Document atoms

[odow]

Just exploring a few options:

https://jump.dev/Convex.jl/previews/PR610/reference/atoms/#Supported-operations

I still have these to go:

• BroadcastMultiplyAtom
• DiagMatrixAtom
• HcatAtom
• IndexAtom
• MatrixFracAtom
• MultiplyAtom
• QolElemAtom
• QuadOverLinAtom
• QuantumEntropyAtom
• QuantumRelativeEntropyAtom
• RationnalNormAtom
• ReshapeAtom
• TraceLogmAtom
• TraceMpowerAtom
• VcatAtom

[codecov]

Codecov Report

Attention: Patch coverage is 98.32636% with 4 lines in your changes missing coverage. Please review.

Project coverage is 98.21%. Comparing base (459f86b) to head (8598f8a).

Files	Patch %	Lines
src/supported_operations.jl	98.19%	4 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #610   +/-   ##
=======================================
  Coverage   98.21%   98.21%           
=======================================
  Files          89       87    -2     
  Lines        5198     5198           
=======================================
  Hits         5105     5105           
  Misses         93       93           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[ericphanson]

the current docs are kind of centered around the constructor functions rather than the structs, since usually users should use the former. I think though it definitely is nice to document the actual primitives that are available. I think maybe it could be most useful to keep the functions as the first thing the users see/interact with, and add docstrings to them, and then have the atoms as more advanced docs.

[odow]

the current docs are kind of centered around the constructor functions rather than the structs, since usually users should use the former

Yeah... I've realized this. I just wondered about adding docstrings for Base.abs, etc. But perhaps it's okay.

I don't really have a good solution yet.

[odow]

There's also: https://jump.dev/Convex.jl/previews/PR610/manual/operations/

But it doesn't have every atom or reformulation (e.g., abs2 is missing).

I also wonder about how important the DCP rules are. I don't think many people are computing the rules in their head from the docs? They probably just try it to see if it works.

[odow]

What are thoughts on https://jump.dev/Convex.jl/previews/PR610/atoms/

Too verbose? What parts are useful?

[ericphanson]

I like it! IMO not too verbose. In fact I think the reformulations could be more verbose. I think the examples are quite useful, and I like the mini table too about convexity properties. I find them useful at least, and I think when folks are debugging stuff it comes in handy.

[odow]

In fact I think the reformulations could be more verbose

Yes, if we're not focusing on atoms, then it makes sense to have these identical.

I even wonder if it makes sense to move all of the Base.abs etc methods to a single operators.jl file. The atoms (as in, the structs) are standalone, and how exactly we implement the operators is independent. We might, for example, decide to rewrite some of the existing "reformulations" to become atoms, but that wouldn't change the public interface.

[ericphanson]

true, and we have done that in the past (though often the other way, deleting slow or buggy atoms in favor of reformulations)

[blegat]

I really like this so I think we should merge before releasing v0.16. I've added it in the milestone

[odow]

I don't know if this needs to block the release

[blegat]

It does not, you can remove it from the milestone if this PR is not ready to merge

[odow]

@blegat thoughts on keeping the operators in a separate file or keep with atom structs?

[odow]

Here's a script to rebuild the doclist

function build_atoms_md(input_filename, filename)
    data = read(input_filename, String);
    reform = joinpath(dirname(input_filename), "reformulations")
    for file in readdir(reform; join = true)
        data *= read(file, String)
    end
    header = """# Atoms

    Convex.jl supports the following functions. These functions may be composed
    according to the [DCP](http://dcp.stanford.edu) composition rules to form new
    convex, concave, or affine expressions.
    """
    open(filename, "w") do io
        return print(io, header)
    end
    block_fn = r"\"\"\"\nfunction (.+?\(.+?\))"ms
    inline_fn = r"\"\"\"\n([a-zA-Z0-9\_\.\:\+\-\*\/\^]+\(.+?\)) = "m
    function fn_name(x)
        x = replace(x, "Base." => "", "LinearAlgebra." => "", ":" => "")
        return String(x[1:(only(findfirst("(", x))-1)])
    end
    args = Dict{String,Vector{String}}()
    block_fns = collect(eachmatch(block_fn, data))
    inline_fns = collect(eachmatch(inline_fn, data))
    for m in vcat(block_fns, inline_fns)
        is_public = occursin("Base.", m[1]) ||
                    occursin("LinearAlgebra.", m[1]) ||
                    Symbol(fn_name(m[1])) in names(Convex)
        if !is_public
            continue # Not exported
        end
        signature = replace(
            m[1], 
            r"[a-zA-Z]+?::" => "::",
            "LinearAlgebra." => "Convex.",
            "AbstractExpr" => "Convex.AbstractExpr",
            "Value" => "Convex.Value",
            "Constant" => "Convex.Constant",
        )
        if !(startswith(signature, "Base.") || startswith(signature, "Convex."))
            signature = "Convex.$signature"
        end
        push!(get!(args, fn_name(m[1]), String[]), signature)
    end
    open(filename, "a") do io
        for name in sort(collect(keys(args)))
            signature = join(sort(args[name]), "\n")
            println(
                io, 
                """
                ## `$name`
                ```@docs
                $signature
                ```
                """,
            )
        end
        return
    end
    return
end

build_atoms_md("src/supported_operations.jl", "docs/src/manual/atoms.md")

[blegat]

@blegat thoughts on keeping the operators in a separate file or keep with atom structs?

I like it in a separate file

[odow]

So what do we think about this?

[ericphanson]

maybe "Operations"?

[ericphanson]

it's too bad these aren't searchable.

Should we use a word also?

Suggested change

	## `*`
	## Multiplication (`*`)

[odow]

I've moved this to the API reference, so I think I'm going to merge, unless there are other comments.