Feature request: add nutpieR as a backend

[potash]

The nutpieR package from @andytimm works great. I am using it with brms models manually via make_stancode() and standata(). Two disadvantages are:

• the posterior has the vanilla brms stan code names as opposed to the informative ones available in brm() results
• I don't take advantage of brms caching of compiled models

Therefore, I would like to request adding nutpieR as a backend to brms. Thanks for considering!

[paul-buerkner]

Why not if the interface is easy enough. Can you show a concrete example for how you currently use brms with nutpie?

[paul-buerkner]

Also, what are the benefits of using nutpieR over rstan and cmdstanr?

[wds15]

I always wanted to try out nutpie, which is written in Python. Reasons for nutpie is the supposedly improved warmup procedure from what I understood. At least this is what I was hoping for. The warmup is supposedly more clever and resulting in a more efficient and faster sampling of models. All Stan features work as BridgeStan is used... but this is just what I know by reading / I would love to try out nutpie via brms.

[andytimm]

Just commenting to say I'd be happy to help here -- there'd likely be some work needed in nutpieR, and I'd be glad to pitch in on the brms side too if useful.

I can answer the question on concrete benefits- in brief, improved mass matrix adaptation leads to pretty broad sampling speedups. Adrian Seyboldt/Eliot Carlson/Bob Carpenter have a good paper + overview post on Gelman's blog of the algorithmic improvement here. The exact performance depends on the problem of course, but their paper found a 1.3x median wall time improvement over Stan's defaults, and I pretty commonly see 3-4x ESS/s speed ups on more challenging models.

Here's their comparison plot over posteriorDB if you're curious:

[potash]

@paul-buerkner: I'm sure @andytimm can provide much better support than me but here's a simple example:

library(brms)
brms_fit <- brm(count ~ zAge + zBase * Trt + (1|patient),
            data = epilepsy, family = poisson())

library(nutpieR)
nutpie_model = nutpie_compile_model(code=stancode(brms_fit))
nutpie_fit = nutpie_sample(nutpie_model, data=standata(brms_fit))

The result is a posterior::draws_array

[wds15]

To get from a draws array to a stanfit object has been hacked up here:

https://discourse.mc-stan.org/t/bayesian-bootstrap-implementation-for-brms/40522?u=wardbrian

However, this lacks all the nice meta info there is in addition - what initial values, what seed, what used warmup stuff (stepsize, mass matrix / its diagonal) and most importantly HMC diagnostic of divergent transitions.

So this is a great start and I am going to try this out, but there is more... I guess nutpier has all of this... I would hope so.

[aseyboldt]

It would be great to see nutpie in brms!

@wds15
nutpie has those (plus a few extra), and I think nutpieR exposes at least most of them?

For typical brms models, I think it would also make sense to just enable the low rank mass matrix adaptation by default.