Raspberry Pi 4 aarch64: ModuleNotFoundError: No module named 'numpy.random.bit_generator'

I’m using the NLU framework Rasa using TensorFlow which is relying on scipy. I have no issues running the code on a amd64 platform, but I’m using a Raspberry Pi 4 aarch64 architecture, where I hit the problem:

ModuleNotFoundError: No module named ‘numpy.random.bit_generator’

I thought recompiling everything would fix this issue with:

pip3 install scipy==1.7.1 numpy==1.18.5 scikit-learn==0.24.2 --no-cache-dir --no-binary :all:

but I get the same error.

I cannot update at least numpy, because version 1.19 is introducing some breaking changes and TensorFlow does not work anymore.

Reproducing code example:

The problem can be also reproduced by executing in a Python 3 shell:

mkdir test
cd test
python3 -m venv venv
source venv/bin/activate
pip3 install scipy==1.7.1 numpy==1.18.5 scikit-learn==0.24.2
python3

Within the Python 3 shell execute:

import scipy.stats

Error message:

>>> import scipy.stats
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/ubuntu/accessibility/a/venv/lib/python3.8/site-packages/scipy/stats/__init__.py", line 441, in <module>
    from .stats import *
  File "/home/ubuntu/accessibility/a/venv/lib/python3.8/site-packages/scipy/stats/stats.py", line 43, in <module>
    from . import distributions
  File "/home/ubuntu/accessibility/a/venv/lib/python3.8/site-packages/scipy/stats/distributions.py", line 11, in <module>
    from . import _discrete_distns
  File "/home/ubuntu/accessibility/a/venv/lib/python3.8/site-packages/scipy/stats/_discrete_distns.py", line 19, in <module>
    from .biasedurn import (_PyFishersNCHypergeometric,
  File "biasedurn.pyx", line 1, in init scipy.stats.biasedurn
ModuleNotFoundError: No module named 'numpy.random.bit_generator'

Scipy/Numpy/Python version information:

1.7.1 1.18.5 sys.version_info(major=3, minor=8, micro=10, releaselevel='final', serial=0)

[rgommers]

What a mess. The attitude of the TensorFlow maintainers is that it’s just not important for them. Both the bug and the unwillingness to fix it are 100% on their side.

[rgommers]

Good point Tirth. If one waits long enough, some problems fix themselves:)

In this case, let’s close it when we drop numpy 1.18.x support, which we can do for the next release. No need to backport to 1.8.1 anymore I’d say.