f-dangel / curvlinops / 8987538117

90%

push

github

[FIX] KFAC scale for `loss_average="batch+sequence"` (#110)

* Fix EF scale for >2d outputs

* Fix scale of MC Fisher for >2d outputs

* Increase tolerance of CG test

* Increase CG precision

* Increase tolerance of CG test (again)

* Increase tolerance of CG test (again)

* Increase tolerance of CG test (again)

* Exclude numerically unstable cases from inverse tests

* Increase tolerance of Jacobian tests

* Adjust loss scale for KFAC-EF

* Add test for KFAC-MC with weight sharing

* Increase tolerance of KFAC (log)det tests

* Add test case for #107 (expand setting scaling issue)

* Fix KFAC scale for batch+sequence loss average

* Fix isort

* Fix test_multi_dim_output

* Change how _num_per_example_loss_terms is inferred and allow setting it explicitly

* Fix darglint and flake8

* Improve docstring for num_per_example_loss_terms

* Minor review fixes

15 of 19 new or added lines in 1 file covered. (78.95%)

1260 of 1427 relevant lines covered (88.3%)

0.88 hits per line