Documentation for NevergradOptimizer
Nevergrad Wrapper for PHOTONAI.
Nevergrad is a gradient-free optimization platform.
Nevergrad usage and implementation details.
Examples:
1 2 3 4 5 6 7 8 9 | |
Source code in photonai/optimization/nevergrad/nevergrad.py
class NevergradOptimizer(PhotonMasterOptimizer):
"""Nevergrad Wrapper for PHOTONAI.
Nevergrad is a gradient-free optimization platform.
Nevergrad [usage and implementation details](
https://facebookresearch.github.io/nevergrad/).
Example:
``` python
import nevergrad as ng
# list of all available nevergrad optimizer
print(list(ng.optimizers.registry.values()))
my_pipe = Hyperpipe('nevergrad_example',
optimizer='nevergrad',
optimizer_params={'facade': 'NGO', 'n_configurations': 30},
...
)
```
"""
def __init__(self, facade='NGO', n_configurations: int = 100, rng: int = 42):
"""
Initialize the object.
Parameters:
facade:
Choice of the Nevergrad backend strategy, e.g. [NGO, ...].
n_configurations:
Number of runs.
rng:
Random Seed.
"""
if not __found__:
msg = "Module nevergrad not found or not installed as expected. " \
"Please install the nevergrad/requirements.txt PHOTONAI provides."
logger.error(msg)
raise ModuleNotFoundError(msg)
if facade in list(ng.optimizers.registry.values()):
self.facade = facade
elif facade in list(ng.optimizers.registry.keys()):
self.facade = ng.optimizers.registry[facade]
else:
msg = "nevergrad.optimizer {} not known. Check out all available nevergrad optimizers " \
"by nevergrad.optimizers.registry.keys()".format(str(facade))
logger.error(msg.format(str(facade)))
raise ValueError(msg.format(str(facade)))
self.n_configurations = n_configurations
self.space = None # Hyperparameter space for nevergrad
self.switch_optiones = {}
self.hyperparameters = []
self.rng = rng
self.maximize_metric = False
self.constant_dictionary = {}
self.objective = None
self.optimizer = None
def prepare(self, pipeline_elements: list, maximize_metric: bool, objective_function: Callable) -> None:
"""Prepare Nevergrad Optimizer.
Parameters:
pipeline_elements:
List of all PipelineElements to create the hyperparameter space.
maximize_metric:
Boolean to distinguish between score and error.
objective_function:
The cost or objective function.
"""
self.space = self._build_nevergrad_space(pipeline_elements)
self.space.random_state.seed(self.rng)
if self.constant_dictionary:
msg = "PHOTONAI has detected some one-valued params in your hyperparameters. Pleas use the kwargs for " \
"constant values. This run ignores following settings: " + str(self.constant_dictionary.keys())
logger.warning(msg)
warnings.warn(msg)
self.maximize_metric = maximize_metric
def nevergrad_objective_function(**current_config):
return objective_function(current_config)
self.objective = nevergrad_objective_function
self.optimizer = self.facade(parametrization=self.space, budget=self.n_configurations)
def optimize(self) -> None:
self.optimizer.minimize(self.objective)
def _build_nevergrad_space(self, pipeline_elements: list):
"""
Build entire Nevergrad hyperparameter space.
Parameters:
pipeline_elements:
List of all pipeline_elements to create the hyperparameter space.
"""
param_dict = {}
for pipe_element in pipeline_elements:
# build conditions for switch elements
if pipe_element.__class__.__name__ == 'Switch':
raise NotImplementedError("Currently PHOTONAIs Switch is not supported by nevergrad.")
if hasattr(pipe_element, 'hyperparameters'):
for name, value in pipe_element.hyperparameters.items():
self.hyperparameters.append(name)
# if we only have one value we do not need to optimize
if isinstance(value, list) and len(value) < 2:
self.constant_dictionary[name] = value[0]
continue
if isinstance(value, PhotonCategorical) and len(value.values) < 2:
self.constant_dictionary[name] = value.values[0]
continue
nevergrad_param = self._convert_photonai_to_nevergrad_param(value)
if nevergrad_param is not None:
param_dict[name] = nevergrad_param
return ng.p.Instrumentation(**param_dict)
@staticmethod
def _convert_photonai_to_nevergrad_param(hyperparam: PhotonHyperparam):
"""
Helper function: Convert PHOTONAI to Nevergrad hyperparameter.
Parameters:
hyperparam:
One of photonai.optimization.hyperparameters.
"""
if isinstance(hyperparam, PhotonCategorical) or isinstance(hyperparam, BooleanSwitch):
return ng.p.Choice(hyperparam.values)
elif isinstance(hyperparam, list):
return ng.p.Choice(hyperparam)
elif isinstance(hyperparam, FloatRange):
if hyperparam.range_type == 'linspace':
return ng.p.Scalar(lower=hyperparam.start, upper=hyperparam.stop)
elif hyperparam.range_type == 'logspace':
return ng.p.Log(lower=hyperparam.start, upper=hyperparam.stop)
msg = str(hyperparam.range_type) + "in your float hyperparameter is not implemented yet."
logger.error(msg)
raise NotImplementedError(msg)
elif isinstance(hyperparam, IntegerRange):
return ng.p.Scalar(lower=hyperparam.start, upper=hyperparam.stop).set_integer_casting()
msg = "Cannot convert hyperparameter " + str(hyperparam) + ". Supported types: Categorical, IntegerRange," \
"FloatRange, list."
logger.error(msg)
raise ValueError(msg)
__init__(self, facade='NGO', n_configurations=100, rng=42)
special
Initialize the object.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
facade |
Choice of the Nevergrad backend strategy, e.g. [NGO, ...]. |
'NGO' |
|
n_configurations |
int |
Number of runs. |
100 |
rng |
int |
Random Seed. |
42 |
Source code in photonai/optimization/nevergrad/nevergrad.py
def __init__(self, facade='NGO', n_configurations: int = 100, rng: int = 42):
"""
Initialize the object.
Parameters:
facade:
Choice of the Nevergrad backend strategy, e.g. [NGO, ...].
n_configurations:
Number of runs.
rng:
Random Seed.
"""
if not __found__:
msg = "Module nevergrad not found or not installed as expected. " \
"Please install the nevergrad/requirements.txt PHOTONAI provides."
logger.error(msg)
raise ModuleNotFoundError(msg)
if facade in list(ng.optimizers.registry.values()):
self.facade = facade
elif facade in list(ng.optimizers.registry.keys()):
self.facade = ng.optimizers.registry[facade]
else:
msg = "nevergrad.optimizer {} not known. Check out all available nevergrad optimizers " \
"by nevergrad.optimizers.registry.keys()".format(str(facade))
logger.error(msg.format(str(facade)))
raise ValueError(msg.format(str(facade)))
self.n_configurations = n_configurations
self.space = None # Hyperparameter space for nevergrad
self.switch_optiones = {}
self.hyperparameters = []
self.rng = rng
self.maximize_metric = False
self.constant_dictionary = {}
self.objective = None
self.optimizer = None
optimize(self)
Start the optimization process based on the underlying objective function.
Source code in photonai/optimization/nevergrad/nevergrad.py
def optimize(self) -> None:
self.optimizer.minimize(self.objective)
prepare(self, pipeline_elements, maximize_metric, objective_function)
Prepare Nevergrad Optimizer.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
pipeline_elements |
list |
List of all PipelineElements to create the hyperparameter space. |
required |
maximize_metric |
bool |
Boolean to distinguish between score and error. |
required |
objective_function |
Callable |
The cost or objective function. |
required |
Source code in photonai/optimization/nevergrad/nevergrad.py
def prepare(self, pipeline_elements: list, maximize_metric: bool, objective_function: Callable) -> None:
"""Prepare Nevergrad Optimizer.
Parameters:
pipeline_elements:
List of all PipelineElements to create the hyperparameter space.
maximize_metric:
Boolean to distinguish between score and error.
objective_function:
The cost or objective function.
"""
self.space = self._build_nevergrad_space(pipeline_elements)
self.space.random_state.seed(self.rng)
if self.constant_dictionary:
msg = "PHOTONAI has detected some one-valued params in your hyperparameters. Pleas use the kwargs for " \
"constant values. This run ignores following settings: " + str(self.constant_dictionary.keys())
logger.warning(msg)
warnings.warn(msg)
self.maximize_metric = maximize_metric
def nevergrad_objective_function(**current_config):
return objective_function(current_config)
self.objective = nevergrad_objective_function
self.optimizer = self.facade(parametrization=self.space, budget=self.n_configurations)