py3 compatibility

[clbarnes]

Shouldn't be too bad as the heavy lifting is done with other libraries

Issues I've hit:

• Relative imports

[tcassou]

Hi @clbarnes
Thanks for reporting! It should work now with 1.0.3, and feel free to reopen the issue if you still experience errors!

[clbarnes]

Thanks for responding so quickly! I didn't expect such a rapid fix.

I'm still having issues with that import, though:

In [1]: from babtest.babtest import BABTest
---------------------------------------------------------------------------
ModuleNotFoundError                       Traceback (most recent call last)
<ipython-input-1-504c4f535c54> in <module>()
----> 1 from babtest.babtest import BABTest

~/.pyenv/versions/miniconda3-4.3.30/envs/heinrich3/lib/python3.6/site-packages/babtest/babtest.py in <module>()
      7 from pymc import MCMC
      8 
----> 9 from .models.bernoulli_model import BernoulliModel
     10 from .models.exponential_model import ExponentialModel
     11 from .models.gaussian_model import GaussianModel

ModuleNotFoundError: No module named 'babtest.models'

[clbarnes]

I have a branch running with tox on travis

https://travis-ci.org/clbarnes/babtest/builds/449090905

Locally, it fails with python 3.6 because of trouble locating libmkl_rt.so, which pymc needs, and on 3.7 because some of your pre-commit hooks here https://github.com/tcassou/python-pre-commit-hooks aren't compatible (astroid manually raising StopIteration rather than just returning from a generator).

On travis, 3.6 passes, so I'm happy to put that down to a problem with my environment. We can't easily run tests against 3.7 on travis yet because their containers are ooooold.

Of course, neither of those are the import error >_< With tests packaged inside the module it's hard to test how imports work. For this reason, I usually have my tests outside the python module, and make sure the library is installed in editable mode (pip install -e .).

(P.S. travis doesn't like doing things with environment variables and command chaining, so I couldn't replicate your "separate builds for lint and test" pattern)

[tcassou]

Ah! Sorry to hear it did not fix it, I'll take another look.
On my machine it worked fine for both 2.7 and 3.6 (and I had to reinstall everything from scratch so it's not an old artefact), but I may have missed something.

I won't be able to push things that quickly this time, and if you find a fix feel free to send a PR!

[clbarnes]

Done a little more digging: I think PyMC2 is incompatible python py3.7, because it defines a function called await which is a reserved name in 3.7+.

Replacing PyMC across an API-breaking version is a much deeper change than a few refactors so no worries if that's not workable within this framework!

As I said above, I think your pre-commit hooks would also need to change for 3.7 because of the raise StopIteration thing.

I'm still trying to figure out the libmkl thing locally, then I'll be able to test the imports.

[tcassou]

Thanks a lot for digging! I'll keep an eye on PyMC, for now everything should work fine with 3.6, that's at least something.

You're right for some reasons I ended up having tests inside the package, which I just changed in e943d51.

Next I'll look into the pre-commit hooks and see if they can be made compatible.

[tcassou]

So pylint was the culprit here, I upgraded it to the latest version (011f7ae) which fixes the StopIteration error, and the code is now compatible with python3.7 .
Only part missing is the support for py3.7 in PyMC, I'll check more recent versions.