ELEX-4413 update dependencies and switch to using the Clarabel solver

[dmnapolitano]

Description

Hi! The changes in this PR update elex-solver's dependencies to their latest versions and silences a warning produced by cvxpy about the fact that we're still using the ECOS solver switches QuantileRegressionSolver's cvxpy solver from ECOS to Clarabel 🎉

Switching from ECOS to cvxpy's recommendation of CLARABEL is pretty easy and will probably result in absolutely no changes in terms of the predictions produced by our model, but IMO that's something we'd want to test pretty extensively before making the switch 😬 So I'm silencing the warning here per cvxpy's recommendation.

Jira Ticket

ELEX-4413

Test Steps

tox and/or pip install -e . and run the elexmodel and/or test bed commands of your choice 🎉

[lennybronner]

oh interesting, do you know why they are switching away from ecos?

[dmnapolitano]

oh interesting, do you know why they are switching away from ecos?

Here's what I found: https://www.cvxpy.org/updates/#ecos-deprecation

CVXPY has used ECOS as the default solver for many years; however, it has known issues with performance and numerical stability in edge cases. Recently, a new solver, Clarabel, that improves the algorithm and implementation of ECOS has been under development.

🤔

[lennybronner]

ok, so I think we should just go for it with clarabel 👀

[dmnapolitano]

ok, so I think we should just go for it with clarabel 👀

You know, I was just thinking the same thing 😂 🎉

[dmnapolitano]

ok, so I think we should just go for it with clarabel 👀

You know, I was just thinking the same thing 😂 🎉

@lennybronner switched! 🎉

[lennybronner]

did you run this locally with the model?

[dmnapolitano]

did you run this locally with the model?

Yup! 🎉 Although, I haven't done a run with the testbed yet, and I do have some results I could compare to...let me see 🤔

[lennybronner]

Also when you run this, just confirm that you are using the cvxpy solver. that only gets used when we run quantile regression with regularization.

[dmnapolitano]

Also when you run this, just confirm that you are using the cvxpy solver. that only gets used when we run quantile regression with regularization.

Oh, right! 🤦🏻‍♀️ Ok, in that case, here's something interesting. This command runs just fine:

elexmodel 2023-11-07_VA_G --estimands=margin --features=baseline_normalized_margin --office_id=Y --geographic_unit_type=county-district --pi_method bootstrap

Now if I want to test the regularization, I switch to pi_method=gaussian (with model_parameters='{"lambda" : whatever}', but):

elexmodel 2023-11-07_VA_G --estimands=margin --features=baseline_normalized_margin --office_id=Y --geographic_unit_type=county-district --pi_method gaussian

And this second command throws a ValueError: Array contains NaN or Infinity exception 🤔 🤔

[lennybronner]

right, margin only works with the bootstrap model

[dmnapolitano]

right, margin only works with the bootstrap model

OMG 🙈 Wait, why is that? 🤔

[lennybronner]

because margin necessitates a whole bunch of math that we never implemented for the other models

[lennybronner]

specifically, margin needs moving from normalized to unnormalized margin space a few times that we haven't implemented in the other models (to go from unit margins to aggregate margins for example). But specifically, this error suggests that a unit is coming through to the model that shouldn't be (we should be converting all margins from nan to zero). might be worth investigating.

[dmnapolitano]

specifically, margin needs moving from normalized to unnormalized margin space a few times that we haven't implemented in the other models (to go from unit margins to aggregate margins for example). But specifically, this error suggests that a unit is coming through to the model that shouldn't be (we should be converting all margins from nan to zero). might be worth investigating.

Oh wow! Yeah, that does sound possible. At some point in the post-GE future lol 🤔

Ok, in that case...how much testing is worth doing here? I can (now) get this to run fine:

elexmodel 2023-11-07_VA_G --estimands=dem --features=median_household_income --office_id=Y --geographic_unit_type=county-district --pi_method gaussian

Adding in regularization works but makes no difference in the output 🤔 Should I run some testbed commands with the current elex-solver vs. this with regularization, just to see? 🤔

[lennybronner]

nope, this is good enough!