Remove Plots & Diagnostics from PyMC Code Base

[AlexAndorra]

Plots and diagnostics are still accessible via PyMC3 but are actually implemented by ArviZ now. Although this was a good call when ArviZ was still nascent, now this causes confusion for new comers and overhead for devs, who have to mirror the functions in ArviZ, which is more work when ArviZ adds or changes functions (e.g pm.hpd vs. az.hdi).

To make all this more robust and easier to maintain, we should completely remove plots and diagnostics from the PyMC code base and rely on ArviZ, since the latter library is now much more known and flies on its own (yeah, I'm a part-time poet).

The coming 4.0 major release is an opportunity to make this breaking change -- plus, it'll make the code base cleaner and leaner 👌 Feel free to comment if you wanna take on this issue 🖖

[MarcoGorelli]

yeah, I'm a part-time poet

Must come from all that listening to Baba Brinkman ;)

[CloudChaoszero]

+1 lol ^^^

---

Hey there @AlexAndorra, is this an Issue I can participate in?

Also, when do you think the ETA/Due Date for the coming 4.0 major release would be?

[AlexAndorra]

Hey there @AlexAndorra, is this an Issue I can participate in?

Of course @CloudChaoszero ! I just assigned it to you. Feel free to ask questions 😉

Also, when do you think the ETA/Due Date for the coming 4.0 major release would be?

We usually don't give hard deadlines but are all commited and working to make that happen as soon as possible 🙂
Getting new contributions is definitely helpful in that regard, so thank you!

[CloudChaoszero]

Great! I'll reach out if I have questions etc @AlexAndorra.

[CloudChaoszero]

@AlexAndorra Hey there!

I was looking looking at the codebase, particularly Pymc's plots/init.py, and I saw functions like compareplot and traceplot borrow functionality from arviz, but do a couple of work arounds.

These workarounds tries to relabel kwargs parameters, and other related items like that.

I was wondering should the kwargs param renaming be copied to arviz repo, or should the expected breakchanges for v4.0 assume users rename the params themselves?

[AlexAndorra]

Yes, you can just remove all this @CloudChaoszero, as the idea is to completely delegate to ArviZ.
As a sidenote, we just switched compact to True by default in az.plot_trace anyway 😉

[CloudChaoszero]

Yes, you can just remove all this @CloudChaoszero, as the idea is to completely delegate to ArviZ.
As a sidenote, we just switched compact to True by default in az.plot_trace anyway 😉

Sounds great, thanks!

[CloudChaoszero]

@AlexAndorra Quick Question: Should customer function plot_posterior_predictive_glm Source be removed as well, and users would try to adopt their function to arviz? I only see arviz's plot_posterior Source right now.

• Note: I can try to make a migration PR for it into Arviz, if necessaary

[AlexAndorra]

Ugh, that's a good question -- we discussed it recently with @MarcoGorelli and decided to keep it around IIRC. But the 4.0 release is a good opportunity to remove it and completely switch to ArviZ.

plot_posterior_predictive_glm can probably be easily replicated with az.plot_ppc 🤔
Curious about what @OriolAbril think: is it worth it to migrate it to ArviZ, or are ArviZ's current methods enough?

[OriolAbril]

I don't know how plot_posterior_predictive_glm works internally, but my guess is that is should be kept in PyMC3 until there are regression plots in ArviZ, that is when arviz-devs/arviz#313 and arviz-devs/arviz#512 are solved. Maybe in this year's GSoc?

How does that sound? It may not match the 3-4 timeline though

[MarcoGorelli]

plot_posterior_predictive_glm is just a few lines of code - if you have e.g. a model of the form y = Normal(a + bx, sigma) then it takes samples from the trace and plots a + bx for the various samples. Granted my knowledge here is really limited, but given that in the above example it disregards the values of sigma and just plots the regression lines a + bx, is it fair to call it "posterior predictive"?

[AlexAndorra]

should be kept in PyMC3 until there are regression plots in ArviZ

Agreed @OriolAbril, let's do that then -- if this function proves hard to maintain because of v4 changes though, we'll probably remove it before ArviZ has regression functionalities.

is it fair to call it "posterior predictive"?

Good catch @MarcoGorelli ! I think these are indeed no posterior predictive samples but rather predictive pushforward samples. As the function is meant to disapear soon though (see my answer above), let's not change the name, as that would be a breaking change for just a few months

[CloudChaoszero]

Sounds good, I'll update the PR accordingly!