progressive restoring of trainer state

[awaelchli]

What does this PR do?

Fixes #7535

Includes the changes from #7955, target branch will be changed to master.

The restore flow before:

call to trainer.fit()

model.setup("fit")  # trainer calls setup
call_configure_sharded_model(model)
accelerator.setup(model) 
...

dispatch() #  _run_train, etc.
restore() # restore everything at once (model, callbacks, optimizers, loop state etc.)

The restore flow now:

call to Trainer.fit()

# check if/what we needs to resume, load the checkpoint file
resume_start() 

model.setup("fit")  # trainer calls setup

# model weights get restored as soon as model is setup
restore_datamodule() # calls datamodule.on_load_checkpoint()
restore_model() # calls model.on_load_checkpoint()
restore_callbacks() # calls callback.on_load_checkpoint()

call_configure_sharded_model(model)
accelerator.setup(model) 

pre_dispatch()

# restore training state as soon as everything
restore_training_state()  # restore optimizer, precision, loop progress

dispatch() #  _run_train, etc.

resume_end() # delete checkpoint file in memory

Before submitting

• Was this discussed/approved via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is free to review the PR once the tests have passed.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

Did you have fun?

Make sure you had fun coding 🙃

Follow up to this PR

Make plugin decide if we want to restore model weights before or after call to configure_sharded_model hook as per #7535 (comment)

if TrainingTypePlugin.restore_before_configure_sharded_model:
    restore_model() # also calls model.on_load_checkpoint()

call_configure_sharded_model(model)

if not TrainingTypePlugin.restore_before_configure_sharded_model:
    restore_model() # also calls model.on_load_checkpoint()

[pep8speaks]

Hello @awaelchli! Thanks for updating this PR.

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2021-06-17 07:49:49 UTC

[codecov]

Codecov Report

Merging #7652 (02c3d54) into master (bc2c2db) will decrease coverage by 0%.
The diff coverage is 100%.

@@           Coverage Diff           @@
##           master   #7652    +/-   ##
=======================================
- Coverage      92%     91%    -0%     
=======================================
  Files         207     207            
  Lines       13374   13485   +111     
=======================================
+ Hits        12246   12317    +71     
- Misses       1128    1168    +40     

[tchaton]

LGTM !