A metadata approach to fitting directories?

[SamuelBrand1]

At the moment we are gathering important information about inference/score runs based on specific locations in a directory structure, and parsing the dir name,

e.g.

pyrenew-hew/hewr/R/directory_utils.R

Lines 81 to 93 in b90dd9f

	parse_model_run_dir_path <- function(model_run_dir_path) {
	batch_dir <- model_run_dir_path |>
	fs::path_dir() |>
	fs::path_dir() |>
	fs::path_file()
	location <- fs::path_file(model_run_dir_path)
	return(c(
	location = location,
	parse_model_batch_dir_path(batch_dir)
	))
	}

This will work well for us at the moment, but I think its inflexible in the medium term because it hard encodes the expected directory structure.

I'm personally against encoding information in dirnames, I'd prefer an autogenerated set of metadata for runs either as a .toml file that gets generated as the inference is done or handled by a work scheduling tool e.g. something like mlflow. I might well be missing something though.

[damonbayer]

This could help with some of the naming collisions I mentioned in #126

[damonbayer]

Related to #62

[dylanhmorris]

I agree and favor a .toml file