How to provide default argument values for a function?

[Yura52]

I have a function where arguments do not have default values. However, I want to generate a CLI for that function with default values for some arguments (for example, based on some config). Something like this:

import json

import tyro


# I don't want this function to have default argument values.
def greet(name: str) -> None:
    print(f'Hello, {name}')


def _load_user_config():
    with open('user_config.json') as f:
        return json.load(f)


if __name__ == '__main__':
    user_config = _load_user_config()
    tyro.cli(greet, defaults={'name': user_config['name']})

Is it possible to achieve this with tyro?

[brentyi]

Hi @Yura52 , that's a good question!

functools.partial can be used to assign defaults to kwargs in Python. This should also work in tyro:

from functools import partial
import tyro

# I don't want this function to have default argument values.
def greet(name: str) -> str:
    print(f"Hello, {name}")

default_kwargs = {"name": "Alice"}
tyro.cli(partial(greet, **default_kwargs))

Here's the helptext you should get:

There are tests here: https://github.com/brentyi/tyro/blob/main/tests/test_partial.py

We should probably add this as an example to the docs.

[Yura52]

Hi @brentyi , thank you for the hint! It helped me to replace my verbose argparse-based code with tyro. I wonder though why function defaults are not supported more "natively", e.g. through the default argument as for dataclasses?

[brentyi]

Great 🙂

Supporting function defaults in the default= kwarg would be easy, I haven't because:

• it doesn't seem like a super common need
• hard to type-check; partial works fine and is special-cased by type checkers
• generally try to be conservative about extending the API

I wouldn't rule it out though!

[amas127]

Hi @brentyi , thank you for the hint! It helped me to replace my verbose argparse-based code with tyro. I wonder though why function defaults are not supported more "natively", e.g. through the default argument as for dataclasses?

functools.partial is the standard way of injecting arguments to functions in python. U can even go deeper by customizing this function if u like.