openboost 0.0.1

GPU-accelerated gradient boosting, written in Python

openboost

GPU-accelerated gradient boosting, written in Python.

Building Blocks

numba-cuda compiles Python to CUDA kernels:

@cuda.jit
def _histogram_kernel(binned, grad, hess, hist_grad, hist_hess):
    feature_idx = cuda.blockIdx.x
    local_grad = cuda.shared.array(256, dtype=float32)
    # ...

The Problem

GBDT research is active:

• DART — dropout for trees
• NGBoost — probabilistic predictions
• GAMLSS — distributional regression
• GOSS — gradient-based sampling
• GBDT-PL — linear leaves
• Oblivious Trees — symmetric splits

OpenBoost lets you implement these in Python.

Goal

GPU GBDT core in Python. Extend it to build variants.

Install

pip install openboost

Status

WIP.

License

Apache-2.0