Move replica axis to front everywhere
Created by: APJansen
This just puts the replica axis first wherever it wasn't yet.
It gives identical results, to the last digit, on:
-
Basic runcard -
feature scaling runcard -
flavour basis -
qed
In the MSR normalization layer, I do transpose them back to their old shape just because having the flavor axis first there makes for cleaner code. I think the performance hit there is negligible, but this can always be optimized later on.