Repository navigation
Use the scheduled epsilon in the discrete barycenter update - #710
Open
shaneraphel wants to merge 1 commit into
Open
shaneraphel wants to merge 1 commit into
shaneraphel wants to merge 1 commit into
Conversation
The entropic step is ε(t) log a - softmin_ε(t)(c - f - g). The vmap accepted the iteration and then dropped it, so the scheduler returned its target ε instead of init * decay^t * target.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
The discrete barycenter step updates the dual potential by
ε(t) log a − softmin_{ε(t)}(c − f − g),
where ε(t) = max(init · decay^t, 1) · target.
update_potentialandupdate_scalingread that value fromepsilon_scheduler(iteration), andNonemeans the target. The vmapped update accepted the iteration and did not forward it, so every step, including the annealed ones, ran at the final ε. The debiased correction already passediteration. The log-sum-exp and scaling updates now do the same.Tests
tests/solvers/linear/discrete_barycenter_test.py: 6 passed on macOS arm64 (the two existing grid cases, the two point-cloud cases, and a new schedule check in both modes).The new check uses ε with target 0.2, init 8, decay 0.5, so ε(0) = 1.6 and the target is 0.2. After one iteration the stored potentials match
update_potential/update_scalingat iteration 0, and they do not match the same update with the iteration omitted.Fixes #705