f-dangel / curvlinops / 12334598491

93%

main: 90%

Build Type

push

github

Committed by web-flow

Commit Message

[ADD] Change KFAC and KFAC inverse to purely PyTorch (#149)

* [ADD] Minimal linear operator interface for PyTorch

* [FIX] Linters

* [ADD] Replicate current base class but inherit from `PyTorchLinearOperator`

* [DEL] Unused import

* [ADD] Implement Hessian as `CurvatureLinearOperator`

* [REF] Combine full and block-diagonal matrix multiply tests

* [ADD] Automatically implement `adjoint` for self-adjoint ops

* [ADD] Implement GGN as `CurvatureLinearOperator`

* [FIX] flake8

* [FIX] Increase tolerance to make test pass

* [DEL] Unused import

* [ADD] Implement `FisherMCLinearOperator` as `CurvatureLinearOperator`

* [ADD] Implement Jacobians as `CurvatureLinearOperator`

* [REF] Extract error test with ``MutableMapping`` and ``batch_size_fn=None``

* [ADD] Class attribute for checking deterministic batches

* [ADD] Extract getters for `_in_shape` and `_out_shape`

* [ADD] Test deterministic batches

* [DOC] Add documentation and type annotations

* [ADD] Implement empirical Fisher as `CurvatureLinearOperator`

* [REF] Extract testing matrix multiplies in expectation

* [FIX] Examples

* [ADD] Check deterministic batches in MC-Fisher (fixes #75)

* [ADD] Implement activation Hessian as `CurvatureLinearOperator`

* [ADD] Implement KFAC as `CurvatureLinearOperator`

* [DEL] Remove original base class

* [FIX] Darglint

* [DEL] Remove unused imports

* [ADD] Declare activation Hessian as self-adjoint

* [FIX] Some tests

* [FIX] KFAC tests

* [FIX] Implement KFAC inverse as Curvature linop

* [FIX] Make tests work on GPU and CPU

* [DOC] Remove `torch_matvec` and `torch_matmat`

* [DEL] Unused import

* [FMT] Apply black

* [REF] Make diff smaller

* [DEL] Unused imports

* [DOC] Update changelog

* [REF] Address review comments

* [DOC] Improve changelog

* [DEL] Remove duplicate entry

Run Details

53 of 53 new or added lines in 2 files covered. (100.0%)

46 existing lines in 2 files now uncovered.

1366 of 1514 relevant lines covered (90.22%)

0.9 hits per line