Rémi Louf
Rémi Louf
There is a third way: you can wrap `blackjax`'s algorithms in whatever data structure you need and have users pass this data structure to `pm.sample` if they want something different...
Seemingly unrelated, but Metropolis-within-Gibbs has been shown to work well (https://github.com/blackjax-devs/blackjax/discussions/275). So this is definitely feasible.
I agree with what @junpenglao said: I would expect that to happen with NUTS; each step can only be as fast as the slowest chain and these delays can add...
No problem! Thank you for letting us know.
You can definitely analyze any `aesara` graph a user provides and see if it matches a certain pattern. May if I knew better how `bambi` works I could tell you...
I general I don't like the "test by example" approach that we've been taking so far. Unless we're testing against examples for which we have theoretical results on convergence, tests...
I am simplifying the tests in #279; SMC is where the test suite currently spend the largest amount of time.
Currently working on a personal project that builds the model's logpdf using `aeppl`. Will add a simplified version to the repo.
Closed with #255
I mean that the blackjax repository should only have notebooks in the [jupytext format](https://jupytext.readthedocs.io/en/latest/formats.html). Notebooks are rendered in CI to display in the documentation.