James Martens

K-FAC shouldn't use centered activation statistics. However if you wanted to do something like this in some general context, the correct thing would probably be to take a decayed average...

Without more information I can't really help you. But if you want to add a custom loss function to K-FAC you would have to write your own class in loss_functions.py...

The type of approximation they use is the same one, essentially. However, you can't use them interchangeably in the code since convs involve special ops and also "reuse" the different...

Hi, Sorry for the late reply. The K-FAC method is based on the generalized Gauss-Newton matrix (GGN), which required a convex loss (as a function of the network's output). If...

Yeah support for recurrent networks is something we have partially implemented internally. If there's interest I guess we could try to get this out sooner.

Sorry, no. Myself and others have been very busy and haven't had time. If you're interested in using a Kronecker-factored method compatible with RNNs out of the box, you could...

Float64 should work. Although maybe a casting to float32 crept in somewhere. What was the error?

Hi. That "14x" figure applies only to a certain architecture, and isn't meant to be universal. However, I can see from the README that the phrasing suggests otherwise, and so...

Thanks for your persistence trying to get the code to work. Unfortunately I don't really know much about Python library versions and this code isn't really being maintained anymore. Regarding...

I've updated the package documentation and hopefully fixed the issues with the installation.