Skip to content
/ gp Public

Differentiable Gaussian Process implementation for PyTorch

License

Notifications You must be signed in to change notification settings

anassinator/gp

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

29 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Gaussian Process

https://travis-ci.org/anassinator/gp.svg?branch=master

This is a differentiable Gaussian Process implementation for PyTorch.

The code is based off of the Gaussian Processes for Machine Learning book and supports both Python 2 and 3.

Install

To install simply clone and run:

python setup.py install

You may also install the dependencies with pipenv as follows:

pipenv install

Finally, you may add this to your own application with either:

pip install 'git+https://github.com/anassinator/gp.git#egg=gp'
pipenv install 'git+https://github.com/anassinator/gp.git#egg=gp'

Usage

After installation, import and use as follows:

from gp import GaussianProcess
from gp.kernels import RBFKernel, WhiteNoiseKernel

k = RBFKernel() + WhiteNoiseKernel()
gp = GaussianProcess(k)
gp.set_data(X, Y)
gp.fit()

where X and Y are your training data's inputs and outputs as torch.Tensor.

You can then use the Gaussian Process's estimates as tensors as follows:

mean = gp(x)
mean, std = gp(x, return_std=True)
mean, covar = gp(x, return_covar=True)
mean, covar, var = gp(x, return_covar=True, return_var=True)
mean, covar, std = gp(x, return_covar=True, return_std=True)

The following is an example of what this Gaussian Process was able to estimate with a few randomly sampled points (in blue) of a noisy sin function. The dotted lines represent the real function that was kept a secret from the Gaussian Process, whereas the red line and the grey area represent the estimated mean and uncertainty.

Gaussian Process estimate of sin(x)

You can see the examples directory for some Jupyter notebooks with more detailed examples. You can also play with the secret functions that the Gaussian Process is attempting to learn and see how well it performs. Depending on the complexity and nature of the function, you might need to sample more data.

Finally, you can also use a custom kernel function instead of the included Radial-Basis Function (RBF) kernel by implementing your own Kernel class as in kernels.py.

Contributing

Contributions are welcome. Simply open an issue or pull request on the matter.

Linting

We use YAPF for all Python formatting needs. You can auto-format your changes with the following command:

yapf --recursive --in-place --parallel .

You can install the formatter with:

pipenv install --dev

License

See LICENSE.