pgmuvi.kernels

Kernel helpers for pgmuvi.

This module previously contained custom GPyTorch kernel subclasses (QuasiPeriodicKernel, SeparableKernel, AchromaticKernel). These have been replaced by compositions of standard GPyTorch kernels in pgmuvi.gps, which requires no custom forward() code and directly benefits from all GPyTorch optimisations.

The helper factory functions make_quasi_periodic_kernel, make_matern_kernel, and make_rbf_kernel are retained for convenience.

pgmuvi.kernels.make_matern_kernel(nu=1.5, lengthscale=1.0, outputscale=1.0)

Create a scaled Matérn kernel.

Parameters:

  • nu (float, optional) – Smoothness parameter (0.5, 1.5, or 2.5), by default 1.5.

  • lengthscale (float, optional) – Correlation length in the same units as the time axis. Observations separated by much less than lengthscale are strongly correlated; observations separated by much more than lengthscale are essentially uncorrelated. By default 1.0.

  • outputscale (float, optional) – Overall amplitude scaling (approximately the standard deviation of the GP output). By default 1.0.

Returns:

Scaled Matérn kernel.

Return type:

gpytorch.kernels.ScaleKernel

pgmuvi.kernels.make_quasi_periodic_kernel(period=1.0, lengthscale=0.5, decay=10.0, outputscale=1.0)

Create a scaled quasi-periodic kernel.

Returns ScaleKernel(ProductKernel(PeriodicKernel, RBFKernel)).

The quasi-periodic kernel is the elementwise product of a PeriodicKernel (oscillatory component) and an RBFKernel (long-term decay). No custom kernel subclass is required.

Parameters:

  • period (float, optional) – Initial period, in the same units as the input time axis. By default 1.0.

  • lengthscale (float, optional) – Periodic lengthscale of the PeriodicKernel (same units as period). Small values produce sharp, narrow peaks within each cycle; large values produce broad, smooth sinusoidal variations. By default 0.5.

  • decay (float, optional) – Lengthscale of the RBFKernel envelope (same units as the time axis). Controls how quickly the periodic pattern decorrelates over time: roughly the timescale over which the amplitude or phase of the oscillation can change. By default 10.0.

  • outputscale (float, optional) – Overall amplitude scaling (approximately the standard deviation of the GP output). By default 1.0.

Returns:

ScaleKernel(ProductKernel(PeriodicKernel(), RBFKernel())).

Return type:

gpytorch.kernels.ScaleKernel

pgmuvi.kernels.make_rbf_kernel(lengthscale=1.0, outputscale=1.0)

Create a scaled RBF (squared-exponential) kernel.

Parameters:

  • lengthscale (float, optional) – Correlation length in the same units as the time axis. Observations separated by much less than lengthscale are strongly correlated; observations separated by much more than lengthscale are essentially uncorrelated. By default 1.0.

  • outputscale (float, optional) – Overall amplitude scaling (approximately the standard deviation of the GP output). By default 1.0.

Returns:

Scaled RBF kernel.

Return type:

gpytorch.kernels.ScaleKernel