Follow-up of MLflowCallback

Motivation

Thanks to the great improvement by https://github.com/optuna/optuna/pull/2670, optuna provides a smoother connection to MLFLow via MLflowCallback. But we still have a TODO list to improve the codebase and its feature mentioned in the review of https://github.com/optuna/optuna/pull/2670.

TODO tasks discussed in https://github.com/optuna/optuna/issues/2788#issuecomment-898890114 to close this issue

• General code/docstring fixes mentioned in #2670 TODO list (https://github.com/optuna/optuna/pull/2855)
• Introduction of multi-objective support (https://github.com/optuna/optuna/pull/2863)
• Refactoring interface to be more similar to WeightsAndBiasesCallback (+ decoupling study name from mlflow experiment) (https://github.com/optuna/optuna/pull/2912)

Please see also the following Description for more detailed contexts.

---

Description

Please see the discussion in the review. But we can separately work on this issue as follows:

• support multi-objective https://github.com/optuna/optuna/pull/2670#discussion_r633204507
• improve the docstring.
• improve the tag truncation https://github.com/optuna/optuna/pull/2670#discussion_r650478298

[nzw0301]

Hi @rafarui thank you for your feedback! Indeed, when we introduce WeightsAndBiasesCallback recently, it has wandb_kwargs that are used for the initialisation of wandb. So we can add a similar argument to MLflowCallback for a more flexible interface as you mentioned.

[rafarui]

If I can suggest one more point.

• Deacoplate the study name from mlflow experiment

Optuna is setting the experiment name for the user via the initialize_experiment(self, study: optuna.study.Study). I think we should let the user choose if he/she wants to create an experiment.

An experiment name is not necessarily the study name. I can have many studies inside the same experiment. If you are not using the in-memory database you will have problems. In my use case, studies are unique identified to be able to track them inside one mlflow experiment. Also, this can be a problem If you are using databricks. The experiment name is defined in databricks and not necessarily is the same as the study name. I can have many studies in one single experiment.

I am using this, so I might have a proposal soon.