Open SamuelBrand1 opened 1 month ago
Worth noting that this forecast approach works for our use case but is not general!
https://github.com/TuringLang/Turing.jl/issues/2239#issuecomment-2144614352
Does this
Worth noting that this forecast approach works for our use case but is not general!
suggest that we move towards not having vectorised calls? e.g. no MvNormal
calls?
Yes. Do you know what implications this has (performance / limiting flexibility etc?)
The main argument against would be AD performance; because known vector-form functions might have adjoints which are accessed rather than calculated.
Flexibility would only be improved.
Current state of code
As per #239 we have a neat forecast function for models which have had inference on some time span $(1, T)$. The posterior chain can be used to condition the sampling of the model extended to a longer time span $(1, T + n)$. This is used in the analysis pipeline.
Lack of generality
The problem here is that under the hood there are hard coded options, which reflect our analysis plan but are too specific to promote this function to
EpiAware
.Pain points:
define_epiprob
: hard codes in discretised Gamma distributions.Proposal
If we can fix these pain points then we can add this nice functionality to
EpiAware
.