Support arbitrary X axis in the plot view using a `ScalarAbscissa`

[abey79]

This discord thread got me thinking and I'd like to suggest this alternative take on:

• Introducing ScalarChart in place of BarChart? #5622

In this discussion, the user wanted to store an entire 2D plot in each frame of their simulation result, and resorted to logging a matplotlib raster:

Note: this proposal address the (possibly time-varying) "instantaneous" plot use case, aka the use case where the entire plot (X and Y axis) is stored at a single time point (modulo latest at). It does not address the case where one time varying scalar must be plotted against another time varying scalar.

Core proposal

The core proposal is to introduce a ScalarAbscissa component (someone please finds a better name) that is essentially a Scalar component to be used as X axis (with tagged components we could actually reuse the same Scalar component).

Then, the existing Timeseries view would be altered as follows:

• Renamed to Plot view.
• When used in Latest At mode: plot the Scalar component (which, in this case, should typically be a n>1 batch) against the ScalarAbscissa component.
• When used in Range mode: no modifications w.r.t to the current behaviour. (For a given time point, if the Scalar is a batch, pick instance at index 0 and ignore others. Also, ignore ScalarAbscissa.)

Scalar x ScalarAbscissa join

• Obviously, we have basic instance join semantics.
• The resulting plot has a number of point which is the min of both components (if Scalaras 10 instance and ScalarAbscissa as 15, then we plot 10 points).
• The fallback provider for ScalarAbscissa is range(n) where n = len(Scalar).
• If multiple entities are added to the plot view, each Scalar are matched against the collocated ScalarAbscissa (if any). In other words, not cross-entity shenanigans (*)

(*) this is a bit of a limitation in the case of many series against the same abscissa. Possible workaround:

• allow 2d tensors to be considered a collection of multiple scalars
• allow multiple scalars to be collocated using tagged components
• allow any vecXD-based component to be plottable

Bonus features

• When in Range mode with n>1 Scalar batches, have some options to display a compound value for each point (mean/media with stddev bars, box plot, etc.)
• …

Related

• view_coordinates for 2D views #1387
• Rerun should be the obvious replacement to imshow, plot, and draw_geometries #1630
• Introducing ScalarChart in place of BarChart? #5622
• Enhance Support for Custom Axes and 2D Plotting #7346

[gavrelina]

• related to Enhance Support for Custom Axes and 2D Plotting #7346

[grtlr]

Very interesting! In general, I would envision Reruns plots + blueprints to evolve into something that resembles Vega-Lite which should give users a lot of flexibility and control over how their data is visualized.

Vega-Lite is a declarative grammar (that should map very well to our concept of blueprints) for creating visualizations.

Here is an example of how plotting a trajectory looks like in VegaLite:

{
  "data": {"url": "data/driving.json"},
  "mark": {"type": "line", "point": true},
  "encoding": {
    "x": {
      "field": "miles", "type": "quantitative",
      "scale": {"zero": false}
    },
    "y": {
      "field": "gas", "type": "quantitative",
      "scale": {"zero": false}
    },
    "order": {"field": "year"}
  }
}

Above code yields the following plot:

For Rerun, and this particular use case, the user could for example log the data to a 2D tensor for each timestamp. In Vega-Lite terms, the order would then be set to use the ordering of the data in the tensor: https://vega.github.io/vega-lite/docs/encoding.html#order

[abey79]

It looks like entity reference would be needed to fully support this model:

• Put more thoughts into entity references #2682

[benliepert]

the use case where the entire plot (X and Y axis) is stored at a single time point

+1 For this.

Specifically, we have an FPGA monitoring an RF envelope that produces a vector per higher level timestep.
It would be nice to visualize this vector as a 2D plot per time point in a way that preserves all the data.