Fix DPP + SyncBN #6838

BloodAxe · 2021-04-05T12:00:20Z

What does this PR do?

This PR ensures that model is already on correct GPU before applying SyncBN conversion. According to PyTorch docs, SyncBN conversion should be applied to model that is already on target GPU. Not following this requirement will lead to hard-tracing errors of some BN layers kept on CPU.
Changing the order of model_to_device call seems to fix this issue without any side-effects to the rest of the code.

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 🙃

Ensure that model is already on correct GPU before applying SyncBN conversion

codecov · 2021-04-05T12:13:13Z

Codecov Report

Merging #6838 (9b4a71d) into master (22a266d) will decrease coverage by 5%.
The diff coverage is 50%.

@@           Coverage Diff           @@
##           master   #6838    +/-   ##
=======================================
- Coverage      91%     87%    -5%     
=======================================
  Files         192     192            
  Lines       12191   12191            
=======================================
- Hits        11145   10595   -550     
- Misses       1046    1596   +550

awaelchli

looks like that's the correct thing to do.
can you apply the same to ddp_spawn.py?

BloodAxe · 2021-04-05T13:19:39Z

Greetings! Thanks for quick response. Added same changes to ddp_spawn in my last commit.

tchaton

LGTM ! Thanks for this catch !

tchaton · 2021-04-06T07:04:08Z

pytorch_lightning/plugins/training_type/ddp.py

        # move the model to the correct device
        self.model_to_device()

+        if self.sync_batchnorm:


Suggested change

if self.sync_batchnorm:

# `sync BatchNorm` requires a model on `cuda` device

if self.sync_batchnorm:

tchaton · 2021-04-06T07:04:26Z

pytorch_lightning/plugins/training_type/ddp_spawn.py

        # move the model to the correct device
        self.model_to_device()

+        if self.sync_batchnorm:


Suggested change

if self.sync_batchnorm:

# `sync BatchNorm` requires a model on `cuda` device

if self.sync_batchnorm:

* Fix DPP + SyncBN Ensure that model is already on correct GPU before applying SyncBN conversion * Fix order of SyncBN for ddp_spawn

* update readme by v1.2.x (#6728) * [bugfix] Add support for omegaconf and tpu (#6741) * fix_hydra * update changelog Co-authored-by: Your Name <you@example.com> * [docs] Update Bolts link (#6743) * Update Bolts link * Update Bolts link * formt Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com> * Update logic for checking TPUs availability (#6767) * Update logic for checking TPUs availability * fix flake8 * add fix * resolve bug (#6781) * Fix validation progress counter with check_val_every_n_epoch > 1 (#5952) Co-authored-by: rohitgr7 <rohitgr1998@gmail.com> Co-authored-by: Carlos Mocholí <carlossmocholi@gmail.com> * Remove extinct parameters from lightning_module.rst (#6801) Fixes #6800 * Update TPU docs for installation (#6794) * fix boolean check on iterable dataset when len not defined (#6828) * fix iterable dataset len check * update predict and validate * add validate to test * add changelog * add predict * Sanitize `None` params during pruning (#6836) * sanitize none params during pruning * amend * Fix `unfreeze_and_add_param_group` expects `modules` rather than `module` (#6822) * Enforce an epoch scheduler interval when using SWA (#6588) Co-authored-by: Carlos Mocholi <carlossmocholi@gmail.com> * Fix DPP + SyncBN (#6838) * Fix DPP + SyncBN Ensure that model is already on correct GPU before applying SyncBN conversion * Fix order of SyncBN for ddp_spawn * [Fix] TPU Training Type Plugin (#6816) * Fix support for symlink save_dir in TensorBoardLogger (#6730) * Add test for symlink support and initial fix * Respond to comment and add docstring * Update CHANGELOG.md * Simplify * Update pytorch_lightning/utilities/cloud_io.py Co-authored-by: Carlos Mocholí <carlossmocholi@gmail.com> * Make `LightningLocalFileSystem` protected Co-authored-by: Carlos Mocholí <carlossmocholi@gmail.com> * Fixed missing arguments in `lr_find` call (#6784) There seem to be 3 arguments missing in the `lr_find` call in the tunining.py file. * Update Changelog & version * Fix TPU tests for checkpoint Skip advanced profiler for torch > 1.8 Skip pytorch profiler for torch > 1.8 Fix save checkpoint logic for TPUs Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com> Co-authored-by: thomas chaton <thomas@grid.ai> Co-authored-by: Your Name <you@example.com> Co-authored-by: Akihiro Nitta <nitta@akihironitta.com> Co-authored-by: Yuan-Hang Zhang <sailordiary@users.noreply.github.com> Co-authored-by: rohitgr7 <rohitgr1998@gmail.com> Co-authored-by: Carlos Mocholí <carlossmocholi@gmail.com> Co-authored-by: Elizaveta Logacheva <elimohl@gmail.com> Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com> Co-authored-by: Karthik Prasad <prasadkr@uci.edu> Co-authored-by: Sadiq Jaffer <sadiq@toao.com> Co-authored-by: Michael Baumgartner <m.baumgartner@ymail.com> Co-authored-by: Eugene Khvedchenya <ekhvedchenya@gmail.com> Co-authored-by: Ethan Harris <ewah1g13@soton.ac.uk> Co-authored-by: Tharindu Hasthika <tharindubathigama@gmail.com>

* Fix DPP + SyncBN Ensure that model is already on correct GPU before applying SyncBN conversion * Fix order of SyncBN for ddp_spawn

Fix DPP + SyncBN

a168079

Ensure that model is already on correct GPU before applying SyncBN conversion

BloodAxe requested review from awaelchli, justusschock, SeanNaren and tchaton as code owners April 5, 2021 12:00

ananthsub approved these changes Apr 5, 2021

View reviewed changes

awaelchli added the distributed Generic distributed-related topic label Apr 5, 2021

awaelchli reviewed Apr 5, 2021

View reviewed changes

awaelchli added the bug Something isn't working label Apr 5, 2021

Fix order of SyncBN for ddp_spawn

9b4a71d

awaelchli approved these changes Apr 5, 2021

View reviewed changes

SeanNaren approved these changes Apr 5, 2021

View reviewed changes

s-rog approved these changes Apr 6, 2021

View reviewed changes

tchaton reviewed Apr 6, 2021

View reviewed changes

tchaton merged commit eafec7d into Lightning-AI:master Apr 6, 2021

kaushikb11 mentioned this pull request Apr 6, 2021

1.2.x cherries 🍒 #6083

Closed

kaushikb11 pushed a commit to kaushikb11/pytorch-lightning that referenced this pull request Apr 6, 2021

Fix DPP + SyncBN (Lightning-AI#6838)

215a9c9

* Fix DPP + SyncBN Ensure that model is already on correct GPU before applying SyncBN conversion * Fix order of SyncBN for ddp_spawn

lexierule pushed a commit that referenced this pull request Apr 7, 2021

Fix DPP + SyncBN (#6838)

edf6289

* Fix DPP + SyncBN Ensure that model is already on correct GPU before applying SyncBN conversion * Fix order of SyncBN for ddp_spawn

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix DPP + SyncBN #6838

Fix DPP + SyncBN #6838

BloodAxe commented Apr 5, 2021

codecov bot commented Apr 5, 2021 •

edited

Loading

awaelchli left a comment

BloodAxe commented Apr 5, 2021

tchaton left a comment

tchaton Apr 6, 2021

tchaton Apr 6, 2021

	if self.sync_batchnorm:
	# `sync BatchNorm` requires a model on `cuda` device
	if self.sync_batchnorm:

Fix DPP + SyncBN #6838

Fix DPP + SyncBN #6838

Conversation

BloodAxe commented Apr 5, 2021

What does this PR do?

Before submitting

PR review

Did you have fun?

codecov bot commented Apr 5, 2021 • edited Loading

Codecov Report

awaelchli left a comment

Choose a reason for hiding this comment

BloodAxe commented Apr 5, 2021

tchaton left a comment

Choose a reason for hiding this comment

tchaton Apr 6, 2021

Choose a reason for hiding this comment

tchaton Apr 6, 2021

Choose a reason for hiding this comment

codecov bot commented Apr 5, 2021 •

edited

Loading