params-proto 3.3.0

Modern Hyper Parameter Management for Machine Learning

params-proto: Modern Declarative Parameters for Machine Learning

params-proto is a lightweight, decorator-based library for defining configurations in Python. Write your parameters once with type hints and inline documentation, and get automatic CLI parsing, validation, and help generation.

Note: This is v3 with a completely redesigned API. For the v2 API, see params-proto-v2.

Why params-proto?

Stop fighting with argparse and click. With params-proto, your configuration is your documentation:

from params_proto import proto

@proto.cli
def train_mnist(
    batch_size: int = 128,  # Training batch size
    lr: float = 0.001,  # Learning rate
    epochs: int = 10,  # Number of training epochs
):
    """Train an MLP on MNIST dataset."""
    print(f"Training with lr={lr}, batch_size={batch_size}, epochs={epochs}")
    # Your training code here...

if __name__ == "__main__":
    train_mnist()

That's it! No argparse boilerplate, no manual help strings, no type conversion logic. Just pure Python functions with type hints and inline comments.

Run it:

$ python train.py --help
usage: train.py [-h] [--batch-size INT] [--lr FLOAT] [--epochs INT]

Train an MLP on MNIST dataset.

options:
  -h, --help           show this help message and exit
  --batch-size INT     Training batch size (default: 128)
  --lr FLOAT           Learning rate (default: 0.001)
  --epochs INT         Number of training epochs (default: 10)

$ python train.py --lr 0.01 --batch-size 256
Training with lr=0.01, batch_size=256, epochs=10

Note: The actual terminal output includes beautiful ANSI colors! See the demo below or check the documentation for colorized examples.

Try It Now

Want to see the colorized help in action? Clone the repo and run the demo:

# Clone and setup
git clone https://github.com/geyang/params-proto.git
cd params-proto
uv sync

# See the colorized help (with bright blue types, bold cyan defaults, bold red required)
uv run python scratch/demo_v3.py --help

# Try running without required --seed (shows error)
uv run python scratch/demo_v3.py
# Error: the following arguments are required: --seed

# Run with required parameter (keyword syntax)
uv run python scratch/demo_v3.py --seed 42

# Or use positional syntax
uv run python scratch/demo_v3.py 42

Installation

pip install params-proto==3.0.0-rc25

Key Features

1. Function-based Configs

Define parameters using type-annotated functions:

@proto.cli
def train(
    model: str = "resnet50",  # Model architecture
    dataset: str = "imagenet",  # Dataset to use
    gpu: bool = True,  # Enable GPU acceleration
):
    """Train a model on a dataset."""
    print(f"Training {model} on {dataset}")

2. Class-based Configs

Or use classes for more structure:

@proto
class Params:    """Training configuration."""

    # Model settings
    model: str = "resnet50"
    pretrained: bool = True  # Use pretrained weights

    # Training settings
    lr: float = 0.001  # Learning rate
    batch_size: int = 32  # Batch size
    epochs: int = 100  # Number of epochs

3. Singleton Prefixed Configs

Create modular, reusable configuration groups:

from params_proto import proto

@proto.prefix
class Environment:
    """Environment configuration."""
    domain: str = "cartpole"  # Domain name
    task: str = "swingup"  # Task name
    time_limit: float = 10.0  # Episode time limit

@proto.prefix
class Agent:
    """Agent hyperparameters."""
    algorithm: str = "SAC"  # RL algorithm
    lr: float = 3e-4  # Learning rate
    gamma: float = 0.99  # Discount factor

@proto.cli
def train_rl(
    seed: int = 0,  # Random seed
    total_steps: int = 1000000,  # Total training steps
):
    """Train RL agent on dm_control."""
    print(f"Training {Agent.algorithm} on {Environment.domain}-{Environment.task}")
    print(f"Agent LR: {Agent.lr}, Gamma: {Agent.gamma}")

Command line:

$ python train_rl.py --Agent.lr 0.001 --Environment.domain walker --seed 42
Training SAC on walker-swingup
Agent LR: 0.001, Gamma: 0.99

4. Multiple Override Patterns

Override parameters in multiple ways:

# 1. Command line
$ python train.py --lr 0.01

# 2. Direct attribute assignment
Params.lr = 0.01

# 3. Function call with kwargs
train(lr=0.01, batch_size=256)

# 4. Using proto.bind() context manager
with proto.bind(lr=0.01, **{"train.epochs": 50}):
    train()

5. Rich Type System

Support for complex types:

from typing import Literal, Union
from enum import Enum, auto

class Optimizer(Enum):
    ADAM = auto()
    SGD = auto()
    RMSPROP = auto()

@proto
class Params:    # Union types
    precision: Literal["fp16", "fp32", "fp64"] = "fp32"

    # Enums
    optimizer: Optimizer = Optimizer.ADAM

    # Tuples
    image_size: tuple[int, int] = (224, 224)

    # Optional types
    checkpoint: str | None = None

Quick Start

1. Define your configuration with a decorated function or class
2. Add type hints for automatic validation
3. Add inline comments for automatic documentation
4. Call your function - params-proto handles the rest!

See our Quick Start Guide for more.

Documentation

• Quick Start - Get started in 5 minutes
• User Guide - Detailed documentation
• Examples - Real-world usage patterns
• API Reference - Complete API documentation
• Migration from v2 - Upgrade guide

What Changed in v3?

v2 (old): Class-based with inheritance

from params_proto import ParamsProto

class Args(ParamsProto):
    lr = 0.001
    batch_size = 32

v3 (new): Decorator-based with type hints

from params_proto import proto

@proto
class Args:
    lr: float = 0.001  # Learning rate
    batch_size: int = 32  # Batch size

Key improvements:

• ✅ Cleaner decorator syntax (no inheritance needed)
• ✅ Full IDE support with type hints
• ✅ Inline documentation becomes automatic help text
• ✅ Support for functions, not just classes
• ✅ Better Union types and Enum support
• ✅ Simplified singleton pattern with @proto.prefix

Contributing

git clone https://github.com/episodeyang/params_proto.git
cd params_proto
make dev test

To publish:

make publish

License

MIT License - see LICENSE for details.