add densityplot

[aloctavodia]

In the same spirit of #2736 a densityplot can be used to compare several traces/models (and also works for a single trace/model).

The KDE plots are truncated at their HPD interval, a point estimate (the mean by default) is represented using... a point. Each group of variables share the same scale (like in a traceplot and unlike a forestplot). At some point I though about implementing a ridgeline plot (AKA joyplot) but I am still not connived that those plots add too much value except for their coolness and fanciness.

Let me know if you think this could be a good addition to the PyMC3 plots, suggestions for alternative names are also welcomed.

[springcoil]

I think it adds something. Not sure about the name though

…

On 3 Dec 2017 14:50, "Osvaldo Martin" ***@***.***> wrote: In the same *spirit* of #2736 <#2736> a densityplot can be used to compare several traces/models (and also works for a single trace/model). [image: index] <https://user-images.githubusercontent.com/1338958/33525822-b2fff0f8-d815-11e7-8c19-ead4e3800ee0.png> The KDE plots are truncated at their HPD interval, a point estimate (the mean by default) is represented using... a point. Each group of variables share the same scale (like in a traceplot and unlike a forestplot). At some point I though about implementing a ridgeline plot (AKA joyplot) but I am still not connived that those plots add too much value except for their coolness and fanciness. Let me know if you think this could be a good addition to the PyMC3 plots, suggestions for alternative names are also welcomed. ------------------------------ You can view, comment on, or merge this pull request online at: #2741 Commit Summary - add densityplot File Changes - *M* docs/source/api/plots.rst <https://github.com/pymc-devs/pymc3/pull/2741/files#diff-0> (2) - *M* docs/source/notebooks/model_averaging.ipynb <https://github.com/pymc-devs/pymc3/pull/2741/files#diff-1> (184) - *M* pymc3/plots/__init__.py <https://github.com/pymc-devs/pymc3/pull/2741/files#diff-2> (1) - *A* pymc3/plots/densityplot.py <https://github.com/pymc-devs/pymc3/pull/2741/files#diff-3> (155) Patch Links: - https://github.com/pymc-devs/pymc3/pull/2741.patch - https://github.com/pymc-devs/pymc3/pull/2741.diff — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#2741>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AA8DiMvEebRpW4Xu0IU8J2Qg-XUnPZB9ks5s8qdBgaJpZM4Qzucb> .

[aloctavodia]

Thanks for the feedback @springcoil, just to add perspective the previous name was aplot :-)

[springcoil]

Maybe Density plot is fine. Compared to aplot. I'd add that to the release notes

…

On 3 Dec 2017 15:06, "Osvaldo Martin" ***@***.***> wrote: Thanks for the feedback @springcoil <https://github.com/springcoil>, just to add perspective the previous name was aplot :-) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#2741 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AA8DiMhtI_pKXWkBfdUxwgFclrc0lujLks5s8qr1gaJpZM4Qzucb> .

[ColCarroll]

Thanks for including the sample! Can you show a snippet of code that would produce that? I am wondering in particular what m_0, m_1, m_2 are (and why the legend appears in only one figure).

[aloctavodia]

Hi @ColCarroll

You will have to do: pm.densityplot(traces) to get that plot where traces is a list of traces (3 in this case). m_0, m_1, m_2 are the names of the 3 models/traces, these names are assigned automatically (the same with the colors) unless the user provides custom values. The legend appear only once to avoid redundancy/cluttering, you are comparing 3 models across all panels (variables), even when a model could not have all variables, like m_0 in this example that is missing the beta parameter.

This PR also adds a very simple usage example to an existing notebook of the pm.densityplot (and of the new forestplot)

[aloctavodia]

Still not happy with this plot, another two options to help see the limits of the distributions

[springcoil]

I like this last one. Do you think it's worth merging now, and iterating on it in the future.

[aloctavodia]

Hi @springcoil, I need to commit the code for the last two figures and I should add a test. I could do it tomorrow.

I guess we don't need to choose between these 3 variations, they can be offered as different options, but which one should be the default one? The last one?

@ColCarroll what do you think?

[springcoil]

Yeah a test would help. I think having three options is good, I think the last one should be the default one.

[aloctavodia]

@springcoil thanks for the feedback! I will work on this.

[ColCarroll]

Probably helpful to mimic the api of https://seaborn.pydata.org/examples/distplot_options.html , since it will be known to users already. In particular, they pass in kde_kwargs to allow shade=True or shade=False. Agree that the last one should probably be default -- it would be very tricky to set the alpha automatically for lots of densities to make them all visible.

Also agree that we can iterate on some of this after merging, so it isn't all on you!

[springcoil]

Yeah I think shipping this, and then people experimenting with it will mean that the load isn't all on you.

[aloctavodia]

Following your suggestions, @springcoil and @ColCarroll, I update the code and set a different default style. I changed the name of argument opacity to shade for internal consistency (this is the name we use for energyplot and kdeplot). BTW we use a float for the shade instead of a boolean, in that way we can avoid having one argument to turn on/off the shade and another one to set the alpha-level.

[springcoil]

LGTM :)