Hi Bjorn,
IIUC, I think I’d try switching from openblas to MKL. openblas is what Numpy uses by default for algebra computations and has a long history of issues involving multiprocessing.
You can do conda list | grep blas to see if you’re using openblas in your virtual dev. The third column should tell which one it is:
libblas 3.8.0 16_openblas conda-forge
libcblas 3.8.0 16_openblas conda-forge
If your env uses openblas, you can try MKL or blis: conda install "libblas=*=*mkl" -c conda-forge . You’ll probably need to install mkl and mkl-service too. Doc about these options is here.
After replacing blas with MKL instead of openblas, you should see:
libblas 3.8.0 14_mkl conda-forge
libcblas 3.8.0 14_mkl conda-forge
...
mkl 2019.5 281 conda-forge
mkl-service 2.3.0 py37h0b31af3_0 conda-forge
Maybe it’s not the whole issue, but this should help with computational speed 