ggplot layer explorer shiny app

[yjunechoe]

I had long planned to make an educational shiny app powered by {ggtrace}. The recent discussion in #87, especially Teun's comment, finally propelled me to take a stab at this.

A prototype is available as a shinylive app: https://yjunechoe.github.io/ggplot2-layer-explorer (link to code)

The current UI is:

1. Define a plot p in the code editor in the middle column (app initializes with an example)
2. Pick a Layer method to inspect the return value of, using the radio buttons in the left column
3. Further inspect and modify the output using arbitrary expressions in the right column ("run expression" button)
4. If modifying the output, optionally use that to replace the return value and see the consequences for the plot ("highjack" button)

Potato quality demo (also, slightly outdated UI):

layer.explorer.demo.mp4

Half-baked thoughts for next steps:

• Whole lot of explanations/documentations/tooltips/etc.
• Option to upload your own ggplot object
• Option to select a lower-level method (nested radio buttons?)
• Currently only targets outputs of ggproto methods. But inspecting and manipulating inputs may be informative as well

[teunbrand]

This is awesome! 🎉
Is that the ye olde new penguins data 🤤?
A couple of quick thoughts, feel free to dismiss as you see fit:

• Didactically it might make sense to show an annotate(), so that you can see it is just a regular layer.
• Also didactically, in you example perhaps stat_count(geom = "label") introduces novelty for novices by using a stat layer instead of geom.
• Is only the Layer class in scope? In the chronology of the layer data, there is also ScalesList$transform_df() , Layout$map_position(), etc. that affect the data.
• I don't know if the shiny apps computes the whole thing first and then shows you stuff, or it dynamically computed; but if it computes the whole thing first: maybe it is possible to highlight the difference in the data with regards to the previous step?
• Is there an intuitive way that the group aesthetic is imputed from compute_aesthetics() and the PANEL variable from the facets and some defaults from compute_geom_2() etc?
• How much of a PITA is this to maintain over time?

Whole lot of explanations/documentations/tooltips/etc.

I think it'd go a long way to just add a 1-2 sentence context to what part of the layer you're in.

Option to upload your own ggplot object

This is asking for nightmares. Saved ggplot objects are only interoperable within the same version. If you are running 3.5.2 but a 3.5.0 plot is uploaded, you're likely to run into some troubles.

[yjunechoe]

Thanks, Teun! All very valid points, and some that I haven't even considered :)

Option to upload your own ggplot object

This is asking for nightmares. Saved ggplot objects are only interoperable within the same version. If you are running 3.5.2 but a 3.5.0 plot is uploaded, you're likely to run into some troubles.

You're totally right. IDK what came over me when I wrote that. I'm trying to be very intentionally about not making the app into a "tweak your plot with some hacks", so that really shouldn't be a feature.

• Didactically it might make sense to show an annotate(), so that you can see it is just a regular layer.
• Also didactically, in you example perhaps stat_count(geom = "label") introduces novelty for novices by using a stat layer instead of geom.

The current example is just a placeholder for now; intentionaly complex so that it touches (almost) each step at least once. Mostly for debugging purposes. The end product will probably have ~3 template snippets of increasing complexity that users can load onto the editor ,and the current will probably be the most complex option 😆

• Is only the Layer class in scope? In the chronology of the layer data, there is also ScalesList$transform_df() , Layout$map_position(), etc. that affect the data.

Figuring that out will be a big part of the next steps! But in theory I think it should work. It's complicated if a method is called more than once per layer (~ once each by_layer() iteration), but AFAIK those are called once per layer?

• I don't know if the shiny apps computes the whole thing first and then shows you stuff, or it dynamically computed; but if it compes the whole thing first: maybe it is possible to highlight the difference in the data with regards to the previous step?
• I think it'd go a long way to just add a 1-2 sentence context to what part of the layer you're in.

Good FRs!

• How much of a PITA is this to maintain over time?

Only as much PITA as maintaining ggtrace 😂. The app just ordinarily runs the code that you see in the editors.

• Is there an intuitive way that the group aesthetic is imputed from compute_aesthetics() and the PANEL variable from the facets and some defaults from compute_geom_2() etc?

TBH idk what's meant by "is there an intuitive way" 😅 -- is this a question about the app or how users intuit the behavior? I don't think I implemented any imputation behavior myself