Repository navigation
fix: Apply a dense Gram after a global projection again - #622
Open
BryanBradfo wants to merge 1 commit into
Open
BryanBradfo wants to merge 1 commit into
BryanBradfo wants to merge 1 commit into
Conversation
This branch has not been deployed
This file contains hidden or 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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
bergson trakstops at step 2/4 onmain:Since #612,
GradientCollectorsends every Hessian given with a global projection toload_projection_preconditioner, which appliesH^-1/2before the projection and only accepts factored Hessians. TRAK's Gram is dense and over the projected gradients, so before #612 the builder applied it after the projection.This only takes the pre-projection path for factored Hessians. A dense Gram goes back to the builder, as before #612. P-SIFT is unchanged.
Tests
tests/test_trak.pyneeds a GPU, so CI skips it. It passes on the commit before SRHT global projection and P-SIFT for the factored preconditioners #612 and has 4 errors from SRHT global projection and P-SIFT for the factored preconditioners #612 on. With this fix it passes again.tests/test_global_projection.py: a dense Gram over a global projection whitens the projected gradients. It fails onmainwith the error above.test_global_kfac,test_psiftandtest_hadamard_projectionstill pass.bergson trakruns all 4 steps. Its scores match the commit before SRHT global projection and P-SIFT for the factored preconditioners #612 to a relative difference of 4e-6, with the same top 10 for every query.