#!/usr/bin/env python3
import warnings
import torch
from linear_operator.operators import BlockDiagLinearOperator
from ... import settings
from ...distributions import MultitaskMultivariateNormal
from ...likelihoods import Likelihood
from ..approximate_gp import ApproximateGP
from ..gp import GP
class _DeepGPVariationalStrategy(object):
def __init__(self, model):
self.model = model
@property
def sub_variational_strategies(self):
if not hasattr(self, "_sub_variational_strategies_memo"):
self._sub_variational_strategies_memo = [
module.variational_strategy for module in self.model.modules() if isinstance(module, ApproximateGP)
]
return self._sub_variational_strategies_memo
def kl_divergence(self):
return sum(strategy.kl_divergence().sum() for strategy in self.sub_variational_strategies)
[docs]class DeepGPLayer(ApproximateGP):
"""
Represents a layer in a deep GP where inference is performed via the doubly stochastic method of
Salimbeni et al., 2017. Upon calling, instead of returning a variational distribution q(f), returns samples
from the variational distribution.
See the documentation for __call__ below for more details below. Note that the behavior of __call__
will change to be much more elegant with multiple batch dimensions; however, the interface doesn't really
change.
:param ~gpytorch.variational.VariationalStrategy variational_strategy: Strategy for
changing q(u) -> q(f) (see other VI docs)
:param int input_dims`: Dimensionality of input data expected by each GP
:param int output_dims: (default None) Number of GPs in this layer, equivalent to
output dimensionality. If set to `None`, then the output dimension will be squashed.
Forward data through this hidden GP layer. The output is a MultitaskMultivariateNormal distribution
(or MultivariateNormal distribution is output_dims=None).
If the input is >=2 dimensional Tensor (e.g. `n x d`), we pass the input through each hidden GP,
resulting in a `n x h` multitask Gaussian distribution (where all of the `h` tasks represent an
output dimension and are independent from one another). We then draw `s` samples from these Gaussians,
resulting in a `s x n x h` MultitaskMultivariateNormal distribution.
If the input is a >=3 dimensional Tensor, and the `are_samples=True` kwarg is set, then we assume that
the outermost batch dimension is a samples dimension. The output will have the same number of samples.
For example, a `s x b x n x d` input will result in a `s x b x n x h` MultitaskMultivariateNormal distribution.
The goal of these last two points is that if you have a tensor `x` that is `n x d`, then
>>> hidden_gp2(hidden_gp(x))
will just work, and return a tensor of size `s x n x h2`, where `h2` is the output dimensionality of
hidden_gp2. In this way, hidden GP layers are easily composable.
"""
def __init__(self, variational_strategy, input_dims, output_dims):
super(DeepGPLayer, self).__init__(variational_strategy)
self.input_dims = input_dims
self.output_dims = output_dims
def forward(self, x):
raise NotImplementedError
def __call__(self, inputs, are_samples=False, **kwargs):
deterministic_inputs = not are_samples
if isinstance(inputs, MultitaskMultivariateNormal):
inputs = torch.distributions.Normal(loc=inputs.mean, scale=inputs.variance.sqrt()).rsample()
deterministic_inputs = False
if settings.debug.on():
if not torch.is_tensor(inputs):
raise ValueError(
"`inputs` should either be a MultitaskMultivariateNormal or a Tensor, got "
f"{inputs.__class__.__Name__}"
)
if inputs.size(-1) != self.input_dims:
raise RuntimeError(
f"Input shape did not match self.input_dims. Got total feature dims [{inputs.size(-1)}],"
f" expected [{self.input_dims}]"
)
# Repeat the input for all possible outputs
if self.output_dims is not None:
inputs = inputs.unsqueeze(-3)
inputs = inputs.expand(*inputs.shape[:-3], self.output_dims, *inputs.shape[-2:])
# Now run samples through the GP
output = ApproximateGP.__call__(self, inputs)
if self.output_dims is not None:
mean = output.loc.transpose(-1, -2)
covar = BlockDiagLinearOperator(output.lazy_covariance_matrix, block_dim=-3)
output = MultitaskMultivariateNormal(mean, covar, interleaved=False)
# Maybe expand inputs?
if deterministic_inputs:
output = output.expand(torch.Size([settings.num_likelihood_samples.value()]) + output.batch_shape)
return output
[docs]class DeepGP(GP):
"""
A container module to build a DeepGP.
This module should contain :obj:`~gpytorch.models.deep.DeepGPLayer`
modules, and can also contain other modules as well.
"""
def __init__(self):
super().__init__()
self.variational_strategy = _DeepGPVariationalStrategy(self)
def forward(self, x):
raise NotImplementedError
class DeepLikelihood(Likelihood):
"""
A wrapper to make a GPyTorch likelihood compatible with Deep GPs
Example:
>>> deep_gaussian_likelihood = gpytorch.likelihoods.DeepLikelihood(gpytorch.likelihood.GaussianLikelihood)
"""
def __init__(self, base_likelihood):
super().__init__()
warnings.warn(
"DeepLikelihood is now deprecated. Use a standard likelihood in conjunction with a "
"gpytorch.mlls.DeepApproximateMLL. See the DeepGP example in our documentation.",
DeprecationWarning,
)
self.base_likelihood = base_likelihood
def expected_log_prob(self, observations, function_dist, *params, **kwargs):
return self.base_likelihood.expected_log_prob(observations, function_dist, *params, **kwargs).mean(dim=0)
def log_marginal(self, observations, function_dist, *params, **kwargs):
return self.base_likelihood.log_marginal(observations, function_dist, *params, **kwargs).mean(dim=0)
def forward(self, *args, **kwargs):
pass
def __call__(self, *args, **kwargs):
return self.base_likelihood.__call__(*args, **kwargs)