Handle `-inf` and `inf` values in `RDBStorage` by xadrianzetx · Pull Request #3238 · optuna/optuna

xadrianzetx · 2022-01-18T19:06:18Z

Motivation

Previously an error could be raised when -inf or inf was registered in rdb storage. Depending on SQL dialect in backend, those special values might be supported or not. This PR introduces a simple way to get around this limitation by storing -inf and inf as minimum and maximum values of 32 bit float and providing thin translation layer between stored and true values on database reads and writes. This gives effective available range for trial values of (-3.4028235e+38, 3.4028235e+38). Everything beyond that is interpreted as infinity. Closes #3206.

Description of the changes

Implement inf handling in RDBStorage
Test

This patch ensures that `RDBBackend` can fully support MySQL specification by representing infinite values as floats and limiting signed floats to 32 bit

xadrianzetx · 2022-01-18T19:18:03Z

Note - this is a simple way of fixing it without a need to modify data models. The downside is that when translating value from storage back to actual value, we can't directly compare two floats. This means some trial values that are close enough to numerical limits might be translated back as infinite. But still, this is for values close to 2e+38 so in reality this is probably a non issue.

Alternative solution is to modify TrialValueModel and TrialIntermediateValueModel to include an integer column which indicates if value is infinite. No need for approximations then. This would also give us a chance to specify double precision for Float in those two models, meaning reported value could be in (-1.7976931348623157e+308, 1.7976931348623157e+308) range and would go in line with python float range (which is C type double in cpython).

toshihikoyanase · 2022-01-19T01:29:19Z

@nzw0301 @HideakiImamura Could you review this PR, please?

nzw0301 · 2022-01-19T09:17:31Z

Sure. I'll review this PR within the next week.

HideakiImamura

Thanks for the PR. Basically, LGTM. Let me add a comment for the test.

codecov-commenter · 2022-01-21T05:12:52Z

Codecov Report

Merging #3238 (cf8a45c) into master (acdf4b7) will increase coverage by 0.06%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #3238      +/-   ##
==========================================
+ Coverage   91.51%   91.58%   +0.06%     
==========================================
  Files         146      145       -1     
  Lines       12011    12011              
==========================================
+ Hits        10992    11000       +8     
+ Misses       1019     1011       -8

Impacted Files	Coverage Δ
optuna/storages/_rdb/storage.py	`92.94% <100.00%> (+0.12%)`	⬆️
optuna/integration/chainermn.py	`95.15% <0.00%> (-0.03%)`	⬇️
optuna/study/_optimize.py	`98.31% <0.00%> (-0.02%)`	⬇️
optuna/multi_objective/samplers/_nsga2.py	`97.14% <0.00%> (-0.02%)`	⬇️
optuna/samplers/_nsga2/sampler.py	`96.61% <0.00%> (-0.02%)`	⬇️
optuna/samplers/_nsga2/crossover.py	`97.83% <0.00%> (-0.01%)`	⬇️
optuna/__init__.py	`100.00% <0.00%> (ø)`
optuna/_callbacks.py	`100.00% <0.00%> (ø)`
optuna/pruners/_successive_halving.py	`100.00% <0.00%> (ø)`
... and 2 more

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update acdf4b7...cf8a45c. Read the comment docs.

HideakiImamura

Thanks for the update! LGTM.

HideakiImamura · 2022-01-28T03:57:23Z

I confirmed that all problems in #3206 are resolved by this PR.

(venv) mamu@HideakinoMacBook-puro 3205-3206 % python inf-objective.py
[I 2022-01-28 12:56:37,375] A new study created in memory with name: no-name-ff107e9e-301c-460b-b9cc-81df43d3968d
[I 2022-01-28 12:56:37,376] Trial 0 finished with value: inf and parameters: {}. Best is trial 0 with value: inf.
In-memory : Trial state is TrialState.COMPLETE
[I 2022-01-28 12:56:37,441] A new study created in RDB with name: no-name-969966ff-fb18-494a-932c-ecddc8a4017b
[I 2022-01-28 12:56:37,502] Trial 0 finished with value: inf and parameters: {}. Best is trial 0 with value: inf.
SQLite    : Trial state is TrialState.COMPLETE
[I 2022-01-28 12:56:37,620] A new study created in RDB with name: no-name-07b13b5f-85bc-46a2-ba72-8b95dbb7eb5f
[I 2022-01-28 12:56:37,763] Trial 0 finished with value: inf and parameters: {}. Best is trial 0 with value: inf.
MySQL     : Trial state is TrialState.COMPLETE
[I 2022-01-28 12:56:37,903] A new study created in RDB with name: no-name-90b645ef-80ff-4730-a1e0-01e2c51011c5
[I 2022-01-28 12:56:38,012] Trial 0 finished with value: inf and parameters: {}. Best is trial 0 with value: inf.
PostgreSQL: Trial state is TrialState.COMPLETE
(venv) mamu@HideakinoMacBook-puro 3205-3206 % python inf-report.py   
[I 2022-01-28 12:56:41,181] A new study created in memory with name: no-name-d9311317-b070-46ea-a22d-39c280b29046
[I 2022-01-28 12:56:41,182] Trial 0 finished with value: 1.0 and parameters: {}. Best is trial 0 with value: 1.0.
In-memory : Trial state is TrialState.COMPLETE
[I 2022-01-28 12:56:41,247] A new study created in RDB with name: no-name-13547701-5097-4ca9-b34a-a7dc1810e732
[I 2022-01-28 12:56:41,315] Trial 0 finished with value: 1.0 and parameters: {}. Best is trial 0 with value: 1.0.
SQLite    : Trial state is TrialState.COMPLETE
[I 2022-01-28 12:56:41,435] A new study created in RDB with name: no-name-c383b025-e13b-4122-9f23-0e1c2b4680ae
[I 2022-01-28 12:56:41,575] Trial 0 finished with value: 1.0 and parameters: {}. Best is trial 0 with value: 1.0.
MySQL     : Trial state is TrialState.COMPLETE
[I 2022-01-28 12:56:41,712] A new study created in RDB with name: no-name-21630f5e-9773-4e9a-bbf9-8011f1c81a53
[I 2022-01-28 12:56:41,829] Trial 0 finished with value: 1.0 and parameters: {}. Best is trial 0 with value: 1.0.
PostgreSQL: Trial state is TrialState.COMPLETE
(venv) mamu@HideakinoMacBook-puro 3205-3206 % python inf-tell.py 
[I 2022-01-28 12:56:47,353] A new study created in memory with name: no-name-66267843-18bf-4c88-8701-2643a782c1d6
In-memory : Trial state is TrialState.COMPLETE
[I 2022-01-28 12:56:47,420] A new study created in RDB with name: no-name-b8cd0fce-6807-4174-8606-b352d214b19d
SQLite    : Trial state is TrialState.COMPLETE
[I 2022-01-28 12:56:47,583] A new study created in RDB with name: no-name-db63b7aa-e06f-4d1d-b67c-1bdc7abe8268
MySQL     : Trial state is TrialState.COMPLETE
[I 2022-01-28 12:56:47,806] A new study created in RDB with name: no-name-121048c0-3d33-408b-b56f-103e83c06e06
PostgreSQL: Trial state is TrialState.COMPLETE

nzw0301

Thank you for your PR! LGTM.

I locally confirmed the code works with

a single objective (inf) with intermediate values (inf)
two objective values (inf, 1.0).

I'm wondering if we could add the explanation of this transformation to Optuna documentation, where storage or study. But it can be addressed as a follow-up.

So I'll merge this PR once.

Handle -inf and 'inf' trial values in RDBStorage

0e28207

This patch ensures that `RDBBackend` can fully support MySQL specification by representing infinite values as floats and limiting signed floats to 32 bit

xadrianzetx changed the title ~~Handle -inf and 'inf' values in RDBStorage~~ Handle -inf and inf values in RDBStorage Jan 18, 2022

github-actions Bot added the optuna.storages Related to the `optuna.storages` submodule. This is automatically labeled by github-actions. label Jan 18, 2022

toshihikoyanase assigned nzw0301 and HideakiImamura Jan 19, 2022

HideakiImamura reviewed Jan 20, 2022

View reviewed changes

Comment thread tests/storages_tests/rdb_tests/test_with_server.py

Add test cases for 64 bit floats

cf8a45c

HideakiImamura approved these changes Jan 24, 2022

View reviewed changes

nzw0301 approved these changes Jan 30, 2022

View reviewed changes

nzw0301 merged commit db71098 into optuna:master Jan 30, 2022

nzw0301 added this to the v3.0.0-a1 milestone Jan 30, 2022

masap mentioned this pull request Jan 30, 2022

Add integration test for RedisStorage #3258

Merged

xadrianzetx deleted the fix-rdb-handle-inf branch January 30, 2022 08:37

HideakiImamura mentioned this pull request Feb 3, 2022

Skip limiting the value if it is nan #3286

Merged

xadrianzetx mentioned this pull request May 7, 2022

Set precision of sqlalchemy.Float in RDBStorage table definition #3327

Merged

3 tasks

This was referenced May 16, 2022

Add intermediate_value_type column to represent inf/-inf on RDBStorage. #3564

Merged

inf handling in Optuna #3206

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Handle `-inf` and `inf` values in `RDBStorage`#3238

Handle `-inf` and `inf` values in `RDBStorage`#3238
nzw0301 merged 2 commits intooptuna:masterfrom
xadrianzetx:fix-rdb-handle-inf

xadrianzetx commented Jan 18, 2022 •

edited

Loading

Uh oh!

xadrianzetx commented Jan 18, 2022 •

edited

Loading

Uh oh!

toshihikoyanase commented Jan 19, 2022

Uh oh!

nzw0301 commented Jan 19, 2022

Uh oh!

HideakiImamura left a comment

Uh oh!

Uh oh!

codecov-commenter commented Jan 21, 2022

Uh oh!

HideakiImamura left a comment

Uh oh!

HideakiImamura commented Jan 28, 2022

Uh oh!

nzw0301 left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Uh oh!

Conversation

xadrianzetx commented Jan 18, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Description of the changes

Uh oh!

xadrianzetx commented Jan 18, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

toshihikoyanase commented Jan 19, 2022

Uh oh!

nzw0301 commented Jan 19, 2022

Uh oh!

HideakiImamura left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

codecov-commenter commented Jan 21, 2022

Codecov Report

Uh oh!

HideakiImamura left a comment

Choose a reason for hiding this comment

Uh oh!

HideakiImamura commented Jan 28, 2022

Uh oh!

nzw0301 left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

xadrianzetx commented Jan 18, 2022 •

edited

Loading

xadrianzetx commented Jan 18, 2022 •

edited

Loading