stryx 0.1.3

Typed configuration compiler for ML experiments

Stryx

Stryx is a typed configuration compiler and experiment manager for Machine Learning. It sits between tools like Sacred and Hydra, offering a lightweight, composable approach without the complexity of Hydra's indirection.

Key Features

• Type-First: Configurations are defined as Pydantic models (or dataclasses), providing IDE support and validation.
• Recipes as Source: Experiment configurations are compiled into static YAML "recipes" that include metadata and lineage.
• Minimal Surface: A single decorator (@stryx.cli) adds powerful CLI capabilities to any script.
• Reproducibility: Defaults are code-driven, overrides are explicit, and recipes track their origin.
• Lifecycle Management: Automatic manifest recording (git state, logs, config) for every run.

Installation

pip install stryx

Quick Start

Define your configuration schema and decorate your entry point:

# train.py
from pydantic import BaseModel
import stryx

class Config(BaseModel):
    lr: float = 1e-4
    epochs: int = 10

@stryx.cli(schema=Config)
def main(cfg: Config):
    print(f"Training with lr={cfg.lr}, epochs={cfg.epochs}")

if __name__ == "__main__":
    main()

CLI Usage

Your script now has a powerful CLI.

1. New Experiments

Create a reusable recipe (configuration file).

# Create from defaults with overrides
python train.py new my_exp lr=1e-3

# Create without name (auto-generated: exp_001.yaml)
python train.py new --message "Baseline run"

2. Experimentation (try)

Run a quick experimental variant without creating a permanent recipe. These are saved to scratches/.

# Try modifying a specific recipe
python train.py try my_exp epochs=5

# Try modifying defaults
python train.py try lr=5e-4

3. Execution (run)

Run a recipe exactly as defined (strict mode).

python train.py run my_exp

4. Lineage (fork)

Create a new experiment based on an existing one.

python train.py fork my_exp better_exp lr=2e-4 --message "Lower LR"

5. Inspection & Management

• List: See all recipes and runs.

  python train.py list configs  # List recipes
  python train.py list runs     # List execution history
  

• Show: Inspect a configuration with annotated sources (default vs recipe vs override).

  python train.py show my_exp
  

• Diff: Compare two recipes.

  python train.py diff my_exp better_exp
  

• Schema: Print the configuration schema (or JSON).

  python train.py schema
  python train.py schema --json
  

• Edit: Open a recipe in an interactive TUI (Terminal UI).

  python train.py edit my_exp
  

CLI Reference

Command	Description
new [name] [overrides...]	Create a fresh experiment recipe from defaults.
fork <src> <name> [ov...]	Fork an existing recipe with modifications.
try [target] [ov...]	Run an experimental variant (saved to scratches).
run <target>	Run an existing recipe exactly (strict).
list {configs,runs}	List saved recipes or execution history.
show [target] [ov...]	Display configuration with source annotations.
diff <A> [B]	Compare two recipes (or A vs defaults).
schema	Show the configuration schema.
edit <recipe>	Edit a recipe interactively (TUI).

Global Options

• --runs-dir: Override directory for run logs and manifests.
• --configs-dir: Override directory for recipe storage.