Replace scan in BetaGeoBetaBinom

[juanitorduz]

trafficstars

Closes https://github.com/pymc-labs/pymc-marketing/issues/703

---

📚 Documentation preview 📚: https://pymc-marketing--707.org.readthedocs.build/en/707/

[juanitorduz]

@ColtAllen Thanks for the nice issue description. I removed the pt.ge(t_x, i) from the unnorm_logp_died_at_tx_plus_i expression and this solved the scalar case, but something is missing in the vectorized one:

tests/clv/test_distributions.py::TestBetaGeoBetaBinom::test_logp_matches_lifetimes[scalar] PASSED                   [ 50%]
tests/clv/test_distributions.py::TestBetaGeoBetaBinom::test_logp_matches_lifetimes[vector] FAILED                   [100%]

with

E           AssertionError: 
E           Not equal to tolerance rtol=1e-07, atol=0
E           
E           Mismatched elements: 4 / 5 (80%)
E           Max absolute difference: 0.0505898
E           Max relative difference: 0.18415497
E            x: array([ -0.360687, -21.684232,  -6.955498,  -0.224123,  -0.360694])
E            y: array([ -0.36069 , -21.696809,  -6.9558  ,  -0.274713,  -0.360694])

Any ideas? The difference is very small ....

[codecov[bot]]

Codecov Report

Attention: Patch coverage is 91.05582% with 133 lines in your changes missing coverage. Please review.

Project coverage is 94.25%. Comparing base (c6be6a9) to head (00efd07). Report is 514 commits behind head on main.

Files with missing lines	Patch %	Lines
pymc_marketing/mmm/delayed_saturated_mmm.py	61.02%	106 Missing :warning:
pymc_marketing/mmm/components/base.py	94.38%	11 Missing :warning:
pymc_marketing/mmm/budget_optimizer.py	91.66%	4 Missing :warning:
pymc_marketing/model_builder.py	90.90%	4 Missing :warning:
pymc_marketing/prior.py	98.83%	3 Missing :warning:
pymc_marketing/mmm/base.py	98.13%	2 Missing :warning:
pymc_marketing/mmm/lift_test.py	97.87%	2 Missing :warning:
pymc_marketing/mmm/transformers.py	90.00%	1 Missing :warning:

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #707      +/-   ##
==========================================
+ Coverage   91.57%   94.25%   +2.68%     
==========================================
  Files          21       29       +8     
  Lines        2172     3101     +929     
==========================================
+ Hits         1989     2923     +934     
+ Misses        183      178       -5     

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

:rocket: New features to boost your workflow:

• :snowflake: Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ColtAllen]

CC'ing @ricardoV94 (he wrote this code).

[juanitorduz]

@ricardoV94 I think I got it and the tests are passing :) . Anything else I am missing?

[ricardoV94]

Changes look fine, although I would suggest a more careful benchmark before deciding on it.

Specifically other backends and cases that should be more favourable to Scan (large variation between T and tx across rows). If there's some reproducible code I don't mind taking a look.

I gave some more details on the original issue.

[ricardoV94]

And more generally when we do performance driven changes I would add a benchmark as part of the CI (pytest-benchmark is pretty nice). That's the conceptual equivalent to a regression test for bugfixes.

[juanitorduz]

Despite the changes, do what we want. We wanna do more benchmarking to decide on the implementation to choose. See https://github.com/pymc-labs/pymc-marketing/issues/703#issuecomment-2141495580