Fix current_epoch value on training end

[carmocca]

What does this PR do?

Fixes #5007, #4385, #11425
Part of #7406

The progress tracking state is now saved/loaded by default.

This PR uses two different epoch values:

We have the “representation epoch”, kept for backwards compatibility

• aka “fake epoch”
• aka trainer.current_epoch
• aka trainer.fit_loop.epoch_progress.current.completed
  Used for the checkpoint name and the naive epoch value saved in the checkpoint

And the “actual epoch”

• aka trainer.fit_loop.epoch_progress.current.processed
  Used to reload on restart.

Does your PR introduce any breaking changes? If yes, please list them.

• With this PR, a Trainer(max_epochs=1, limit_train_batches=1) which saves a checkpointon_train_epoch_end will have the current_epoch=0, global_step=1 values saved. This is because we consider on_train_epoch_end to be part of the epoch itself:

for epoch in epochs:
  ...
  trainer.fit_loop.epoch_progress.current.processed++
  on_train_epoch_end()
  current_epoch++ # (aka trainer.fit_loop.epoch_progress.current.completed++)
on_train_end()

• current_epoch is now increased by 1 on_train_end. This means that if a model is run for 3 epochs (0, 1, 2), trainer.current_epoch will return 3. This is consistent with the fact that a new trainer returns trainer.current_epoch == 0, as in, the 0-th (first) epoch needs to run. This will be breaking for anybody accessing the trainer.current_epoch attribute after fit is finished

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 list all the breaking changes introduced by this pull request?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is welcome to review the PR.
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?

🤪
this was painful

cc @awaelchli @ananthsub @ninginthecloud @rohitgr7 @Borda @carmocca @justusschock

[codecov]

Codecov Report

Merging #8578 (c26988e) into master (7914e49) will increase coverage by 0%.
The diff coverage is 0%.

@@          Coverage Diff           @@
##           master   #8578   +/-   ##
======================================
  Coverage      45%     45%           
======================================
  Files         218     218           
  Lines       14396   14393    -3     
======================================
  Hits         6450    6450           
+ Misses       7946    7943    -3     

[awaelchli]

I think one of the most important things to remember from this PR is the fact that on_train_epoch_end checkpoints are different from the regular checkpoint since they have a different "fake epoch" (current epoch) value saved.

[rohitgr7]

just one quick ques:

checkpoints saved during checkpointing... let's say during the last epoch, be it inside on_validation_end or on_train_epoch_end, the epoch values saved and reloaded will be different from what will be saved if trainer.save_checkpoint is called after training?

[carmocca]

Correct. There are several tests testing both cases.

[rohitgr7]

Correct. There are several tests testing both cases.

thanks for clarifying.. yeah I saw... Just wanted a final confirmation to keep this change in mind.

but global_step will be the same in both cases, right?

[carmocca]

Yes.

I'm looking at global step in #11805 (don't look inside yet!)

[tchaton]

@carmocca TODO: Remove checkpointing fault tolerant states.

[tchaton]

LGTM !