#!/usr/bin/env python3
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
# pyre-strict
from __future__ import annotations
from typing import Any
import torch
from ax.exceptions.core import AxError
from gpytorch.constraints import Interval
from gpytorch.kernels import PeriodicKernel
from gpytorch.kernels.matern_kernel import MaternKernel
from gpytorch.kernels.scale_kernel import ScaleKernel
from gpytorch.priors.torch_priors import Prior
[docs]
class ScaleMaternKernel(ScaleKernel):
def __init__(
self,
ard_num_dims: int | None = None,
batch_shape: torch.Size | None = None,
lengthscale_prior: Prior | None = None,
outputscale_prior: Prior | None = None,
lengthscale_constraint: Interval | None = None,
outputscale_constraint: Interval | None = None,
**kwargs: Any,
) -> None:
r"""
Args:
ard_num_dims: The number of lengthscales.
batch_shape: The batch shape.
lengthscale_prior: The prior over the lengthscale parameter.
outputscale_prior: The prior over the scaling parameter.
lengthscale_constraint: Optionally provide a lengthscale constraint.
outputscale_constraint: Optionally provide a output scale constraint.
Returns: None
"""
base_kernel = MaternKernel(
nu=2.5,
ard_num_dims=ard_num_dims,
lengthscale_constraint=lengthscale_constraint,
lengthscale_prior=lengthscale_prior,
batch_shape=batch_shape,
)
super().__init__(
base_kernel=base_kernel,
outputscale_prior=outputscale_prior,
outputscale_constraint=outputscale_constraint,
**kwargs,
)
[docs]
class TemporalKernel(ScaleKernel):
"""A product kernel of a periodic kernel and a Matern kernel.
The periodic kernel computes the similarity between temporal
features such as the time of day.
The Matern kernel computes the similarity between the tunable
parameters.
"""
def __init__(
self,
dim: int,
temporal_features: list[int],
matern_ard_num_dims: int | None = None,
batch_shape: torch.Size | None = None,
lengthscale_prior: Prior | None = None,
temporal_lengthscale_prior: Prior | None = None,
period_length_prior: Prior | None = None,
fixed_period_length: float | None = None,
outputscale_prior: Prior | None = None,
lengthscale_constraint: Interval | None = None,
outputscale_constraint: Interval | None = None,
temporal_lengthscale_constraint: Interval | None = None,
period_length_constraint: Interval | None = None,
**kwargs: Any,
) -> None:
r"""
Args:
dim: The input dimension.
temporal_features: The features to pass to the periodic kernel.
matern_ard_num_dims: The number of lengthscales. This must be
equal to the total number of parameters (excluding
temporal parameters)
batch_shape: The batch shape.
lengthscale_prior: The prior over the lengthscale parameters.
temporal_lengthscale_prior: The prior over the lengthscale parameters
for the periodic kernel.
period_length_prior: The prior over the period length.
fixed_period_length: A fixed period length for the periodic kernel.
If provided, the period length will not be tuned with the other
hyperparameters.
outputscale_prior: The prior over the scaling parameter.
lengthscale_constraint: Optionally provide a lengthscale constraint.
outputscale_constraint: Optionally provide a output scale constraint.
temporal_lengthscale_constraint: Optionally provide a
lengthscale constraint for the periodic kernel.period_length_constraint:
Optionally provide a constraint for the period length.
"""
if len(temporal_features) == 0:
raise AxError(
"The temporal kernel should only be used if there "
"are temporal features."
)
if fixed_period_length is not None and (
period_length_prior is not None or period_length_constraint is not None
):
raise ValueError(
"If `fixed_period_length` is provided, then `period_length_prior` "
"and `period_length_constraint` are not used."
)
non_temporal_dims = sorted(set(range(dim)) - set(temporal_features))
matern_kernel = MaternKernel(
nu=2.5,
ard_num_dims=matern_ard_num_dims,
lengthscale_prior=lengthscale_prior,
active_dims=non_temporal_dims,
batch_shape=batch_shape,
lengthscale_constraint=lengthscale_constraint,
)
periodic_kernel = PeriodicKernel(
ard_num_dims=len(temporal_features),
active_dims=temporal_features,
lengthscale_prior=temporal_lengthscale_prior,
period_length_prior=period_length_prior,
lengthscale_constraint=temporal_lengthscale_constraint,
period_length_constraint=period_length_constraint,
batch_shape=batch_shape,
)
if fixed_period_length is not None:
periodic_kernel.raw_period_length.requires_grad_(False)
periodic_kernel.period_length = fixed_period_length
super().__init__(
base_kernel=matern_kernel * periodic_kernel,
outputscale_prior=outputscale_prior,
outputscale_constraint=outputscale_constraint,
**kwargs,
)
