gpvecchia

Vecchia approximation routines for Gaussian Process and Data Assimilation. This code is a lightweight adaption of code from this package.

Please see the examples for particular use cases.

Documentation

Documentation is available on read the docs.

Installation

pip

pip install gpvecchia

To install a local development version

pip install -e ./

To install latest from github

pip install git+https://github.com/tide-itrh/gpvecchia.git

Features

• Vecchia-approximated Gaussian Process likelihood, prediction, prior/posterior sampling and conditional sampling, scaling to large datasets by conditioning each point on a small set of nearest neighbours instead of the full covariance matrix.
• A small library of covariance kernels (Matern 1/2, 3/2, 5/2, squared-exponential, cosine) in gpvecchia.cov, plus anisotropic coordinate scaling and rotation.
• Numba-jitted core routines, parallelised across CPU cores.

Quick Usage

import numpy as np
from gpvecchia import GPtideVecchia
from gpvecchia.cov import matern32

# Training data locations and observed values
xd = np.linspace(0, 100, 1000)[:, None]
yd = np.sin(xd[:, 0] / 5) + 0.1 * np.random.randn(1000)

# covparams = (marginal std dev, length scale)
covparams = (1.0, 5.0)
noise = 0.1

GP = GPtideVecchia(
    xd, xd, noise, matern32, covparams,
    order_func=np.random.permutation, order_params=len(xd),
)

mean, std = GP(yd)

See the examples for anisotropic/scaled kernels, MLE and MCMC parameter estimation, and conditional sampling.