Fix `BatchSizeFinder` leaving model in train state #18826

tanaymeh · 2023-10-19T14:28:54Z

What does this PR do?

This PR patches the bug where BatchSizeFinder would leave the model in train state if used with trainer.validate

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:

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--18826.org.readthedocs.build/en/18826/

for more information, see https://pre-commit.ci

BoringDonut

@tanaymeh probably missed new is_training arg in on_fit_start

  File "lightning/pytorch/callbacks/batch_size_finder.py", line 187, in on_fit_start
    self.scale_batch_size(trainer, pl_module)
TypeError: scale_batch_size() missing 1 required positional argument: 'is_training'

BoringDonut · 2023-10-19T14:43:24Z

Other than aforementioned missing arg, this PR seems to fix the issue for val/test/prediction, thanks a lot

…der_trainer_state

tanaymeh · 2023-10-19T14:55:41Z

@BoringDonut Thanks for the review! I have added the requested changes, can you please review it again?

Thanks again!

BoringDonut

Seems to work well and fixed the issue. Tested with trainer.fit .predict .validate and .test
Presumably this PR doesn't change any behavior withing BatchSizeFinder internals, since it only affects the model after _scale_batch_size already did its main task anyway

codecov · 2023-10-19T15:18:17Z

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 49%. Comparing base (b8a96fe) to head (591eb4e).
Report is 492 commits behind head on master.

❗ There is a different number of reports uploaded between BASE (b8a96fe) and HEAD (591eb4e). Click for more details.

HEAD has 190 uploads less than BASE

Flag BASE (b8a96fe) HEAD (591eb4e)

lightning 42 16

cpu 71 24

pytest 53 2

python3.10 21 9

python3.9 6 3

lightning_fabric 10 0

python3.8 12 6

app 9 0

examples 9 0

python3.11 15 6

gpu 4 2

tpu 2 0

lightning_app 4 0

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #18826      +/-   ##
==========================================
- Coverage      84%      49%     -35%     
==========================================
  Files         443      435       -8     
  Lines       36123    35974     -149     
==========================================
- Hits        30231    17579   -12652     
- Misses       5892    18395   +12503

BoringDonut · 2023-10-19T17:16:29Z

src/lightning/pytorch/callbacks/batch_size_finder.py

+    def scale_batch_size(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule", is_training: bool) -> None:
        new_size = _scale_batch_size(
            trainer,
            self._mode,
            self._steps_per_trial,
            self._init_val,
            self._max_trials,
            self._batch_arg_name,
+            is_training,
        )


@tanaymeh On a second thought, I think we can avoid adding a new argument here

pl_module passed to scale_batch_size anyway, so there are no need to read pl_module.training earlier

Example:

def scale_batch_size(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None: is_training = pl_module.training new_size = _scale_batch_size( trainer, self._mode, self._steps_per_trial, self._init_val, self._max_trials, self._batch_arg_name, is_training, ) ... def on_fit_start(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None: self.scale_batch_size(trainer, pl_module) ...

Doing it that way would keep compatibility for people who might have called to scale_batch_size in their own code (e.g. classes inherited from BatchSizeFinder)

awaelchli · 2023-10-21T14:22:09Z

@tanaymeh @BoringDonut have you found where the batch size tuner calls model.train()? Ideally, we would want to make the finder not change it in the first place. This would be the real fix IMO. We should leave the responsibility of calling model.train()/eval() to the respective loop IMO.

tanaymeh · 2023-10-21T14:24:23Z

@tanaymeh @BoringDonut have you found where the batch size tuner calls model.train()? Ideally, we would want to make the finder not change it in the first place. This would be the real fix IMO. We should leave the responsibility of calling model.train()/eval() to the respective loop IMO.

@awaelchli I didn't find out the place where the bug is happening but I am working on it!

BoringDonut · 2023-10-21T22:29:38Z

have you found where the batch size tuner calls model.train()?

@tanaymeh @awaelchli

I think I found it
Call trace looks like that

1.https://github.com/Lightning-AI/lightning/blob/e7afe04ee86b64c76a5446088b3b75d9c275e5bf/src/lightning/pytorch/tuner/batch_size_scaling.py#L333
which redirects to
2. https://github.com/Lightning-AI/lightning/blob/e7afe04ee86b64c76a5446088b3b75d9c275e5bf/src/lightning/pytorch/loops/evaluation_loop.py#L108
Last line of which calls to return self.on_run_end()
3. https://github.com/Lightning-AI/lightning/blob/e7afe04ee86b64c76a5446088b3b75d9c275e5bf/src/lightning/pytorch/loops/evaluation_loop.py#L246
Aaaaand this methods assumes it runs within the train loop, so it converts model to the train state
4.
https://github.com/Lightning-AI/lightning/blob/e7afe04ee86b64c76a5446088b3b75d9c275e5bf/src/lightning/pytorch/loops/evaluation_loop.py#L270
This _on_evaluation_model_train ends up in
5. https://github.com/Lightning-AI/lightning/blob/e7afe04ee86b64c76a5446088b3b75d9c275e5bf/src/lightning/pytorch/core/hooks.py#L166

Basically it seems BatchSizeFinder should either not call _EvaluationLoop through .run(), as this method would unconditionally turn model to a train state
Or it indeed should compensate by calling to model.eval if it understands that it works with an _EvaluationLoop
I don't see an easy way to modify _EvaluationLoop.run() to address it without too much of a hustle
As such, i think solution by @tanaymeh looks reasonable.
I guess to be more consistent with existing BatchSizeFinder style this call to .eval can be done in BatchSizeFinder.__scale_batch_restore_params

tanaymeh · 2023-10-22T19:07:19Z

@BoringDonut I added your previously suggested changes of moving the is_training inside the scale_batch_size function.

Please let me know if I should add any more changes

cc: @awaelchli

src/lightning/pytorch/callbacks/batch_size_finder.py

src/lightning/pytorch/tuner/batch_size_scaling.py

for more information, see https://pre-commit.ci

carmocca

Do you think you can add a test to tests/tests_pytorch/tuner/test_scale_batch_size.py?

src/lightning/pytorch/tuner/batch_size_scaling.py

tanaymeh · 2023-10-25T14:07:30Z

Do you think you can add a test to tests/tests_pytorch/tuner/test_scale_batch_size.py?

@carmocca Should the test be similar to @BoringDonut's script (here) that is used to test if the model produces the same output in train and eval modes?

carmocca · 2023-10-27T17:56:10Z

I would suggest calling _scale_batch_size directly with some unittest Mocks, and assert that the original training state is preserved

tanaymeh · 2023-10-30T20:24:57Z

I would suggest calling _scale_batch_size directly with some unittest Mocks, and assert that the original training state is preserved

@carmocca Should create a mock for the trainer and lightning module as well?

from unittest.mock import Mock

...

trainer = Mock()
lightning_module = Mock()
trainer.lightning_module = lightning_module

...

…tate

gitguardian · 2024-01-16T09:24:43Z

⚠️ GitGuardian has uncovered 2 secrets following the scan of your pull request.

Please consider investigating the findings and remediating the incidents. Failure to do so may lead to compromising the associated services or software components.

🔎 Detected hardcoded secrets in your pull request

GitGuardian id	Secret	Commit	Filename
-	Generic High Entropy Secret	`78fa3af`	tests/tests_app/utilities/test_login.py	View secret
-	Base64 Basic Authentication	`78fa3af`	tests/tests_app/utilities/test_login.py	View secret

🛠 Guidelines to remediate hardcoded secrets

Understand the implications of revoking this secret by investigating where it is used in your code.
Replace and store your secret safely. Learn here the best practices.
Revoke and rotate this secret.
If possible, rewrite git history. Rewriting git history is not a trivial act. You might completely break other contributing developers' workflow and you risk accidentally deleting legitimate data.

To avoid such incidents in the future consider

following these best practices for managing and storing secrets including API keys and other credentials
install secret detection on pre-commit to catch secret before it leaves your machine and ease remediation.

^{🦉 GitGuardian detects secrets in your source code to help developers and security teams secure the modern development process. You are seeing this because you or someone else with access to this repository has authorized GitGuardian to scan your pull request.

Our GitHub checks need improvements? Share your feedbacks!}

add: checks to fix the train and eval state

9a82f39

tanaymeh requested review from williamFalcon, awaelchli, carmocca and justusschock as code owners October 19, 2023 14:28

github-actions bot added the pl Generic label for PyTorch Lightning package label Oct 19, 2023

[pre-commit.ci] auto fixes from pre-commit.com hooks

c8a7bd8

for more information, see https://pre-commit.ci

BoringDonut suggested changes Oct 19, 2023

View reviewed changes

tanaymeh added 2 commits October 19, 2023 15:51

Merge branch 'Lightning-AI:master' into bugfix/18813_fix_BatchSizeFin…

438010c

…der_trainer_state

fix: add is_training arg to on_fit_start fn

b2fa58a

tanaymeh requested a review from BoringDonut October 19, 2023 14:55

BoringDonut approved these changes Oct 19, 2023

View reviewed changes

BoringDonut reviewed Oct 19, 2023

View reviewed changes

BoringDonut mentioned this pull request Oct 20, 2023

BatchSizeFinder limits number of validation batches for the whole training process #18834

Closed

awaelchli added the community This PR is from the community label Oct 21, 2023

awaelchli added this to the 2.1.x milestone Oct 21, 2023

awaelchli added tuner bug Something isn't working labels Oct 21, 2023

fix: is_training now moved to scale_batch_size function

ca86bc4

awaelchli reviewed Oct 23, 2023

View reviewed changes

src/lightning/pytorch/callbacks/batch_size_finder.py Outdated Show resolved Hide resolved

src/lightning/pytorch/tuner/batch_size_scaling.py Outdated Show resolved Hide resolved

src/lightning/pytorch/tuner/batch_size_scaling.py Outdated Show resolved Hide resolved

tanaymeh and others added 2 commits October 23, 2023 09:47

fix: moved the train state capture to _scale_batch_size fn

6ae318e

[pre-commit.ci] auto fixes from pre-commit.com hooks

830f98b

for more information, see https://pre-commit.ci

BoringDonut approved these changes Oct 23, 2023

View reviewed changes

carmocca reviewed Oct 24, 2023

View reviewed changes

src/lightning/pytorch/tuner/batch_size_scaling.py Show resolved Hide resolved

carmocca mentioned this pull request Nov 12, 2023

Remember the eval mode of submodules when switching trainer stages #18951

Merged

awaelchli self-assigned this Nov 16, 2023

Merge branch 'master' into bugfix/18813_fix_BatchSizeFinder_trainer_s…

591eb4e

…tate

awaelchli modified the milestones: 2.1.x, 2.2.x Feb 8, 2024

awaelchli modified the milestones: 2.2.x, 2.3.x Jun 13, 2024

tanaymeh closed this Jun 17, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix `BatchSizeFinder` leaving model in train state #18826

Fix `BatchSizeFinder` leaving model in train state #18826

tanaymeh commented Oct 19, 2023 •

edited by github-actions bot

Loading

BoringDonut left a comment

BoringDonut commented Oct 19, 2023 •

edited

Loading

tanaymeh commented Oct 19, 2023

BoringDonut left a comment

codecov bot commented Oct 19, 2023 •

edited

Loading

BoringDonut Oct 19, 2023

awaelchli commented Oct 21, 2023

tanaymeh commented Oct 21, 2023

BoringDonut commented Oct 21, 2023 •

edited

Loading

tanaymeh commented Oct 22, 2023

carmocca left a comment

tanaymeh commented Oct 25, 2023

carmocca commented Oct 27, 2023

tanaymeh commented Oct 30, 2023 •

edited

Loading

gitguardian bot commented Jan 16, 2024 •

edited

Loading

Fix BatchSizeFinder leaving model in train state #18826

Fix BatchSizeFinder leaving model in train state #18826

Conversation

tanaymeh commented Oct 19, 2023 • edited by github-actions bot Loading

What does this PR do?

PR review

BoringDonut left a comment

Choose a reason for hiding this comment

BoringDonut commented Oct 19, 2023 • edited Loading

tanaymeh commented Oct 19, 2023

BoringDonut left a comment

Choose a reason for hiding this comment

codecov bot commented Oct 19, 2023 • edited Loading

Codecov Report

BoringDonut Oct 19, 2023

Choose a reason for hiding this comment

awaelchli commented Oct 21, 2023

tanaymeh commented Oct 21, 2023

BoringDonut commented Oct 21, 2023 • edited Loading

tanaymeh commented Oct 22, 2023

carmocca left a comment

Choose a reason for hiding this comment

tanaymeh commented Oct 25, 2023

carmocca commented Oct 27, 2023

tanaymeh commented Oct 30, 2023 • edited Loading

gitguardian bot commented Jan 16, 2024 • edited Loading

⚠️ GitGuardian has uncovered 2 secrets following the scan of your pull request.

Fix `BatchSizeFinder` leaving model in train state #18826

Fix `BatchSizeFinder` leaving model in train state #18826

tanaymeh commented Oct 19, 2023 •

edited by github-actions bot

Loading

BoringDonut commented Oct 19, 2023 •

edited

Loading

codecov bot commented Oct 19, 2023 •

edited

Loading

BoringDonut commented Oct 21, 2023 •

edited

Loading

tanaymeh commented Oct 30, 2023 •

edited

Loading

gitguardian bot commented Jan 16, 2024 •

edited

Loading