Adding Plot ESS Evolution

[imperorrp]

First commit for plot_ess_evolution(), as a branched off plot from legacy plot_ess for kind = 'evolution' (#5). The line_xy and scatter_xy visual elements were used for the plotting, with rug plot, axes labelling and title labelling functionality taken from ESS plot functionality. Since the ESS plot PR hasn't been merged yet, the scatter_xy visual element was duplicated for use here.

WIP:

Marking the 'bulk' and 'tail' artists with a legend

Current plot outputs:

azp.plot_ess_evolution(data)

azp.plot_ess_evolution(data, var_names=["mu", "tau"])

---

📚 Documentation preview 📚: https://arviz-plots--71.org.readthedocs.build/en/71/

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 94.76440% with 10 lines in your changes missing coverage. Please review.

Project coverage is 86.47%. Comparing base (3cc46aa) to head (ea60c47).
Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/arviz_plots/plots/evolutionplot.py	94.73%	10 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #71      +/-   ##
==========================================
+ Coverage   85.65%   86.47%   +0.81%     
==========================================
  Files          22       23       +1     
  Lines        2545     2736     +191     
==========================================
+ Hits         2180     2366     +186     
- Misses        365      370       +5     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[OriolAbril]

lists in rST need an empty line before and after to be rendered correctly.

[OriolAbril]

Suggested change

	Passed to :class:`arviz_plots.PlotCollection.wrap`
	Returns
	Passed to :class:`arviz_plots.PlotCollection.wrap`
	Returns

this also needs a white line to separate the sections

[OriolAbril]

Suggested change

	x_diff = 1 / n_points
	if "x" not in pc_kwargs:
	pc_kwargs["x"] = np.linspace(-x_diff / 3, x_diff / 3, distribution.sizes["model"])
	n_models = distribution.sizes["model"]
	x_diff = min(1 / n_points / 3, 1 / n_points * n_models / 10)
	pc_kwargs.setdefault("x", np.linspace(-x_diff, x_diff, n_models))

Seeing the plot now, 1/3 would probably look good for several models, but for only 2 of them the points end up equispaced (the model1 point corresponding to quantile 0 is as far from the model0 point corresponding to quantile 0 than from the model0 point corresponding to quantile 0.05). This change will result in smaller spacing with small number of models, but never going over 1/3 of the available space.

[OriolAbril]

I thought I was reviewing the plot_ess PR. Having x with an aesthetic mapping here makes no sense and should be completely removed.

[OriolAbril]

I think by default we don't want any aesthetic mapping to be applied to either mean nor sd, so I would remove these 4 lines. Also note the lines 2 and 4 are not doing anything because the lines right before already define that key, but if they did, it would be incorrect. Values in the aes_map dict should be sets of strings that represent aesthetics.

This seems a confusion with setting a default linestyle in plot_kwargs. If you want an aesthetic property to always have the same value, use plot_kwargs if instead you want it to depend on a dimension/variable use aes in pc_kwargs to define the mapping and aes_map to indicate to which visual elements should the mapping be applied too.

There is nothing about this concepts in https://arviz-plots.readthedocs.io/en/latest/contributing/new_plot.html because that doc is for writing plots and these are usage questions which are therefore out of scope. The usage docs are instead https://arviz-plots.readthedocs.io/en/latest/tutorials/plots_intro.html and to a much lesser extent https://arviz-plots.readthedocs.io/en/latest/tutorials/intro_to_plotcollection.html

[OriolAbril]

This key should be "rug" instead. The most common thing to use (and therefore default for rug_kind are divergences) but they are not the only option. As we agreed on smaller set of fixed keys in #66, I think this key should be "rug" instead of rug_kind.

Note this will require multiple changes in the function code

[OriolAbril]

We should add a check like in plot_forest to raise an error if plot_kwargs["rug"] is False given there is a top level kwarg for it

[OriolAbril]

careful with this

[OriolAbril]

Not sure, you'll have to check both backends behaviour. I seem to remember in some cases the default was 0 which meant nothing got plotted, if that were the case then we'd need a default here (which you can get the same way you get the default color).

[imperorrp]

I tried this with bokeh, seems like there definitely is some default that isn't zero when no linewidth is used-

[OriolAbril]

then no need to set it, you can delete these lines

[OriolAbril]

Suggested change

	xlabel = "Total Number of Draws"
	xlabel = sample_dims[0] if len(sample_dims) == 1 else "Total Number of Draws"

Plus, I would add that into label_kwargs as text key so it can be modified if desired. Consequently a single "label" key can be used anymore, otherwise it would be impossible to set different names

[imperorrp]

So "xlabel", "ylabel" as two separate keys, right?

[OriolAbril]

yes

[OriolAbril]

labels also need the "False" value to be available and have an effect

[imperorrp]

Updated plot_ess_evolution. The points and lines for both 'bulk' and 'tail' methods can now be turned on and off via plot_kwargs, the mean, sd, and min_ess were added following the current plot_ess logic, and xlabel and ylabel texts can also now be set separately.

Also added a common ess_dataset computing function since it can be called upto 4 times (one for each of the cartesian product of 'bulk', 'tail' and points, lines).

The rug plot element is still WIP.

The outputs without the rug are now like:

An aesthetics issue with the color coding of the mean/sd plot elements:

For the mean/sd plotting (and this'd similarly apply to plot_ess too) it looks like setting mean and sd to one color by default as I'm doing currently makes it hard to interpret that they have been overlapped when they coincide:

When they're not set to the same color, it's visible that they coincide and are overlapped but then the colors might be confusing as they correspond to the 'bulk' and 'tail' plotting elements' colors too:

[imperorrp]

Rebased plot ess evolution commits

[imperorrp]

The mean and sd lines are now annotated:

Should the min_ess line also be annotated? (This also applies for ESS plot #58 ). Current Arviz ESS plot doesn't have it annotated (https://github.com/arviz-devs/arviz/blob/main/arviz/plots/backends/matplotlib/essplot.py#L150) but it would probably be nice to have and is easy to add.

[imperorrp]

Fixed hypothesis tests errors and documentation build