OTT 0.4.7

What's Changed

• add dropout to VelocityField and OTFM/GENOT by @MUCDK in #521
• Update icnn.py by @marcocuturi in #524
• Update Codecov Action to v4 by @michalk8 in #523
• Add h-transform for TI costs by @michalk8 in #525
• Update univariate.py by @marcocuturi in #527
• Fix/h legendre tests by @michalk8 in #529
• Update costs.py by @marcocuturi in #532
• Debiased property change by @michalk8 in #536
• Don't stop grads w.r.t. the parametrized costs by @michalk8 in #544
• univariate dual vectors for arbitrary sizes by @marcocuturi in #505
• Fix low-rank convergence criterion by @michalk8 in #547
• Fixing typos in documentation by @pkassraie in #548
• add MMSinkhorn solver by @marcocuturi in #549
• Force ICNN to adopt default initialization of its own layers by @Algue-Rythme in #551
• Deprecate Python 3.8 by @michalk8 in #554
• New neural OT solver ENOT by @nazarblch in #503
• Refactor regularized TI costs by @michalk8 in #553

New Contributors

• @Algue-Rythme made their first contribution in #551
• @nazarblch made their first contribution in #503

Full Changelog: 0.4.6...0.4.7

OTT 0.4.6

What's Changed

• Enable M1 runner by @michalk8 in #486
• Potentials as tuple by @michalk8 in #488
• Remove bool option for scale_cost by @michalk8 in #492
• Use more iterations for LR-GW by @michalk8 in #494
• update sinhorn divergence gradient flow nb by @marcocuturi in #497
• add barycenter operator for (convex) translation invariant costs by @marcocuturi in #498
• Addition of gangbo-mccann map estimators using twist operator by @marcocuturi in #500
• Add sparse Chebyshev approximation by @guillaumehu in #502
• Fix/io callback by @michalk8 in #506
• Maint/third lower bound by @michalk8 in #507
• Remove epsilon scheduler in GW by @michalk8 in #508
• Fix SCOT tutorial imports and data by @michalk8 in #511
• restructuring neural models + addition of OT-FM and GENOT by @MUCDK in #468
• compute_sparse_laplacian gives int32 vs int64 index mismatch by @selmanozleyen in #510
• Update One_Sinkhorn.ipynb by @pkassraie in #515
• Update actions' versions by @michalk8 in #516
• Fix/batched non sym cost by @michalk8 in #518
• bug fix: avoid mixing up linear and quadratic in genot by @soerenab in #517

New Contributors

• @selmanozleyen made their first contribution in #510
• @pkassraie made their first contribution in #515
• @soerenab made their first contribution in #517

Full Changelog: 0.4.5...0.4.6

Neural module refactoring

What's Changed

• Re-run and refactoring of some docs/tutorials + minor changes by @marcocuturi in #431
• Feature/quadratic solve by @michalk8 in #433
• Fix n_iters by @michalk8 in #437
• Fix/rank2 init by @michalk8 in #442
• nn.dataset.GaussianMixture: minor fix to the docstring and standard deviation parameter name by @bamos in #445
• introduce multivariate cdf / quantiles by @marcocuturi in #447
• Feature/update jax version by @michalk8 in #448
• Histogram Transport Implementation by @Daniel-Packer in #444
• Doc refresh by @marcocuturi in #450
• Fix/jax prng deprecation by @michalk8 in #455
• Added CDF distance to Univariate Solver by @Daniel-Packer in #451
• Use PyPI trusted publishers by @michalk8 in #459
• first lb -> third lb by @Daniel-Packer in #462
• Restructure neural module by @michalk8 in #466
• refactoring UnivariateSolver by @marcocuturi in #472
• Feature/geodesic sinkhorn by @guillaumehu in #457
• Enable 3.12 CI by @michalk8 in #475
• Feature/apply lse mode by @michalk8 in #474
• Fix/neural docs by @michalk8 in #473
• Add quadratic layers and enhance ICNNs, update tutorial by @nvesseron in #477
• Fix heat scaling by @guillaumehu in #481
• Add low-rank kernel geometry by @michalk8 in #440

New Contributors

• @Daniel-Packer made their first contribution in #444
• @guillaumehu made their first contribution in #457
• @nvesseron made their first contribution in #477

Full Changelog: 0.4.4...0.4.5

Unbalanced low-rank Gromov-Wasserstein

What's Changed

• introduce new norm operator that can differentiate through lambda x: norm(x-x) + lighter tests across the board by @marcocuturi in #411
• Feature/better io callbacks by @michalk8 in #413
• Feature/optional dependencies by @michalk8 in #412
• Update pre-commits by @michalk8 in #416
• Feature/move solve by @michalk8 in #418
• Feature/lineax ridge by @michalk8 in #424
• Feature/ulrgw by @michalk8 in #410

Full Changelog: 0.4.3...0.4.4

Add top-k mask and some minor fixes

What's Changed

• A few changes to increase Coverage by @marcocuturi in #390
• rename ent_reg_cost in sinkhorn solver by @marcocuturi in #391
• Increase coverage of NN modules by including init in tests by @marcocuturi in #392
• add tests for plots in potentials by @marcocuturi in #393
• test Plot class in tools module to increase coverage by @marcocuturi in #394
• Add soft topk_mask operator, fix ranks docs and tests by @marcocuturi in #396
• Add relative_epsilon option to GromovWasserstein by @michalk8 in #355
• [ENH]: Fix misc typos in soft_sort doc by @pierreablin in #397
• Update docs by @michalk8 in #398
• add ability to add titles on animation plots by @marcocuturi in #400
• Update contributing visibility by @michalk8 in #399
• update nb by @marcocuturi in #403
• Fix not being able to jit the soft quantile by @michalk8 in #408
• Test default_progress_fn, skip types.py by @michalk8 in #404

Full Changelog: 0.4.2...0.4.3

Unbalanced LR Sinkhorn + New MBO Estimators + minor fixes

What's Changed

• Add a tuto about sparse monge displacements by @pierreablin in #374
• introduce ent_reg_cost / kl_reg_cost in Sinkhorn to clarify which is returned to the user in SinkhornOutput objects. by @marcocuturi in #376
• more precise pydocs for #376 by @marcocuturi in #377
• Update sinkhorn.py by @marcocuturi in #378
• Compute multiple soft-quantiles in one execution without using vmap by @marcocuturi in #382
• incorporate feedback in #382 by @marcocuturi in #385
• Update soft_sort.py by @marcocuturi in #387
• Add elastic-net type costs with an orthogonal linear transformation of the input before it is fed into the regularizer by @michalk8 in #383
• Add monge gap by @theouscidda6 in #361
• Feature/unbalanced LR sinkhorn by @michalk8 in #379
• changes in definition of Monge gap by @marcocuturi in #389
• fix #375, converged flag when predefined number of iterations. by @marcocuturi in #386

Full Changelog: 0.4.1...0.4.2

0.4.1 Various new features (kernel steps for LR Sinkhon, lineax implicit diff), solvers must now be jitted.

Highlights:

• reformatted tutorial NBs
• progressbars for Sinkhorn
• low-rank Sinkhorn solvers can run in kernel mode
• discrete_barycenter is refactored using the Solver/Problem model, rather than a simple function.
• soft-DTW added as a cost function.
• Sparsity inducing (in displacements) costs, see https://arxiv.org/abs/2302.04065

What's Changed

• Fix typo: Wasserstein-4 -> Wasserstein-2 by @bamos in #249
• Fix/warnings lint by @michalk8 in #260
• Addresses #229: Copyrights unification by @AWehenkel in #261
• fixed the docs about the implicit_diff argument by @gjhuizing in #270
• Correct package requirements for tests. by @AWehenkel in #271
• Add pre-commit CI cache by @michalk8 in #279
• Fix flake8 D10{1,2,3} errors by @giovp in #269
• Added CostFn and Plot to the documentation by @gjhuizing in #288
• add primal_cost property to output of GW, to compute cost without any regularization. by @marcocuturi in #286
• add ruff and remove redundant pre-commits by @giovp in #278
• Remove matplotlib dependency by @michalk8 in #304
• Fix/epsilon regularization by @michalk8 in #310
• Address #252 - Make soft ranking notebook self contained by @AWehenkel in #301
• Fix small TODOs by @michalk8 in #311
• Add CI tests for upstream JAX by @michalk8 in #313
• Subsample initializer by @JTT94 in #309
• feat: add progress reporting to Sinkhorn by @bosr in #228
• Refactor discrete_barycenter as solver, cf. issue #255 by @marcocuturi in #316
• Make random seeding consistent with PRNG keys everywhere instead of seeds by @othmanesebbouh in #290
• Add Soft-DTW cost function by @michalk8 in #318
• Fix/sinkdiv hessian by @michalk8 in #321
• Enable D* checks by @michalk8 in #319
• Fix spelling by @michalk8 in #323
• Remove ignoring passed params in ICNN by @michalk8 in #325
• Fix spelling by @michalk8 in #326
• neuraldual: fix issues with objective-based amortization and no-fine-tuning by @bamos in #327
• Parameter to make line strength proportional to coupling strength in Plot by @gjhuizing in #280
• Remove sparse option in Graph geometry by @michalk8 in #329
• gh: refactor community files by @SauravMaheshkar in #331
• Have a simpler "getting started" notebook by @pierreablin in #296
• update NBs by @marcocuturi in #332
• Docs/fix minor issues by @michalk8 in #334
• Small fix in soft_sort.quantile docstring by @pierreablin in #336
• Fix/jit inside jit by @michalk8 in #335
• typos by @marcocuturi in #339
• Remove assertions in cost functions by @michalk8 in #340
• Fix EntropicPotentials's test by @michalk8 in #343
• Added tutorial on Sinkhorn divergence by @gjhuizing in #282
• Pin chex version in GPU CI by @michalk8 in #346
• Feature/unscaled gw cost by @michalk8 in #348
• Fix typos in notebook by @michalk8 in #350
• docs/progress bar by @bosr in #347
• Fix/default prng by @michalk8 in #353
• tracking progress of gromov-wasserstein by @bosr in #351
• Store individual costs to in the GW barycenter by @michalk8 in #354
• Update brain notebook by @michalk8 in #358
• add dtype to kernel by @MUCDK in #365
• Neural OT notebook: add more 2d datasets by @bamos in #303
• New kernel lr by @meyerscetbon in #324
• Fine grained control of Bures barycenters by @marcocuturi in #366
• use lineax to solve linear system in implicit diff by @marcocuturi in #370
• Fix .readthedocs.yaml by @michalk8 in #372

New Contributors

• @AWehenkel made their first contribution in #261
• @gjhuizing made their first contribution in #270
• @giovp made their first contribution in #269
• @bosr made their first contribution in #228
• @othmanesebbouh made their first contribution in #290
• @pierreablin made their first contribution in #296

Full Changelog: 0.4.0...0.4.1

0.4.0: GPU tests, re-adding jit, improved NN OT solvers, and misc.

What's Changed

• Fix badges by @michalk8 in #211
• Feature/faster gw init by @michalk8 in #213
• Fix/device array check by @michalk8 in #216
• Remove support for Python 3.7 by @michalk8 in #227
• chore: add yapf config to pyproject.toml by @SauravMaheshkar in #230
• Remove functional API by @michalk8 in #222
• Docs/notebooks myst by @michalk8 in #233
• [ci skip] Add environment.yml for binder by @michalk8 in #236
• Use lax.cond when computing error in Sinkhorn by @michalk8 in #235
• Fix/debiased entropic map by @michalk8 in #238
• Misc/tox by @michalk8 in #239
• Re-add jit argument by @michalk8 in #221
• Add GPU test runner by @michalk8 in #242
• Remove old Transport interface by @michalk8 in #243
• Remove shape check in LRCGeometry by @michalk8 in #247
• Add regularized costs by @michalk8 in #244
• Check only whether problem can be LR by @michalk8 in #248
• Improvements to the neural dual solver by @bamos in #219

New Contributors

• @SauravMaheshkar made their first contribution in #230

Full Changelog: 0.3.1...0.4.0

0.3.1 with new cost features, different code layout, changes in potentials handling. Some solvers in function form deprecated.

What's Changed

• Fix jax.device_put for potentials by @michalk8 in #183
• Introduce primal_cost and dual_cost for Sinkhorn outputs (only primal for LR) for arbitrary geometries. by @marcocuturi in #184
• Feature/src layout by @michalk8 in #188
• Handle better inf with p parameter in SqPNorm by @marcocuturi in #189
• Feature/graph normalized laplacian by @michalk8 in #191
• Update installation instruction by @michalk8 in #195
• Feature/center potentials by @michalk8 in #194
• Remove jit from solvers by @michalk8 in #192
• Accelerating unbalanced OT by @michalk8 in #197
• Feature/depracate functional api by @michalk8 in #204
• Fixes numerical errors in Bures barycenter, and sqrtm, due to low default precision. by @marcocuturi in #205
• Update docstrings for deprecated functions by @michalk8 in #207

Full Changelog: 0.3.0...0.3.1

New organization for code; Changes in `costs` for `PointClouds`

The main changes in this version are twofold:

• Changes in PointCloud geometry, and more specifically handling of cost function. The power parameter that was used to optionally pass a c(x,y) = CostFn(x,y) ** power is now deprecated. One can add it manually by defining a custom CostFn. To compensate this change, a new class of translation invariant costs (TICost) has been created, from which most costs now inherit, defined as c(x,y) = h(x-y). Additionally, to handle Brenier's theorem, the user has the option of passing on the legendre transform of h, h_legendre.
• The core folder was too horizontal, containing various modules. It has been reorganized and split into 3 modules that make more sense, problems (to describe OT problems), solvers (solve them) and initializers (as optional modules to help solvers). The latter two have an arborescence that reflects that in problems.

What's Changed

• fix point_clouds.ipynb by @marcocuturi in #159
• Deprecate power in PointCloud, introduce TICost and use it to compute Entropic (Brenier) maps. by @marcocuturi in #167
• Misc/project structure by @michalk8 in #176

Full Changelog: 0.2.11...0.3.0