def gen_noise(dims, points, random_state=None):
"""
Generates an ndarray representing random noise.
This array has dims dimensions and points points per dimension. Each
element is between 0 and 1.
Parameters
----------
dims : int
The dimensionality of the noise.
points : int
The number of points per dimension.
random_state : numpy RandomState
The random state to generate the noise.
Returns
-------
noise_gen : ndarray
The ndarray containing the noise.
"""
random_state = check_random_state(random_state)
dimension_tuple = (points, ) * dims
noise_gen = random_state.rand(*dimension_tuple)
return noise_gen
def gen_noise(dims, points, random_state=None):
"""
Generates an ndarray representing random noise.
This array has dims dimensions and points points per dimension. Each
element is between 0 and 1.
Parameters
----------
dims : int
The dimensionality of the noise.
points : int
The number of points per dimension.
random_state : numpy RandomState
The random state to generate the noise.
Returns
-------
noise_gen : ndarray
The ndarray containing the noise.
"""
random_state = check_random_state(random_state)
dimension_tuple = (points,)*dims
noise_gen = random_state.rand(*dimension_tuple)
return noise_gen
def __init__(self, optimizer_arguments=None):
"""
Initializes a bayesian optimizer.
Parameters
----------
optimizer_arguments: dict of string keys
Sets the possible arguments for this optimizer. Available are:
"initial_random_runs" : int, optional
The number of initial random runs before using the GP. Default
is 10.
"random_state" : scipy random state, optional
The scipy random state or object to initialize one. Default is
None.
"acquisition_hyperparameters" : dict, optional
dictionary of acquisition-function hyperparameters
"num_gp_restarts" : int
GPy's optimization requires restarts to find a good solution.
This parameter controls this. Default is 10.
"acquisition" : AcquisitionFunction
The acquisition function to use. Default is
ExpectedImprovement.
"num_precomputed" : int
The number of points that should be kept precomputed for faster
multiple workers.
"""
self.logger = get_logger(self)
if optimizer_arguments is None:
optimizer_arguments = {}
self.initial_random_runs = optimizer_arguments.get(
'initial_random_runs', self.initial_random_runs)
self.random_state = check_random_state(
optimizer_arguments.get('random_state', None))
self.acquisition_hyperparams = optimizer_arguments.get(
'acquisition_hyperparams', None)
self.num_gp_restarts = optimizer_arguments.get('num_gp_restarts',
self.num_gp_restarts)
if not isinstance(optimizer_arguments.get('acquisition'),
AcquisitionFunction):
self.acquisition_function = optimizer_arguments.get(
'acquisition',
ExpectedImprovement)(self.acquisition_hyperparams)
else:
self.acquisition_function = optimizer_arguments.get("acquisition")
self.kernel_params = optimizer_arguments.get("kernel_params", {})
self.kernel = optimizer_arguments.get("kernel", "matern52")
self.random_searcher = RandomSearch(
{"random_state": self.random_state})
if mcmc_imported:
self.mcmc = optimizer_arguments.get("mcmc", False)
else:
self.mcmc = False
self.num_precomputed = optimizer_arguments.get('num_precomputed', 10)
self.logger.info("Bayesian optimization initialized.")
def __init__(self, optimizer_arguments=None):
"""
Initializes a bayesian optimizer.
Parameters
----------
optimizer_arguments: dict of string keys
Sets the possible arguments for this optimizer. Available are:
"initial_random_runs" : int, optional
The number of initial random runs before using the GP. Default
is 10.
"random_state" : scipy random state, optional
The scipy random state or object to initialize one. Default is
None.
"acquisition_hyperparameters" : dict, optional
dictionary of acquisition-function hyperparameters
"num_gp_restarts" : int
GPy's optimization requires restarts to find a good solution.
This parameter controls this. Default is 10.
"acquisition" : AcquisitionFunction
The acquisition function to use. Default is
ExpectedImprovement.
"num_precomputed" : int
The number of points that should be kept precomputed for faster
multiple workers.
"""
self.logger = get_logger(self)
if optimizer_arguments is None:
optimizer_arguments = {}
self.initial_random_runs = optimizer_arguments.get(
'initial_random_runs', self.initial_random_runs)
self.random_state = check_random_state(
optimizer_arguments.get('random_state', None))
self.acquisition_hyperparams = optimizer_arguments.get(
'acquisition_hyperparams', None)
self.num_gp_restarts = optimizer_arguments.get(
'num_gp_restarts', self.num_gp_restarts)
if not isinstance(optimizer_arguments.get('acquisition'), AcquisitionFunction):
self.acquisition_function = optimizer_arguments.get(
'acquisition', ExpectedImprovement)(self.acquisition_hyperparams)
else:
self.acquisition_function = optimizer_arguments.get("acquisition")
self.kernel_params = optimizer_arguments.get("kernel_params", {})
self.kernel = optimizer_arguments.get("kernel", "matern52")
self.random_searcher = RandomSearch({"random_state": self.random_state})
if mcmc_imported:
self.mcmc = optimizer_arguments.get("mcmc", False)
else:
self.mcmc = False
self.num_precomputed = optimizer_arguments.get('num_precomputed', 10)
self.logger.info("Bayesian optimization initialized.")
def demo_branin(steps=50, random_steps=10, cv=5, disable_auto_plot=False):
logging.basicConfig(level=logging.DEBUG)
#produce the same random state
random_state_rs = check_random_state(42)
param_defs = {
"x": MinMaxNumericParamDef(-5, 10),
"y": MinMaxNumericParamDef(0, 15)
}
LAss = ValidationLabAssistant(cv=cv, disable_auto_plot=disable_auto_plot)
LAss.init_experiment("RandomSearch",
"RandomSearch",
param_defs,
minimization=True,
optimizer_arguments={"random_state": random_state_rs})
optimizers = ["RandomSearch", "BayOpt_EI"]
optimizer_arguments = [{
"random_state": random_state_rs
}, {
"initial_random_runs": random_steps
}]
#evaluate random search for 10 steps use these steps as init value for bayesian
for i in range(random_steps * cv):
evaluated_candidate = single_branin_evaluation_step(
LAss, 'RandomSearch')
#now clone experiment for each optimizer
for j in range(1, len(optimizers)):
LAss.clone_experiments_by_name(
exp_name=optimizers[0],
new_exp_name=optimizers[j],
optimizer="BayOpt",
optimizer_arguments=optimizer_arguments[j])
logger.info("Random Initialization Phase Finished.")
logger.info("Competitive Evaluation Phase starts now.")
#from there on go step by step all models
for i in range(random_steps * cv, steps * cv):
for optimizer in optimizers:
single_branin_evaluation_step(LAss, optimizer)
#plot results comparatively
# a very detailed plot containing all points
LAss.plot_result_per_step(optimizers)
#a plot only showing the evaluation of the best result
LAss.plot_validation(optimizers)
def _gen_one_candidate(self):
"""
Generates a single candidate.
This is done by generating parameter values for each of the
available parameters.
Returns
-------
candidate : Candidate
The generated candidate
"""
self.random_state = check_random_state(self.random_state)
value_dict = {}
for key, param_def in self._experiment.parameter_definitions.iteritems():
value_dict[key] = self._gen_param_val(param_def)
return Candidate(value_dict)
def _gen_one_candidate(self):
"""
Generates a single candidate.
This is done by generating parameter values for each of the
available parameters.
Returns
-------
candidate : Candidate
The generated candidate
"""
self._logger.debug("Generating single candidate.")
self.random_state = check_random_state(self.random_state)
value_dict = {}
for key, param_def in self._experiment.parameter_definitions.iteritems(
):
value_dict[key] = self._gen_param_val(param_def)
generated_candidate = Candidate(value_dict)
self._logger.debug("Generated candidate: %s", generated_candidate)
return generated_candidate
def demo_MNIST_MCMC(steps,
random_steps,
percentage,
cv,
plot_at_end=True,
disable_auto_plot=False):
logging.basicConfig(level=logging.DEBUG)
regressor = SVC(kernel="poly")
param_defs = {
"C": MinMaxNumericParamDef(0, 10),
#"degree": FixedValueParamDef([1, 2, 3]),
"gamma": MinMaxNumericParamDef(0, 1),
"coef0": MinMaxNumericParamDef(0, 1)
}
random_state_rs = check_random_state(42)
optimizer_names = ["RandomSearch", "BayOpt_EI", "BayOpt_EI_MCMC"]
optmizers = ["RandomSearch", "BayOpt", "BayOpt"]
optimizer_args = [{
"random_state": random_state_rs
}, {
"initial_random_runs": random_steps
}, {
"initial_random_runs": random_steps,
"mcmc": True
}]
evaluate_on_mnist(optimizer_names,
optmizers,
param_defs,
optimizer_args,
regressor,
cv,
percentage,
steps=steps * cv,
random_steps=random_steps,
plot_at_end=True,
disable_auto_plot=disable_auto_plot)
def demo_MNIST_LibSVM(steps,
random_steps,
percentage,
cv,
plot_at_end=True,
disable_auto_plot=False):
logging.basicConfig(level=logging.DEBUG)
param_defs = {
"C": MinMaxNumericParamDef(0, 10),
"gamma": MinMaxNumericParamDef(0, 1),
"nu": MinMaxNumericParamDef(0, 1)
}
random_state_rs = check_random_state(42)
regressor = libsvm()
optimizer_names = ["RandomSearch", "BayOpt_EI"]
optmizers = ["RandomSearch", "BayOpt"]
optimizer_args = [{
"random_state": random_state_rs
}, {
"initial_random_runs": random_steps
}]
evaluate_on_mnist(optimizer_names,
optmizers,
param_defs,
optimizer_args,
regressor,
cv,
percentage,
steps=steps,
random_steps=random_steps,
plot_at_end=plot_at_end,
disable_auto_plot=disable_auto_plot)
def _get_one_candidate(self, experiment):
self.random_state = check_random_state(self.random_state)
value_dict = {}
for key, param_def in experiment.parameter_definitions.iteritems():
value_dict[key] = self._gen_param_val(param_def)
return Candidate(value_dict)
def __init__(self, experiment, optimizer_params=None):
"""
Initializes a bayesian optimizer.
Parameters
----------
experiment : Experiment
The experiment for which to optimize.
optimizer_arguments: dict of string keys
Sets the possible arguments for this optimizer. Available are:
"initial_random_runs" : int, optional
The number of initial random runs before using the GP. Default
is 10.
"random_state" : scipy random state, optional
The scipy random state or object to initialize one. Default is
None.
"acquisition_hyperparameters" : dict, optional
dictionary of acquisition-function hyperparameters
"num_gp_restarts" : int
GPy's optimization requires restarts to find a good solution.
This parameter controls this. Default is 10.
"acquisition" : AcquisitionFunction
The acquisition function to use. Default is
ExpectedImprovement.
"num_precomputed" : int
The number of points that should be kept precomputed for faster
multiple workers.
"""
self._logger = get_logger(self)
self._logger.debug(
"Initializing bayesian optimizer. Experiment is %s,"
" optimizer_params %s", experiment, optimizer_params)
if optimizer_params is None:
optimizer_params = {}
self.random_state = optimizer_params.get("random_state", None)
self.initial_random_runs = optimizer_params.get(
'initial_random_runs', self.initial_random_runs)
self.random_state = check_random_state(
optimizer_params.get('random_state', None))
self.acquisition_hyperparams = optimizer_params.get(
'acquisition_hyperparams', None)
self.num_gp_restarts = optimizer_params.get('num_gp_restarts',
self.num_gp_restarts)
self._logger.debug(
"Initialized relevant parameters. "
"initial_random_runs is %s, random_state is %s, "
"acquisition_hyperparams %s, num_gp_restarts %s",
self.initial_random_runs, self.random_state,
self.acquisition_hyperparams, self.num_gp_restarts)
if not isinstance(optimizer_params.get('acquisition'),
AcquisitionFunction):
self.acquisition_function = optimizer_params.get(
"acquisition", ExpectedImprovement)
self.acquisition_function = check_acquisition(
acquisition=self.acquisition_function,
acquisition_params=self.acquisition_hyperparams)
self._logger.debug(
"acquisition is no AcquisitionFunction. Set "
"it to %s", self.acquisition_function)
else:
self.acquisition_function = optimizer_params.get("acquisition")
self._logger.debug(
"Loaded acquisition function from "
"optimizer_params. Is %s", self.acquisition_function)
self.kernel_params = optimizer_params.get("kernel_params", {})
self.kernel = optimizer_params.get("kernel", "matern52")
self._logger.debug("Kernel details: Kernel is %s, kernel_params %s",
self.kernel, self.kernel_params)
self.random_searcher = RandomSearch(experiment, optimizer_params)
self._logger.debug("Initialized required RandomSearcher; is %s",
self.random_searcher)
Optimizer.__init__(self, experiment, optimizer_params)
self._logger.debug("Finished initializing bayOpt.")
def _get_one_candidate(self, experiment):
self.random_state = check_random_state(self.random_state)
value_dict = {}
for key, param_def in experiment.parameter_definitions.iteritems():
value_dict[key] = self._gen_param_val(param_def)
return Candidate(value_dict)
