Honegumi Fully Bayesian generated code may be deprecated

[WinC3]

Honegumi generates code importing classes from deprecated modules. Honegumi tries to import from ax.modelbridge, which has been renamed to adapter (see recent ax commits on GitHub). Models has also been renamed to Generators.

Honegumi may need a quick update to reflect this change with ax as generated code with the Fully Bayesian parameter no longer runs without error fresh from Honegumi.

[sgbaird]

Thanks for catching. I think I need to pin to ax-platform<0.5, or at least to ax-platform<1 for the assignments. Could you try this?

[zweaung1014]

Also having the same problem! The following line that Honegumi spits out doesn't work:
from ax.modelbridge.factory import Models

Downgrading to 0.4.0 works.

[sgbaird]

0.4.3 is likely the latest compatible for this one. Which assignments is this affecting?

[WinC3]

It directly affects 2.3-2.6 as these are the assignments that feature the need for setting generation strategies, which require the import of Models. Indirectly, having the latest version of ax-platform may mean some tasks in assignments fail due to (likely) differences in function outputs. In some assignments, the optimal parameters and values are unable to be found (using the provided seeding of 42) in the given amount of trials. Especially in 2.5 multi-task optimization, I was only able to find the optimal values and parameters to pass the corresponding pytest after fiddling with the set seeding.

I have been using the latest version so far but should I revert to older versions going forward?

[sgbaird]

Probably using 0.4.3 (or at least <1) would be best for the assignments. If you make changes, I can have a look and incorporate some of those upstream.

[sgbaird]

In terms of the assignments you've already done, no worries if you used 1.0 (actually, I'd be more interested to see what you've done there since I'm planning to learn/migrate to 1.0)

[sgbaird]

Updated Honegumi so it gives instructions for %pip install ax-platform==0.4.3

While I'm planning to migrate to 1.0, there's quite a bit of validation I need to do in terms of performance and cross-compatibility of various characteristics (e.g., multi-objective + multi-task).