[ENH] Gradient-based soft barycentre averaging#3526
Open
chrisholder wants to merge 2 commits into
Open
Conversation
aeon-actions-bot
Bot
added
clustering
Contributor
Thank you for contributing to
|
Add `soft_barycenter_average` (in `_ba_soft.py`): a gradient-based barycentre that minimises a soft elastic objective with `scipy.optimize.minimize` (L-BFGS-B) using the private soft-DTW / soft-MSM gradients. Wired into `elastic_barycenter_average` as `method="soft"` and into the averaging callable registry, exported alongside `VALID_SOFT_BA_METHODS`. Clean separation (no soft distances in discrete BA): soft distances are only averaged via `method="soft"`. `soft_dtw` is removed from `VALID_BA_DISTANCE_METHODS` and from `_get_alignment_path`, and early validation rejects `method="soft"` with a hard distance and discrete methods with a soft distance. The discrete-BA test list drops `soft_dtw` and the temporary `test_kasba` skip is reverted. Fixes over the prototype: - `return_distances_to_center` now recomputes distances at the optimised barycentre instead of returning the initial values. - Multivariate soft-MSM averaging accumulates the gradient per channel (independent), matching the soft-MSM distance; the prototype assigned a 1D gradient into a 2D slot and never worked for soft-MSM. - Threading uses `numba_thread_handler`; docstring covers both soft distances. - `_ba_setup` gains `compute_previous_cost` so the soft path skips the upfront pairwise-cost computation. Tests parametrise over soft-DTW and soft-MSM, and validate the soft-DTW barycentre against tslearn's `softdtw_barycenter`.
…dd tests - `_soft_barycenter_one_iter` now threads `window` and `itakura_max_slope` through to both gradient branches. Previously `window` was silently ignored for `distance="soft_msm"` (and `itakura_max_slope` for both), despite being listed in the docstring kwargs. - document `method="soft"` and the `minimise_method` parameter in `elastic_barycenter_average` (both were undocumented). - replace the unused `_ba_setup` unpack targets with `_`. - add tests: the optimiser moves the barycentre away from its init, and a MULTITHREAD_TESTING-gated determinism test (n_jobs must not change the result), mirroring the discrete-BA threaded tests.
chrisholder
force-pushed
the
soft-msm-distance
branch
from
97e3af5 to
4cfc8c2
Compare
chrisholder
force-pushed
the
soft-barycentre-averaging
branch
from
13fcd3c to
4a175f4
Compare
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
1 participant
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.
Reference Issues/PRs
Third of the soft-distance stack. Stacked on #3508 (soft-MSM distance), which is stacked on #3483. Please review/merge those first; this branch targets #3508 and the diff shrinks as they merge.
What does this implement/fix?
Adds gradient-based soft barycentre averaging.
soft_barycenter_average(new_ba_soft.py) minimises a differentiable soft elastic objective withscipy.optimize.minimize(L-BFGS-B), using the private soft-DTW / soft-MSM gradients from the earlier PRs. It is wired intoelastic_barycenter_averageasmethod="soft", added to the averaging callable registry, and exported alongsideVALID_SOFT_BA_METHODS.Clean separation — soft distances only via
method="soft". Soft distances are not used by the discrete BA methods (a soft distance has an alignment distribution, not a single hard path, so DBA-style discrete realignment is ill-defined and was a no-op).soft_dtwis removed fromVALID_BA_DISTANCE_METHODSand from_get_alignment_path; early validation rejectsmethod="soft"with a hard distance and the discrete methods with a soft distance. The discrete-BA test list dropssoft_dtwand the temporarytest_kasbaskip added in #3483 is reverted.Fixes over the prototype:
return_distances_to_centernow recomputes the distances at the optimised barycentre instead of returning the initial (inf) values from setup.numba_thread_handler(the prototype imported a module that does not exist onmain); the docstring covers both soft distances._ba_setupgains acompute_previous_costflag so the soft path skips the upfront pairwise-cost computation.Does your contribution introduce a new dependency?
No. (
tslearn, already optional, is used only as a guarded test oracle.)Test plan
test_soft_ba.pyparametrised over soft-DTW and soft-MSM: shape/finiteness (uni + multivariate),return_cost/return_distances_to_center, weights change the result, single-series passthrough, themethod="soft"validation matrix, and dispatch-equals-direct.softdtw_barycenter(guarded withimportorskip).aeon/clustering/averagingsuite +test_k_means(402 passed) with the discrete-BA changes, and the soft BA tests. Pre-commit passes.