bayesify takes the pain out of Bayesian performance-influence modeling. It can be used like a linear regression to determine the influence of individual software configuration options and to predict the performance of unseen configurations. Bayesify draws from a scientific paper (see below) as it adopts the major pre-processing and modeling steps. However, Bayesify uses numpyro as its Probabilistic Programming backend, greatly accelerating model fitting. As a result, it may be used for new experiments, but cannot reproduce the original paper results. For a replication package, please refer to the original supplementary website.
- probability distribution for each configuration option
- confidence intervals of custom confidence for the influence of each configuration option
- probability distribution as prediction
- custom confidence intervals as prediction
- robust data pre-processing module
- scikit-learn interface both for pre-processing and modeling
This Python package can be installed using pip:
pip install git+ssh://[email protected]/SWS/bayesify
For now, you can refer to the unit tests to see how to use bayesify.