Move away from auto-detecting models

[SamuelBrand1]

At the moment,

pyrenew-hew/hewr/R/process_loc_forecast.R

Lines 355 to 375 in 115c9ef

	#' Detect model type from model name
	#'
	#' Internal helper function to infer model type from naming conventions
	#'
	#' @param model_name Character string with model name
	#' @return Character string with model type ("pyrenew", "timeseries", etc.)
	#' @keywords internal
	detect_model_type <- function(model_name) {
	if (
	grepl("^ts_", model_name) ||
	grepl("timeseries", model_name, ignore.case = TRUE)
	) {
	return("timeseries")
	} else if (grepl("epiautogp", model_name, ignore.case = TRUE)) {
	return("epiautogp")
	} else {
	# Default to pyrenew for backward compatibility
	return("pyrenew")
	}
	}

is a pretty weak pattern since the point of introducing an S3 generic to handle model specific post-processing was to avoid needing to add new parameters to all the required post-processing functions for any new model.

I think we should move to passing the model_type to dispatch on, assuming that we need a model_name which is distinct e.g. model_name = pyrenew_he with model_type = pyrenew.

This is connected to #789 since it is effectively a code organisation decision.

[SamuelBrand1]

This is one way to fulfil #789.

Eventually, for any model we offer a range of utilities in both pipelines and hewr. To add a new model you need to offer:

• As required, extensions to pipeline classes (these don't exist yet but cf Issue 780: Adds forecast_epiautogp #783 for my pass on that) that extend existing methods where the default methods don't work
• As required, a new S3 method for the post-processing (if the default method doesn't work).

There can then exist an entrypoint script forecast_model.py that can read something like a config file and do the model forecasting via dispatch rather than having to have a different entrypoint for each model.