Circular import from utils

[mdmould]

Imports from sbi.neural_nets.* result in a circular import due to the interactions between sbi.utils.__init__.py and sbi.utils.get_nn_models with sbi.neural_nets. E.g., from sbi.neural_nets.flow import build_nsf gives:

>>> from sbi.neural_nets.flow import build_nsf
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/matthew/.pyenv/versions/sbi/lib/python3.8/site-packages/sbi/neural_nets/flow.py", line 14, in <module>
    from sbi.utils.sbiutils import (
  File "/home/matthew/.pyenv/versions/sbi/lib/python3.8/site-packages/sbi/utils/__init__.py", line 4, in <module>
    from sbi.utils.get_nn_models import classifier_nn, likelihood_nn, posterior_nn
  File "/home/matthew/.pyenv/versions/sbi/lib/python3.8/site-packages/sbi/utils/get_nn_models.py", line 14, in <module>
    from sbi.neural_nets.flow import build_made, build_maf, build_nsf
ImportError: cannot import name 'build_made' from partially initialized module 'sbi.neural_nets.flow' (most likely due to a circular import) (/home/matthew/.pyenv/versions/sbi/lib/python3.8/site-packages/sbi/neural_nets/flow.py)

This is because the sbi.neural_nets.* import sbi.utils, in which sbi.utils.get_nn_models then imports sbi.neural_nets.

One can instead use, e.g., from sbi.utils.get_nn_models import build_nsf or import sbi.utils; from sbi.neural_nets.flow import build_nsf, or the documented interface sbi.utils.get_nn_models.posterior_nn. But the first two options are less intuitive and the last obfuscates the parameters of the chosen model.

[jsvetter]

TODO for Hackathon:

Circular import

Fix by either

• Restructuring codebase
• Directly importing the relevant objects and removing the circular import from __init__.py

[alvorithm]

Maybe relevant:

• Better ways to name your utils module
• Stop naming your python modules "utils"

[famura]

I plan to have a look at it during the hackathon