[2/2] Remove outputs from evaluation epoch end hooks

[ananthsub]

What does this PR do?

This addresses part of #6865 and undoes #6120.

This is inclusive of changes in #7339 .

Because outputs as part of the evaluation epoch end hooks were never part of a stable release, we can safely change the signature and remove the arguments without worrying about backwards compatibility or breaking changes.

We do so because we no longer have a mechanism to not store step outputs during training if any callback or module needs to run something at the end of the epoch which doesn't rely on the outputs. Traditionally, this was the differentiator between validation_epoch_end defined as a method on the LightningModule vs on_validation_epoch_end which takes no arguments and is available in both the callbacks and lightning module via hooks.

As a result, we see increased memory pressure which both slows down training and puts us at greater risk of OOMs, especially for large scale training.

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 🙃

[codecov]

Codecov Report

Merging #7338 (c489e8b) into master (6104a63) will decrease coverage by 0%.
The diff coverage is 100%.

@@          Coverage Diff           @@
##           master   #7338   +/-   ##
======================================
- Coverage      88%     87%   -0%     
======================================
  Files         200     200           
  Lines       12985   12982    -3     
======================================
- Hits        11400   11357   -43     
- Misses       1585    1625   +40     

[pep8speaks]

Hello @ananthsub! Thanks for updating this PR.

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

Comment last updated at 2021-05-05 15:29:06 UTC

[maximsch2]

If you are doing deprecation anyway, would it make sense to also consolidate naming? Right now we have on_train_epoch_end, on_validation_epoch_end, on_test_epoch_end
training_step, validation_step, test_step

Would be nice if *_step and _epoch_end would match up?

[ananthsub]

@maximsch2 I'd prefer to tackle naming separately: #6094 lays out a number of inconsistencies which we need to address

[edgarriba]

LGTM. Just fix failing test

[ananthsub]

@awaelchli the faiilng test is related to the worker init fn check. I'm seeing if this was fixed in pytorch 1.9

[awaelchli]

we added 1.9 just today. I'm sure it's not related to your PR.
Will take a look, I have 1.9 installed somewhere

[Borda]

@ananthsub shall we add the docs about cache here?

[ananthsub]

@ananthsub shall we add the docs about cache here?

Sent out #7389 for this