curifactory 0.20.1

An experiment workflow and organization tool

Curifactory

Curifactory is a library and CLI tool designed to help organize and manage research experiments in python.

Experiment management must fulfill several tasks, including experiment orchestration, parameterization, caching, reproducibility, reporting, and parallelization. Existing projects such as MLFlow, MetaFlow, Luigi, and Pachyderm support these tasks in several different ways and to various degrees. Curifactory provides a different opinion, with a heavier focus on supporting general research experiment workflows for individuals or small teams working primarily in python.

You can read more about these design principles in our paper in the SciPy 2022 proceedings.

Features

• Adds a CLI layer on top of your codebase, a single entrypoint for running experiments
• Automatic caching of intermediate data and lazy loading of stored objects
• Jupyter notebook output for further exploration of an experiment run
• Docker container output with copy of codebase, conda environment, full experiment run cache, and jupyter run notebook
• HTML report output from each run with graphviz-rendered diagram of experiment
• Easily report plots and values to HTML report
• Configuration files are python scripts, allowing programmatic definition, parameter composition, and parameter inheritance
• Output logs from every run
• Run experiments directly from CLI or other python code, notebooks, etc.

Installation

pip install curifactory

Graphviz is required for certain features and can be installed through conda via:

conda install python-graphviz

Tab completion

For tab-completion in bash/zsh, install the argcomplete package (if using curifactory inside a conda environment, you'll need to install this in your system python.)

pip install argcomplete

To enable, you can either use argcomplete's global hook activate-global-python-argcomplete, which will enable tab complete on all argcomplete-enabled python packages (e.g. pytest), or you can add eval "$(register-python-argcomplete experiment)" to your shell's rc file. Curifactory can add this line for you automatically with:

curifactory completion [--bash|--zsh]  # use the shell flag appropriate

Once enabled, the experiment command will provide tab complete for experiment names, parameter names, and flags.

Requirements

OS: We primarily develop and test Curifactory on Linux, but it runs on Windows and MacOS as well.

Python: 3.9-3.11

Optional:

• Conda/Mamba
• Graphviz
• Docker
• Jupyter notebook/lab

Documentation

The documentation for the latest version of Curifactory can be found at: https://ornl.github.io/curifactory/stable/index.html.

Examples

Several small example can be found in the examples folder. examples/notebooks includes walkthroughs demonstrating usage of curifactory solely in Jupyter.

Citation

Please use the following BibTeX if citing this project:

@article{Martindale2023,
  doi = {10.21105/joss.05793},
  url = {https://doi.org/10.21105/joss.05793},
  year = {2023},
  publisher = {The Open Journal},
  volume = {8},
  number = {90},
  pages = {5793},
  author = {Nathan Martindale and Scott L. Stewart and Jason Hite and Mark B. Adams},
  title = {Curifactory: A research experiment manager},
  journal = {Journal of Open Source Software}
}

Similar Projects

Curifactory is one tool and one opinion among many, other projects that have similar goals and/or approaches:

• Kedro
• Tango
• Sacred
• MLFlow