gpr

Simple Python package for Gaussian process (GP) regression problems. Depends on scipy. Just do

python setup.py install

To have python run setup.py which installs.

Basic Workflow

The basic workflow this package supports looks like this. More details are further down in this README.

Get data.
Wash hands.
Put data into instantiation of a subclass of gpr.interface.DataSet.
Construct a mean function and a covariance function.
Construct a GP taking the mean and covariance functions.
Construct a model based on the GP and the (training) data.
Construct a trainer and train the model.
Predict data.

Standardization of the training data and test data are up to you, but the DataSet class may be able to handle this for you.

Imports

This package exposes useful interfaces but also provides a number of pre-built parts to combine out of the box.

import gpr

provides the main Gaussian process class. Another important import is of interfaces which users may subclass. Most important in here is DataSet. Just do

import gpr.interface

to get the interface module. Next,

import gpr.data_sets

provides pre-built data set classes. Users should probably subclass gpr.interface.DataSet instead of using a pre-built data set class for production. This could also handle standardization. Next there is

import gpr.functions

which contains mean functions, covariance functions, and operator functions that can be used to combine them together. Then,

import gpr.models

provides pre-built model classes. Users may wish to construct their own by subclassing gpr.interface.Model. Finally,

import gpr.trainers

provides trainers for models. This maximizes the log likelihood of the model. How this is done may change in the future.

Data Set

Now let's create a fake data set using one of the pre-built data set classes. Just a basic data set. We will construct it from a sine function plus a linear function with a dash of noise.

inputs     = scipy.linspace( -1.0, 1.0 )
responses  = 1.0 + inputs + scipy.sin( 2.5 * scipy.pi * inputs ) + 
    scipy.random.normal( scale = 0.1, size = len( inputs ) )
covariance = 0.01 * scipy.eye( len( inputs ) )

data_set = gpr.data_sets.BasicDataSet( inputs[ :, None ], responses, covariance )

Note that inputs here is a 2D array, each row corresponds to one independent variable. This is a 1D input example, but you can do multi-dimensional input problems with this package.

Functions

There are mean functions, covariance functions, and operator functions for combining them. Don't combine mean functions with covariance functions using operator functions. Mean functions can be combined with mean functions, and covariance functions combined with other covariance functions, but don't mix them.

Try a polynomial mean function and squared exponential covariance plus noise.

mean_response = data_set.responses.mean()
std_response  = data_set.responses.std()
input_spacing = scipy.mean( inputs[ 1 : ] - inputs[ : -1 ] )

mean_function = gpr.functions.PolynomialMean( [ 1.0, mean_response ] )
covariance_1  = gpr.functions.IsotropicSquaredExponentialCovariance( std_response, input_spacing )
covariance_2  = gpr.functions.GaussianWhiteNoiseCovariance( 0.01 * std_response )
cov_function  = gpr.functions.Sum( covariance_1, covariance_2 )

Note the mean and covariance functions are initialized with some hyperparameter settings.

Gaussian Process

By itself a Gaussian process is kind of useless but it is possible to generate functions from it (i.e., treat it like a prior over functions).

gaussian_process = gpr.GaussianProcess( mean_function, cov_function )

A Gaussian process is just our mean function and the covariance function.

Model

The model is a Gaussian process plus the training data set.

model = gpr.models.BasicModel( gaussian_process, data_set )

All too easy.

Trainer

This gets the model in shape for the big game. It's kind of stupidly defined and I need a better interface.

trainer = gpr.trainers.BasinHoppingTrainer()
print trainer( model )

Predictions

Now we are ready to predict. We can see what different parts of our trained model actually look like. For instance, let's separate the Gaussian white noise piece out and just look at the squared exponential.

pred_gaussian_process = gpr.GaussianProcess( mean_function, covariance_1 )
pred_model            = gpr.models.BasicModel( pred_gaussian_process, data_set )

pred_responses, pred_covariance = pred_model( data_set.inputs )

pred_sigma = scipy.sqrt( pred_covariance.diagonal() )

# Plot.

for i in range( 1, 4 ) :
    lower = pred_responses - i * pred_sigma
    upper = pred_responses + i * pred_sigma
    plt.fill_between( inputs, lower, upper, alpha = 0.1 )
plt.plot( inputs, pred_responses )
plt.scatter( inputs, responses )
plt.show()

The Code

Refer to the example codes. The code presented here is based on the 1D example script.

Hyperparameters

Hyperparameters are passed to individual functions py placing them into a collections.deque object and passing that to the function's take_hyperparameters() method. Hyperparameters are read from the front of the deque. Hyperparameter settings for a function may be read by calling the hyperparameters property, which returns a collections.deque object.

If a function is an operator, the hyperparameters are distributed to the operands functions such that the first operand takes its hyperparameters and then the second operand takes its hyperparameters.

For the Gaussian process and model, the mean function's hyperparameters are read first, then those of the covariance function are read. Gaussian processes and models do not have a take_hyperparameters() method and they do not consume the hyperparameters. Rather, one just does

thing.hyperparameters = iterable_of_hyperparameters

The iterable doesn't have to be a deque. It could be a list or a scipy array.

rcthomas / gpr

gpr