Stream: helpdesk (published)

Topic: Advice on signaling which part of struct to solve for?

Alec (Jun 13 2023 at 03:18):

In a rewrite of a package, part of what I'm trying to do is to define a generic interface that looks like fit(type_of_struct_to_solve_for,data) and then I use multiple dispatch to define the different logic for different combinations of arguments.

I'm trying to design for the case where type_of_struct_to_solve_for looks like:

struct Params{A,B,C}
   a::A
   b::B
  c::C
end

and C is an abstract type with three subtypes X, Y, and Z.

What I am thinking about is that fitting this object would look like:

fit(Params(1.,1.,X),data) where X is the type X instead of an initialized X(...)

I'm curious if anyone has thoughts or advice on the matter?

Michael Fiano (Jun 13 2023 at 03:40):

Alec said:

and C is an abstract type

Then Params can never be statically known in size, and performance may be an issue in some cases. Otherwise, seems fine?

Alec (Jun 13 2023 at 03:56):

Hmm, after fitting I would want to use Params in high performance contexts. After fitting I would return aParams where C is concrete for further use. Does that avoid the performance problems downstream? Or is the point you are making that fit will have to sacrifice performance as a result?

Daniel González (Jun 13 2023 at 08:14):

For performance, write

struct Params{A,B,C<:AbstractC}
    a::A
    b::B
    c::C
end

So that the type is parameterized by the concrete subtype X, Y or Z (assuming the abstract type is named AbstractC and C is just the placeholder). As for the fit function, you can dispatch on a type (with ::Type{T} where T or ::Type{Params{A, B, C} where {A, B, C}), but you can't mix values and subtypes in an instance in that way directly. I'd suggest passing a, b and C as arguments to fit and build Params there or defining an initializer structure, eg

struct ParamsInitializer{C, A, B}
    a::A
    b::B
    ParamsInitializer{C}(a::A, b::B) where {A, B, C} = new{C, A, B}(a, b)
end

and then fetch the specific subtypes in the dispatch.

Alexander (Jun 13 2023 at 08:44):

I admit I don't fully understand the usecase, but sounds generally similar to what Accessors.jl do. See AccessibleOptimization as a recent example of how it allows specifying parts of a struct to vary in fitting.

Alec (Jun 14 2023 at 14:20):

@Alexander AccessibleOptimizaiton looks awesome! I think it will provide the functionality I'm looking to expose in my package. I'm experimenting with it and have a question about nonbounded solvers. I created an issue on the repo with a MWE.

Last updated: Oct 02 2023 at 04:34 UTC