In-place version of knn about nearestneighbors.jl HOT 6 CLOSED

Yeah, I've thought about this too.

Just to check, you know that you can query knn against multiple points in one call, which is faster than one at a time?

from nearestneighbors.jl.

Yes, I know. It is just that I need to do it point by point in this case. In places where I can pass a bunch of points, I am already doing it. 👍

from nearestneighbors.jl.

Do you have some performance benchmarks? I think I tried this a while ago and while there was a speedup I don't recall it as "huge".

from nearestneighbors.jl.

I don't have a benchmark yet, can try to run it locally and report here 👍

from nearestneighbors.jl.

seems like this issue was closed? I wonder what's the progress of in-place KNN, in particular in-place KDTree?

I have a problem that needs to do KNN (for density estimation) millions of times on matrices with the same size.
Ideally, I could create a KDTree, then re-use the memory for new data, e.g.

kdtree = KDTree(data)

set!(kdtree, newdata)    # in-place update of kdtree object with new data

Thanks.

from nearestneighbors.jl.

I think this feature will be useful. Querying knn against multiple points certainly helps, but if we do it $n$ times on $m$ querying points, it will allocate $n$ vectors of length $m$. When $n$ and $m$ are large, this will cause more frequency garbage collection.

from nearestneighbors.jl.