Loop Refactor 1/N - Training Loop

[awaelchli]

What does this PR do?

Part of #7700

Introduces the new loop interface.
Old classes will be removed asap in a follow up.

co-author: @justusschock

How to review this PR

1. Review the new loop base class found in pytorch_lightning/loops/base.py
2. Review the new training loop classes
   • EpochLoop: runs over epochs
   • TrainingLoop: runs over all batches in a single epoch
   • BatchLoop: runs over a single batch (could include looping over several optimizers or truncated backprop steps).
3. Review changes in the Trainer class file. Here we have some re-routing from old to new. This will be cleaned up

The new training loop classes are re-organizing the old TrainLoop class.
Important: Changes requested on code lines that are 1-1 from the old TrainLoop class will be dismissed. We cannot address these comments here. More refactor iterations will be required to simplify all existing helper functions.

TODO:

• typing (@justusschock wants to do it)
• docstrings (@justusschock wants to do it)
• [x[ get all tests to pass

More follow-up work is listed in this issue tracker: #7938

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 🙃

Fixes #7034

[kaushikb11]

What a kickass PR!

[tchaton]

Hey @awaelchli,

It is not a blocker for this PR, but there is a small inconsistency in the loop right now.
"on_train_epoch_start" is being called in FitLoop `on_advance_start
and "on_train_epoch_end" is being called in TrainellopLoop in "on_run_end".

Best,
T.C

[awaelchli]

great observation @tchaton, moved to on_run_start!

[carmocca]

@awaelchli do you remember why this was necessary?

edit: checking here #9298

[awaelchli]

there is a good reason, something along those lines of my other comment. pretty sure without this a test will fail. I would not have added it otherwise. But I should have added a comment. De-cyphering handling of these edge cases was definitely a center point of frustration for these loop refactors.

[awaelchli]

pretty sure this: https://pytorch-lightning.readthedocs.io/en/latest/common/lightning_module.html?highlight=skip%20epoch#on-train-batch-start

returning -1 here skipping the epoch

[carmocca]

@awaelchli do you remember why this was necessary?

edit: checking here #9298

[awaelchli]

yes, basically if the epoch is empty we don't want to run epoch_end hooks or invoke any logging. when does this happen? a user can stop training or skip an epoch by returning a signal from one of the hooks. I don't remember the exact use case. Pretty sure if you remove this line a test will fail and inform you exactly the reason.