用Go编写的黑盒函数的贝叶斯优化框架

CHANGES

Support a new sampler based on Sobol's Quasi-random Sequence

Goptuna now supports Sobol based sampler. See the paper S. Joe and F. Y. Kuo, Remark on Algorithm 659: Implementing Sobol's quasirandom sequence generator, ACM Trans. Math. Softw. 29, 49-57 (2003). for details.

np.random.rand() (seed=0)	Sobol (Goptuna)

Usage is like this:

import (
	"github.com/c-bata/goptuna"
	"github.com/c-bata/goptuna/sobol"
)

func main() {
	study, err := goptuna.CreateStudy(
		"goptuna-example",
		goptuna.StudyOptionRelativeSampler(sobol.NewSampler()),  // Set Sobol based sampler.
	)
	...
}

Add github.com/c-bata/gorm/rdb.v2

Add rdb.v2 package built on gorm v2.
So github.com/c-bata/goptuna/rdb package is now deprecated.

• Use gorm v2 package (#124).
• Fix duplicate trial numbers are assigned on parallel execution (#130).
• Add CASCADE ON DELETE constraints (#170).
• Drop schema compatibility with Optuna (#166).

Dashboard

• Add datetime_start column on top page of Dashboard (#169).
• Put select box to change the reload interval time on AppBar (#176).

Minor compatibility changes

• Include the upper bound of uniform/loguniform distributions (#205)

New features

Dashboard: filtering trials by state

You can now walk through trials by filtering and sorting.

Dashboard: Add button to delete a study

You can delete a study from dashboard.

New feature

Built-in Web Dashboard (#157)

You can download standalone binaries (JavaScript files are also embedded in these binaries) via GitHub release, then you can open Web dashboard like following command.

$ goptuna --version
goptuna version 0.7.0 (rev: e8db5dc)
$ goptuna dashboard --storage sqlite:///example.db --host 127.0.0.1 --port 8000

Goptuna's RDB storage backend is compatible with Optuna. So this CLI is able to visualize Optuna's optimization results.

Improvements

• Faster TPE categorical sampling (#152)

Bug fixes

• Fix log message of pruned trials (#150)
• Calculate logLikeLihoodsAbove from samplesBelow (#151)

New feature

• Support IPOP-CMA-ES and BI-POP-CMA-ES algorithm (#131, #135, #136, #141)
  • Please check the benchmark results on two-dimensional Rastrigin function at #141.
  • Auger, A., Hansen, N.: A restart CMA evolution strategy with increasing population size. In: Proceedings of the 2005 IEEE Congress on Evolutionary Computation (CEC’2005), pp. 1769–1776 (2005a)
  • Hansen N. Benchmarking a BI-Population CMA-ES on the BBOB-2009 Function Testbed. In the workshop Proceedings of the Genetic and Evolutionary Computation Conference, GECCO, pages 2389–2395. ACM, 2009.
• Add BlackHoleStorage towards 100k+ evaluations (#105)
• Add TPE sampler options (#127)

API changes

Pruning (Early-stopping) APIs are changed without backward compatibility.

While I added an example which uses successive halving pruner at #120, I found a lot of issues. I'm sure that nobody uses pruning functionalities (I also confirmed that Kubeflow/Katib and sile/kurobako-go does not use pruning functionalities). So I fixes those issues. And I merge it without backward compatibility.

• Fix typo of OptionSetReductionFactor
• Alter the argument of trial.ShouldPrune() and remove trial.Report().

Some APIs are renamed without breaking changes.

These changes are added with backward compatibilities. Older functions are deprecated.

• Rename RandomSearchSampler to RandomSampler (#137).
• Rename StudyOptionSetDirection to StudyOptionDirection (#138).
• Rename StudyOptionSetTrialNotifyChannel to StudyOptionTrialNotifyChannel (#138).

Minor bug fixes

• Fix IntUniformDistribution.ToExternalRepr() (#102)
• Fix IntUniformDistributions.Contains() (#103)
• Fix a bug when set attrs multiple times (#121)
• CMA-ES: fix numerical overflow errors (#126)
• Fix a bug of successive halving pruner (#120)

Example

• Add an example of Gorgonia hyperparameter optimization with early stopping (ASHA) (#118, #119 and #120)

In this release, some internal methods made public for Kubeflow/Katib.

New feature

• Add study option to define search space (#99)

For Developers

• Public API to call relative sampler for Kubeflow/Katib (#100)

Others

• Generate stringer code for trial state (#101)

This release supports CMA-ES sampler and compatible with the RDB schema of Optuna v1.3.0.

New feature

• CMA-ES sampler (#93)

Internal chages

• Make StudyAttrs/UserAttrs more readable (#89)
• Add number field in trials table (#88)

Bug fixes

• Fix concurrency problems of EnqueuTrial (#92)

Compatibility with Optuna RDB storage

This release has compatibility with Optuna v1.3.0.

This release supports enqueue_trial and has compatibility with the RDB schema of Optuna v1.2.0.

New feature

• Experimental support of EnqueueTrial (#86)

Bug fixes

• Fix trial number on RDB storage (#87)

Compatibility with Optuna RDB storage

This release has compatibility with Optuna v1.2.0.

This release contains ASHA support, experimental GP-BO support and a serious bug fix of the TPE optimization on DiscreteUniformDistribution.

New feature

• Published an integration sampler using go-bayesopt ( https://github.com/c-bata/goptuna-bayesopt ).
• Support an Asynchronous Successive Halving Algorithm (#52, #62)
• Add RelativeSampler interface for Gaussian Process and CMA-ES (#75, #76)
• Add a subcommand to delete study (#56)
• Add a function to delete study (#57)
• Add study option for loading if exists (#61)

API changes

• Remove ID suffix from storage.CreateNew{Study,Trial}ID methods (#53).
• Avoid to update value when set intermediate values (#84)

Bug fixes

• Set a seed number to the inner random sampler of TPE sampler (#66)
• Fix an optimization of DiscreteUniform on TPE sampler (#79)

For Developers

• Use Go v1.13 on GitHub Actions (#59)
• Parse SQLAlchemy engine format using regexp (#67, #68, #69)
• Add kurobako-go benchmark for TPE sampler (#71)

Compatibility with Optuna RDB storage

This release has the compatibility with Optuna v1.1.0 or lower.

New feature

• Add logger interface (#46)
• Add validation for trial.SuggestXXX methods (#41)

Bug fixes

• Fix inconsistent state of trials (#50)
• Add transaction for creating trial (#36)
• Fix 'Too many SQL variables' error by separate queries to get all trials (#33)
• Encode attr values to the format which is loadable from Python's json module. (#35)

API changes

• New pruner interface (#47)
• Remove FrozenTrial.ParamsInIR field (#20)

New feature

• Support Optuna compatible RDB storage backend (#18)
• Add missing SuggestLogUniform and SuggestDiscreteUniform methods in Trial (#23)
• Provide SetUserAttr, GetUserAttrs, SetSystemAttr and GetSystemAttrs methods in Trial and Study (#24)
• Provide GetContext method in Trial for handling signals (#29)

Others

• Print more detail error logs (#28).

New feature

• Add MedianPruner and PercentilePruner which implements median stopping rule of Vizier (Google) (#11)
• Support DiscreteUniformDistribution (#14)
• Support LogUniformDistribution (#16)

Internal change

• Make TPE 30% faster (#9)

New feature

• Accept a go's context via study.WithContext (#4)
• Support CategoricalDistribution (#8)

Internal change

• Redesign the distribution interface. (#6)
• Add about 20 methods in storage interface and InMemoryStorage. (#7)

For developers

• Improve test coverages.
• Fix all golint errors.

ʕ◔ϖ◔ʔ Initial release 🎉

• Design components of goptuna
  • Study: optimization based on an objective function
  • Trial: a single execution of the objective function
  • Storage: store the result of each trials
  • Distribution: a representation of the parameter search space
• Sampler
  • Tree of Parzen Estimators (TPE)
  • Random Search
• Distribution
  • UniformDistribution: a uniform distribution in the linear domain
  • IntUniformDistribution: a uniform distribution on integers
• Add original notification system of finished trials.
• Add examples
  • Simple example using TPE
  • Simple example of using Random Search
  • Using notification of each trials
  • Using multiple goroutine workers