Plot function against unit

[gustaphe]

I looked into the possibility of making function plots (plot(f)) work automatically, but since we don't (and can't) set the limits to unitful quantities, I think that's not going to happen.

To me, the second best thing is being able to plot a function against all of the x axis, while supplying the unit of the x axis. Heuristically, "plot this function against 'seconds'". This is easier than figuring out how to modify a function so that it works in a unitful plot (actually just plot(x->f(x*u)), but that always takes me a while to wrap my head around).

model(x) = 1u"V"*exp(-((x-0.5u"s")/0.7u"s")^2)
t = randn(10)u"s" # Sample points
U = model.(t) + randn(10)u"dV" .|> u"V" # Noisy acquicisions
plot(t, U; xlabel="t", ylabel="U", st=:scatter, label="Samples")
plot!(model, t; st=:scatter, label="Noise removed")
plot!(model, u"s"; label="True function")

Also documents the existing plot(f, x).

Theoretically, we could use any unoccupied signature and use the axis unit parameter, but I couldn't really think of a simpler signature than this. Maybe it should ignore the value of the unit argument, as it stands f(model, u"ms") doesn't make any sense (the x axis will be off by a factor 1e3), and f(model, u"kg") will straight up fail, so you're not really communicating anything more than that you're expecting a unitful axis anyway.

[gustaphe]

Currently, calling plot!(f, u) on an already unitful figure will do the right thing. Calling plot(f, u) sets no unit for the x axis. To do so would require us to either replicate the behavior of plotting unitless functions (including finding a suitable x vector), or to pretty much run the entire vector-vector recipe. I'm not in love with either approach, but I don't think we can really merge like this.

The first approach is probably going to be better, but I've looked into the way Plots plots functions before, and there's a bit of voodoo going on in there. I'll see what I can do.

[codecov-io]

Codecov Report

Merging #50 (7164808) into master (d0508b7) will decrease coverage by 0.45%. The diff coverage is 75.00%.

@@            Coverage Diff             @@
##           master      #50      +/-   ##
==========================================
- Coverage   85.71%   85.26%   -0.46%     
==========================================
  Files           1        1              
  Lines          91       95       +4     
==========================================
+ Hits           78       81       +3     
- Misses         13       14       +1     

Impacted Files	Coverage Δ
src/UnitfulRecipes.jl	85.26% <75.00%> (-0.46%)	:arrow_down:

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 42a9a37...7164808. Read the comment docs.

[jw3126]

Currently, calling plot!(f, u) on an already unitful figure will do the right thing. Calling plot(f, u) sets no unit for the x axis. To do so would require us to either replicate the behavior of plotting unitless functions (including finding a suitable x vector), or to pretty much run the entire vector-vector recipe. I'm not in love with either approach, but I don't think we can really merge like this.

The first approach is probably going to be better, but I've looked into the way Plots plots functions before, and there's a bit of voodoo going on in there. I'll see what I can do.

Instead of replicating the approach, maybe there can be some upstream refactoring be done? For instance, expose the machinery that guesses a reasonable domain for f?

[gustaphe]

My guess is, if these lines

https://github.com/JuliaPlots/RecipesPipeline.jl/blob/19dd30b3f68d29476448829141f42b2372fdd6ba/src/user_recipe.jl#L244-L255

were in a function _functionlimits(f, plotattributes), then we could do

@recipe function f(f::Function, u::Units)
    (xmin, xmax) = RecipesPipeline._functionlimits(f, plotattributes)
    f, xmin*u, xmax*u
end

and that would do everything we want.

I guess that's kind of what _apply_recipe does. How's this for an idea:

[gustaphe]

That test failure is bizarre to me. On macOS-latest, the "m^2" label is turned to "m²". It runs on a slightly older version of Julia than ubuntu-latest, but I checked the diff between those two commits and there is nothing that seems even related. Maybe there's a difference in a package, I think I'll just ask it to rerun in a couple of days.

[briochemc]

On macOS-latest, the "m^2" label is turned to "m²"

I don't have lots of time (and skill 😅) to properly investigate, but FWIW, I recall that in https://github.com/PainterQubits/Unitful.jl/pull/297 the Unitful.jl dev set the default to use Unicode superscripts for OSX and Linux and use "^" for Windows, so maybe that's what fails there?

[gustaphe]

Yeah, that was probably it. This test circumvents that.

Only the @recipe lines uncovered now, so RFC.

[briochemc]

Very little time to comment on the changes but LGTM. And you can safely ignore the missed lines because I think that's a Julia/Coverage.jl issue! 🙂

[gustaphe]

Sorry about the noise, I thought of another case. Functions where the domain must be unitful (like exp(x/3u"m")) didn't work before. It's sorted now, but it requires this more complicated approach.