gplite 2.1.3

Lightweight Gaussian Process regression library

gplite

A lightweight Gaussian Process regression library built on NumPy and SciPy, originally designed for delta machine learning in computational chemistry.

Features

• Gaussian Process regression with automatic hyperparameter optimization
• Multiple kernels: RBF, Matérn, Periodic, Constant (combinable via + and *)
• Active learning with multiple selection strategies (uncertainty, MAE, expected improvement)
• Model saving and loading with pickle
• Anisotropic kernels (support for per-dimension hyperparameters)
• String export for integration with external tools

Installation

pip install gplite

or for optional dependencies to run the example files

pip install "gplite[examples]"

Quick Start (more detailed examples can be found here)

Gaussian Process Regression

import numpy as np
from gplite import GaussianProcess, RBFKernel

X_train = np.linspace(0, 10, 20).reshape(-1, 1)
y_train = np.sin(X_train).ravel() + 0.1 * np.random.randn(20)

kernel = RBFKernel(length_scale=1.0)
gp = GaussianProcess(kernel)
gp.fit(X_train, y_train, optimize=True)

X_test = np.linspace(0, 10, 100).reshape(-1, 1)
y_mean, y_std = gp.predict(X_test, return_std=True)

Matérn Kernel

from gplite import MaternKernel

# Matérn 5/2 (twice differentiable, default)
kernel = MaternKernel(length_scale=1.0, nu=2.5)

# Matérn 3/2 (once differentiable, rougher functions)
kernel = MaternKernel(length_scale=1.0, nu=1.5)

Combining Kernels

from gplite import RBFKernel, PeriodicKernel, ConstantKernel
import numpy as np

# additive: smooth trend + periodic pattern
kernel = RBFKernel(length_scale=2.0) + PeriodicKernel(length_scale=1.0, period=2*np.pi)

# product: scaled periodic
kernel = ConstantKernel(constant=2.0) * PeriodicKernel(length_scale=1.0, period=2*np.pi)

Anisotropic Kernels

from gplite import RBFKernel

# separate length scale per input dimension
kernel = RBFKernel(length_scale=[1.0, 5.0, 0.5], isotropic=False)

Active Learning

from gplite import ActiveLearner, PeriodicKernel
import numpy as np

X_full = np.linspace(0, 2*np.pi, 100).reshape(-1, 1)
y_full = np.sin(X_full).ravel()

kernel = PeriodicKernel(length_scale=1.0, period=2*np.pi)
learner = ActiveLearner(
    kernel=kernel,
    x_full=X_full,
    y_full=y_full,
    rmse_threshold=0.1,
    max_points=50
)

learner.learn(learning_strategy="uncertainty")
y_pred = learner.gp.predict(X_full)

Save and Load

# save a fitted model
gp.save("model.pkl")

# load it back
from gplite import GaussianProcess
gp_loaded = GaussianProcess.load("model.pkl")
y_pred = gp_loaded.predict(X_test)

String Export

expression = gp.to_str(variable_names=["x"])

Modules

• GaussianProcess - core GP regression
• Kernels - RBF, Matérn, Periodic, Constant, composites
• ActiveLearning - data sampling strategies
• Optimization - hyperparameter optimization

Requirements

• Python >= 3.10
• NumPy >= 2.2
• SciPy >= 1.15

License

GPLv3