Again quoting straight from the blog post (we really should port all these tips into an FAQ, shouldn’t we?):
The number of effective samples is smaller than XYZ for some parameters.
- Quoting Junpeng Lao on discourse.pymc3.io: “A low number of effective samples is usually an indication of strong autocorrelation in the chain.”
- Make sure you’re using an efficient sampler like NUTS. (And not, for instance, Metropolis-Hastings. (I mean seriously, it’s the 21st century, why would you ever want Metropolis-Hastings?))
- Tweak the acceptance probability (
target_accept) - it should be large enough to ensure good exploration, but small enough to not reject all proposals and get stuck.
Also, I think that prior visualization will be moved to ArviZ, see this issue. But we appreciate any code you think would be good to contribute!