Use shallow copy in _get_latest_trial

[not522]

Motivation

The deep copy in _get_latest_trial takes a long time when the parameters are many. This PR changes it to the shallow copy.

The returned FrozenTrial is passed to samplers and pruners, but they must not modify the trial object. This behavior is described in the base classes.

optuna/optuna/samplers/_base.py

Lines 136 to 138 in b00caef

	trial:
	Target trial object.
	Take a copy before modifying this object.

benchmark

import optuna

def objective(trial):
    return sum([trial.suggest_float(f"x_{i}", -100, 100) ** 2 for i in range(1000)])

study = optuna.create_study(sampler=optuna.samplers.RandomSampler(seed=42))
study.optimize(objective, n_trials=10)

• master

real	1m16.989s
user	1m15.750s
sys	0m0.710s

• PR

real	0m2.370s
user	0m2.277s
sys	0m0.307s

Description of the changes

• Use shallow copy in _get_latest_trial

[codecov-commenter]

Codecov Report

Merging #4774 (84d71ba) into master (b00caef) will not change coverage.
The diff coverage is 100.00%.

❗ Your organization is not using the GitHub App Integration. As a result you may experience degraded service beginning May 15th. Please install the Github App Integration for your organization. Read more.

@@           Coverage Diff           @@
##           master    #4774   +/-   ##
=======================================
  Coverage   90.81%   90.81%           
=======================================
  Files         189      189           
  Lines       14416    14416           
=======================================
  Hits        13092    13092           
  Misses       1324     1324           

Impacted Files	Coverage Δ
optuna/trial/_trial.py	95.67% <100.00%> (ø)

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[not522]

@toshihikoyanase @HideakiImamura Could you review this PR?

[HideakiImamura]

LGTM.

[toshihikoyanase]

I'm sorry for the delayed response.

The returned FrozenTrial is passed to samplers and pruners, but they must not modify the trial object. This behavior is described in the base classes.

Let me check the implementation of the built-in samplers/pruners.

[toshihikoyanase]

LGTM.

The trial objects returned by _get_latest_trial are used in the following methods:

• BaseSampler.sample_independent
• BasePruner._filter_study

sampler	check	comment
BoTorchSampler	OK	pass to independent sampler
PyCmaSampler	OK	pass to independent sampler
SkoptSampler	OK	pass to independent sampler
BruteForceSampler	OK	read trial.number
CmaEsSampler	OK	pass to independent sampler
GridSampler	OK	read trial.system_attrs
NSGAIIISampler	OK	pass to independent sampler
PartialFixedSampler	OK	pass to independent sampler
QMCSampler	OK	pass to independent sampler
RandomSampler	OK	not used
TPESampler	OK	pass to independent sampler. not used in _sample ?
NSGAIISampler	OK	pass to random sampler
DeterministicSampler	OK	not used
FirstTrialOnlyRandomSampler	OK	not used
(tests) FixedSampler	OK	not used
(tests) SamplerStubForTestRelativeParameters	OK	not used
(tutorial) SimulatedAnnealingSampler	OK	pass to independent sampler

pruner	check	comment
HyperbandPruner	OK	read trial.number in _get_bracket_id