Adding support for FunctionChains.jl?

[oschulz]

In the spirit of #199, which resulted in InverseFunctions and ChangesOfVariables: I've been looking for a lightweight package that defines composed functions that go beyond Base.ComposedFunction, so that allow for arrays/iterators/tuples of functions, type-stable and performant. I couldn't really find anything lightweight that provides this (, though obviously the ML packages all have their own equivalent, and Bijectors has Composed.

Update: There's FunctionChains.jl now.

(@giordano, did I overlook something?)

@willtebbutt, @devmotion, @torfjelde, @cscherrer would there be interest in creating such a common lightweight package like that? I'd volunteer to draft it.

[cscherrer]

Sounds good! I've often hoped for a common interface for this sort of thing, especially the currently-incompatible diffeomorphism implementations that are spread across a few packages like this one.

One thing I think it's important to account for is that high-dimensional transforms are themselves often implemented in terms of iteration. So we need ways to compose these iterations "vertically" and "horizontally". I've mostly worked with TransformVariables; I'm not sure how similar or different Bijectors approach might be

[oschulz]

I've registered FunctionChains.jl. It should, in principle, be able to replace Bijectors.Composed, though the result will obviously not be a subtype of Bijector (is that important, though?).

[oschulz]

I could add support for FunctionChains to Bijectors, if there's interest, e.g. via bi-directioinal conversion in a package extension?

[devmotion]

More a general question, not an actual objection at this point, but similar to the discussion in JuliaDiff/AbstractDifferentiation.jl#88 it's unclear to me if bi-directional conversions should be defined here or if only Base.convert(::Type{T}, ...) for types that Bijectors owns should be defined in a possible extension of it.

[oschulz]

bi-directional conversions should be defined here or if only Base.convert(::Type{T}, ...) for types that Bijectors owns should be defined in a possible extension of it.

I agree, it would be nice to have a kind of convention here, to avoid "extension piracy". On the other hand, maintenance wise it would be much easier (versioning, CI, etc.) to not split and extension that "bridges" two package in respect to a specific topic, but to host the complete extension in one of them. Ideally the package maintainers would come to some agreement where to put the extensions in cases where the choice is not obvious.

If the two packages is very lightweight/abstract, I think would would technically better to host the extension in the "heavy" package. That way, one can use a non-weak depencency instead of Requires on Julia <v1.9, which will improve load times there. In this specific case here, for example, FunctionChains can be very lightweight (I'll move most of it's dependencies into extensions, and Bijectors already shares pretty much all of them already anyway). So Bijectors depending on FunctionChains on Julia <= v1.8 and using and extension on >= v1.9 seems better, load-time wise, than FunctionChains using Requires on Julia <= v1.8 (since Bijectors is much heavier) and using an extension on >= v1.9.

[torfjelde]

I could add support for FunctionChains to Bijectors, if there's interest, e.g. via bi-directioinal conversion in a package extension?

Shouldn't FunctionChains just "work" with Bijectors? Or am I missing something?

[oschulz]

I think you're right @torfjelde - I hadn't noticed that Bijectors.Composed <: Bijector has gone. So Bijectors wouldn't require FunctionChain to be a subtype of Bijector anymore and no conversion is necessary?

[torfjelde]

Exactly:) Now it should "just work" 👍

[oschulz]

Thanks @torfjelde !