Skip to content

Repository files navigation

Multi-Output Gaussian Process Toolkit

Paper - API Documentation - Tutorials & Examples - Code of Conduct

MOGPTK (Multi-Output Gaussian Process Toolkit) is an open-source Python library for interpretable probabilistic modelling of multichannel time series. Built on PyTorch, it provides Gaussian process models with a range of covariance architectures, spectral analysis tools, GPU acceleration, and visualisation utilities for scientific applications underpinned by temporal observations.

The project was initiated in 2020 at the Center for Mathematical Modelling, Universidad de Chile, and has been hosted at Imperial College London since October 2024. Development has benefited from research supported by:

  • Center for Mathematical Modelling, Universidad de Chile (2020–2024)
  • ANID Fondecyt research grants, Chile (2020–2024)
  • Google Research Awards (2020–2024)

Installation

With Anaconda installed on your system, open a command prompt and create a virtual environment:

conda create -n myenv python=3.14
conda activate myenv

where myenv is the name of your environment, and where the version of Python could be 3.6 or above. Next we will install this toolkit and automatically install the necessary dependencies such as PyTorch.

pip install mogptk

In order to upgrade to a new version of MOGPTK or any of its dependencies, use --upgrade as follows:

pip install --upgrade mogptk

For developers of the library or for users who need the latest changes, we recommend cloning the git master or develop branch and using the following command inside the repository folder:

pip install --upgrade -e .

See Tutorials & Examples to get started.

Introduction

This repository provides a toolkit to perform multi-output GP regression with kernels that are designed to utilise correlation information among channels in order to better model signals. The toolkit is mainly targeted at time series, and includes plotting functions for the case of single input with multiple outputs (time series with several channels).

The main kernel corresponds to Multi Output Spectral Mixture Kernel, which correlates every pair of data points (irrespective of their channel of origin) to model the signals. This kernel is specified in detail in the following publication: G. Parra, F. Tobar, Spectral Mixture Kernels for Multi-Output Gaussian Processes, Advances in Neural Information Processing Systems, 2017. Available here

The kernel learns the cross-channel correlations of the data, so it is particularly well-suited for the task of signal reconstruction in the event of sporadic data loss. All other included kernels can be derived from the Multi Output Spectral Mixture kernel by restricting some parameters or applying some transformations.

One of the main advantages of the present toolkit is the GPU support, which enables the user to train models through PyTorch, speeding computations significantly. It also includes sparse-variational GP regression functionality to decrease computation time even further.

See MOGPTK: The Multi-Output Gaussian Process Toolkit for our publication in Neurocomputing.

Features

Implemented inference models:

  • Exact Gaussian process (maximum likelihood)
  • Snelson (E. Snelson, Z. Ghahramani, "Sparse Gaussian Processes using Pseudo-inputs", 2005)
  • OpperArchambeau (M. Opper, C. Archambeau, "The Variational Gaussian Approximation Revisited", 2009)
  • Titsias (Titsias, "Variational learning of induced variables in sparse Gaussian processes", 2009)
  • Hensman (J. Hensman, et al., "Scalable Variational Gaussian Process Classification", 2015)

Implemented likelihoods:

  • Gaussian
  • Student-T
  • Exponential
  • Laplace
  • Bernoulli
  • Beta
  • Gamma
  • Poisson
  • Weibull
  • Log-Logistic
  • Log-Gaussian
  • Chi
  • Chi-Squared

Tutorials

00 - Quick Start: Short notebook showing the basic use of the toolkit.

01 - Data Loading: Functionality to load CSVs and DataFrames while using formatters for dates.

02 - Data Preparation: Handle data, removing observations to simulate sensor failure and apply tranformations to the data.

03 - Parameter Initialization: Parameter initialization using different methods, for single output regression using spectral mixture kernel and multioutput case using MOSM kernel.

04 - Model Training: Training of models while keeping certain parameters fixed.

05 - Error Metrics Obtain different metrics in order to compare models.

06 - Custom Kernels and Mean Functions Use or create custom kernels as well as training custom mean functions.

07 - Sparse Multi Input Use 8 input dimensions to train the Abalone data set using sparse GPs.

08 - Multi Likelihood Classification Use a different likelihood for each channel, one Bernoulli for classification and one StudentT's for regression.

Examples

Airline passengers: Regression using a single output spectral mixture on the yearly number of passengers of an airline.

Seasonal CO2 of Mauna-Loa: Regression using a single output spectral mixture on the CO2 concentration at Mauna-Loa throughout many years.

Currency Exchange: Model training, interpretation and comparison on a dataset of 11 currency exchange rates (against the dollar) from 2017 and 2018. These 11 channels are fitted with the MOSM, SM-LMC, CSM, and CONV kernels and their results are compared and interpreted.

Gold, Oil, NASDAQ, USD-index: The commodity indices for gold and oil, together with the indices for the NASDAQ and the USD against a basket of other currencies, we train multiple models to find correlations between the macro economic indicators.

Human Activity Recognition: Using the Inertial Measurement Unit (IMU) of an Apple iPhone 4, the accelerometer, gyroscope and magnetometer 3D data were recorded for different activities resulting in nine channels.

Bramblemet tidal waves: Tidal wave data set of four locations in the south of England. We model the tidal wave periods of approximately 12.5 hours using different multi-output Gaussian processes.

Documentation

See the API documentation for documentation of our toolkit, including usage and examples of functions and classes.

Development Roadmap

MOGPTK is developed through a long-term roadmap that advances the toolkit as a general platform for probabilistic modelling of multichannel time series. While the software remains broadly applicable across scientific and engineering domains, recent collaborations and interactions with researchers working in the life sciences have motivated some of the scheduled improvements to the toolbox.

Current development is organised around four complementary themes:

  • Scale: scalable probabilistic modelling for larger datasets and more channels.
  • Represent: richer probabilistic models and interpretable latent representations.
  • Learn: integration with modern AI methods for learning from limited or partially labelled data.
  • Apply: reproducible scientific workflows, benchmarking, visualisation, and high-quality tutorials.

Contributing

We welcome contributions across all areas of the toolkit, including bug fixes, documentation, tutorials, performance improvements, new models, and scientific applications. We accept contributions through pull requests (PRs), bug reports, GitHub Discussions, and feature requests. Please consider starting an open discussion before proposing substantial new features. For smaller PRs, we recommend that each addresses a single issue or introduces one new feature. All PRs should keep documentation and notebooks up to date. For more details, see our Contribution Guidelines.

Citing MOGPTK

Please refer to the publication in Neurocomputing MOGPTK: The Multi-Output Gaussian Process Toolkit. We recommend the following BibTeX entry:

@article{mogptk,
    author = {T. {de Wolff} and A. {Cuevas} and F. {Tobar}},
    title = {{MOGPTK: The Multi-Output Gaussian Process Toolkit}},
    journal = "Neurocomputing",
    year = "2020",
    issn = "0925-2312",
    doi = "https://doi.org/10.1016/j.neucom.2020.09.085",
    url = "https://github.com/GAMES-UChile/mogptk"
}

Authors

  • Taco de Wolff
  • Alejandro Cuevas
  • Felipe Tobar

License

Released under the MIT license.

Citations

Books

Used in code

About

Multi-Output Gaussian Process Toolkit

Resources

Code of conduct

Contributing

Stars

189 stars

Watchers

8 watching

Forks

Releases

Packages

Used by

Contributors

Languages