Add profiling to dataloader next()

[akashkw]

What does this PR do?

Adds profiling to next(dataloader_iter) in train, eval, and prediction loops.

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

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?

Make sure you had fun coding 🙃

[carmocca]

Hi @akashkw!

We cannot profile next() like this as this would not cover prefetched batches or inter-batch paralelism.

I recently had a call with @daniellepintz where I went over possible solutions. Here's some code pointers:

The concrete lines to profile:
https://github.com/PyTorchLightning/pytorch-lightning/blob/a0655611de460f1659b46150cda256d2e3fa974e/pytorch_lightning/utilities/fetching.py#L267

https://github.com/PyTorchLightning/pytorch-lightning/blob/a0655611de460f1659b46150cda256d2e3fa974e/pytorch_lightning/utilities/fetching.py#L336

Option 1:
Pass a profiler reference to the data fetchers, maybe also the stage and dataloader idx for more fine-grained profiling, then use these directly inside the fetching functions
Option 2:
Inject the profiler usage through callables that could be part of these hooks:
https://github.com/PyTorchLightning/pytorch-lightning/blob/a0655611de460f1659b46150cda256d2e3fa974e/pytorch_lightning/utilities/fetching.py#L61-L65
with a pattern similar to the one used in this PR to inject logic around the optimizer.step()
4cc05b2

[akashkw]

Option 1:
Pass a profiler reference to the data fetchers, maybe also the stage and dataloader idx for more fine-grained profiling, then use these directly inside the fetching functions
Option 2:
Inject the profiler usage through callables that could be part of these hooks:

@carmocca do you have any preference between these two approaches?

[carmocca]

I prefer option (2) so the fetching does not need to keep a reference to the trainer. (the profiler is owned by the trainer)

[carmocca]

Good progress!

[carmocca]

I'm unsure if this is a known issue with the profiler on Windows or a bug...

[rohitgr7]

we should open an issue. It just uses time module. Nothing fancy.