Commands like fairseq-train currently does not display the full list of options.
This can be improved upon by using sub-parsers.
By refactoring the codebase with sub-parsers, we can
fairseq-cli though command-line help messages.parse_args_and_arch.Convert arguments like --arch to use sub-parsers instead.
There are other examples like --criterion etc.
Alternatively, we can stick with the current code-style.
But personally, I think the current level of smell is sufficient to warrant a refactor.
The smellier it gets, the more effort will be needed in the future.
I'm a willing contributor! I work as a software engineer on our machine translation experimentation platform at work.
Yes, I completely agree. We have an active effort to migrate away from argparse to instead use dataclasses and hydra. We hope to share more details soon.
cc @kevinmtian @alexeib
Most helpful comment
Yes, I completely agree. We have an active effort to migrate away from argparse to instead use dataclasses and hydra. We hope to share more details soon.
cc @kevinmtian @alexeib