Rasmus Henningsson (Jun 07 2024 at 10:38): Rasmus Henningsson (Jun 07 2024 at 10:38):Quite often I want to optionally pass a kwarg based on some condition, like this:
if condition
f(; a=1)
else
f()
end
But this quickly becomes verbose if there are many other args to f (and potentially other "optional" kwargs).
The above example can be turned into a one-liner, for instance like this:
f(; (condition ? (;a=1) : (;))...)
But that doesn't read very nicely.
Another approach, which scales better to situations where f takes more arguments, would be:
kwargs = Pair{Symbol,Any}[]
condition && push!(kwargs, :a=>1)
f(; kwargs...)
But it involves creating a temporary a Vector (or Dict) with Anys. So it doesn't feel great either.
Is there a better way?
Júlio Hoffimann (Jun 07 2024 at 11:10):I usually define the kwargs in all branches of the code, and always pass them. If a function accepts kwargs, it probably has a default that can be passed anyways:
a = condition ? 1 : default_a()
f(; a) # always forward some value
Yes, that's a good solution and it often works.
But sometimes there is no default value, or I don't want to specify the default (it can be complicated or defined in a dependency). So it would be nice to be able handle these situations as well.
jar (Jun 07 2024 at 18:18):g(a) = f(1,2,3,4,5; a)
condition ? g(1) : g(2)
No, that's not what I want. That could also be written compactly as:
f(1,2,3,4,5; a=condition ? 1 : 2)
What I want is to not pass the kwarg at all in one of the cases.
jar (Jun 07 2024 at 20:26):julia> f(;a=1) = a;
julia> f(;(rand(Bool) ? () : (:a=>2,))...)
2
julia> f(;(rand(Bool) ? () : (:a=>2,))...)
1
julia> f(;(rand(Bool) ? (;) : (;a=2))...)
2
julia> f(;(rand(Bool) ? (;) : (;a=2))...)
1
if there are many other args to
f(and potentially other "optional" kwargs).
This happens indeed, and I personally prefer to fully build the kwargs namedtuple and only then pass it to the function. This makes it more convenient to have eg some keywords depending on others.
Here's a nice simple and readable way to do that:
julia> using Accessors, DataPipes
julia> kwargs = @p let
(;a=1)
@insert __.b = 2
rand(Bool) ? __ : @insert __.c = __.a * 10
end
(a = 1, b = 2, c = 10)
julia> f(;kwargs...)
I hate to be critical but to me that's neither nice, nor simple, nor readable.
For these situations I generally just resort to some variation of
kwargs = (;)
if condition
kwargs = (; a = 1)
end
f(; kwargs...)
I usually just use an if statement like Gunnar's as well. I personally don't have no idea how to read that Accessors syntax.. If you have kwargs in the function signature (where the other function is being called, not the f in these examples) you can also work with that object directly and delete/add keys to it.
aplavin (Jun 08 2024 at 13:37):resort to some variation of ...
This is not type-stable though, so only suitable for pieces when performance doesn't matter.
aplavin (Jun 08 2024 at 13:41):I personally don't have no idea how to read that Accessors syntax..
Of course, any syntax requires some familiarity to be readable and understandable :)
In this specific case, code looks reasonably nice even without Accessors:
julia> kwargs = @p let
(;a=1)
(;__..., b = 2)
rand(Bool) ? __ : (;__..., c = __.a * 10)
end
although the intent is arguably more clean with @insert. Also, it has @delete/@set/@modify for more complex usecases.
Gunnar Farnebäck (Jun 08 2024 at 14:04):This is not type-stable though, so only suitable for pieces when performance doesn't matter.
Yes, of course. Is that any different with your proposal?
aplavin (Jun 08 2024 at 22:56):Ehm... Surely it's type stable for reasonable conditions (not rand(Bool) :) ). I wouldn't have suggested a non-type-stable solution here without a major caveat/disclaimer.
julia> f(x, y) = @p let
(;a=1)
@insert __.b = 2
ndims(x) == 1 ? __ : @insert __.c = __.a * 10
haskey(y, :somekey) ? __ : @insert __.d = y
end
julia> @code_warntype f([1], (mykey=1,))
<...>
Body::@NamedTuple{a::Int64, b::Int64, d::@NamedTuple{mykey::Int64}}
<...>
julia> @code_warntype f([1;;], (mykey=1, somekey=2))
<...>
Body::@NamedTuple{a::Int64, b::Int64, c::Int64}
<...>
There's some middle ground between rand() and conditions that can be determined from the input types though. Anyway I don't think it's that often it's critical because
kwargs will be a small union and union splitting will apply.But sure, if this happens on the way to a frequently called function, you should review what it does to performance.
Eric Hanson (Jun 10 2024 at 10:03):I don't see how this f is helping if the user needs to pass in a namedtuple in the first place? If you had that, you would just splat it in. If you don't have that, and you need to construct a namedtuple with variable number of keys depending on a Bool, well then that's by definition type-unstable (the type of the output depends on a runtime value).
Last updated: Nov 06 2024 at 04:40 UTC