When I recently benchmarked the core inversion loop, it seems that a significant (maybe the majority? I don't quite remember) amount of time is spent running the simulation by emcee instead of calculating the posterior probabilities. It would be good to investigate whether there are any other MCMC backends that have less overheads than emcee that could be employed by this package.
When I recently benchmarked the core inversion loop, it seems that a significant (maybe the majority? I don't quite remember) amount of time is spent running the simulation by
emcee
instead of calculating the posterior probabilities. It would be good to investigate whether there are any other MCMC backends that have less overheads thanemcee
that could be employed by this package.