nice work, may I use it

[JeffreySarnoff]

I would like to replace my default rolling median (which is rolling(VectorizedStatistics.vmedian, ..)) with a call to yours that may be post-adjusted to comport with my api.

[Firionus]

Hey Jeffrey, great to see you dropping by :blush:

And yes, you may use it. The code is MIT-licensed, so if you really want to copy it over or fork, just include the license from this repo and you are good to go. I've always envisioned that FastRunningMedian.jl would one day be part of a bigger package, it always felt a bit too small for me. However, I don't think the algorithm here is done yet. I still want to explore BTrees and better performance for small windows before I'll tag the 1.0. And such experimentation is certainly easier in a smaller, more contained package. Progress has been slow though, I only work on it occasionally when I find some time and motivation.

For the moment I would prefer if you could use FastRunningMedian as a dependency to RollingFunctions. Is there anyting I can do to ease that integration? running_median with tapering=:none should already be really close to rollmedian. I'd even consider adding an option tapering=:beginning_only or tapering=:start to give the same boundary conditions as runmedian. Would that help?

[JeffreySarnoff]

I am interested in using your work in version 1 of RollingWindows (whatever its name).

The API is different and more flexible see Introducing Windowed Functions

One feature yet to be implemented is to roll a function over a data vector where the window width may vary as the function rolls, the successive widths are given as an integer vector.

I am happy to take a look toward easy integration.

[Firionus]

That's cool! Some thoughts of mine:

• rolling you may be able to implement with running_median and tapering=:none
• running we may be able to add as a boundary condition to this package. I've opened an issue (#26).
• tiling, other special boundary conditions, step sizes or changing window sizes you can implement with the stateful API. For examples, see how I've done it in this package: https://github.com/Firionus/FastRunningMedian.jl/blob/master/src/FastRunningMedian.jl
• roll! currently errors when you haven't grown the windows to its declared maximum size, which is a problem for varying window sizes. You can use shrink! followed by grow!, though, where the performance should only be lower by a constant factor. I've opened an issue on this (#25), since there is no inherent reason for the current behavior.
• This package is not very generic, customizable or polished in its output datatype yet. median yields a Float64 with even window size, but the element data type with odd window size, which is usually converted to Float64 since that is what the output is allocated as (https://github.com/Firionus/FastRunningMedian.jl/blob/93402ad4300d9fd4af4f16c1f68d254676d3e925/src/FastRunningMedian.jl#L81). EDIT: The output eltype is now customizable :heavy_check_mark:
• The input only accepts Real data types, and no missing. At least NaNs in float inputs are finally supported. (see #22)