnx_cpp2py_export.cpython-35m-
 arm-linux-gnueabihf.so: undefined symbol: 
 _ZNK6google8protobuf7Message11GetTypeNameB5cxx11Ev

Can anyone help us determine the possible cause?

About this issue

Original URL
State: closed
Created 7 years ago
Comments: 16 (4 by maintainers)

Most upvoted comments

The cxx11 in the missing symbol name suggests maybe that you are compiling with the new C++11 ABI, but the protobuf you are using is somehow on the old ABI. One possibility is that pip is downloading and installing a wheel that was built with the old C++11 ABI. To see if this is the case, I need to see the entire install log; e.g., blow away anything pip installed previously (pip uninstall numpy six setuptools protobuf onnx) and then report the FULL log when you run pip install -v onnx. You can also disable wheels by passing --no-binary (of course, now you will have to build these deps.)

Another possibility is that ONNX is being built with the C++11 ABI when it shouldn’t. You can disable this by passing CFLAGS="-D_GLIBCXX_USE_CXX11_ABI=0" as an environment variable.

Please let us know what the problem is in the end so we can help other users. Thanks.

@linkerzhang Note that conda is not an option here, because conda-forge does not currently do ARM builds: https://github.com/conda-forge/conda-forge.github.io/issues/269

ezyang on Dec 28, 2017

I had the exact same error when trying to build onnx on raspberry pi. I managed to solve the issue by adding the following to setup.py so that the compiler uses the old ABI:

extra_compile_args += [ '-D_GLIBCXX_USE_CXX11_ABI=0']

Thanks @ezyang for pointing me in the direction.

maitek on Jan 8, 2018