def fit(self,
train_inputs,
train_outputs,
optimization_config = None,
optimize_stochastic=True,
param_threshold=1e-3,
objective_threshold=5,
max_iterations=200):
"""
train_inputs : ndarray
A matrix containing the input training data. Dimensions: num_data_points * input_dim.
train_outputs : ndarray
A matrix containing the output training data. Dimensions: num_data_points * num_latent.
optimization_config : dict
The maximum number of function evaluations per subset of parameters for each local
optimization. Valid keys are: 'mog' (variational parameters), 'hyp' (kernel
hyperparameters), 'll' (likelihood parameters), 'inducing' (inducing point locations).
If a set of parameters is not included it will not get optimized.
optimize_stochastic : bool
Whether to optimize the model stochastically. If set to false batch optimization will
be used.
param_threshold : float
The tolerance threshold for convergence of the variational parameters. The model
is considered to have converged if (means_difference + covariances_difference) / 2 <
param_threshold.
objective_threshold : float
The tolerance threshold for convergence of the objective function.
max_iterations : int
The maximum number of global optimisations. If max_iterations is None the model will
keep getting optimized until convergence.
"""
# Initialize the model. We should move this to init once we fix the API design issue.
if self.posterior == 'full':
self.model = FullGaussianProcess(train_inputs,
train_outputs,
self.num_inducing,
self.num_samples,
self.kernels,
self.likelihood,
self.latent_noise,
False,
self.random_inducing,
num_threads=1,
partition_size=self.partition_size)
elif self.posterior == 'mix1' or self.posterior == 'mix2':
num_components = 1 if method == 'mix1' else 2
self.model = DiagonalGaussianProcess(train_inputs,
train_outputs,
self.num_inducing,
num_components,
self.num_samples,
self.kernels,
self.likelihood,
self.latent_noise,
False,
self.random_inducing,
num_threads=1,
partition_size=partition_size)
else:
assert False
# Setup default optimization values.
if optimization_config is None:
if optimize_stochastic:
optimization_config = {'mog': 15, 'hyp': 5, 'll': 5, 'inducing': 2}
else:
optimization_config = {'mog': 50, 'hyp': 15, 'll': 15, 'inducing': 5}
optimize_func = (optimizer.stochastic_optimize_model if optimize_stochastic else
optimizer.batch_optimize_model)
# Optimize the model.
_, total_evals = optimize_func(self.model,
optimization_config,
max_iterations,
param_threshold,
objective_threshold)
return total_evals
def fit(self,
train_inputs,
train_outputs,
optimization_config=None,
optimize_stochastic=True,
param_threshold=1e-3,
objective_threshold=5,
max_iterations=200):
"""
train_inputs : ndarray
A matrix containing the input training data. Dimensions: num_data_points * input_dim.
train_outputs : ndarray
A matrix containing the output training data. Dimensions: num_data_points * num_latent.
optimization_config : dict
The maximum number of function evaluations per subset of parameters for each local
optimization. Valid keys are: 'mog' (variational parameters), 'hyp' (kernel
hyperparameters), 'll' (likelihood parameters), 'inducing' (inducing point locations).
If a set of parameters is not included it will not get optimized.
optimize_stochastic : bool
Whether to optimize the model stochastically. If set to false batch optimization will
be used.
param_threshold : float
The tolerance threshold for convergence of the variational parameters. The model
is considered to have converged if (means_difference + covariances_difference) / 2 <
param_threshold.
objective_threshold : float
The tolerance threshold for convergence of the objective function.
max_iterations : int
The maximum number of global optimisations. If max_iterations is None the model will
keep getting optimized until convergence.
"""
# Initialize the model. We should move this to init once we fix the API design issue.
if self.posterior == 'full':
self.model = FullGaussianProcess(
train_inputs,
train_outputs,
self.num_inducing,
self.num_samples,
self.kernels,
self.likelihood,
self.latent_noise,
False,
self.random_inducing,
num_threads=1,
partition_size=self.partition_size)
elif self.posterior == 'mix1' or self.posterior == 'mix2':
num_components = 1 if method == 'mix1' else 2
self.model = DiagonalGaussianProcess(train_inputs,
train_outputs,
self.num_inducing,
num_components,
self.num_samples,
self.kernels,
self.likelihood,
self.latent_noise,
False,
self.random_inducing,
num_threads=1,
partition_size=partition_size)
else:
assert False
# Setup default optimization values.
if optimization_config is None:
if optimize_stochastic:
optimization_config = {
'mog': 15,
'hyp': 5,
'll': 5,
'inducing': 2
}
else:
optimization_config = {
'mog': 50,
'hyp': 15,
'll': 15,
'inducing': 5
}
optimize_func = (optimizer.stochastic_optimize_model
if optimize_stochastic else
optimizer.batch_optimize_model)
# Optimize the model.
_, total_evals = optimize_func(self.model, optimization_config,
max_iterations, param_threshold,
objective_threshold)
return total_evals
class Savigp(object):
"""
A scikit-like class which allows for predictions using scalable automated variational
inference on Gaussian process models.
Parameters
----------
likelihood : subclass of Likelihood
An object representing the likelihood function. See likelihood.py for the methods
such an object needs to implement.
kernels : list of kernels
A list containing a kernel for each latent process. See kernel.py for the methods each
element in the list needs to implement.
posterior : str
A string denoting the shape of the posterior distribution. Possible values include:
'full' (a single Gaussian with a full covariance matrix), 'mix1' (a single Gaussian
with a diagonal covariance matrix), and 'mix2' (a mixture of two Gaussians with
diagonal covariance matrices).
num_inducing : int
The number of inducing points to use in training.
random_inducing: boolean
Whether to put inducing points randomly on training data. If False, inducing points will be
determined using clustering.
num_samples : int
The number of samples used to approximate gradients and objective functions.
latent_noise : float
The amount of latent noise that will be added to each kernel.
partition_size : int
How large each partition of the training data will be when calculating expected log
likelihood. In the case of batch optimization this is simply used to save on memory;
smaller partitions lead to less memory usage but slower runtime. In the case of
stochastic optimization, this will also denote the size of the mini-batches used in
training.
denbug_output : bool
Whether to print the value of the objective function as it gets optimized.
"""
def __init__(self,
likelihood,
kernels,
posterior='full',
num_inducing=100,
random_inducing=False,
num_samples=2000,
latent_noise=0.001,
partition_size=100,
debug_output=False):
# Init the logger and disable output if necessary.
model_logging.init_logger('savigp', False)
if not debug_output:
model_logging.disable_logger()
# Save the model configuration options.
self.likelihood = likelihood
self.kernels = kernels
self.posterior = posterior
self.num_inducing = num_inducing
self.random_inducing = random_inducing
self.num_samples = num_samples
self.latent_noise = latent_noise
self.partition_size = partition_size
# We delay initializing the model until the user fits the data since
# it unfortunately needs access to training data.
self.model = None
def fit(self,
train_inputs,
train_outputs,
optimization_config = None,
optimize_stochastic=True,
param_threshold=1e-3,
objective_threshold=5,
max_iterations=200):
"""
train_inputs : ndarray
A matrix containing the input training data. Dimensions: num_data_points * input_dim.
train_outputs : ndarray
A matrix containing the output training data. Dimensions: num_data_points * num_latent.
optimization_config : dict
The maximum number of function evaluations per subset of parameters for each local
optimization. Valid keys are: 'mog' (variational parameters), 'hyp' (kernel
hyperparameters), 'll' (likelihood parameters), 'inducing' (inducing point locations).
If a set of parameters is not included it will not get optimized.
optimize_stochastic : bool
Whether to optimize the model stochastically. If set to false batch optimization will
be used.
param_threshold : float
The tolerance threshold for convergence of the variational parameters. The model
is considered to have converged if (means_difference + covariances_difference) / 2 <
param_threshold.
objective_threshold : float
The tolerance threshold for convergence of the objective function.
max_iterations : int
The maximum number of global optimisations. If max_iterations is None the model will
keep getting optimized until convergence.
"""
# Initialize the model. We should move this to init once we fix the API design issue.
if self.posterior == 'full':
self.model = FullGaussianProcess(train_inputs,
train_outputs,
self.num_inducing,
self.num_samples,
self.kernels,
self.likelihood,
self.latent_noise,
False,
self.random_inducing,
num_threads=1,
partition_size=self.partition_size)
elif self.posterior == 'mix1' or self.posterior == 'mix2':
num_components = 1 if method == 'mix1' else 2
self.model = DiagonalGaussianProcess(train_inputs,
train_outputs,
self.num_inducing,
num_components,
self.num_samples,
self.kernels,
self.likelihood,
self.latent_noise,
False,
self.random_inducing,
num_threads=1,
partition_size=partition_size)
else:
assert False
# Setup default optimization values.
if optimization_config is None:
if optimize_stochastic:
optimization_config = {'mog': 15, 'hyp': 5, 'll': 5, 'inducing': 2}
else:
optimization_config = {'mog': 50, 'hyp': 15, 'll': 15, 'inducing': 5}
optimize_func = (optimizer.stochastic_optimize_model if optimize_stochastic else
optimizer.batch_optimize_model)
# Optimize the model.
_, total_evals = optimize_func(self.model,
optimization_config,
max_iterations,
param_threshold,
objective_threshold)
return total_evals
def predict(self, test_inputs, test_outputs=None):
"""
Make predictions on test inputs and computes the negative log predictive density for the
test outputs if they are given.
Parameters
----------
test_inputs : ndarray
Dimensions: num_test * input_dim.
test_outputs : ndarray
Dimensions: num_test * output_dim.
Returns
-------
predicted_values : ndarray
Predicted values for each test point. Dimensions : num_test * output_dim.
predicted_variance : ndarray
Predicted variance of the values. Dimensions : num_test * output_dim
nlpd : ndarray
The negative log predictive density for each test point or None if test outputs aren't
provided. Dimensions: num_test * output_dim.
"""
return self.model.predict(test_inputs, test_outputs)
class Savigp(object):
"""
A scikit-like class which allows for predictions using scalable automated variational
inference on Gaussian process models.
Parameters
----------
likelihood : subclass of Likelihood
An object representing the likelihood function. See likelihood.py for the methods
such an object needs to implement.
kernels : list of kernels
A list containing a kernel for each latent process. See kernel.py for the methods each
element in the list needs to implement.
posterior : str
A string denoting the shape of the posterior distribution. Possible values include:
'full' (a single Gaussian with a full covariance matrix), 'mix1' (a single Gaussian
with a diagonal covariance matrix), and 'mix2' (a mixture of two Gaussians with
diagonal covariance matrices).
num_inducing : int
The number of inducing points to use in training.
random_inducing: boolean
Whether to put inducing points randomly on training data. If False, inducing points will be
determined using clustering.
num_samples : int
The number of samples used to approximate gradients and objective functions.
latent_noise : float
The amount of latent noise that will be added to each kernel.
partition_size : int
How large each partition of the training data will be when calculating expected log
likelihood. In the case of batch optimization this is simply used to save on memory;
smaller partitions lead to less memory usage but slower runtime. In the case of
stochastic optimization, this will also denote the size of the mini-batches used in
training.
denbug_output : bool
Whether to print the value of the objective function as it gets optimized.
"""
def __init__(self,
likelihood,
kernels,
posterior='full',
num_inducing=100,
random_inducing=False,
num_samples=2000,
latent_noise=0.001,
partition_size=100,
debug_output=False):
# Init the logger and disable output if necessary.
model_logging.init_logger('savigp', False)
if not debug_output:
model_logging.disable_logger()
# Save the model configuration options.
self.likelihood = likelihood
self.kernels = kernels
self.posterior = posterior
self.num_inducing = num_inducing
self.random_inducing = random_inducing
self.num_samples = num_samples
self.latent_noise = latent_noise
self.partition_size = partition_size
# We delay initializing the model until the user fits the data since
# it unfortunately needs access to training data.
self.model = None
def fit(self,
train_inputs,
train_outputs,
optimization_config=None,
optimize_stochastic=True,
param_threshold=1e-3,
objective_threshold=5,
max_iterations=200):
"""
train_inputs : ndarray
A matrix containing the input training data. Dimensions: num_data_points * input_dim.
train_outputs : ndarray
A matrix containing the output training data. Dimensions: num_data_points * num_latent.
optimization_config : dict
The maximum number of function evaluations per subset of parameters for each local
optimization. Valid keys are: 'mog' (variational parameters), 'hyp' (kernel
hyperparameters), 'll' (likelihood parameters), 'inducing' (inducing point locations).
If a set of parameters is not included it will not get optimized.
optimize_stochastic : bool
Whether to optimize the model stochastically. If set to false batch optimization will
be used.
param_threshold : float
The tolerance threshold for convergence of the variational parameters. The model
is considered to have converged if (means_difference + covariances_difference) / 2 <
param_threshold.
objective_threshold : float
The tolerance threshold for convergence of the objective function.
max_iterations : int
The maximum number of global optimisations. If max_iterations is None the model will
keep getting optimized until convergence.
"""
# Initialize the model. We should move this to init once we fix the API design issue.
if self.posterior == 'full':
self.model = FullGaussianProcess(
train_inputs,
train_outputs,
self.num_inducing,
self.num_samples,
self.kernels,
self.likelihood,
self.latent_noise,
False,
self.random_inducing,
num_threads=1,
partition_size=self.partition_size)
elif self.posterior == 'mix1' or self.posterior == 'mix2':
num_components = 1 if method == 'mix1' else 2
self.model = DiagonalGaussianProcess(train_inputs,
train_outputs,
self.num_inducing,
num_components,
self.num_samples,
self.kernels,
self.likelihood,
self.latent_noise,
False,
self.random_inducing,
num_threads=1,
partition_size=partition_size)
else:
assert False
# Setup default optimization values.
if optimization_config is None:
if optimize_stochastic:
optimization_config = {
'mog': 15,
'hyp': 5,
'll': 5,
'inducing': 2
}
else:
optimization_config = {
'mog': 50,
'hyp': 15,
'll': 15,
'inducing': 5
}
optimize_func = (optimizer.stochastic_optimize_model
if optimize_stochastic else
optimizer.batch_optimize_model)
# Optimize the model.
_, total_evals = optimize_func(self.model, optimization_config,
max_iterations, param_threshold,
objective_threshold)
return total_evals
def predict(self, test_inputs, test_outputs=None):
"""
Make predictions on test inputs and computes the negative log predictive density for the
test outputs if they are given.
Parameters
----------
test_inputs : ndarray
Dimensions: num_test * input_dim.
test_outputs : ndarray
Dimensions: num_test * output_dim.
Returns
-------
predicted_values : ndarray
Predicted values for each test point. Dimensions : num_test * output_dim.
predicted_variance : ndarray
Predicted variance of the values. Dimensions : num_test * output_dim
nlpd : ndarray
The negative log predictive density for each test point or None if test outputs aren't
provided. Dimensions: num_test * output_dim.
"""
return self.model.predict(test_inputs, test_outputs)
