def build_deep_gp(input_dim, num_data):
layers = [input_dim, 2, 2, 1]
# Below are different ways to build layers
# 1. Pass in Lists:
kernel_list = [RBF(), Matern12()]
num_inducing = [25, 25]
l1_kernel = construct_basic_kernel(kernels=kernel_list)
l1_inducing = construct_basic_inducing_variables(num_inducing=num_inducing, input_dim=layers[0])
# 2. Pass in kernels, specificy output dims (shared hyperparams/variables)
l2_kernel = construct_basic_kernel(kernels=RBF(), output_dim=layers[2], share_hyperparams=True)
l2_inducing = construct_basic_inducing_variables(
num_inducing=25, input_dim=layers[1], share_variables=True
)
# 3. Pass in kernels, specificy output dims (independent hyperparams/vars)
# By default and the constructor will make indep. copies
l3_kernel = construct_basic_kernel(kernels=RBF(), output_dim=layers[3])
l3_inducing = construct_basic_inducing_variables(
num_inducing=25, input_dim=layers[2], output_dim=layers[3]
)
# Assemble at the end
gp_layers = [
GPLayer(l1_kernel, l1_inducing, num_data),
GPLayer(l2_kernel, l2_inducing, num_data),
GPLayer(l3_kernel, l3_inducing, num_data, mean_function=Zero()),
]
return DeepGP(gp_layers, Gaussian(0.1))
def make_inducing_variables(num_latent_iv):
return [
construct_basic_inducing_variables(
num_inducing=[num_inducing for _ in range(num_latent_iv)],
input_dim=input_dim,
),
construct_basic_inducing_variables(num_inducing=num_inducing,
input_dim=input_dim,
output_dim=num_latent_iv),
]
def build_gp_layers(layer_sizes, num_data):
gp_layers = []
for input_dim, output_dim in zip(layer_sizes[:-1], layer_sizes[1:]):
kernel = construct_basic_kernel(kernels=RBF(), output_dim=output_dim)
inducing_vars = construct_basic_inducing_variables(
num_inducing=25, input_dim=input_dim, output_dim=output_dim)
layer = GPLayer(kernel, inducing_vars, num_data)
gp_layers.append(layer)
gp_layers[-1].mean_function = Zero()
return gp_layers
def setup_gp_layer_and_data(num_inducing: int, **gp_layer_kwargs):
input_dim = 30
output_dim = 5
num_data = 100
data = make_data(input_dim, output_dim, num_data=num_data)
kernel = construct_basic_kernel(RBF(), output_dim)
inducing_vars = construct_basic_inducing_variables(num_inducing, input_dim,
output_dim)
mean_function = Zero(output_dim)
gp_layer = GPLayer(kernel,
inducing_vars,
num_data,
mean_function=mean_function,
**gp_layer_kwargs)
return gp_layer, data
def test_construct_inducing_separate_independent_custom_list(z_init):
num_inducing = [25, 35, 45]
input_dim = 5
if z_init:
z_init = [
xavier_initialization_numpy(m, input_dim) for m in num_inducing
]
else:
z_init = None
moiv = construct_basic_inducing_variables(num_inducing,
input_dim,
z_init=z_init)
assert isinstance(moiv, SeparateIndependentInducingVariables)
assert isinstance(moiv, MultioutputInducingVariables)
for i, iv in enumerate(moiv.inducing_variable_list):
assert len(iv) == num_inducing[i]
def test_construct_inducing_shared_independent_duplicates(z_init):
num_inducing = 25
input_dim = 5
output_dim = 7
if z_init:
z_init = np.random.randn(num_inducing, input_dim)
else:
z_init = None
moiv = construct_basic_inducing_variables(num_inducing,
input_dim,
output_dim=output_dim,
share_variables=True,
z_init=z_init)
assert isinstance(moiv, SharedIndependentInducingVariables)
assert isinstance(moiv, MultioutputInducingVariables)
assert len(moiv.inducing_variable) == num_inducing
def test_construct_inducing_separate_independent_duplicates(z_init):
num_inducing = 25
input_dim = 5
output_dim = 7
if z_init:
z_init = np.random.randn(output_dim, num_inducing, input_dim)
else:
z_init = None
moiv = construct_basic_inducing_variables(num_inducing,
input_dim,
output_dim=output_dim,
z_init=z_init)
assert isinstance(moiv, SeparateIndependentInducingVariables)
assert isinstance(moiv, MultioutputInducingVariables)
for iv in moiv.inducing_variable_list:
assert len(iv) == num_inducing
def test_verify_compatibility_type_errors():
valid_inducing_variable = construct_basic_inducing_variables([35],
input_dim=40)
valid_kernel = construct_basic_kernel([Matern52()])
valid_mean_function = Zero(
) # all gpflow mean functions are currently valid
with pytest.raises(GPLayerIncompatibilityException
): # gpflow kernels must be MultioutputKernels
verify_compatibility(Matern52(), valid_mean_function,
valid_inducing_variable)
Z = valid_inducing_variable.inducing_variable_list[0].Z
inducing_variable = InducingPoints(Z)
with pytest.raises(
GPLayerIncompatibilityException
): # gpflow inducing_variables must be MultioutputInducingVariables
verify_compatibility(valid_kernel, valid_mean_function,
inducing_variable)
def __init__(
self,
likelihood: gpflow.likelihoods.Likelihood = gpflow.likelihoods.
Gaussian(0.01)
) -> None:
kernel = construct_basic_kernel(gpflow.kernels.SquaredExponential(),
output_dim=1,
share_hyperparams=True)
inducing_var = construct_basic_inducing_variables(
num_inducing=5,
input_dim=1,
share_variables=True,
z_init=tf.random.normal([5, 1], dtype=gpflow.default_float()),
)
gp_layer = GPLayer(kernel, inducing_var, 10)
super().__init__(
[gp_layer], # not actually used
likelihood,
)
# ## Constructing the layers
#
# Note that we give the `full_cov=True` argument to `GPLayer` so that we obtain correlated samples.
# We give the last layer a `gpflow.mean_functions.Zero` mean function (the GPflux default is an Identity mean function).
# %%
num_samples = 5
# %%
Z = X.copy()
M = Z.shape[0]
# Layer 1
inducing_var1 = construct_basic_inducing_variables(M,
D,
D,
share_variables=True,
z_init=Z.copy())
kernel1 = construct_basic_kernel(
gpflow.kernels.SquaredExponential(lengthscales=0.15),
output_dim=D,
share_hyperparams=True,
)
layer1 = GPLayer(kernel1,
inducing_var1,
num_data,
full_cov=True,
num_samples=num_samples)
# Layer 2
inducing_var2 = construct_basic_inducing_variables(M,
def build_constant_input_dim_deep_gp(X: np.ndarray, num_layers: int,
config: Config) -> DeepGP:
r"""
Build a Deep GP consisting of ``num_layers`` :class:`GPLayer`\ s.
All the hidden layers have the same input dimension as the data, that is, ``X.shape[1]``.
The architecture is largely based on :cite:t:`salimbeni2017doubly`, with
the most notable difference being that we keep the hidden dimension equal
to the input dimensionality of the data.
.. note::
This architecture might be slow for high-dimensional data.
.. note::
This architecture assumes a :class:`~gpflow.likelihoods.Gaussian` likelihood
for regression tasks. Specify a different likelihood for performing
other tasks such as classification.
:param X: The training input data, used to retrieve the number of datapoints and
the input dimension and to initialise the inducing point locations using k-means. A
tensor of rank two with the dimensions ``[num_data, input_dim]``.
:param num_layers: The number of layers in the Deep GP.
:param config: The configuration for (hyper)parameters. See :class:`Config` for details.
"""
num_data, input_dim = X.shape
X_running = X
gp_layers = []
centroids, _ = kmeans2(X, k=config.num_inducing, minit="points")
for i_layer in range(num_layers):
is_last_layer = i_layer == num_layers - 1
D_in = input_dim
D_out = 1 if is_last_layer else input_dim
# Pass in kernels, specify output dim (shared hyperparams/variables)
inducing_var = construct_basic_inducing_variables(
num_inducing=config.num_inducing,
input_dim=D_in,
share_variables=True,
z_init=centroids)
kernel = construct_basic_kernel(
kernels=_construct_kernel(D_in, is_last_layer),
output_dim=D_out,
share_hyperparams=True,
)
assert config.whiten is True, "non-whitened case not implemented yet"
if is_last_layer:
mean_function = gpflow.mean_functions.Zero()
q_sqrt_scaling = 1.0
else:
mean_function = construct_mean_function(X_running, D_in, D_out)
X_running = mean_function(X_running)
if tf.is_tensor(X_running):
X_running = X_running.numpy()
q_sqrt_scaling = config.inner_layer_qsqrt_factor
layer = GPLayer(
kernel,
inducing_var,
num_data,
mean_function=mean_function,
name=f"gp_{i_layer}",
)
layer.q_sqrt.assign(layer.q_sqrt * q_sqrt_scaling)
gp_layers.append(layer)
likelihood = Gaussian(config.likelihood_noise_variance)
return DeepGP(gp_layers, LikelihoodLayer(likelihood))
