Fast implementation of the MuyGPs scalable Gaussian process algorithm

MuyGPs is a scalable approximate Gaussian process (GP) model that achieves fast prediction and model optimization while retaining high-accuracy predictions and uncertainty quantification. The MuyGPyS implementation allows the user to easily create GP models that can quickly train and predict on million-scale problems on a laptop or scale to billions of observations on distributed memory systems using the same front-end code.

What is MuyGPyS?

MuyGPyS is a general-purpose Gaussian process library, similar to GPy, GPyTorch, or GPflow.

MuyGPyS differs from the other options in that it constructs approximate GP models using nearest neighbors sparsification, conditioning predictions only on the most relevant training data to drastically improve training time and time-to-solution on large-scale problems. Indeed, MuyGPyS is intended for GP problems with millions or more observations, and supports a distributed memory backend for smoothly scaling to billion-scale problems.

MuyGPs uses nearest neighbors sparsification and performs leave-one-out cross validation using regularized loss functions to rapidly optimize a GP model without evaluating a much more expensive likelihood, which is required by similar scalable methods.

Getting Started

See the illustration tutorial to see an illustration of why the neighborhood sparsification approach of MuyGPs works.

Next, see the univariate regression tutorial for a full description of the API and an end-to-end walkthrough of a simple regression problem.

The full documentation, including several additional tutorials with code examples, can be found at readthedocs.io.

Read further in this document for installation instructions.

Backend Math Implementation Options

In addition to the default basic numpy backend, as of release v0.6.6, MuyGPyS supports three additional backend implementations of all of its underlying math functions:

• MPI - distributed memory acceleration
• PyTorch - GPU acceleration and neural network integration
• JAX - GPU acceleration

It is possible to include the dependencies of any, all, or none of these additional backends at install time. Please see the below installation instructions.

MuyGPyS uses the MUYGPYS_BACKEND environment variable to determine which backend to use at import time. It is also possible to manipulate MuyGPyS.config to switch between backends programmatically. This is not advisable unless the user knows exactly what they are doing (and must occur before importing any other MuyGPyS components).

MuyGPyS will default to the numpy backend. It is possible to switch back ends by manipulating the MUYGPYS_BACKEND environment variable in your shell, e.g.

$ export MUYGPYS_BACKEND=jax    # turn on JAX backend
$ export MUYGPYS_BACKEND=torch  # turn on Torch backend
$ export MUYGPYS_BACKEND=mpi    # turn on MPI backend

Distributed memory support with MPI

The MPI version of MuyGPyS performs all tensor manipulation in distributed memory. The tensor creation functions will in fact create and distribute a chunk of each tensor to each MPI rank. This data and subsequent data such as posterior means and variances remains partitioned, and most operations are embarassingly parallel. Global operations such as loss function computation make use of MPI collectives like allreduce. If the user needs to reason about all products of an experiment, such the full posterior distribution in local memory, it is necessary to employ a collective such as MPI.gather.

The wrapped KNN algorithms are not distributed, and so MuyGPyS does not yet have an internal distributed KNN implementation. Future versions will support a distributed memory approximate KNN solution.

The user can run a script myscript.py with MPI using, e.g. mpirun (or srun if using slurm) via

$ export MUYGPYS_BACKEND=mpi
$ # mpirun version
$ mpirun -n 4 python myscript.py
$ # srun version
$ srun -N 1 --tasks-per-node 4 -p pbatch python myscript.py

PyTorch Integration

The torch version of MuyGPyS allows for construction and training of complex kernels, e.g., convolutional neural network kernels. All low-level math is done on torch.Tensor objects. Due to PyTorch's lack of support for the Bessel function of the second kind, we only support special cases of the Matern kernel, in particular when the smoothness parameter is $\nu = 1/2, 3/2,$ or $5/2$. The RBF kernel is supported as the Matern kernel with $\nu = \infty$.

The MuyGPyS framework is implemented as a custom PyTorch layer. In the high-level API found in examples/muygps_torch, a PyTorch MuyGPs model is assumed to have two components: a model.embedding which deforms the original feature data, and a model.GP_layer which does Gaussian Process regression on the deformed feature space. A code example is provided below.

Most users will want to use the MuyGPyS.torch.muygps_layer module to construct a custom MuyGPs model. The model can then be calibrated using a standard PyTorch training loop. An example of the approach based on the low-level API is provided in docs/examples/torch_tutorial.ipynb.

In order to use the MuyGPyS torch backend, run the following command in your shell environment.

$ export MUYGPYS_BACKEND=torch

One can also use the following workflow to programmatically set the backend to torch, although the environment variable method is preferred.

from MuyGPyS import config
MuyGPyS.config.update("muygpys_backend","torch")

...subsequent imports from MuyGPyS

Just-In-Time Compilation with JAX

MuyGPyS supports just-in-time compilation of the underlying math functions to CPU or GPU using JAX since version v0.5.0. The JAX-compiled versions of the code are significantly faster than numpy, especially on GPUs. In order to use the MuyGPyS torch backend, run the following command in your shell environment.

$ export MUYGPYS_BACKEND=jax

NOTE: There is a known conflict between recent versions of MuyGPyS and JAX on Python $\geq$ 3.9. The current fix is to downgrade to Python 3.8.

Precision

JAX and torch use 32 bit types by default, whereas numpy tends to promote everything to 64 bits. For highly stable operations like matrix multiplication, this difference in precision tends to result in a roughly 1e-8 disagreement between 64 bit and 32 bit implementations. However, MuyGPyS depends upon matrix-vector solves, which can result in disagreements up to 1e-2. Hence, MuyGPyS forces all back end implementations to use 64 bit types by default.

However, the 64 bit operations are slightly slower than their 32 bit counterparts, and limit throughput on GPUs. MuyGPyS accordingly supports 32 bit types, but this feature is experimental and might have sharp edges. For example, MuyGPyS might throw errors or otherwise behave strangely if the user passes arrays of 64 bit types while in 32 bit mode. Be sure to set your data types appropriately.

A user can have MuyGPySuse 32 bit types by setting the MUYGPYS_FTYPE environment variable to "32", e.g.

$ export MUYGPYS_FTYPE=32  # use 32 bit types in MuyGPyS functions

It is also possible to manipulate MuyGPyS.config to switch between types programmatically. This is not advisable unless the user knows exactly what they are doing.

Installation

Installation using Pip: CPU

The index muygpys is maintained on PyPI and can be installed using pip. muygpys supports many optional extras flags, which will install additional dependencies if specified. If installing CPU-only with pip, you might want to consider the following flags:
These extras include:

• hnswlib - install hnswlib dependency to support fast approximate nearest neighbors indexing
• jax_cpu - install JAX dependencies to support just-in-time compilation of math functions on CPU (see below to install on GPU CUDA architectures)
• torch - install PyTorch dependencies to employ GPU acceleration and the use of the MuyGPyS.torch submodule
• mpi - install MPI dependencies to support distributed memory parallel computation. Requires that the user has installed a version of MPI such as mvapich or open-mpi.

$ # numpy-only installation. Functions will internally use numpy.
$ pip install --upgrade muygpys

$ # The same, but includes hnswlib.
$ pip install --upgrade muygpys[hnswlib]

$ # CPU-only JAX installation. Functions will be jit-compiled using JAX.
$ pip install --upgrade muygpys[jax_cpu]

$ # The same, but includes hnswlib.
$ pip install --upgrade muygpys[jax_cpu,hnswlib]

$ # MPI installation. Functions will operate in distributed memory.
$ pip install --upgrade muygpys[mpi]

$ # The same, but includes hnswlib.
$ pip install --upgrade muygpys[mpi,hnswlib]

$ # pytorch installation. MuyGPyS.torch will be usable.
$ pip install --upgrade muygpys[torch]

Installation using Pip: GPU (CUDA)

JAX GPU Instructions

JAX also supports just-in-time compilation to CUDA, making the compiled math functions within MuyGPyS runnable on NVidia GPUS. This requires you to install CUDA and CuDNN in your environment, if they are not already installed, and to ensure that they are on your environment's $LD_LIBRARY_PATH. See scripts for an example environment setup.

MuyGPyS no longer supports automated GPU-supported JAX installation using pip extras. To install JAX as a dependency for MuyGPyS to be deployed on cuda-capable GPUs, please read and follow the JAX installation instructions. After installing JAX, the user will also need to install Tensorflow Probability with a JAX backend via

pip install tensorflow-probability[jax]>=0.16.0

PyTorch GPU Instructions

MuyGPyS does not and most likely will not support installing CUDA PyTorch with an extras flag. Please install PyTorch separately.

Installation From Source

This repository includes several extras_require optional dependencies.

• tests - install dependencies necessary to run tests
• docs - install dependencies necessary to build the docs
• dev - install dependencies for maintaining code style, running performance benchmarks, linting, and packaging

For example, follow these instructions to install from source for development purposes with CPU JAX support:

$ git clone [email protected]:LLNL/MuyGPyS.git
$ cd MuyGPyS
$ pip install -e .[dev,jax_cpu]

If you would like to perform a GPU installation from source, you will need to install the JAX dependency directly.

Additionally check out the develop branch to access the latest features in between stable releases. See CONTRIBUTING.md for contribution rules.

Full list of extras flags

• hnswlib - install hnswlib dependency to support fast approximate nearest neighbors indexing
• jax_cpu - install JAX dependencies to support just-in-time compilation of math functions on CPU (see below to install on GPU CUDA architectures)
• torch - install PyTorch
• mpi - install MPI dependency to support parallel computation
• tests - install dependencies necessary to run tests
• docs - install dependencies necessary to build the docs
• dev - install dependencies for maintaining code style, linting, and packaging

Building Docs

In order to build the docs locally, first pip install from source using either the docs or dev options and then execute:

$ sphinx-build -b html docs docs/_build/html

Finally, open the file docs/_build/html/index.html in your browser of choice.

Testing

In order to run tests locally, first pip install MuyGPyS from source using the tests option. All tests in the tests/ directory are then runnable as python scripts, e.g.

$ python tests/kernels.py

Individual absl unit test classes can be run in isolation, e.g.

$ python tests/kernels.py DistancesTest

It is also possible to run a single method from a test case:

$ python tests/kernels.py DistancesTest.test_l2

The user can run most tests in all backends. Some tests use backend-dependent features, and will fail with informative error messages when attempting an unsupported backend. The user needs to set MUYGPYS_BACKEND and possibly MUYGPYS_FTYPE prior to running the desired test, e.g.,

$ export MUYGPYS_BACKEND=jax
$ python tests/kernels.py

or

$ export MUYGPYS_BACKEND=torch
$ export MUYGPYS_FTYPE=32
$ python tests/backends/torch_correctness.py

If the MPI dependencies are installed, the user can also run absl tests using MPI, e.g. using mpirun

$ export MUYGPYS_BACKEND=mpi
$ mpirun -n 4 python tests/kernels.py

or using srun

$ export MUYGPYS_BACKEND=mpi
$ srun -N 1 --tasks-per-node 4 -p pdebug python tests/kernels.py

About

Authors

• Benjamin W. Priest (priest2 at llnl dot gov)
• Amanda L. Muyskens (muyskens1 at llnl dot gov)
• Imène Goumiri (goumiri1 at llnl dot gov)

Papers

MuyGPyS has been used the in the following research papers (newest first):

1. A Robust Approach to Gaussian Process Implementation
2. Enhancing Electrocardiography Data Classification Confidence: A Robust Gaussian Process Approach (MuyGPs)
3. Stellar Blend Image Classification Using Computationall Efficient Gaussian Processes
4. Closely-Spaced Object Classification Using MuyGPyS
5. Light Curve Forecasting and Anomaly Detection Using Scalable, Anisotropic, and Heteroscedastic Gaussian Process Models
6. Scalable Gaussian Process Hyperparameter Optimization via Coverage Regularization
7. Bayesian Hyperparameter Optimization in Gaussian Processes using Statistical Coverage
8. Light Curve Completion and Forecasting Using Fast and Scalable Gaussian Processes (MuyGPs)
9. Fast Gaussian Process Posterior Mean Prediction via Local Cross Validation and Precomputation
10. Gaussian Process Classification of Galaxy Blend Identification in LSST
11. MuyGPs: Scalable Gaussian Process Hyperparameter Estimation Using Local Cross-validation
12. Star-Galaxy Image Separation with Computationally Efficient Gaussian Process Classification
13. Genetic Algorithm for Hyperparameter Optimization in Gaussian Process Modeling
14. Star-Galaxy Separation via Gaussian Processes with Model Reduction

Citation

If you use MuyGPyS in a research paper, please reference our article:

@article{muygps2021,
  title={MuyGPs: Scalable Gaussian Process Hyperparameter Estimation Using Local Cross-Validation},
  author={Muyskens, Amanda and Priest, Benjamin W. and Goumiri, Im{\`e}ne and 
  Schneider, Michael},
  journal={arXiv preprint arXiv:2104.14581},
  year={2021}
}

License

MuyGPyS is distributed under the terms of the MIT license. All new contributions must be made under the MIT license.

See LICENSE-MIT, NOTICE, and COPYRIGHT for details.

SPDX-License-Identifier: MIT

Release

LLNL-CODE-824804