[Feature request]: project from inference results

[pearsonca]

Label

config, gempyor

Priority Label

medium priority

Is your feature request related to a problem? Please describe.

Users requested the ability to just be able to project from inference results, ideally using the same config that one could also project from before having done an inference operation.

Is your feature request related to a new application, scenario round, pathogen? Please describe.

Forecasting from any kind of synthetic posterior (such as was needed for SMH flu 2024) needs something like this capability.

Describe the solution you'd like

For a hypothetical configuration file (maybe a new approach to syntax like this or sticking with the currenter more involved syntax this reflects):

parameter:
  a: ~unif(min=0, max=1)
  b: ~normal(mean=5, sd=0.1)
  c: 0.5

inference:
  targets: [a, b]

• simulating without doing inference should take a value draw for a, b (and c, but its drawn value is always the same).
• inference should know its varying a, b AND should use the distribution as a prior
• if asking to simulate after inference has been run, it should refer to the inference results for a and b
• ...but also there should be someway to do simulation using the priors (or some other value to override inference results)

So:

$ flepimop simulate config.yml # does simulation from priors, with some expected stochastic behavior
$ flepimop calibrate config.yml # does inference with priors, and as a side effect writes inference results
$ flepimop simulate config.yml # now does simulation from priors
$ flepimop simulate config.yml --ignore-inference # ... or some such, to simulate as if no posterior available

Not clear if we should do this before or after general overhaul of parameter specification (#437).

Thoughts @saraloo @MacdonaldJoshuaCaleb @alsnhll?

[pearsonca]

This is related to, but distinct from: #416.

Earlier conversation at dev meeting 13 Jan, suggests something like

$ flepimop calibrate someconfig.yml
$ flepimop calibrate --render=posterior someconfig.yml
$ # or maybe flepimop process someconfig.yml --target=posterior

where the first version is sampling only and "only" produces the intermediate fitting outputs, and the second version (or something similar) gives the best possible (based on available calibration results) posterior object, which can then be used for projection.

Similarly in #416, the idea would be to produce via other ad hoc / manual means something that complies with the spec for an posterior object.