getindex not implemented on ProductGroup and is there a way to view components?

[Affie]

julia> M = ProductManifold(SpecialEuclidean(2), SpecialEuclidean(2));

julia> G = ProductGroup(M);

julia> p = identity_element(G);

julia> p[G,1]
ERROR: MethodError: no method matching getindex(::ProductRepr{Tuple{ProductRepr{Tuple{Vector{Float64}, Matrix{Float64}}}, ProductRepr{Tuple{Vector{Float64}, Matrix{Float64}}}}}, ::ProductGroup{ℝ, Tuple{GroupManifold{ℝ, ProductManifold{ℝ, Tuple{TranslationGroup{Tuple{2}, ℝ}, SpecialOrthogonal{2}}}, Manifolds.SemidirectProductOperation{RotationAction{TranslationGroup{Tuple{2}, ℝ}, SpecialOrthogonal{2}, LeftAction}}}, GroupManifold{ℝ, ProductManifold{ℝ, Tuple{TranslationGroup{Tuple{2}, ℝ}, SpecialOrthogonal{2}}}, Manifolds.SemidirectProductOperation{RotationAction{TranslationGroup{Tuple{2}, ℝ}, SpecialOrthogonal{2}, LeftAction}}}}}, ::Int64)
Closest candidates are:
  getindex(::ProductRepr, ::ProductManifold, ::Union{Colon, Integer, Val, AbstractVector}) at ~/.julia/packages/Manifolds/WsRwu/src/manifolds/ProductManifold.jl:627
  getindex(::ProductRepr, ::VectorBundle, ::Symbol) at ~/.julia/packages/Manifolds/WsRwu/src/manifolds/VectorBundle.jl:511
Stacktrace:
 [1] top-level scope
   @ REPL[6]:1

I added this to carry on and it seems to work.

Base.@propagate_inbounds function Base.getindex(
    p::ProductRepr,
    M::ProductGroup,
    i::Union{Integer,Colon,AbstractVector,Val},
  )
  return get_component(base_manifold(M), p, i)
end

EDIT: put in the correct stack trace

[kellertuer]

Thanks, we should definetly do something about this. The one thing we could check is, whether its reasonable to make this more general, i.e. to @trait_function both getindex and setindex! Sith M as first parameter. What do you think @mateuszbaran ?

[mateuszbaran]

I'll fix it, this definitely should work. I wouldn't use @trait_function though, getindex and setindex! have way too many loosely typed methods for that.

[mateuszbaran]

Regarding viewing components, do you mean viewing parts of an affine or screw matrix for SE(n)? Because I don't really see how viewing would be useful for ProductRepr or ArrayPartition.

[Affie]

Thank you very much for the quick fix.

Regarding viewing components, do you mean viewing parts of an affine or screw matrix for SE(n)? Because I don't really see how viewing would be useful for ProductRepr or ArrayPartition.

I'm trying to improve optimization performance and currently I copy values from a container to the cost function based on the structure of the graph.

I experimented with creating a product manifold with all the variables (tired with 100 for now) but I don't want to copy the data to the cost function every iteration. It works something like this:
With a structure:
x1---f1---x2---f1----x3----f2----x4----f2----x5

the cost is build from the structure, so i would evaluate:
cost = f1(x1,x2) + f1(x2,x3) + f2(x3,x4) + f2(x4,x5)

I previously used (before upgrading to Manifolds.jl) a view, f1(view(vars[x1]), view(vars[x2]), etc.

I hope that is clear enough, that's why I asked how to create a view of the components.

There is a lot of room for improvement, but I'm trying to take small steps at a time.

[mateuszbaran]

In ProductRepr and ArrayPartition representations getindex doesn't need a view because you get the component without any copying. For the affine matrix representation, you should already get a view.

By the way, could you should me a small-ish example of code you are benchmarking? You don't need to remove AMP/IncrementalInference stuff. Achieving good performance with Manifolds.jl is not always intuitive.

One more tip, if you have a product of something like 100 manifolds, you should try factoring it into a product of a few PowerManifolds by grouping components with the same manifold together. ProductManifold is actually designed to perform the best with up to maybe 10 manifolds? I'm not even sure because I've never needed more than 3.

[mateuszbaran]

Also keep in mind that you can have things like a product manifold of power manifolds of product manifolds.

[Affie]

In ProductRepr and ArrayPartition representations getindex doesn't need a view because you get the component without any copying. For the affine matrix representation, you should already get a view.

Yes, of course, thanks. I was stuck in the way of thinkging pre Manifolds.jl code where I just had a vector and I had to do a view for slices to avoid copies. But when I read your answer I realized it.

By the way, could you should me a small-ish example of code you are benchmarking? You don't need to remove AMP/IncrementalInference stuff. Achieving good performance with Manifolds.jl is not always intuitive.

So far I'm working here JuliaRobotics/IncrementalInference.jl#1546 (comment) (this is one of the factors i'm testing with), so we can perhaps continue the discussion there. I'm still experimenting with how exactly the implementation should look. Therefore, I'll send you an example as soon as I've made enough progress there. I first want to refactor into using this tip you gave:

One more tip, if you have a product of something like 100 manifolds, you should try factoring it into a product of a few PowerManifolds by grouping components with the same manifold together. ProductManifold is actually designed to perform the best with up to maybe 10 manifolds? I'm not even sure because I've never needed more than 3.

That is a great idea. Thanks, I will try it out.