implements best legend place holder

[lmiq]

Implemented the best legend position placeholder: It will choose between bottomleft/right or topleft/right, depending on a sample of the data, trying to maximize the difference to the closest point. The function is fast (5-10 microseconds), and should be good enough for a guess of the best position.

Tries to address: #3399

Test it with:

using Plots
scatter(randn(100), randn(100))

[t-bltg]

Nice, I'm going to try it now.

Don't forget to run julia -e 'using JuliaFormatter; format(["src", "test"])' and commit for the format check to pass.

[t-bltg]

Here are the changed reference images (25 out of 65).

Looking weird: refs 7, 15, 19.

ref4

ref7

ref8

ref12

ref13

ref15

ref17

ref18

ref19

ref23

ref25

ref26

ref29

ref30

ref36

ref37

ref44

ref46

ref48

ref51

ref53

ref56

ref58

ref61

ref62

[lmiq]

Here are the changed reference images (25 out of 65).

Actually most of them look weird to me, it is not what I was expecting. But now that I know those tests exist, I can check each of them before committing changes.

(The histogram is case is a case apart, as the plot does not show the data series, and what I do know is based on the data series).

[t-bltg]

But now that I know those tests exist, I can check each of them before committing changes.

The refs files are checked when running test Plots.

You can set julia> ENV["VISUAL_REGRESSION_TESTS_AUTO"] = true before pkg> test Plots (see

Plots.jl/test/runtests.jl

Line 53 in e6a0689

if get(ENV, "VISUAL_REGRESSION_TESTS_AUTO", "false") == "true" && name != "backends"

) in order to update them automatically (no prompt), and partially run the test-suite (faster).

See also https://docs.juliaplots.org/stable/contributing/#VisualRegressionTests.

[t-bltg]

To quickly run an example, you can use Plots.test_examples(3).

[codecov]

Codecov Report

Base: 90.89% // Head: 90.95% // Increases project coverage by +0.06% 🎉

Coverage data is based on head (569912d) compared to base (816d0f4).
Patch coverage: 100.00% of modified lines in pull request are covered.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #4536      +/-   ##
==========================================
+ Coverage   90.89%   90.95%   +0.06%     
==========================================
  Files          40       40              
  Lines        7707     7740      +33     
==========================================
+ Hits         7005     7040      +35     
+ Misses        702      700       -2     

Impacted Files	Coverage Δ
src/examples.jl	97.26% <ø> (ø)
src/backends/gr.jl	91.66% <100.00%> (-0.01%)	⬇️
src/utils.jl	92.93% <100.00%> (+0.41%)	⬆️
RecipesPipeline/src/type_recipe.jl	94.44% <0.00%> (+5.55%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[t-bltg]

Actually most of them look weird to me, it is not what I was expecting.

Maybe we need to add some weights in the algorithm to favor the :topright position.

[lmiq]

I'll close the PR for now to work on it further without disturbing you.

[t-bltg]

You could also have marked it as a draft (it disables notifications ;)).

[lmiq]

Corresponding image PR: JuliaPlots/PlotReferenceImages.jl#142

[t-bltg]

I'd like @BeastyBlacksmith to comment before merging the reference images.

To me, there is no ambiguity on the automatic placement (at least on the ref images).

[lmiq]

I just updated a comment. Nothing will change.

[lmiq]

To me, there is no ambiguity on the automatic placement (at least on the ref images).

I don't think there is any ambiguity. The position is deterministic, given the data set (which is a difference from the original implementation I posted on the discourse forum, which used a random sample from the data set).

[BeastyBlacksmith]

I'd be more happy if it kept :topright in the cases where it is impossible not to overlap with data (or maybe even choose :outertopright then).
But I like it either way. Good work!

[lmiq]

where it is impossible not to overlap

If we could obtain the coordinates of the legend box we could try that. But I didn't find how to do it, so we are just examining the plot edges relationship with the data. We can try some different heuristics, such as summing some distances, but we would need a case where the current strategy is clearly bad. Also biased examples against any other strategy should be tested.

[lmiq]

I'll post the results of this different placement strategy, which does better if there is a bunch of points in one corner, but the closest overlap is in the other (should I have created a new branch? I'm still confused with this)

[lmiq]

Here they are: https://github.com/JuliaPlots/PlotReferenceImages.jl/pull/143/files

It actually looks good, IMO. Maybe better, particularly in the scatter plots.

(I'll be away for a couple of hours now)

[t-bltg]

Thanks, merged, let's try that now.
There are indeed less changes w.r.t. the original ref imgs, making this PR less controversial.

[t-bltg]

I've have cleaned up reference images between versions 1.36.3 and 1.36.4 ("only" 21 images changed by this PR).

[t-bltg]

Should we merge ?
EDIT: merging now, since it blocks other PRs, because of the reference images.

[t-bltg]

Thanks @lmiq, this is a really nice contribution 🎉 .