Small compatibility fix for master of Theano-PyMc

[semohr]

• Import of pymc3 failed with the most recent master branch of Theano-PyMC
• I could nail down the problem to this commit on the Theano-PyMC GitHub
• Changed the __init__.py to use the new config style

[twiecki]

We probably need a new Theano release and pin.

[twiecki]

CC @brandonwillard @dfm

[MarcoGorelli]

Maybe after this is fixed we also need a theano-pymc-compatibility CI job which installs the latest commit to master (similarly to the arviz-compatibility one)

[brandonwillard]

There is already an open Theano-PyMC compatibility update PR that's waiting for the next release of Theano-PyMC (the PR title is too specific to one of the updates, though). All these updates are better off in a single PR corresponding to a single new release of Theano-PyMC.

The changes addressed in this PR were merged just last night, so they haven't been added to the existing PR, and the version of Theano-PyMC used by PyMC3 should already be pinned in this repo (from a commit merged a while back). In other words, the relevant Theano-PyMC changes aren't an active issue, since one would need to install an explicitly unsupported version of Theano-PyMC for them to appear.

[semohr]

Sorry if I wasted your time here a bit. I just tried to get the new jax sampler working, I noticed this problem and found a easy fix for it.

Thought I should share this, since maybe someone else could have the same problem.

I'm not too sure how you want to handle such small fixes.
Feel free to delete this PR.

Best Sebastian

[brandonwillard]

If you want to take on these recent updates, you can pull in the changes from the existing PR and make this one the cumulative update PR. There shouldn't be much in that PR, so you could probably just look at the diffs and apply them manually.

We simply need those changes in a single PR, because we can't merge them one at a time—at least not without temporarily putting the repo in a truly broken state.

Otherwise, did you install a development version of PyMC3 (e.g. master) and it pulled in the most recent Theano-PyMC? If that's the case, we should probably find out why and fix it, because it implies that our pinned version is being ignored/circumvented.

[semohr]

I totally unterstand, will look into that.

Yes I installed the master branch. Tried the jax branch beforehand but that one also installed the master theano-pymc branch.

[brandonwillard]

Tried the jax branch beforehand but that one also installed the master theano-pymc branch.

Ah, that's probably it; if you're referring to the pymc3jax branch, it might need to be rebased onto master in order to get the commit that pinned Theano-PyMC (and some previous compatibility updates, as well).

@twiecki, wasn't the pymc3jax merged? If so, we might want to delete it; otherwise, people could accidently use it and run into problems.

[brandonwillard]

The updates in this PR will also close aesara-devs/aesara#216.

[codecov]

Codecov Report

Merging #4293 (8ac3e3a) into master (0431292) will decrease coverage by 0.00%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #4293      +/-   ##
==========================================
- Coverage   87.56%   87.56%   -0.01%     
==========================================
  Files          88       88              
  Lines       14270    14270              
==========================================
- Hits        12496    12495       -1     
- Misses       1774     1775       +1     

Impacted Files	Coverage Δ
pymc3/distributions/multivariate.py	83.33% <100.00%> (ø)
pymc3/math.py	68.47% <100.00%> (ø)
pymc3/ode/ode.py	93.81% <100.00%> (ø)
pymc3/theanof.py	89.61% <0.00%> (-0.44%)	⬇️

[brandonwillard]

This won't pass until conda-forge/theano-pymc-feedstock#8 is merged and theano-pymc==1.0.12 is made available via conda-forge.

[twiecki]

Merged now.