Fixed periodicity

[adamjstewart]

I'm working with weather data, which I know will exhibit approximately sinusoidal trends with a periodicity of exactly 24 hours and 365 days. I would like to fix the periodicity of PeriodicKernel so that I only need to learn the lengthscale. In scikit-learn, this corresponds to:

ExpSineSquared(periodicity=24, periodicity_bounds="fixed")

Is it possible to do this in GPyTorch? I tried:

PeriodicKernel(period_length_constraint=Interval(24, 24, initial_value=24))

but it seems like this isn't allowed:

  File "/Users/ajstewart/.spack/.spack-env/view/lib/python3.9/site-packages/gpytorch/constraints/constraints.py", line 30, in __init__
    raise RuntimeError("Got parameter bounds with empty intervals.")
RuntimeError: Got parameter bounds with empty intervals.

[Balandat]

You could just implement a new kernel that uses the forward logic from PeriodicKernel but gets rid of the period as a learnable parameter:

class FixedPeriodPeriodicKernel(Kernel):

    def __init__(
        self,
        period_length: Tensor,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.period_length = period_length

    def forward(self, x1, x2, diag=False, **params):
        return PeriodicKernel.forward(self, x1, x2, diag=False, **params)

[adamjstewart]

Thanks, this looks promising, I'll try this out!

[adamjstewart]

Also had to add has_lengthscale = True, with that it works like a charm!