Speed up Bootstrapping Computation #409
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
|
Hi @JoelNiklaus , thanks a lot for this PR, I'll take a deeper look hopefully bf Monday |
|
(it's looking good from a first glance but I want to take some time to test it deeply) |
|
Hi! You get quite a huge difference in the bootstrap from the results we hardcoded in our test suite (like, an order of magnitude) - can you check why? (I was expecting a diff on a few decimal points, not stg this huge) |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
I am trying to run your tests with |
|
It's should be taking some time (around 30 min if you're on CPU), as it first needs to generate a bunch of predictions using a gpt2 model. It will be way faster if you have a GPU available |
|
I aborted it after 30min on an A100 GPU and I only ran the lite version. |
|
Which metrics did you check? BLEU or CHRF? For these the original version computes corpus level metrics. I switched to sample level metrics to speed up computation. There it would make sense to me that the stderr differs since the corpora are different for the different samples distributions in bootstrapping. For the metrics that are already sample level, I don't see a reason why they should be different. |
|
30min on an A100 is not normal. I wonder if there's an issue with the command you're running. Let me share the raw logs with you. (Do you have the rights to access them, by clicking on "Details" next to the failing test in the check list?) |
|
Thanks for the logs! Unfortunately, for me the test is just hanging and doesn't show any useful output at all. |
|
I think I may have found the problem. The parallelization was happening one level too high so it didn't have any effect. For me it happens way faster now and the stderrs are very similar. @clefourrier wdyt? |
|
Btw. corpus ter is always hanging for me in the middle of the execution. |
|
Launching the tests on our CLI - if we're not too far, I'm OK with updating the numbers there |
clefourrier
approved these changes
Dec 4, 2024
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.
Fixes Issue 408.
Added fix for heavy recomputation of sample level metrics.
No change for corpus level metrics, but for sample level metrics we can just look up the values instead of recomputing them for each sample, which can be prohibitively expensive for heavier metrics such as XCOMET-XXL.