build(autoconf): Managing library dependencies in custom experiments. Use-case: Autogluon models

[danielelotito]

Managing library dependencies in custom experiments. Use-case: Autogluon models

Some artifacts used in custom experiments may be strictly tied to specific library versions. The user may face the case in which to use different artifacts might install and uninstall specific libraries. We need to think about how to address this scenario, e.g. spinning a dynamic environment with ray.

Immediate use-case

Autogluon 1.5 has been published. Release notes: https://auto.gluon.ai/stable/whats_new/v1.5.0.html

Models trained with the old version will not be compatible with the new one. Citing the release notes:

Loading models trained on older versions of AutoGluon is not supported. Please re-train models using AutoGluon 1.5.0.

We have in plugins/custom_experiments/autoconf/autoconf/AutoGluonModels v1.4 models. Since we may face this situation again not in the immediate future and no one consumes the models atm except for ourselves (is this correct?). I would opt for the patch.

My proposed solution is:

• Producing using AutoGluon 1.5.0 a v3 model and put it in the folder,
• change a line in plugins/custom_experiments/autoconf/pyproject.toml to "autogluon.tabular[catboost,xgboost]==1.5.0"
• Update plugins/custom_experiments/autoconf/autoconf/AutoGluonModels/README.md and plugins/custom_experiments/autoconf/autoconf/AutoGluonModels/changelog.md accordingly

Optionally we could also:

• Delete old models to enhance clarity in the repo.

[danielelotito]

I have ticked the three boxes and marked the PR #338 ready for review.
@srikumar003