sampling times in bolfi.sample about elfi HOT 4 CLOSED

Yes, I can observe the same behaviour, though the order can be the opposite as well, and vary from sample to sample.

The number of evidence points for BOLFI affects the shape of the fitted GP, which in turn affects the shape of the posterior. The NUTS sampler is a self-adjusting, recursive and random algorithm, and possibly chooses different internal settings in these cases. Due to the recursion the differences in runtime can be significant. Also, the default initialization points are taken from the evidence with minimal discrepancy, and these can have a big effect as well.

from elfi.

oh great, makes sense! I forgot about the NUTS peculiarities....
Thank you!

from elfi.

Is NUTS the only available sampler within bolfi.sample? How do I find the list of parameters/options for bolfi.sample in the documentation?

from elfi.

There's an open PR #296 to add the existing Metropolis sampler to the BOLFI.sample interface, but for the time being, NUTS is your only built-in option. Basically you could of course give the BOLFI posterior logpdf to any external sampler (such as elfi.methods.mcmc.metropolis).

The standard Python docstring shows the available arguments for BOLFI.sample, available also here:

https://elfi.readthedocs.io/en/latest/api.html#elfi.BOLFI.sample

from elfi.