def __init__(
self,
num_dimensions,
variance_prior=None,
offset_prior=None,
eps=1e-5,
active_dims=None,
variance_bounds=None,
offset_bounds=None,
):
super(LinearKernel, self).__init__(active_dims=active_dims)
self.eps = eps
variance_prior = _bounds_to_prior(prior=variance_prior,
bounds=variance_bounds,
log_transform=False)
self.register_parameter(name="variance",
parameter=torch.nn.Parameter(torch.zeros(1)),
prior=variance_prior)
offset_prior = _bounds_to_prior(prior=offset_prior,
bounds=offset_bounds,
log_transform=False)
self.register_parameter(name="offset",
parameter=torch.nn.Parameter(
torch.zeros(1, 1, num_dimensions)),
prior=offset_prior)
def __init__(
self,
num_dimensions,
variance_prior=None,
offset_prior=None,
active_dims=None,
variance_bounds=None,
offset_bounds=None,
):
"""
Args:
num_dimensions (int): Number of data dimensions to expect. This is necessary to create the offset parameter.
variance_prior (:obj:`gpytorch.priors.Prior`): Prior over the variance parameter (default `None`).
offset_prior (:obj:`gpytorch.priors.Prior`): Prior over the offset parameter (default `None`).
active_dims (list): List of data dimensions to operate on. `len(active_dims)` should equal `num_dimensions`.
variance_bounds (tuple, deprecated): Min and max value for the variance parameter. Deprecated, and now
creates a :obj:`gpytorch.priors.SmoothedBoxPrior`.
offset_bounds (tuple, deprecated): Min and max value for the offset parameter. Deprecated, and now creates a
:obj:'gpytorch.priors.SmoothedBoxPrior'.
"""
super(LinearKernel, self).__init__(active_dims=active_dims)
variance_prior = _bounds_to_prior(prior=variance_prior, bounds=variance_bounds, log_transform=False)
self.register_parameter(name="variance", parameter=torch.nn.Parameter(torch.zeros(1)), prior=variance_prior)
offset_prior = _bounds_to_prior(prior=offset_prior, bounds=offset_bounds, log_transform=False)
self.register_parameter(
name="offset", parameter=torch.nn.Parameter(torch.zeros(1, 1, num_dimensions)), prior=offset_prior
)
def __init__(
self,
has_lengthscale=False,
ard_num_dims=None,
batch_size=1,
active_dims=None,
log_lengthscale_bounds=None,
log_lengthscale_prior=None,
eps=1e-6,
):
super(Kernel, self).__init__()
if active_dims is not None and not torch.is_tensor(active_dims):
active_dims = torch.tensor(active_dims, dtype=torch.long)
self.active_dims = active_dims
self.ard_num_dims = ard_num_dims
self.batch_size = batch_size
self.__has_lengthscale = has_lengthscale
if has_lengthscale:
lengthscale_num_dims = 1 if ard_num_dims is None else ard_num_dims
log_lengthscale_prior = _bounds_to_prior(
prior=log_lengthscale_prior, bounds=log_lengthscale_bounds)
self.register_parameter(
name="log_lengthscale",
parameter=torch.nn.Parameter(
torch.zeros(batch_size, 1, lengthscale_num_dims)),
prior=log_lengthscale_prior,
)
def __init__(self, log_noise_prior=None, log_noise_bounds=None):
# TODO: Remove deprecated log_noise_bounds kwarg
log_noise_prior = _bounds_to_prior(prior=log_noise_prior,
bounds=log_noise_bounds)
super(GaussianLikelihood, self).__init__()
self.register_parameter(name="log_noise",
parameter=torch.nn.Parameter(torch.zeros(1)),
prior=log_noise_prior)
def __init__(self, prior=None, batch_size=None, constant_bounds=None):
# TODO: Remove deprecated bounds kwarg
prior = _bounds_to_prior(prior=prior, bounds=constant_bounds, batch_size=batch_size, log_transform=False)
super(ConstantMean, self).__init__()
self.batch_size = batch_size
self.register_parameter(
name="constant", parameter=torch.nn.Parameter(torch.zeros(batch_size or 1, 1)), prior=prior
)
def __init__(
self,
log_lengthscale_prior=None,
log_period_length_prior=None,
eps=1e-5,
active_dims=None,
log_lengthscale_bounds=None,
log_period_length_bounds=None,
):
log_period_length_prior = _bounds_to_prior(
prior=log_period_length_prior, bounds=log_period_length_bounds)
super(PeriodicKernel, self).__init__(
has_lengthscale=True,
active_dims=active_dims,
log_lengthscale_prior=log_lengthscale_prior,
log_lengthscale_bounds=log_lengthscale_bounds,
)
self.eps = eps
self.register_parameter(name="log_period_length",
parameter=torch.nn.Parameter(
torch.zeros(1, 1, 1)),
prior=log_period_length_prior)
def __init__(
self,
has_lengthscale=False,
ard_num_dims=None,
log_lengthscale_prior=None,
active_dims=None,
batch_size=1,
log_lengthscale_bounds=None,
):
"""
The base Kernel class handles both lengthscales and ARD.
Args:
has_lengthscale (bool): If True, we will register a :obj:`torch.nn.Parameter` named `log_lengthscale`
ard_num_dims (int): If not None, the `log_lengthscale` parameter will have this many entries.
log_lengthscale_prior (:obj:`gpytorch.priors.Prior`): Prior over the log lengthscale
active_dims (list): List of data dimensions to evaluate this Kernel on.
batch_size (int): If training or testing multiple GPs simultaneously, this is how many lengthscales to
register.
log_lengthscale_bounds (tuple): Deprecated min and max values for the lengthscales. If supplied, this
now registers a :obj:`gpytorch.priors.SmoothedBoxPrior`
"""
super(Kernel, self).__init__()
if active_dims is not None and not torch.is_tensor(active_dims):
active_dims = torch.tensor(active_dims, dtype=torch.long)
self.active_dims = active_dims
self.ard_num_dims = ard_num_dims
if has_lengthscale:
lengthscale_num_dims = 1 if ard_num_dims is None else ard_num_dims
log_lengthscale_prior = _bounds_to_prior(
prior=log_lengthscale_prior, bounds=log_lengthscale_bounds)
self.register_parameter(
name="log_lengthscale",
parameter=torch.nn.Parameter(
torch.zeros(batch_size, 1, lengthscale_num_dims)),
prior=log_lengthscale_prior,
)
def test_bounds_to_prior(self):
prior = GammaPrior(1, 1)
self.assertEqual(prior, _bounds_to_prior(prior=prior, bounds=None))
self.assertIsInstance(_bounds_to_prior(prior=None, bounds=(-10, 10)),
SmoothedBoxPrior)
