Closed tfjgeorge closed 9 months ago
All modified and coverable lines are covered by tests :white_check_mark:
Comparison is base (
d7225d9
) 94.13% compared to head (ade2a1d
) 94.89%.
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
BUGFIX: KFAC for convolutions was scaled by the # of spatial positions but it should not.actually the former implementation corresponds to what is derived in the KFC paper. However arguably the scaling by 1/S with S the number of spatial positions might not be optimal (see discussion in
derivations/kfc.pdf
) thus I instead scaled the test by S and reverted back to KFC expressions.