Code

Refactor dual cone projections#237

Code

Merged

ValerianRey merged 36 commits intomainTorchJD/torchjd:mainfrom factorize_dual_cone_projectionsTorchJD/torchjd:factorize_dual_cone_projectionsCopy head branch name to clipboardFeb 4, 2025

+164-94

Commits

Commits on Feb 1, 2025

• Factorize projection onto the dual cone.

  PierreQuinton

  committed
  aa9897e

  Copy full SHA for aa9897e

• Move regularization out of projection onto the dual cone.

  PierreQuinton

  committed
  f534006

  Copy full SHA for f534006

• Improve code coverage

  PierreQuinton

  committed
  864d7a4

  Copy full SHA for 864d7a4

• Add link to paper in weights_of_projection

  PierreQuinton

  committed
  b0739fe

  Copy full SHA for b0739fe

• Rename functions in _dual_cone_utils

  PierreQuinton

  committed
  cd43f65

  Copy full SHA for cd43f65

• Rename lagrangian to lagrangian_multipliers to match the maths

  PierreQuinton

  committed
  402024c

  Copy full SHA for 402024c

• Separate _get_projection_weights into a tensorization of _get_lagrange_multipliers_array and a sum with the weights

  PierreQuinton

  committed
  a7ae47f

  Copy full SHA for a7ae47f

• call the cast to numpy (and thus CPU) only once on the weights.

  PierreQuinton

  committed
  895b123

  Copy full SHA for 895b123

• Improve the tensorization of _get_lagrange_multipliers_array

  PierreQuinton

  committed
  25c5666

  Copy full SHA for 25c5666

• Improve readability of tensorization by using a for comprehension.

  PierreQuinton

  committed
  8c52aa5

  Copy full SHA for 8c52aa5

• Exchange _compute_normalized_regularized_gramian and _regularize_gramian

  PierreQuinton

  committed
  3591577

  Copy full SHA for 3591577

• Add comment explaining the goal of regularization

  PierreQuinton

  committed
  c631f30

  Copy full SHA for c631f30

• Rename:

  Show description for 4aa0cbd
  PierreQuinton

  committed
  4aa0cbd

  Copy full SHA for 4aa0cbd

• Exchange _normalize and _normalize_and_regularize

  PierreQuinton

  committed
  193ff9c

  Copy full SHA for 193ff9c

• rename epsilons in _normalize and _regularize

  PierreQuinton

  committed
  a62c708

  Copy full SHA for a62c708

Commits on Feb 2, 2025

• Set explicitly the dtype and device of regularization matrix

  PierreQuinton

  committed
  fb07035

  Copy full SHA for fb07035

• Rename functions & variables, add _to_array

  ValerianRey

  committed
  756e8c5

  Copy full SHA for 756e8c5

• Fix bug of transposed output

  ValerianRey

  committed
  601bc6e

  Copy full SHA for 601bc6e

• Delete outdated tests

  ValerianRey

  committed
  b3615cd

  Copy full SHA for b3615cd

• Rename functions _normalize etc

  ValerianRey

  committed
  c9afb41

  Copy full SHA for c9afb41

• Improve docstring and naming of test of kkt conditions.

  PierreQuinton

  committed
  e141600

  Copy full SHA for e141600

• Improve docstring of _get_projection_weights

  PierreQuinton

  committed
  9c7eaac

  Copy full SHA for 9c7eaac

• Add test of tensorization of _get_projection_weights

  PierreQuinton

  committed
  f04581e

  Copy full SHA for f04581e

Commits on Feb 3, 2025

• Clarify the QP problem that is solved, it now matches exactly that of Proposition 1 instead of the dual of projection onto the dual cone.

  PierreQuinton

  committed
  a5f562a

  Copy full SHA for a5f562a

• Improve docstring of _get_projection_weight_vector

  PierreQuinton

  committed
  627adf0

  Copy full SHA for 627adf0

Commits on Feb 4, 2025

• Add changelog entry

  PierreQuinton

  committed
  bb411c9

  Copy full SHA for bb411c9

• Add no_grad context to the computation of UPGrad and DualProj as we cannot infer any grad from our solve using qp_solvers.

  PierreQuinton

  committed
  782aad6

  Copy full SHA for 782aad6

• Improve dual cone utility functions

  Show description for ffca7d0
  ValerianRey

  committed
  ffca7d0

  Copy full SHA for ffca7d0

• Improve variable names:

  Show description for 60a3851
  ValerianRey

  committed
  60a3851

  Copy full SHA for 60a3851

• Improve docstring of test_solution_weights

  ValerianRey

  committed
  c7a2fe8

  Copy full SHA for c7a2fe8

• Remove torch.no_grad in DualProj and UPGrad

  ValerianRey

  committed
  633b286

  Copy full SHA for 633b286

• Add docstring to test_tensorization_shape

  ValerianRey

  committed
  4f3e48a

  Copy full SHA for 4f3e48a

• Improve docstring formatting

  ValerianRey

  committed
  0c224fc

  Copy full SHA for 0c224fc

• Simplify UPGrad

  Show description for c759201
  ValerianRey

  committed
  c759201

  Copy full SHA for c759201

• Add docstring to _regularize

  ValerianRey

  committed
  c604c80

  Copy full SHA for c604c80

• Improve formatting

  ValerianRey

  committed
  20aa85c

  Copy full SHA for 20aa85c