A kernel is causal if it only utilizes current and previous inputs in the rolling dimension. In most cases the rolling dimension is interpreted as a time dimension.
CausalKernels.jl is a small package that allows you to apply causal kernels to arrays, mainly for time series applications. This package is a performant replacement for MapSlide.jl, however "expanding" operations are currently unimplemented here.
Install this package to your Julia project environment as you would any other package from a Git repo.
From the Julia REPL:
julia> ]
(MyProject) pkg> add https://github.com/HaedronResearch/CausalKernels.jl
The package exports the functions roll and rollp.
We don't use any extension support from StaticKernels.jl.
In the future wrappers for extension across non-rolling dimensions may be included.
We export a function rollp that prepends padding to the roll dimension after the rolling map is applied.
StaticKernels.jl is currently the backend for this package.
The code is fairly simple. We limit the kernel bounds from a negative lookback -(w-1) (where w is the window size, w≥0) to 0 (current time) and fully expand across the other dimensions.
This means the kernel function fn for roll/rollp can access data across all non-time slices in each time window.
By contrast using the StaticKernels.jl @kernel macro will always slide a minimal kernel across all dimensions. Not acessing future values in
the kernel function will make the kernel causal, but for tabular data / time series we often do not want to stride across non-roll dimensions.
If you want this behavior it is easy to define and run a kernel with StaticKernels.jl.
In the future part or all of this package may be reimplemented with LoopVectorization.jl.