Discussion: syntax for plotting from a DataFrame about statsplots.jl HOT 7 CLOSED

To make the discussion more complete, here is an example of a simplified @with macro called @given:

macro given(d, x)
    esc(_given(d,x))
end

_given(d, x) = x

function _given(d, x::Expr)
    (x.head == :quote) && return :(haskey($d, $x) ? $d[$x] : $x)
    (x.head == :call) && x.args[1] == :^ && length(x.args) == 2 && return x.args[2]
    return Expr(x.head, _given.(d, x.args)...)
end

Usage:

using StatPlots, DataFrames,RDatasets
iris = RDatasets.dataset("datasets","iris")
@given iris scatter(:PetalLength, :PetalWidth, group = :Species, legend = :topleft)
iris[:topleft] = rand(150)
@given iris scatter(:SepalLength, :SepalWidth, group = :Species, legend = ^(:topleft))

Advantages over @with :

• No need to escape the symbols unless there actually is ambiguity
• No DataFramesMeta dependency
• It's easy to adjust it to include other datasources, provided they allow extracting one column (which I'm assuming all of them do)

from statsplots.jl.

I am sure I had written this already here, but apparently I forgot to post it: as a first non-breaking step, if there is consensus, I'd be happy to make a PR with the @given macro, mentioning it instead of @with in the README. Of course we should maybe spend some time figuring out what's the best name for @given.

Even if we get the plot(df,...) syntax to work, it's good to also have the macro version because it retains some advantages that are just impossible to achieve with a function:

• Manipulating columns doesn't require wrapping everything in a :(). For example one can write

@given iris scatter(log.(:PetalLength), :PetalWidth .+ 1, group = :SepalWidth .> 3, legend = :topleft)

and it would just work (I'm even tempted to add a where kw to also select what data to plot in the plot command)

• Manipulating columns doesn't require a runtime eval

@mkborregaard : what's your take on this?

from statsplots.jl.

Sorry for not responding - I'm reading and thinking carefully but haven't had a chance to do it justice - please have a little patience :-)

from statsplots.jl.

My gut reaction is that adding additional syntax would be unfortunate given the goals of the Plots.jl ecosystem. Then again, the functionality is more important, and if this would fix some of the issues with using DataFrames to their full extent...

Alas, I don't understand enough of the backend stuff to comment on implementation.

from statsplots.jl.

I think this looks really cool. It's certainly a lot simpler than the current non-macro framework!! If you'd make a PR with the given macro I'll try it out in some of the other cases I've been considering. This involves e.g. passing a part of a DataFrame as a matrix but that should be possible with [:a :b] here I think.

I think we should be sure to be compatible not just with DataFrames but with abstract table formats such as DataTables, JuliaDB, IndexedTables etc but I'll need to check the current status of those to see if that could work.

Another thing I'd like the macro to do was to replace all NAs with NaNs for number fields, and "" / Symbol() for string/symbol fields, so that the plot function doesn't automatically fail when passing it a DataFrame with NAs in it.

In short, I agree that a macro approach is the best way to deal with table symbols.

from statsplots.jl.

Your intuition is correct, using the horizontal concatenation syntax:
@given iris corrplot([:SepalLength :SepalWidth :PetalLength :PetalWidth])

I'll add the missing value part and make a PR tomorrow.

As for the support of any table-like structure, maybe the easiest is to support IterableTables? I'm also not sure what the state of affairs is in the data echosystem...

from statsplots.jl.

This should be fully resolved now.

from statsplots.jl.