Upgrading Optuna from v2.x.x to v3.0.0

[himkt]

Sorry but it is not a feature request.

Related: optuna/optuna#2941

Summary

Thank you for integrating Optuna, hyperparameter optimization library. We, Optuna dev-team, are working on the next major version of the library Optuna v3. In Optuna v3, we will deprecate distributions used in:

hydra/plugins/hydra_optuna_sweeper/hydra_plugins/hydra_optuna_sweeper/_impl.py

Lines 33 to 42 in 61ab29a

	from optuna.distributions import (
	BaseDistribution,
	CategoricalChoiceType,
	CategoricalDistribution,
	DiscreteUniformDistribution,
	IntLogUniformDistribution,
	IntUniformDistribution,
	LogUniformDistribution,
	UniformDistribution,
	)

We will introduce FloatDistribution and IntDistribution. So distributions in hydra_optuna_sweeper will have to be updated:

• UniformDistribution(low, high) => FloatDistribution(low, high)
• DiscreteUniformDistribution(low, high, q) => FloatDistribution(low, high, step)
• LogUniformDistribution(low, high) => FloatDistribution(low, high, log=True)
• IntUniformDistribution(low, high) => IntDistribution(low, high)
• IntUniformDistribution(low, high, step) => IntDistribution(low, high, step)
• IntLogUniformDistribution(low, high) => IntDistribution(low, high, log=True)

Now hydra optuna plugin installs the latest Optuna. Users will see the deprecation warnings in the Optuna v3 release. #2163 restricts the constraint to avoid such warnings. Note that it doesn't mean Optuna v3 is imcompatible with Optuna v2. It only shows warnings and deprecated distributions are internally converted to appropriate distributions.

root@b9ffe3ebef50:/work/plugins/hydra_optuna_sweeper# python example/sphere.py --multirun
/work/plugins/hydra_optuna_sweeper/hydra_plugins/hydra_optuna_sweeper/_impl.py:109: FutureWarning: DiscreteUniformDistribution has been deprecated in v3.0.0. This feature will be removed in v6.0.0. See https://github.com/optuna/optuna/releases/tag/v3.0.0. Use :class:`~optuna.distributions.FloatDistribution` instead.
  return DiscreteUniformDistribution(value.start, value.stop, value.step)
[I 2022-04-20 06:13:41,041] A new study created in memory with name: sphere
[2022-04-20 06:13:41,042][HYDRA] Study name: sphere
[2022-04-20 06:13:41,043][HYDRA] Storage: None
[2022-04-20 06:13:41,043][HYDRA] Sampler: TPESampler
[2022-04-20 06:13:41,043][HYDRA] Directions: ['minimize']

Motivation

I open this issue to let hydra developers know about our next major release.
And also, I will open a PR to add a version constraint to avoid installing Optuna v3.

We will back here when the release is available.

If you have some concerns or questions, please feel free to contact Optuna dev-team.
🔗 optuna/optuna#2941

Additional context

Add any other context or screenshots about the feature request here.

[Jasha10]

Thanks @himkt! I'm labeling this issue as "blocked" until Optuna 3.0 is released.

[atgilles]

Any update on this request, as optuna v3 has been released?

[himkt]

So sorry for that, and thank you for reminding. We will proceed that.

[atgilles]

Thank you

[himkt]

I just asked @keisuke-umezawa to get #2360 back to active.

[ridvan-eksi]

Optuna is now at version 3.6.1, but optuna_sweeper dependency is still at 2.10.0. Is there any plan to update this requirement?

[odelalleau]

Optuna is now at version 3.6.1, but optuna_sweeper dependency is still at 2.10.0. Is there any plan to update this requirement?

I'm afraid some third-party plugins like this one currently lack proper support -- ideally someone who uses it would need to step up and help maintain it.