Hi everyone,
My name is Hugh McDougall, I’m a PhD student working in astrophysics at the University of Queensland; some of you may have seen me posting around here a few months ago. Bayesian analysis tools like emcee are ubiquitous in my field, but knowledge about the joys of NumPyro is in sadly short supply. To that end, I’ve started writing a set of tutorials aimed specifically at bringing people like myself into the fold.
These are still in early stages, and I’d love any suggestions people can give about changes to make or what areas to focus on next.
Thanks,
Hugh
5 Likes
Awesome work, Hugh! Thanks for sharing! Those tutorials are on my reading list this week. 
(a comment: some links in the getting started page are not working.)
looks great! very extensive.
it would be great to link to this from the numpyro docs or readme. although in the best possible world some of these tutorials would be merged into the repo. this is because any api changes that happen in jax or numpyro might break your blog post code and it will become stale. if it’s in the repo tests will catch those changes and the code can be kept up-to-date. so if you’re up for it…
didn’t look in detail but from a quick skim:
- DAIS is best initialized with a guide learned from a vanilla ELBO run. yes DAIS is expensive but it can work pretty well if you take care.
- on the same topic note that the DAIS ELBO is looser than a vanilla ELBO so the loss is not a great proxy for how well it’s doing (it will tend to underestimate how well it’s actually doing)
Thanks for the heads up: I’d re-arranged some folders and hadn’t updated the links. Should be working now
These examples are super nice! I’m also doing something to have more explicit examples! See NumPyro time series examples
Great job! I’m exactly at the same position. Will check your GitHub page and try to contribute!
I also found very useful the Numpyro translations of the exercises/examples of from “Statistical Rethinking”: Overview | Statistical Rethinking (2nd ed.) with NumPyro