Enforce custom initval in SMC

[tvwenger]

Description

Per the discussion here, it is apparent that SMC does not respect custom initval. This is a problem when a custom initval is required in order to support a transformation, such as Ordered.

Related Issue

• Closes BUG: SMC does not respect initval #7438

Checklist

• Checked that the pre-commit linting/style checks pass
• Included tests that prove the fix is effective or that the new feature works
• Added necessary documentation (docstrings and/or example notebooks)
• If you are a pro: each commit corresponds to a relevant logical change

Type of change

• New feature / enhancement
• Bug fix
• Documentation
• Maintenance
• Other (please specify):

---

📚 Documentation preview 📚: https://pymc--7439.org.readthedocs.build/en/7439/

[tvwenger]

@ricardoV94 thanks for the review! Before I revert the formatting changes (thanks VSCode!), I want to address the comment

Won't this override the default behavior of sampling from the prior, to using the moment?

I guess this is true, and we'll need a different strategy for this fix! If you have any ideas, please let me know.

[ricardoV94]

@ricardoV94 thanks for the review! Before I revert the formatting changes (thanks VSCode!), I want to address the comment

Won't this override the default behavior of sampling from the prior, to using the moment?

I guess this is true, and we'll need a different strategy for this fix! If you have any ideas, please let me know.

Ordered transform is an edge case that keeps posing problems all over the place. Much of this pain would be gone with: #7297

[ricardoV94]

What is model.rvs_to_initvals when those are not specified though? I'm not sure the default is moment, my comment was a question not a statement

[tvwenger]

What is model.rvs_to_initvals when those are not specified though? I'm not sure the default is moment, my comment was a question not a statement

Thanks for the clarification @ricardoV94 ! Indeed, the default value for model.rvs_to_initvals is None unless initval is set. Thus, sampling from the prior distribution will be respected unless initval is set. I've updated the documentation to make this explicit!

I've also reverted those pesky formatting changes.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 92.18%. Comparing base (8eaa9be) to head (439bffd).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #7439   +/-   ##
=======================================
  Coverage   92.18%   92.18%           
=======================================
  Files         103      103           
  Lines       17263    17263           
=======================================
  Hits        15914    15914           
  Misses       1349     1349           

Files	Coverage Δ
pymc/smc/kernels.py	98.16% <ø> (ø)

[aloctavodia]

Thanks @tvwenger!

[tvwenger]

@aloctavodia Actually, I think this "fix" causes other issues. When sampling a model with sample_smc and this change, I get errors like:

Traceback (most recent call last):
  File "/home/twenger/miniconda3/envs/bayes_spec-dev/lib/python3.12/concurrent/futures/process.py", line 263, in _process_worker
    r = call_item.fn(*call_item.args, **call_item.kwargs)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/twenger/science/python/pymc/pymc/smc/sampling.py", line 352, in _sample_smc_int
    smc.mutate()
  File "/home/twenger/science/python/pymc/pymc/smc/kernels.py", line 428, in mutate
    pearson_r = corr.get(self.tempered_posterior)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/twenger/science/python/pymc/pymc/smc/kernels.py", line 466, in get
    return np.abs(ab / (self.aa * bb))

Presumably, these errors are due to the fact that all of the starting positions are equal, thus there is no variance in the starting particle positions and the denominator of this pearson_r derivation is zero. Yikes!

@jessegrabowski had another idea here that I will try to implement...

[aloctavodia]

Ohh, I see. Thanks for keep trying to find a good solution

[ricardoV94]

@aloctavodia the PR should have been squashed+merge, since it had useless / badly titled commits.