Welcome to MLXP’s documentation!¶
MLXP (Machine Learning eXperimentalist for Python) package is an open-source Python framework for managing multiple experiments with a flexible option structure from launching, and logging to querying results. The source code is available in MLXP’s GitHub repository.
Key functionalities¶
Launching several jobs automatically using hydra and hierarchical configs by adding a single decorator to the main task function.
Logging outputs (metrics, artifacts, checkpoints) of a job in a uniquely assigned directory along with all metadata and configuration options to reproduce the experiment.
Code version management by automatically generating a deployment version of the code based on the latest git commit.
Submitting jobs to a cluster using a job scheduler.
Exploiting the results of several experiments by easily reading, querying, grouping, and aggregating the output of several jobs.
Features |
MLXP |
MLFlow tracking |
Weights & Biases |
|---|---|---|---|
Configuration management |
|
|
|
Hyperparameter Sweeps |
|
|
|
Launching/Job submission |
|
|
|
Logging/Tracking |
|
|
|
Experiment Search/Filtering |
|
|
|
Experiment Comparison |
|
|
|
Lazy loading results |
|
|
|
Code versioning |
|
|
|
Code version checking |
|
|
|
Model Versioning |
|
|
|
Interactive Visualizations |
|
|
|
Collaboration and Sharing |
|
|
|
License¶
MLXP is distributed under MIT license.
Citing MLXP¶
Even though this is non-legally binding, the author kindly ask users to cite MLXP in their publications if they use it in their research as follows:
@Misc{mlxp2024,
author = {Michael Arbel, Alexandre Zouaoui},
title = {MLXP: A framework for conducting replicable Machine Learning eXperiments in Python},
howpublished = {arXiv preprint arXiv:2402.13831},
year = {2024},
url = {https://arxiv.org/abs/2402.13831}
}
Acknowledgments¶
The authors would like to thank the following people for their valuable feedback and contributions to the development of MLXP:
Juliette Marrie tested an early version of MLXP. We are grateful for her feedback which was extremetly helpful for shaping and improving MLXP’s functionalities.
Romain Ménégaux and Pierre Wolinski provided valuable feedback and suggestions to improve MLXP. They also found and reported several bugs in the software which helped improve its quality and stability.
Thomas Ryckeboer for providing guidance on continuous integration of MLXP.