Add load rewriter

[yhna940]

This is for when the trial is restarted by the scheduler. In other cases, some results from previous training may be used selectively. e.g. pbt

[nijkah]

For mmsegmentation case, an argument for load-from checkpoints already exists.
https://github.com/SIAnalytics/mmtune/blob/4d64042c66fbb418e9dc5218f7bc7aead2bc292c/mmtune/mm/tasks/mmseg.py#L35-L36

Is there any reason to support it as Rewriter?
See related discussions open-mmlab/mmdetection#8144 (comment), open-mmlab/mmdetection#8189

[nijkah]

The parameter resume-from also exists. Can I ask you to explain the use case briefly?

[github-actions]

Coverage report

The coverage rate is 84.97%

The branch rate is 65.92%

80% of new lines are covered.

Diff Coverage details (click to unfold)

mmtune/mm/context/rewriters/init.py

100.0% of new lines are covered

mmtune/mm/context/rewriters/resume.py

100.0% of new lines are covered

mmtune/mm/hooks/checkpoint.py

55.55555555555556% of new lines are covered

[yhna940]

https://docs.ray.io/en/latest/tune/examples/includes/pbt_function.html

https://docs.ray.io/en/latest/train/user_guide.html#saving-checkpoints

This is an example of checkpointing.

[nijkah]

OK, I understand its purpose. But, there is still one thing left to worry about.
Is there any side effect possibility that some searches, which do not use checkpoint_dir, try to load the latest checkpoint?

[yhna940]

As far as I know there are no side effects. Search algorithm that does not use checkpoint_dir puts None in that place by default.