Do not run the GC after every trial by default.

[hvy]

Motivation

Fixes #1307. See this issue for discussions.

Description of the changes

Changes the default GC collection behavior of Study.optimize to not run a collection.

[codecov-commenter]

Codecov Report

Merging #1380 into master will decrease coverage by 0.02%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #1380      +/-   ##
==========================================
- Coverage   88.08%   88.06%   -0.03%     
==========================================
  Files          97       97              
  Lines        7313     7297      -16     
==========================================
- Hits         6442     6426      -16     
  Misses        871      871              

Impacted Files	Coverage Δ
optuna/study.py	94.71% <100.00%> (-0.17%)	⬇️
optuna/_study_summary.py	100.00% <0.00%> (ø)

---

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 a652d04...43a4821. Read the comment docs.

[toshihikoyanase]

Thank you for your PR.
I think we need to set gc_after_trial=True when we use ChainerMNStudy (c.f., #325). I'm not sure if this issue affects the GitHub Actions, though.

[hvy]

Is it not addressed by https://github.com/optuna/optuna/blob/master/optuna/integration/chainermn.py#L146? I was also wondering whether this line should be kept or controllable similar to the standard Study.

[toshihikoyanase]

The line is executed by non-rank-0 nodes, but I think we need to call GC for rank-0 node. The rank-0 node calls Study.optimize without gc_after_trial option at https://github.com/optuna/optuna/blob/master/optuna/integration/chainermn.py#L122. And I think it may cause similar issues to #325.

this line should be kept or controllable similar to the standard Study.

Depending on the user code and environments, GC calls can be omitted. I support the latter approach, but I think we can work on it in a new PR.

[hvy]

Thanks for your explanation. I now see what you meant, that the behavior of ChainerMNStudy will be different for different workers, that only the first worker will not run the GC collection. However, as the underlying motivation behind this change is to leave the memory handling to the users, I suggest we stop running the GC in the ChainerMNStudy as well. Let me change it that way (discussed with @toshihikoyanase).

[hvy]

PTAL.

[toshihikoyanase]

Thank you for the fruitful discussion about ChainerMN. I understood the motivation of this PR better.

LGTM.

[HideakiImamura]

Thanks for the PR and the reasonable discussion. LGTM!