Kernel layers

UnlockNN provides Keras-compatible implementations of a few kernel functions for use with the Gaussian process, as well as an extensible API for implementing your own kernels.

Built-in kernels

RBFKernelFn(*args, **kwargs)	A radial basis function implementation that works with keras.
MaternOneHalfFn(*args, **kwargs)	A Matern kernel with parameter 1/2 implementation that works with keras.

Each of these kernels can be saved using their KernelLayer.save() method and then loaded using load_kernel():

unlocknn.kernel_layers.load_kernel(directory: PathLike, kernel_type: Optional[Type[KernelLayer]] = None) → KernelLayer

Load a kernel from the disk.

Parameters

• directory – Path to the kernel save folder.

• kernel_type – A specific class to force the model to load. Useful for custom kernels which don’t appear in this module.

Examples

Create an RBFKernelFn, save it, then reload from disk:

>>> from tempfile import TemporaryDirectory
>>> rbf = RBFKernelFn()
>>> with TemporaryDirectory() as tmpdirname:
...     rbf.save(tmpdirname)
...     new_rbf = load_kernel(tmpdirname)

Extending the kernels API

All kernels should:

• Inherit from the KernelLayer abstract base class

• Implement the KernelLayer.kernel() property

• Add appropriate “weights” (kernel parameters) during KernelLayer.__init__()

• Implement KernelLayer.config() for compatibility with load_kernel()

class unlocknn.kernel_layers.KernelLayer(*args, **kwargs)

An ABC for kernel function implementations that work with keras.

__init__(trainable=True, name=None, dtype=None, dynamic=False, **kwargs)

Initialize kernel parameters.

Subclasses should use add_weight() to initialize kernel layer weights during initialization.

call(x)

Do nothing – a placeholder for keras.

abstract property config: dict

Get configuration information for the class.

This configuration information is used by load_kernel() to determine what type of kernel to load. For built-in kernels, this configuration dictionary specifies {"type": type}, where type is one of the keys in KERNEL_TYPES.

For custom kernels, this dictionary should return any keyword arguments to be passed during kernel initialization.

abstract property kernel: PositiveSemidefiniteKernel

Get a callable kernel.

save(directory: PathLike)

Save the kernel to disk.

Subclasses should update this method to save configuration information as well (c.f. config()). See also load_kernel().

Both of the current built-in kernels are parameterized by an amplitude and a length scale. The AmpAndLengthScaleFn abstract base class initializes these weights during instantiation and other kernels that are parameterized as such may inherit from this class also:

class unlocknn.kernel_layers.AmpAndLengthScaleFn(*args, **kwargs)

An ABC for kernels with amplitude and length scale parameters.

_amplitude_basis

The basis for the kernel amplitude, which is passed through a softplus to calculate the actual amplitude.

Type

tf.Tensor

_length_scale_basis

The basis for the length scale of the kernel. which is passed through a softplus to calculate the actual amplitude.

Type

tf.Tensor

__init__(**kwargs)

Initialize the layer, its amplitude and its length scale.

property amplitude: Tensor

Get the current kernel amplitude.

property length_scale: Tensor

Get the current kernel length scale.