There is a crashing bug with fmod on Windows 10 2004. KB4598291 of February 2021 fixes this issue.

To prevent unexpected runtime behavior, opencv (and numpy 1.19.4 through 1.19.5) contains a check for this:

from numpy import arange, linalg, float64
try:
    a = arange(13 * 13, dtype=float64).reshape(13, 13)
    a = a % 17  # calls fmod
    linalg.eig(a)
except Exception:
    raise RuntimeError("The current Numpy installation (...) fails to pass a sanity check due to a bug in the windows runtime...")

See Pull Request https://github.com/numpy/numpy/pull/17553. You can verify the behavior yourself using the script above too.

Workaround

If you are unable to upgrade Windows, you can:

• Pinning against NumPy 1.19.3 should help (it uses a newer OpenBLAS version, but this caused other problems). This can be achieved using e.g. with pip install numpy==1.19.3 or similar depending on your setup. To be clear: The only difference between NumPy 1.19.3 and 1.19.4 is the OpenBLAS version it ships.
  • Or you can go to earlier versions like 1.18.5. Expect more differences.
• 32bit Python does not have these issues
• In principle you could revert the buggy windows update or deactivate the _win_os_check in NumPy (if you are lucky, your code is unaffected by the bug).