Importance sampler column labels when running on a chain with log_weight column
Closed this issue · 2 comments
Just flagging something that could be unexpected behavior: I've recently run an importance sampler, using an input chain that was run with nautilus. One way the nautilus chains differ from things run with other samplers I've worked with, is that it has a log_weight
column instead of just weight
.
Usually when running the importance sampler on a nested sampler chain results in an output file which has one column called old_weight
(the weight from the original chain), and one called weight
, for the importance sampling weight from the ratio of the two posteriors in question.
For the run on this nautilus chain, the last handful of column labels are log_weight prio r old_post log_weight post
. That is to say, there is no old_weight
, there is just two copies of log_weight
. It looks like the data in the columns are probably what they're supposed to be, so I think this is just an issue with the labels.
With apologies for the delay, I've opened PR #150 which I think addresses this. Do you have a simple test case lying around that should work for it?
This is now in the main branch.