Periodic Metric about nearestneighbors.jl HOT 11 CLOSED

If it satisfies the requirements for being a metric then the ball tree should be able to handle it.

from nearestneighbors.jl.

ok, then I will give it a try. Can we leave this open so I can ask questions if needed.

from nearestneighbors.jl.

Sure

from nearestneighbors.jl.

I get the following error message; the data I am passing in is a Array{FixedSizeArrays.Vec{3,Float64},1}; I am guessing this fails because FixedSizeArrays.Vec is not a subtype of AbstractArray? Can you confirm?

LoadError: MethodError: no method matching NearestNeighbors.BallTree{V<:AbstractArray{T,1},N,T,M<:Distances.Metric}(::Array{FixedSizeArrays.Vec{3,Float64},1}, ::PerMet.PeriodicEuclidean)
Closest candidates are:
  NearestNeighbors.BallTree{V<:AbstractArray{T,1},N,T,M<:Distances.Metric}{V<:AbstractArray{T,N},M<:Distances.Metric}(!Matched::Array{V<:AbstractArray,1}, ::M<:Distances.Metric; leafsize, reorder, storedata, indicesfor, reorderbuffer) at /Users/ortner/.julia/v0.5/NearestNeighbors/src/ball_tree.jl:39
  NearestNeighbors.BallTree{V<:AbstractArray{T,1},N,T,M<:Distances.Metric}{T<:AbstractFloat,M<:Distances.Metric}(!Matched::Array{T<:AbstractFloat,2}, ::M<:Distances.Metric; leafsize, storedata, reorder, indicesfor, reorderbuffer) at /Users/ortner/.julia/v0.5/NearestNeighbors/src/ball_tree.jl:83
  NearestNeighbors.BallTree{V<:AbstractArray{T,1},N,T,M<:Distances.Metric}{T}(::Any) at sysimg.jl:53
while loading In[12], in expression starting on line 2

from nearestneighbors.jl.

Yes, that is the reason. There are two solutions to this. Either, I relax the constraint of AbstractVector or alternatively you convert / reinterpret your vector to something like https://github.com/andyferris/StaticArrays.jl (which I think has a cleaner interface).

julia> using FixedSizeArrays

julia> using StaticArrays

julia> a = [Vec{3}(1.0,2.0,3.0), Vec{3}(1.0,2.0,3.0)]
2-element Array{FixedSizeArrays.Vec{3,Float64},1}:
 Vec(1.0,2.0,3.0)
 Vec(1.0,2.0,3.0)

julia> reinterpret(SVector{3, Float64}, a, (2,))
2-element Array{StaticArrays.SVector{3,Float64},1}:
 [1.0,2.0,3.0]
 [1.0,2.0,3.0]

from nearestneighbors.jl.

Yes I just checked that staticarrays works. I've been thinking of switching for a while. Anyhow now I get a stack overflow and need to do some debugging. Thank you.

from nearestneighbors.jl.

Finally found time to look at it. My problem is a generic Julia problem - I hope you don't mind that I post this here, but feel free to ask me to move this to Julia users.

So I am trying to define PeriodicEuclidean <: Metric not in the Distances module, but in my own module. I was hoping that your function Distances.evaluate(d::Distances.PreMetric, ...) would correctly call my implementation of evaluate, but it turns out it doesn't and instead calls itself 80_000 times causing a stack overflow. I think there is still something I don't fully understand about dispatch. Probably there are multiple versions of the Distances module?

using Distances, NearestNeighbors, StaticArrays
module PerMet
using StaticArrays, Distances
type PeriodicEuclidean <: Metric end   # at the moment just do Euclidean
function Distances.evaluate(dist::PeriodicEuclidean, a::SVector{3, Float64}, b::SVector{3, Float64}) 
    print("*")
    norm(a-b)
end 
end 
dist = PerMet.PeriodicEuclidean()
X =  reinterpret(SVector{3, Float64}, rand(3, 10), (10,))
# Distances.evaluate(dist, X[1], X[2])   # this calls the correct `evaluate`
pertree = BallTree(X, dist)    # this never calls the "*" implementation but instead creates a stack overflow

from nearestneighbors.jl.

Your evaluate function is too strongly typed. Change the types of a and b to AbstractVector and it will at least work to create the tree. The problem is that evaluate is not only called with SVectors during the creation of the ball tree.

from nearestneighbors.jl.

that seems to have done it - thanks. I'll now go back to trying periodicity.

from nearestneighbors.jl.

Any updated. Should this be closed?

from nearestneighbors.jl.

I didn't find time to work on it - sorry. Best close it and I'll either reopen it or find you on Gitter if I ever get back to it. Thank yu

from nearestneighbors.jl.