Ancestors

Written by Nick Tierney on 2024-12-17 at 23:13

{greta.gam} is now on CRAN!

{greta.gam} is another addition to the {greta} universe, that lets you use {mgcv}’s smoother functions and formula syntax to define smooth terms for use in a {greta} model. You can then define your own likelihood to complete the model, and fit it by MCMC.

The {greta.gam} package has been around for some time, and was developed by @millerdl and Nick Golding. Thanks to Nick and David for their hard work on this!

See https://github.com/greta-dev/greta.gam for an example of how to use {greta.gam}

[#]rstats

=> View attached media

=> More informations about this toot | More toots from njtierney@aus.social

Toot

Written by Noam Ross on 2024-12-17 at 23:21

@njtierney @millerdl I never got on the greta train but this of course piques my interest. What's the pitch for why or when I would do this in greta over, say, a Stan-based approach using brms? Faster parallelization/GPU usage or big data support? More distributions to make use of?

=> More informations about this toot | More toots from noamross@ecoevo.social

Descendants

Written by David Lawrence Miller on 2024-12-18 at 13:33

@noamross @njtierney some quick thoughts:

1. it's TensorFlow under the hood, meaning that we can use not just HMC (like Stan) but also other optimizers. I guess if you have access to TPUs maybe it's faster, I don't know.

2. no C++ nonsense

3. extension to more complicated models is a bit easier (take yr pop dyn model, parameterise, say K in yr Ricker model using a spline), which in my (not recent) experience is somewhat painful in Stan/TMB/JAGS/Nimble.

hth :)

=> More informations about this toot | More toots from millerdl@mathstodon.xyz

Written by Nick Tierney on 2024-12-18 at 23:58

@millerdl @noamross Thanks Dave!

Yup so what Dave said about extending to more complicated models - there's dynamical systems support in greta with https://github.com/greta-dev/greta.dynamics

The TensorFlow back end also means it has nice speedups for some types of model, and scales well.

We are working on making more distributions accessible in greta (https://github.com/greta-dev/greta.distributions) - as well as thinking about making it easier for users to add and test new distributions. I haven't done a side-by-side comparison of what is in STAN and greta, but if there are any that are missing for you I'd be happy to work on adding them :)

Obviously I'm biased, but the thing that I really enjoy with greta is you get to write R code to fit and design models, and we have tried hard to provide informative error messages for the user.

There's more we are going to do with greta in 2025, one thing I'm particularly excited about is the marginalisation interface that Golding has designed.

Let me know if there's things you want to do or if I can help, Noam!

=> More informations about this toot | More toots from njtierney@aus.social

Written by Aki Vehtari on 2024-12-19 at 11:50

@millerdl @noamross @njtierney I don't understand the ”no C++ nonsense" comment. Based on GitHub stats, both Greta and TensorFlow are mostly C++

=> More informations about this toot | More toots from avehtari@bayes.club

Written by David Lawrence Miller on 2024-12-19 at 13:50

@avehtari (that was probably an overly dismissive comment but...) I was really meaning from the user perspective, in terms of not having to write a template. Especially in comparison to TMB and to some extent for Stan. You are building models in R, rather than in a separate language with separate syntax etc.

@noamross @njtierney

=> More informations about this toot | More toots from millerdl@mathstodon.xyz

Written by David Lawrence Miller on 2024-12-19 at 13:50

@avehtari perhaps I should have said "less C++ nonsense"

@noamross @njtierney

=> More informations about this toot | More toots from millerdl@mathstodon.xyz