Hierarchical-Hyperplane Kernels for Actively Learning Gaussian Process Models of Nonstationary Systems
This code is belonging to the AISTATS 2023 paper called "Hierarchical-Hyperplane Kernels for Actively Learning Gaussian Process Models of Nonstationary Systems". It can be used as a library to utilize our method/kernel and to run our algorithm with different configurations.
After cloning the repo switch to the base folder and build the conda environment and the package via
conda env create --file environment.yml
conda activate hhk
pip install -e .
Our method is build on top of gpflow. Our proposed Hierarchical-Kyperplane-Kernel class inherits from gpflow.kernels.Kernel. It can therefore directly be used in gpflow. If one wants to use it in this repo we offer the following setup. We use the factory patern and config classes to build all respective obejcts with the right configuration. An instance of the kernel can be easily retrieved via:
kernel_config = HHKFourLocalDefaultConfig(input_dimension=2)
kernel = KernelFactory.build(kernel_config)
For all HHK configs see file hhk_configs.py.
In case a GP should be configured with our kernel, we also provide a wrapper class that can be initiated via
model_config = GPModelWithNoisePriorConfig(kernel_config=kernel_config)
model = ModelFactory.build(model_config)
For the standard GP model configs see file gp_model_config.py and for the fully-Bayesian GP model see file gp_model_marginalized_config.py.
Inference for a given dataset with np.arrays x_data and y_data and test points x_test can than be done via
model.infer(x_data,y_data)
pred_mu_test,pred_sigma_test = model.predictive_dist(x_test)
To test our kernel in the active learning setting we provide the main.py script that allows the configuration of the different run settings that are used in the paper.
Hierarchical-Hyperplane-Kernels is open-sourced under the AGPL-3.0 license. See the LICENSE file for details.