RPi 上的 Numpy 导入错误“未定义符号：PyFPE_jbuf”

Question

Context

I'm attempting to install TensorFlow on my Raspberry Pi but numpy is having trouble working.

My RPi installation is fresh from the microSD. I am running Raspbian Buster and have not done anything prior on this system other than sudo apt update && sudo apt upgrade , install python3.5, and try to install the TensorFlow wheel file (the very last one) using python -m pip install <tensorflow>.whl .

I am using a clean virtualenv and pip to install python packages (both pip install ... and python -m pip install ... while in the env). The installations have no errors or warnings; things look fine until I try to use/import numpy.

The error referred to in the title is as follows:

Python 3.5.4 (default, Sep  5 2017, 18:32:10) 
[GCC 7.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/pi/tensorflow/lib/python3.5/site-packages/numpy/__init__.py", line 152, in <module>
    from . import random
  File "/home/pi/tensorflow/lib/python3.5/site-packages/numpy/random/__init__.py", line 181, in <module>
    from . import _pickle
  File "/home/pi/tensorflow/lib/python3.5/site-packages/numpy/random/_pickle.py", line 1, in <module>
    from .mtrand import RandomState
ImportError: /home/pi/tensorflow/lib/python3.5/site-packages/numpy/random/mtrand.cpython-35m-arm-linux-gnueabihf.so: undefined symbol: PyFPE_jbuf

Reproduce

I can consistently reproduce this by starting a clean virtualenv using python3.5, pip install numpy or python -m pip ... , then trying to import numpy like above.

Attempts

Most of the other questions I have seen pertaining to this have to do with cython installation paths or having another numpy on the computer, so this is what I have done so far:

1. Uninstalled the default numpy installation on the pi

2. Installed cython, reinstalled numpy

3. Installing numpy with pip install --no-cache-dir numpy

Note

numpy installation works on a python3.7 virtualenv

Update

I followed phd's comment and tried to build from source ( git clone ... , cd numpy , python setup.py all in the virtualenv on python3.5). This led to the same missing symbol, though in a different place. It looks Cython related?

Running from numpy source directory.
Cythonizing sources
Processing numpy/random/bounded_integers.pxd.in
Processing numpy/random/bounded_integers.pyx.in
Traceback (most recent call last):
  File "/usr/lib/python3.5/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/usr/lib/python3.5/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/home/pi/tensorflow/lib/python3.5/site-packages/cython.py", line 17, in <module>
    main(command_line = 1)
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/Main.py", line 858, in main
    result = compile(sources, options)
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/Main.py", line 780, in compile
    return compile_multiple(source, options)
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/Main.py", line 750, in compile_multiple
    context = options.create_context()
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/Main.py", line 597, in create_context
    self.cplus, self.language_level, options=self)
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/Main.py", line 80, in __init__
    from . import Builtin, CythonScope
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/CythonScope.py", line 5, in <module>
    from .UtilityCode import CythonUtilityCode
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/UtilityCode.py", line 3, in <module>
    from .TreeFragment import parse_from_strings, StringParseContext
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/TreeFragment.py", line 17, in <module>
    from .Visitor import VisitorTransform
  File "Cython/Compiler/Visitor.py", line 17, in init Cython.Compiler.Visitor
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/ExprNodes.py", line 4724, in <module>
    class SliceIndexNode(ExprNode):
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/ExprNodes.py", line 4922, in SliceIndexNode
    "SliceObject", "ObjectHandling.c", context={'access': 'Get'})
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/Code.py", line 404, in load
    return cls(**kwargs)
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/Code.py", line 648, in __init__
    proto = sub_tempita(proto, context, file, name)
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Compiler/Code.py", line 640, in sub_tempita
    from ..Tempita import sub
  File "/home/pi/tensorflow/lib/python3.5/site-packages/Cython/Tempita/__init__.py", line 4, in <module>
    from ._tempita import *
ImportError: /home/pi/tensorflow/lib/python3.5/site-packages/Cython/Tempita/_tempita.cpython-35m-arm-linux-gnueabihf.so: undefined symbol: PyFPE_jbuf

After this it just tells me that running cythonize failed in setup.py.

Answer 1

You need to compile numpy from the source. Get numpy 1.18.5 :

wget https://github.com/numpy/numpy/archive/refs/tags/v1.18.5.tar.gz

Uninstall the current Python 3.5 's numpy and install cython module:

pip3.5 uninstall numpy
pip3.5 install cython==0.29.22

Change directory to numpy 1.18.5 's source code, install it:

python3.5 setup.py install

Once finished, your code will be able to successfully run import numpy