BART vectoriZed¶
An implementation of Bayesian Additive Regression Trees (BART) in JAX.
If you don’t know what BART is, but know XGBoost, consider BART as a sort of Bayesian XGBoost. bartz makes BART run as fast as XGBoost.
BART is a nonparametric Bayesian regression technique. Given training predictors \(X\) and responses \(y\), BART finds a function to predict \(y\) given \(X\). The result of the inference is a sample of possible functions, representing the uncertainty over the determination of the function.
This Python module provides an implementation of BART that runs on GPU, to process large datasets faster. It is also good on CPU. Most other implementations of BART are for R, and run on CPU only.
On CPU, bartz runs at the speed of dbarts (the fastest implementation I know of) if n > 20,000, but using 1/20 of the memory. On GPU, the speed premium depends on sample size; it is convenient over CPU only for n > 10,000. The maximum speedup is currently 200x, on an Nvidia A100 and with at least 2,000,000 observations.
This Colab notebook runs bartz with n = 100,000 observations, p = 1000 predictors, 10,000 trees, for 1000 MCMC iterations, in 5 minutes.
Links¶
Citing bartz¶
Article: Petrillo (2024), “Very fast Bayesian Additive Regression Trees on GPU”, arXiv:2410.23244.
To cite the software directly, including the specific version, use zenodo.