Stepwise LR scheduler

[01AbhiSingh]

What does this PR do?

Fixes #<17544>

Hii @awaelchli. Can you please verify the changes I made. If they are correct then i will take up and correct any failing tests also.

Before submitting

• Was this discussed/agreed 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 minor internal changes/refactors)

PR review

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

hii

Reviewer checklist

- [ ] 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

---

📚 Documentation preview 📚: https://pytorch-lightning--20211.org.readthedocs.build/en/20211/

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 79%. Comparing base (ea59e40) to head (337c1c2).
Report is 43 commits behind head on master.

❗ There is a different number of reports uploaded between BASE (ea59e40) and HEAD (337c1c2). Click for more details.

HEAD has 2247 uploads less than BASE

Flag	BASE (ea59e40)	HEAD (337c1c2)
cpu	525	24
lightning_fabric	67	0
pytest	266	0
python3.9	132	6
lightning	394	18
python3.10	66	3
python3.11	132	6
python3.12.7	195	9
gpu	2	0
pytorch2.1	99	9
pytest-full	261	24
pytorch2.2.2	33	3
pytorch_lightning	66	6
pytorch2.3	33	3
pytorch2.4.1	31	3
pytorch2.5.1	65	6

Additional details and impacted files

@@            Coverage Diff            @@
##           master   #20211     +/-   ##
=========================================
- Coverage      88%      79%     -9%     
=========================================
  Files         267      264      -3     
  Lines       23380    23325     -55     
=========================================
- Hits        20481    18366   -2115     
- Misses       2899     4959   +2060     

[01AbhiSingh]

Hii @Borda. Do I need to make any kind of changes in the PR ?

[lantiga]

This looks good, thank you for the contribution @01AbhiSingh

Ideally we could add a test to verify the behavior described in #17544. The current test suite can't detect the current change and this is usually a sign of insufficient coverage. Would you be willing to contribute such test?

[01AbhiSingh]

Yes, sure let me look into it.

[01AbhiSingh]

Hi @lantiga , Do you want a new test written from scratch or need me to make the necessary changes in a preexisting file? All the tests have been passed. If the changes need to be made in a preexisting file, it would be very helpful if you could point out the test in which I need to make the changes, as all the tests have been passed, and due to that, I can't find the test.

[lantiga]

hey @01AbhiSingh sorry for the wait

You can take inspiration from:

pytorch-lightning/tests/tests_pytorch/trainer/optimization/test_optimizers.py

Line 459 in b0aa504

def test_lr_scheduler_epoch_step_frequency(mocked_sched, check_val_every_n_epoch, tmp_path):

and add a new test where scheduling goes across epoch boundaries. Maybe @falckt can help too?

[lantiga]

Hey @01AbhiSingh can you import LightningModule here?

https://github.com/Lightning-AI/pytorch-lightning/pull/20211/files#diff-3c3f104dbdd06271c9e6e6d4fdf61398458148412401dd55a9bac1e9b5f913a8R19

Change:

from lightning.pytorch import Trainer

to

from lightning.pytorch import Trainer, LightningModule

this should fix the failing test

[01AbhiSingh]

Yeah, my bad. Forgot to add it even after seeing it. Done, please check.

[01AbhiSingh]

https://github.com/Lightning-AI/pytorch-lightning/actions/runs/12291356552/job/34299991507?pr=20211#:~:text=FAILED%20utilities/test_data.py%3A%3Atest_update_dataloader_typerror_custom_exception%20%2D%20AssertionError%3A%20Regex%20pattern%20did%20not%20match.

This is the test that is currently failing.

def train_dataloader(self):
           # Create a simple dataset for testing
           x = torch.randn(21, 32)  # 7 batches of size 3
           y = torch.randn(21, 2)
           return DataLoader(TensorDataset(x, y), batch_size=3)

should I add this and try to run the test again ?

[lantiga]

Go for it : )

You can also run this kind of test locally with pytest tests/tests_pytorch/<test_file>.py::<name_of_test> to make things quicker on your end. This test in particular can be ran on any machine (and you can use Lightning Studios for free if you want to run on GPUs ofc)

[01AbhiSingh]

Go for it : )

You can also run this kind of test locally with pytest tests/tests_pytorch/<test_file>.py::<name_of_test> to make things quicker on your end. This test in particular can be ran on any machine (and you can use Lightning Studios for free if you want to run on GPUs ofc)

I actually tried to run the test locally with the method you suggested but this error keeps showing up ERROR: file or directory not found: tests/tests_pytorch/test_optimizers.py anyway I am trying to solve this problem on my local env.

Edit: I've solved this problem, will now update the PR only when it's running perfectly on my local environment. Thanks :)

Another Edit 😝 : updated the PR please check

[01AbhiSingh]

Test passing on my local environment but not in the PR in the repo.

[01AbhiSingh]

I think this time it is all done. Can you please check once ? @lantiga

[lantiga]

Looks good, added a couple of comments

[lantiga]

Please remove the debug statements, I'd just convert them to asserts that compare the values with expected ones.

[lantiga]

Print statements in tests are not super helpful, just use asserts so the test will break if we don't get the expected value here.

[lantiga]

I'm not sure why there should be rounding discrepancies. Shouldn't this be fully deterministic?

[01AbhiSingh]

Actually the test was passing in my local environment but not in the CI / CD pipeline for some reason. I forgot to change it later. Let me correct it asap.

[stale]

This pull request has been automatically marked as stale because it has not had recent activity. It will be closed in 7 days if no further activity occurs. If you need further help see our docs: https://lightning.ai/docs/pytorch/latest/generated/CONTRIBUTING.html#pull-request or ask the assistance of a core contributor here or on Discord. Thank you for your contributions.

[stale]

This pull request is going to be closed. Please feel free to reopen it or create a new one based on top of the 'master' branch.