Functors.jl provides tools to express a powerful design pattern for dealing with large / nested structures, as in machine learning and optimisation. For large machine learning models it can be cumbersome or inefficient to work with parameters as one big, flat vector, and structs help manage complexity; but it is also desirable to easily operate over all parameters at once, e.g. for changing precision or applying an optimiser update step.
Functors.jl provides fmap to make those things easy, acting as a 'map over parameters':
julia> using Functors
julia> struct Foo
x
y
end
julia> model = Foo(1, [1, 2, 3])
Foo(1, [1, 2, 3])
julia> fmap(float, model)
Foo(1.0, [1.0, 2.0, 3.0])It works also with deeply-nested models:
julia> struct Bar
x
end
julia> model = Bar(Foo(1, [1, 2, 3]))
Bar(Foo(1, [1, 2, 3]))
julia> fmap(float, model)
Bar(Foo(1.0, [1.0, 2.0, 3.0]))Note
Up to to v0.4, Functors.jl's functionality had to be opted in on custom types via the @functor Foo macro call.
With v0.5 instead, this is no longer necessary: by default any type is recursively traversed up to the leaves
and ConstructionBase.constructorof is used to reconstruct it.
In order to opt-out of this behaviour and make a type non traversable you can use @leaf Foo.
Most users should be unaffected by the change and could remove @functor from their custom types.
The workhorse of fmap is actually a lower level function, functor:
julia> children, reconstruct = Functors.functor(Foo(1, [1, 2, 3]))
((x = 1, y = [1, 2, 3]), Functors.var"#3#6"{DataType}(Foo))
julia> reconstruct(map(float, children))
Foo(1.0, [1.0, 2.0, 3.0])functor returns the parts of the object that can be inspected, as well as a reconstruct function that takes those values and restructures them back into an object of the original type.
To include only certain fields, pass a tuple of field names to @functor:
julia> struct Baz
x
y
end
julia> @functor Baz (x,)
julia> model = Baz(1, 2)
Baz(1, 2)
julia> fmap(float, model)
Baz(1.0, 2)Any field not in the list will not be returned by functor and passed through as-is during reconstruction. This is done by invoking the default constructor, so structs that define custom inner constructors are expected to provide one that acts like the default.
It is also possible to implement functor by hand when greater flexibility is required. See here for an example.
For a discussion regarding the need for a cache in the implementation of fmap, see here.
Use exclude for more fine-grained control over whether fmap descends into a particular value (the default is exclude = Functors.isleaf):
julia> using CUDA
julia> x = ['a', 'b', 'c'];
julia> fmap(cu, x)
3-element Array{Char,1}:
'a': ASCII/Unicode U+0061 (category Ll: Letter, lowercase)
'b': ASCII/Unicode U+0062 (category Ll: Letter, lowercase)
'c': ASCII/Unicode U+0063 (category Ll: Letter, lowercase)
julia> fmap(cu, x; exclude = x -> CUDA.isbitstype(eltype(x)))
3-element CuArray{Char,1}:
'a': ASCII/Unicode U+0061 (category Ll: Letter, lowercase)
'b': ASCII/Unicode U+0062 (category Ll: Letter, lowercase)
'c': ASCII/Unicode U+0063 (category Ll: Letter, lowercase)