f-dangel / curvlinops / 24531067658

92%

Build Type

push

github

Committed by web-flow

Commit Message

[ADD] Support pre-conditioner in CG and Neumann inverses (#254)

* Add preconditioner support for CG and Neumann inverse operators

* Add CUDA-gated tests for inverse preconditioner support

* Add more CUDA inverse-operator tests and register cuda marker

* Add CPU efficiency stress tests and optimize inverse preconditioner paths

* Add CPU test asserting preconditioned Neumann outperforms naive

* Add overparameterized single-linear CPU inverse operator test

* Add CUDA smoke tests for KFAC/EKFAC on real Conv/Linear nets

* remove unnecessary tests

* adding example for preconditioning

* remove unecessary modification

* remove the duplicate test in test_inverse.py

* update the precondition logic and update the tests

* add the docstring example and the test

* fix the unnecessary difference

* update Neuamnn preconditioner type

* update the tests, add one test case to test preconditioning with KFAC/EKFAC

* update the inverse.py

* fix the test

* make the parameter compatible with new backend

* update docstring for inverse.py, fix the typos

* update the changelog

* fix ruff lint errors in test_inverse.py

* fix the ruff

---------

Co-authored-by: liao-disen <dyson.liao@gmail.com>

Coverage Stats

22 of 23 new or added lines in 1 file covered. (95.65%)

24 existing lines in 4 files now uncovered.

2520 of 2745 relevant lines covered (91.8%)

0.92 hits per line