shenhanqian / tyro Goto Github PK

Documentation   •   pip install tyro

tyro is a tool for building command-line interfaces and configuration objects in Python.

Our core interface, tyro.cli(), generates command-line interfaces from type-annotated callables.

To summarize how tyro.cli() can be used, let's consider a script based on argparse. We define two inputs and print the sum:

"""Sum two numbers from argparse."""
import argparse

parser = argparse.ArgumentParser()
parser.add_argument("--a", type=int, required=True)
parser.add_argument("--b", type=int, default=3)
args = parser.parse_args()

total = args.a + args.b

print(total)

This pattern is dramatically cleaner than manually parsing sys.argv, but has several issues: it lacks type checking and IDE support (consider: jumping to definitions, finding references, docstrings, refactoring and renaming tools), requires a significant amount of parsing-specific boilerplate, and becomes difficult to manage for larger projects.

The basic goal of tyro.cli() is to provide a wrapper for argparse that solves these issues.

(1) Command-line interfaces from functions.

We can write the same script as above using tyro.cli():

"""Sum two numbers by calling a function with tyro."""
import tyro

def add(a: int, b: int = 3) -> int:
    return a + b

# Populate the inputs of add(), call it, then return the output.
total = tyro.cli(add)

print(total)

Or, more succinctly:

"""Sum two numbers by calling a function with tyro."""
import tyro

def add(a: int, b: int = 3) -> None:
    print(a + b)

tyro.cli(add)  # Returns `None`.

(2) Command-line interfaces from config objects.

A class in Python can be treated as a function that returns an instance. This makes it easy to populate explicit configuration structures:

"""Sum two numbers by instantiating a dataclass with tyro."""
from dataclasses import dataclass

import tyro

@dataclass
class Args:
    a: int
    b: int = 3

args = tyro.cli(Args)
print(args.a + args.b)

Unlike directly using argparse, both the function-based and dataclass-based approaches are compatible with static analysis; tab completion and type checking will work out-of-the-box.

(3) Additional features.

These examples only scratch the surface of what's possible. tyro aims to support all reasonable type annotations, which can help us define things like hierachical structures, enums, unions, variable-length inputs, and subcommands. See documentation for examples.

tyro is still a new library, but being stress tested in several projects!

• nerfstudio-project/nerfstudio provides a set of tools for end-to-end training, testing, and rendering of neural radiance fields.
• Sea-Snell/JAXSeq is a library for distributed training of large language models in JAX.
• kevinzakka/obj2mjcf is an interface for processing composite Wavefront OBJ files for Mujoco.
• blurgyy/jaxngp is a CUDA-accelerated implementation of instant-ngp, implemented in JAX.
• NVIDIAGameWorks/kaolin-wisp is a PyTorch library for working with neural fields.
• openrlbenchmark/openrlbenchmark is a comprehensive collection of tracked experiments for reinforcement learning.