Decorate your functions to save your experiments

What it does

Sacred is a Python library that wraps experiments in decorators to capture configuration, inject parameters by name into functions, and log runs via pluggable observers. It generates a command-line interface for every experiment and can store configurations, dependencies, host details, and results in MongoDB. The idea is to remove the tedious bookkeeping around your actual experiment code so you can rerun or reproduce a specific setup later.

The interesting bit

The core mechanic is the @ex.config decorator: local variables inside a function become your experiment’s hyperparameters, and Sacred automatically injects those values into any function that asks for them by name. It is a distinctly Pythonic approach to experiment management—born at IDSIA and published at SciPy 2017—that relies on decorators and MongoDB rather than external configuration files.

Key highlights

• Config scopes turn ordinary function-local variables into structured experiment parameters
• Automatic config injection passes parameters into functions by matching argument names
• Built-in observers log configurations, dependencies, host details, and results to MongoDB
• Automatic seeding controls randomness to aid reproducibility
• A small ecosystem of third-party frontends (Omniboard, Incense, Sacredboard) provides dashboards and Jupyter integration

Caveats

• Dashboard frontends are externally maintained and fragmented; at least one early prototype (Prophet) is explicitly discontinued
• MongoDB is the primary documented storage backend, with support for other databases unclear

Verdict

Sacred suits researchers and Python developers who want self-hosted experiment tracking without subscribing to a managed service, especially if they already run MongoDB. Teams looking for a polished, collaborative SaaS experience will likely find modern alternatives more convenient.