def __init__(self, layer_lower, dim_up, X=None, X_variance=None, Z=None, num_inducing=10, kernel=None, inference_method=None, noise_var=1e-2, init='rand', mpi_comm=None, mpi_root=0, back_constraint=True, encoder=None, auto_update=True, name='hiddenlayer'):
self.dim_up, self.dim_down = dim_up, layer_lower.X.shape[1]
likelihood = likelihoods.Gaussian(variance=noise_var)
self.variationalterm = NormalEntropy()
super(HiddenLayer, self).__init__(layer_lower, self.dim_down, dim_up, likelihood, init=init, X=X, X_variance=X_variance, Z=Z,
num_inducing=num_inducing, kernel=kernel, inference_method=inference_method, mpi_comm=mpi_comm, mpi_root=mpi_root, back_constraint=back_constraint, encoder=encoder, auto_update=auto_update, name=name)
def __init__(self, dim_down, dim_up, Y, X=None, X_variance=None, Z=None, num_inducing=10, kernel=None, inference_method=None, likelihood=None, init='rand',
mpi_comm=None, mpi_root=0, back_constraint=True, encoder=None, auto_update=True, repeatX=False, repeatXsplit=0, name='obslayer'):
self.dim_up, self.dim_down = dim_up, dim_down
self._Y = Y
self.repeatX = repeatX
self.repeatXsplit = repeatXsplit
if likelihood is None: likelihood = likelihoods.Gaussian()
self._toplayer_ = False
self.variationalterm = NormalEntropy()
super(ObservedLayer, self).__init__(None, self.dim_down, dim_up, likelihood, init=init, X=X, X_variance=X_variance, Z=Z,
num_inducing=num_inducing, kernel=kernel, inference_method=inference_method, mpi_comm=mpi_comm, mpi_root=mpi_root, back_constraint=back_constraint, encoder=encoder, auto_update=auto_update, name=name)
def gen_pred_layer(self,
layer_lower=None,
Y=None,
X=None,
binObserved=False):
from .pred_layers import PredLayer, BinaryPredLayer
from deepgp.encoder.mlp import MLP
from ..inference import SVI_Ratio, SVI_Ratio_Binary
from copy import deepcopy
X = self.X.copy() if X is None else X
Y = self.Y.copy() if Y is None else Y
Z = self.Z.values.copy()
X_var = self.X_var.values.copy() if self.back_constraint else None
encoder = MLP.clone(self.encoder) if self.back_constraint else None
kernel = self.kern.copy()
likelihood = self.likelihood.copy()
posterior = deepcopy(self.posterior)
variationalterm = NormalPrior() if isinstance(
self.variationalterm, NormalPrior) else NormalEntropy()
if binObserved:
layer = BinaryPredLayer(X,
Y,
kernel,
Z,
posterior,
likelihood=likelihood,
layer_lower=layer_lower,
inference_method=SVI_Ratio_Binary(),
variationalterm=variationalterm,
X_var=X_var,
encoder=encoder,
name=self.name)
else:
layer = PredLayer(X,
Y,
kernel,
Z,
posterior,
likelihood=likelihood,
layer_lower=layer_lower,
inference_method=SVI_Ratio(),
variationalterm=variationalterm,
X_var=X_var,
encoder=encoder,
name=self.name)
return layer
def __init__(self,
dim_down,
dim_up,
Y,
X=None,
X_variance=None,
Z=None,
num_inducing=10,
kernel=None,
inference_method=None,
likelihood=None,
init='rand',
mpi_comm=None,
mpi_root=0,
MLP_dims=None,
name='obslayer',
back_constraint=False):
self.layer_lower = None
self.dim_up, self.dim_down = dim_up, dim_down
self.Y = Y
self._toplayer_ = False
self.variationalterm = NormalEntropy()
if not back_constraint:
if X is None:
# Can change self to super if we want init_X to be even for non-observed MRD layers
X, fracs = self._init_X(Y, dim_up, init)
if X_variance is None:
X_variance = np.random.uniform(0, .1, X.shape)
super(ObservedMRDLayer,
self).__init__(dim_down,
dim_up,
likelihood,
init=init,
X=X,
X_variance=X_variance,
Z=Z,
MLP_dims=MLP_dims,
num_inducing=num_inducing,
kernel=kernel,
inference_method=inference_method,
mpi_comm=mpi_comm,
mpi_root=mpi_root,
name=name,
back_constraint=back_constraint)
def set_as_toplayer(self, flag=True):
if flag:
self.variationalterm = NormalPrior()
else:
self.variationalterm = NormalEntropy()
self._toplayer_ = flag