Fix DDP Reduction in PyTorch 1.7 when the model has unused parameters #586
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
facebook-github-bot
added
the
CLA Signed
Skylion007
approved these changes
Feb 4, 2021
mathfac
approved these changes
Feb 5, 2021
habitat_baselines/rl/ppo/ppo.py
Outdated
| @@ -152,6 +152,13 @@ def update(self, rollouts: RolloutStorage) -> Tuple[float, float, float]: | |||
|
|
|||
| return value_loss_epoch, action_loss_epoch, dist_entropy_epoch | |||
|
|
|||
| def evaluate_actions( | |||
There was a problem hiding this comment.
Add a little doc string and made it internal since it just exists to be overridden with inheritance.
|
@mathfac @Skylion007 does this test look reasonable? It's a little ugly but I wanted it to be as close to testing the actual code path used in training as possible. |
Skylion007
approved these changes
Feb 8, 2021
|
Yes, looks good. |
dannymcy
pushed a commit
to dannymcy/habitat-lab
that referenced
this pull request
Jul 8, 2024
…facebookresearch#586) * Hacky thing to interface with ddp.forward * State guard * Add unit test and comment in readme about it.
HHYHRHY
pushed a commit
to SgtVincent/EMOS
that referenced
this pull request
Aug 31, 2024
…facebookresearch#586) * Hacky thing to interface with ddp.forward * State guard * Add unit test and comment in readme about it.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Motivation and Context
PyTorch 1.7 slightly changed the way Reducer should be used when the model has unused parameters, leading to DD-PPO breaking when the model does, why this is certainly an edge case, it is a problem.
This PR solve this and attempts to solve future issues by using a wrapper to interface with DDP's forward call instead of using the undocumented Reducer APIs. While I am not sure if this won't cause it's own issues since it is kinda hacky, this seems like it has the potential to be better.
How Has This Been Tested
Tested on a model with unused parameters. I will write a test for this before merging this, but don't have time today.
Types of changes