[RLlib] Make torch PPO regression test longer #31892
Merged
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
Signed-off-by: Artur Niederfahrenhorst <artur@anyscale.com>
|
I've "resolved" the mystery -- it just seems that tf and torch are better at different tasks. Varying the batch and minibatch size I was able to make torch models faster than tf eager models. After talking with Sven I think the main issue is that the hyperparameters for the tuned examples were optimized for tf. Messing w/ batch size, minibatch size, num SGD iters to find suitable torch hyperparameters could likely make it as fast as tf. |
Signed-off-by: Artur Niederfahrenhorst <artur@anyscale.com>
|
@smorad This PR separates the tests from each other. Can you re-tune torch's train_batch_size etc so that it comes closer to TF's performance? |
|
If you state your intuition in terms of what these paramters should be in this test, I can also do it. |
gjoliver
approved these changes
Jan 28, 2023
edoakes
pushed a commit
to edoakes/ray
that referenced
this pull request
Mar 22, 2023
…roject#31892) Signed-off-by: Artur Niederfahrenhorst <artur@anyscale.com> Signed-off-by: Edward Oakes <ed.nmi.oakes@gmail.com>
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Outdated suggestions cannot be applied.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.
Signed-off-by: Artur Niederfahrenhorst artur@anyscale.com
Why are these changes needed?
Because of a throughput difference between torch and tf (tf being 2x faster), we should give torch more time in this test.
I've run a couple of experiments and observed that torch has the same sample efficiency and that the only difference appears to be throughput. Until @smorad has resolved this mystery, we should make this test longer and simply make it shorter when we are able to resolve this so that this test can properly fail/succeed for the time being.
Checks
git commit -s) in this PR.scripts/format.shto lint the changes in this PR.