params-proto v3.2.1

Declarative hyperparameter management for ML experiments with automatic CLI generation.

Installation

pip install params-proto==3.2.0

Three Decorators

Quick Start

Simple CLI Script

from params_proto import proto

@proto.cli
def train(
    lr: float = 0.001,  # Learning rate (inline comment = help text)
    batch_size: int = 32,  # Batch size
    epochs: int = 100,  # Number of epochs
):
    """Train a model."""  # Docstring = CLI description
    print(f"Training with lr={lr}")

if __name__ == "__main__":
    train()

python train.py --lr 0.01 --batch-size 64
python train.py --help

Multi-Namespace Configuration

@proto.prefix
class Model:
    name: str = "resnet50"  # Architecture
    dropout: float = 0.5  # Dropout rate

@proto.prefix
class Training:
    lr: float = 0.001  # Learning rate
    epochs: int = 100  # Epochs

@proto.cli
def main(seed: int = 42):
    """Train with namespaced config."""
    print(f"Model: {Model.name}, LR: {Training.lr}")

# CLI: python train.py --model.name vit --training.lr 0.01

Environment Variables

from params_proto import proto, EnvVar

@proto.cli
def train(
    lr: float = EnvVar @ "LEARNING_RATE" | 0.001,  # Env var with default
    api_key: str = EnvVar @ "API_KEY",  # Required env var (no default)
    # OR operation: try multiple env vars in order
    token: str = EnvVar @ "API_TOKEN" @ "AUTH_TOKEN" | "default",
): ...

Union Types (Subcommand Pattern)

from dataclasses import dataclass

@dataclass
class Adam:
    lr: float = 0.001
    beta1: float = 0.9

@dataclass
class SGD:
    lr: float = 0.01
    momentum: float = 0.9

@proto.cli
def train(optimizer: Adam | SGD):
    """Train with selected optimizer."""
    print(f"Using {type(optimizer).__name__}")

# CLI: python train.py adam --lr 0.001
# CLI: python train.py sgd --momentum 0.95

Hyperparameter Sweeps with piter

from params_proto.hyper import piter

# Zip (default): pairs values element-wise
configs = piter @ {"lr": [0.001, 0.01], "batch_size": [32, 64]}
# 2 configs: (0.001, 32), (0.01, 64)

# Cartesian product with * (only first needs piter @)
configs = piter @ {"lr": [0.001, 0.01]} * {"batch_size": [32, 64]}
# 4 configs: all combinations

# Override with fixed values using %
configs = piter @ {"lr": [0.001, 0.01]} * {"batch_size": [32, 64]} % {"seed": 42}

# Repeat for multiple trials using **
configs = (piter @ {"lr": [0.001, 0.01]}) ** 3  # 2 configs x 3 trials

for config in configs:
    train(**config)

Type Annotations

Boolean Flags

@proto.cli
def train(
    verbose: bool = False,  # --verbose sets True
    cuda: bool = True,      # --no-cuda sets False
): ...

Override Priority (highest to lowest)

1. CLI arguments
2. Direct assignment (Config.lr = 0.01)
3. Context manager (with proto.bind(Config, lr=0.01): ...)
4. Environment variables
5. Default values

Getting a Clean Dict

Config._dict      # → {'lr': 0.001, 'batch_size': 32}
dict(Config)      # → same (works for classes and functions)

Reference Files

For detailed documentation, see:

• cli-and-types.md - @proto.cli, @proto.prefix, type system
• environment-vars.md - EnvVar with templates and inheritance
• sweeps.md - piter and Sweep for hyperparameter search
• patterns.md - Common ML patterns and examples