Provide moving average smoother

[isaacgerg]

Currently, tensorboard has the ability to smooth scalar plots using an exponentially moving average (EMA). Although EMA is useful for visualizing trends in the scalar time series, many times a more appropriate smoother is simply a moving average (MA). Propose to implement MA smoother in tensorboard and have slider to control filter width. MA has the advantage of being a "memory-less" smoother which is quite useful when plotting.

[bmd3k]

Thanks Isaac. Is this for the Scalars dashboard? Time Series? Other?

This could be interesting but we aren't able to prioritize taking it on now. If you are interested, you are welcome to put together a PR with the proposed changes. It would also be nice, as part of that PR, to highlight some scenarios where MA is preferable over EMA so we can have a better idea of the value!

[isaacgerg]

Scalar dashboard. In most cases of averaging, the moving average is preferred over the EMA as the EMA has an infinite memory which is usually an undesirable property. When you expect a "moving" average you expect it to forget over time, the EMA does not do that.

…

On Fri, Apr 22, 2022 at 4:25 PM Brian Dubois ***@***.***> wrote: Thanks Isaac. Is this for the Scalars dashboard? Time Series? Other? This could be interesting but we aren't able to prioritize taking it on now. If you are interested, you are welcome to put together a PR with the proposed changes. It would also be nice, as part of that PR, to highlight some scenarios where MA is preferable over EMA so we can have a better idea of the value! — Reply to this email directly, view it on GitHub <#5680 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AC3KXK4EIVS63W5QNLLAHC3VGMDLZANCNFSM5T45DKRQ> . You are receiving this because you authored the thread.Message ID: ***@***.***>