f-dangel / curvlinops / 22869106971

91%

Build Type

push

github

Committed by web-flow

Commit Message

[ADD] Support non-binary labels for `BCEWithLogitsLoss` (#257)

* [DEL] Remove binary-only restriction for BCEWithLogitsLoss targets (#257)

The BCE loss Hessian diag(σ(f)·(1-σ(f))) is independent of targets,
so the exact and MC GGN are valid for any target in [0, 1]. Remove the
_check_binary_if_BCEWithLogitsLoss safeguard and related warning
infrastructure. Update test cases to use soft labels via rand().

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* [DEL] Remove binary_classification_targets, use soft labels everywhere

Migrate KFAC/EKFAC tests to rand() targets and remove the
pass-through wrapper in KFAC grad outputs setup. Delete the
now-unused binary_classification_targets helper.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* [FIX] Use float64 in gradient_and_loss test, convert BCE targets in change_dtype

The gradient_and_loss test was using float32 which caused flaky
precision failures. Convert to float64 like other tests. Also fix
change_dtype to convert BCEWithLogitsLoss targets (now soft labels).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: f-dangel <felix.dangel@concordia.ca>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Run Details

6 of 6 new or added lines in 4 files covered. (100.0%)

23 existing lines in 5 files now uncovered.

2189 of 2400 relevant lines covered (91.21%)

0.91 hits per line