Inference engine input interface

[eerolinna]

Currently in this R version of bayesbench inference engines are defined like this

stan_vb <- function(cfg){
  model <- stan_model(cfg)
  data <- jsonlite::read_json(data_file_path(cfg), simplifyVector = TRUE)
  inference_engine_args <- inference_engine_arguments(cfg)
  ...
}

(simplified from ie_stan.R)

In python version of bayesbench I have used something like

stan_vb <- function(model, data, inference_engine_args=NULL, diagnostics=NULL){
  ...
}

Which one should we use?

Things I like with the python version

• It only passes in what is required. In contrast with R cfg for example includes a field inference_engine, which is never used by the inference engine itself, only to select which inference engine to call
• We can eliminate some duplicated code, for example we can call jsonlite::read_json directly in bayesbench core instead of having to always call it in inference engines. This also makes it easier to write inference engines when some of the work has already been taken care of

Downsides to the python version

• The argument list can become long
• Some validations can be easier to do when all the arguments are in one object (however if the validation can be performed in bayesbench core then there's no difference because the python version also has everything as one object in bayesbench core)
• Filling default values for inference engine arguments is a bit more complex (we need to return them instead of just modifying cfg in place)

[MansMeg]

Hmmm. I think the R-version is a little more clear with regards to our abstractions. We have the PDB and a cfg that contain everything. So for an IE creator the R version is more inline with abstractions. But this is not clear cut how to do.

[eerolinna]

We could also have

stan_vb <- function(cfg){
  model <- stan_model(cfg)
  data <- cfg$data
  inference_engine_args <- inference_engine_arguments(cfg)
  ...
}

Where we would still have everything in one object but bayesbench core would do some preprocessing, such as loading the json.

Slightly unrelated: how does stan_model(cfg) find the model as the config only specifies a posterior name and not the file paths etc? Wouldn't it need the posterior database also as a parameter? Or is stan_model(cfg) essentially an alias for (stan_model(cfg, posterior_database = get_default_posterior_database())?