I recently started using Julia, and am at the moment struggling with some aspects of typing. I am trying to define a type TensorTrain (https://www.researchgate.net/profile/Ivan_Oseledets2/publication/220412263_Tensor-Train_Decomposition/links/5bbfb5c5299bf1004c5a56e3/Tensor-Train-Decomposition.pdf), but I cannot get one of the methods I define to be stable:
abstract type AbstractTensorTrain <: Any end
struct TensorTrain{T<:AbstractFloat} <: AbstractTensorTrain
cores::Vector{Array{T,3}}
end
Base.size(t::TensorTrain, d::Int) = size(t.cores[d],2)
Base.size(t::TensorTrain) = ntuple(d->size(t,d),length(t.cores))
However if I run:
@code_warntype TensorTrain([rand(2,3,4)]);
I get:
Variables
#self#::Core.Compiler.Const(size, false)
t::TensorTrain{Float64}
#19::var"#19#20"{TensorTrain{Float64}}
d::Int64
Body::Tuple{Vararg{Int64,N} where N}
1 ─ %1 = Base.getproperty(t, :cores)::Array{Array{Float64,3},1}
│ (d = Main.length(%1))
│ %3 = Main.:(var"#19#20")::Core.Compiler.Const(var"#19#20", false)
│ %4 = Core.typeof(t)::Core.Compiler.Const(TensorTrain{Float64}, false)
│ %5 = Core.apply_type(%3, %4)::Core.Compiler.Const(var"#19#20"{TensorTrain{Float64}}, false)
│ (#19 = %new(%5, t))
│ %7 = #19::var"#19#20"{TensorTrain{Float64}}
│ %8 = Main.ntuple(%7, d)::Tuple{Vararg{Int64,N} where N}
└── return %8
Why is this?
Because
d = length(t.cores)
is only known at runtime, so there's no wayntuple
could statically infer theN
parameter. There is a method that takes aVal
, which is useful for type stability in other cases, but it does not change anything here -- thed
is still dynamic.You could make
cores
anNTuple
orStaticVector
, though (which are really the same things), then this will work. But do check whether it really makes a difference, before delving into type-level programming.