It would be good to add a method exponential_moving_average that takes two KFACLinearOperators and produces another KFAC whose Kronecker matrices were combined using an exponential moving average. This 'smoothing' scheme is extremely common in practise.
It would be good to add a method
exponential_moving_average
that takes twoKFACLinearOperator
s and produces another KFAC whose Kronecker matrices were combined using an exponential moving average. This 'smoothing' scheme is extremely common in practise.