Using CircularArrayBuffers for non-iterable data

[3rdCore]

Hello, I have an issue regarding the latest updates of the package CircularArrayBuffers.jl

I am using the package ReinforcementLearning.jl to train a RL-based agent to make decisions on graphs. The trajectory is stored in a CircularArraySARTTrajectory object, in which environment states are stored in a CircularArrayBuffers.

In my case, my environment state is a custom class called DefaultTrajectoryState which is a non-tabular data. However, since the latest update of the package, CircularArrayBuffers expects the data to be iterable in order to be broadcasted for the case we want to push multiple states in the CircularArrayBuffers, which is not my case.
I getthe following error :
Got exception outside of a @test MethodError: no method matching iterate(::SeaPearl.DefaultTrajectoryState)

The error comes from the following function of CircularArrayBuffers.jl :

function Base.push!(cb::CircularArrayBuffer{T,N}, data) where {T,N}
    if cb.nframes == capacity(cb)
        cb.first = (cb.first == capacity(cb) ? 1 : cb.first + 1)
    else
        cb.nframes += 1
    end
    if N == 1
        i = _buffer_frame(cb, cb.nframes)
        if ndims(data) == 0
            cb.buffer[i:i] .= data[]
        else
            cb.buffer[i:i] .= data                    <----- ERROR
        end
    else
        cb.buffer[ntuple(_ -> (:), N - 1)..., _buffer_frame(cb, cb.nframes)] .= data
    end
    cb
end

It is designed "on purpose"? Knowing that should I consider encapsulating my state into a Vector containing one single element?

[3rdCore]

One solution I found was to virtually overwrite the base function iterate and directly return the element.

[findmyway]

Hi @3rdCore ,

Thanks for reporting.

The relevant change seems to be the broadcasting as shown below:

+        i = _buffer_frame(cb, cb.nframes)
        if ndims(data) == 0
-            cb[cb.nframes] = data[]
+            cb.buffer[i:i] .= data[]
        else
-            cb[cb.nframes] = data
+            cb.buffer[i:i] .= data
        end

v0.1.5...master

I think here it was meant to overcome the scalar operation when processing CuArrays. @darsnack may confirm this.

[findmyway]

One solution I found was to virtually overwrite the base function iterate and directly return the element.

Yeah, this will do the trick.

Previously, the data is assumed to be an AbstractArray or a Number. So a customized data type is not taken into account. Let me have a think of how to remove this assumption.

[findmyway]

A MWE:

julia> b = CircularArrayBuffer([:a, :b])
CircularVectorBuffer(::Vector{Symbol}) with eltype Symbol:
 :a
 :b

julia> push!(b, :c)
ERROR: MethodError: no method matching ndims(::Symbol)
Closest candidates are:
  ndims(::Base.Generator) at /mnt/E4E0A9C0E0A998F6/software/julia-1.7.1/share/julia/base/generator.jl:53
  ndims(::Base.Broadcast.Broadcasted{<:Any, <:Tuple{Vararg{Any, N}}}) where N at /mnt/E4E0A9C0E0A998F6/software/julia-1.7.1/share/julia/base/broadcast.jl:247
  ndims(::Base.Iterators.ProductIterator) at /mnt/E4E0A9C0E0A998F6/software/julia-1.7.1/share/julia/base/iterators.jl:983
  ...
Stacktrace:
 [1] push!(cb::CircularVectorBuffer{Symbol, Vector{Symbol}}, data::Symbol)
   @ CircularArrayBuffers /mnt/E4E0A9C0E0A998F6/github/CircularArrayBuffers.jl/src/CircularArrayBuffers.jl:95
 [2] top-level scope
   @ REPL[4]:1

[darsnack]

I think here it was meant to overcome the scalar operation when processing CuArrays. @darsnack may confirm this.

That's correct, and I did not consider about the non-numeric case here.

[findmyway]

@JuliaRegistrator register

[JuliaRegistrator]

Registration pull request created: JuliaRegistries/General/56748

After the above pull request is merged, it is recommended that a tag is created on this repository for the registered package version.

This will be done automatically if the Julia TagBot GitHub Action is installed, or can be done manually through the github interface, or via:

git tag -a v0.1.9 -m "<description of version>" 73ccf07ecb35ab9f0b4024f35e209154af303a45
git push origin v0.1.9

[3rdCore]

Thanks for the support guys !

Amazing work !