numpyro
numpyro copied to clipboard
pyro-ppl
HMC Gibbs Example
Hi all, I've been working with the HMC Gibbs sampler.
Thanks for your work on this, I think it's a really unique and powerful feature. I've written an example notebook (a ttached). This uses HMC Gibbs and HMC on the same hierarchical problem, and shows significant speed up with HMCGibbs as the dimension of the conditionally conjugate part of the problem increases. (vary the number of series n to see this). In my experiments HMCG scales more or less linearly with n, HMC is better on small problems, but orders of magnitude worse as dimension increases.
A few questions and queries...
- Some more examples and guidance / documentation on this would be super useful.
- Whats the idiomatic way of transferring values computed within the model function to the Gibbs function? - I've done this here using 'deterministic' statements, and this seems to work... though it feels a bit of a hack... undocumented functionality!
- Any further refinements (or errors) in the attached please let me know!
a few corrections to this:
in the gibbs_fn, the scale needs a sqrt: betai = dist.Normal(beta_post_mean, jnp.sqrt(beta_post_var)).sample(rng_key)
and the data frame results should be: pd.DataFrame(np.asarray([[10,11],[58,30],[1920,55]]),index=[50,500,1000],columns=['hmc','hmc/gibbs'])
Whats the idiomatic way of transferring values computed within the model function to the Gibbs function
I think using deterministic is intended here. Could you mention this usage in the docs for gibbs_fn?
You can find some examples in the tests but I think that there's nothing new to you. Given your experience with this class, it would be nice if you can contribute an example or tutorial for users. :)
(just a note that you can set group_by_chain=True in mcmc.get_samples() to add an additional singleton dimension to the samples)
Thanks for hint with group by chain. Very happy to contribute based an example / tutorial on this notebookLet me know what else I need to do!
Is this tutorial / some similar tutorial for HMCGibbs now officially in the numpyro tutorial yet?
Hi @edwinnglabs, we don't have a tutorial yet. But we have an example in the doc https://num.pyro.ai/en/stable/mcmc.html#numpyro.infer.hmc_gibbs.HMCGibbs. Hope that it's helpful.
