Pin SciPy to <1.8.0

[michaelosthege]

See tracking issue #5448.

[codecov]

Codecov Report

Merging #5449 (af4896d) into main (18346ac) will increase coverage by 1.18%.
The diff coverage is n/a.

❗ Current head af4896d differs from pull request most recent head 060cfdf. Consider uploading reports for the commit 060cfdf to get more accurate results

@@            Coverage Diff             @@
##             main    #5449      +/-   ##
==========================================
+ Coverage   80.23%   81.41%   +1.18%     
==========================================
  Files          82       82              
  Lines       13941    14218     +277     
==========================================
+ Hits        11185    11576     +391     
+ Misses       2756     2642     -114     

Impacted Files	Coverage Δ
pymc/bart/bart.py	94.33% <0.00%> (-1.58%)	⬇️
pymc/distributions/simulator.py	86.11% <0.00%> (-0.66%)	⬇️
pymc/util.py	75.62% <0.00%> (-0.48%)	⬇️
pymc/bart/pgbart.py	95.23% <0.00%> (-0.37%)	⬇️
pymc/bart/tree.py	100.00% <0.00%> (ø)
pymc/initial_point.py	100.00% <0.00%> (ø)
pymc/distributions/bound.py	100.00% <0.00%> (ø)
pymc/smc/smc.py	97.90% <0.00%> (+<0.01%)	⬆️
pymc/step_methods/mlda.py	96.39% <0.00%> (+<0.01%)	⬆️
pymc/distributions/discrete.py	99.76% <0.00%> (+0.01%)	⬆️
... and 30 more

[OriolAbril]

rtd is still failing, from trying to install numpy==1.19 which I can only assume is a strange requirement of some testing library. We might need to split test/docs... requirements for this to work (and therefore address #5074)

[michaelosthege]

Running code --diff .github\workflows\pytest.yml .github\workflows\jaxtests.yml and code --diff .github\workflows\pytest.yml .github\workflows\arviz_compat.yml it looks like the environments between those workflows are quite different:

• pytest.yml vs. jaxtests.yml/arviz-compat.yml
• ubuntu-18-04 vs. ubuntu-latest/maxos-latest
• environment-dev-py37.yml vs. environment-dev-py39.yml
• no extras vs. pip install "numpyro>=0.8.0"/pip install git+git://github.com/arviz-devs/arviz.git

[OriolAbril]

I guess there are significant differences between ubuntu 18 and latest (should be 20 I think). But also with python 3.7! IIRC, the latest versions of numpy/scipy... don't support it anymore, there was a NEP about this but don' t have the link at hand, I'll try to come back and help a bit more tonight.

[twiecki]

This is probably fine as a hot fix, but technically this is not a pymc issue but an aesara one and that's where it should be fixed (e.g. by pinning the version there), rather than here.

[twiecki]

Just seeing it's already fixed there: aesara-devs/aesara#812 so shouldn't we just wait for a new aesara release and pin to that?

[twiecki]

If you pin aesara=2.4.0 we shouldn't need to pin anything else here.

[twiecki]

Ah, that's what you did in 498dc43. That actually looks right, not sure why numpy even gets compiled in the arviz installation?

[ricardoV94]

Just quick check to see if problems come from the recent change to mamba...

It wasn't that

[ricardoV94]

Okay installation seems to have worked now!

[ricardoV94]

Also weirdly... arviz compat and jax workflows are using the "dev" environments, and not the "test" ones... Why is that?

And some environments pin the pandas version but not others...

[ricardoV94]

Reverting the Aesara upgrade, because that fails due to #5446

[twiecki]

Also weirdly... arviz compat and jax workflows are using the "dev" environments, and not the "test" ones... Why is that?

It shouldn't, oversight.

[ricardoV94]

Ok, so scipy appears in the list of packages that should be excluded from requirements... not sure why this is:

pymc/scripts/generate_pip_deps_from_conda.py

Line 53 in b54c8af

"scipy",