Implementing KernelsΒΆ

Implementing an own kernel is easy. Just create a class which is derived from Kernel. Then implement the Kernel.params attribute and the Kernel.full() method. Here is some example code to get you started:

from goppy import Kernel

class MyKernel(Kernel):
    def __init__(self, param1, param2):
        self.param1 = param1
        self.param2 = param2

    def params(self):
        return np.array([self.param1, self.param2])

    def params(self, values):
        self.param1 = values[0]
        self.param2 = values[1]

    def full(x, y, what=('y',)):
        # Evaluate your kernel

By implementing Kernel.params as a property it is possible to access the parameters as an array (which is needed for evaluating log likelihood derivatives of Gaussian processes), but also by more expressive names like k.param1.

The Kernel.full() full method should by default evaluate the kernel normally and return the result in a dictionary with the key 'y'. This is sufficient for the basic functionality when used in conjunction with OnlineGP. For more advanced usage involving predicting derivatives Kernel.full() has also be able to return the derivatives of the kernel. See the documentation of Kernel.full() for more information.

Sometimes only the diagonal of the Gram matrix is needed. The diagonal can usually be calculated more efficiently than evaluating the full Gram matrix and just using the diagonal. Thus, it might be a good idea to add code for this special case by implementing Kernel.diag().

Look at the source of the kernel module to see some complete implementations of kernels.