gpie 0.2.2

GPie: Gaussian Process tiny explorer

GPie

Gaussian Process tiny explorer

• simple: an intuitive syntax inspired by scikit-learn
• powerful: a compact core of expressive abstractions
• extensible: a modular design for effortless composition
• lightweight: a minimal set of dependencies {standard library, numpy, scipy}

This is a ongoing research project with many parts currently under construction - please expect bugs and sharp edges.

Features

• several "avant-garde" kernels such as spectral kernel and neural kernel allow for exploration of new ideas
• each kernel implements anisotropic variant besides isotropic one to support automatic relevance determination
• a full-fledged toolkit of kernel operators enables all sorts of "kernel engineering", for example, handcrafting composite kernels based on expert knowledge or exploiting special structure of datasets
• core computations such as likelihood and gradient are carefully formulated for speed and stability
• sampling inference embraces a probabilistic perspective in learning and prediction to promote robustness
• Bayesian optimizer offers a principled strategy to optimize expensive and black-box objectives globally

Functionality

• kernel functions
  • white kernel
  • constant kernel
  • radial basis function kernel
  • rational quadratic kernel
  • Matérn kernel
    • Ornstein-Uhlenbeck kernel
  • periodic kernel
  • spectral kernel
  • neural kernel
• kernel operators
  • Hadamard (element-wise)
    • sum
    • product
    • exponentiation
  • Kronecker
    • sum
    • product
• Gaussian process
  • regression
  • classification
• t process
  • regression
  • classification
• Bayesian optimizer
  • surrogate: Gaussian process, t process
  • acquisition: PI, EI, LCB, ES, KG
• sampling inference
  • Markov chain Monte Carlo
    • Metropolis-Hastings
    • Hamiltonian + no-U-turn
  • simulated annealing
• variational inference

Note: parts of the project in italic font are under construction.

Examples

Gaussian process regression on Mauna Loa CO₂

In this example, we use Gaussian process to model the concentration of CO₂ at Mauna Loa as a function of time.

# handcraft a composite kernel based on expert knowledge
# long-term trend
k1 = 30.0**2 * RBFKernel(l=200.0)
# seasonal variations
k2 = 3.0**2 * RBFKernel(l=200.0) * PeriodicKernel(p=1.0, l=1.0)
# medium-term irregularities
k3 = 0.5**2 * RationalQuadraticKernel(m=0.8, l=1.0)
# noise
k4 = 0.1**2 * RBFKernel(l=0.1) + 0.2**2 * WhiteKernel()
# composite kernel
kernel = k1 + k2 + k3 + k4
# train GPR on data
gpr = GaussianProcessRegressor(kernel=kernel)
gpr.fit(X, y)

In the plot, scattered dots represent historical observations, and shaded area shows the predictive interval (μ - σ, μ + σ) prophesied by a Gaussian process regressor trained on the historical data.

Sampling inference for Gaussian process regression

Here we use a synthesized dataset for ease of illustration and investigate sampling inference techniques such as Markov chain Monte Carlo. As a Gaussian process defines the predictive distribution, we can get a sense of it by sampling from its prior distribution (before seeing training set) and posterior distribution (after seeing training set).

# with the current hyperparameter configuration,
# ... what is the prior distribution p(y_test)
y_prior = gpr.prior_predictive(X, n_samples=6)
# ... what is the posterior distribution p(y_test|y_train)
y_posterior = gpr.posterior_predictive(X, n_samples=4)

We can also sample from the posterior distribution of a hyperparameter, which characterizes its uncertainty beyond a single point estimate such as MLE or MAP.

# invoke MCMC sampler to sample hyper values from its posterior distribution
hyper_posterior = gpr.hyper_posterior(n_samples=10000)

Bayesian optimization

We demonstrate a simple example of Bayesian optimization. It starts by exploring the objective function globally and shifts to exploiting "promising areas" as more observations are made.

# number of evaluations
n_evals = 10
# surrogate model (Gaussian process)
surrogate = GaussianProcessRegressor(1.0 * MaternKernel(d=5, l=1.0) +
                                     1.0 * WhiteKernel())
# bayesian optimizer
bayesopt = BayesianOptimizer(fun=f, bounds=b, x0=x0, n_evals=n_evals,
                             acquisition='lcb', surrogate=surrogate)
bayesopt.minimize(callback=callback)

Backend

GPie makes extensive use of de facto standard scientific computing packages in Python:

• numpy: linear algebra, stochastic sampling
• scipy: gradient-based optimization, stochastic sampling

Installation

GPie requires Python 3.6 or greater. The easiest way to install GPie is from a prebuilt wheel using pip:

pip install --upgrade gpie

You can also install from source to try out the latest features (requires pep517>=0.8.0 and setuptools>=40.9.0):

pip install --upgrade git+https://github.com/zackxzhang/gpie

Roadmap

• implement Hamiltonian Monte Carlo and no-U-turn
• add a demo on characteristics of different kernels
• add a demo of quantified Occam's razor
• implement Kronecker operators for scalable learning on grid data
• replace Cholesky decomposition with Krylov subspace methods for speed
• ...