NPenas

Bases: MetaOptimizer

Implements the Npenas optimizer.

Npenas is a Neural Predictor Guided Evolution for Neural Architecture Search optimization algorithm which uses machine learning to guide the search process of neural architecture search (NAS). This algorithm employs predictive models, acquisition functions, and zero-cost proxies to speed up the search process.

Attributes:

Name	Type	Description
config	obj	Configuration object containing setup parameters.
zc_api	obj	Zero cost API object to query zero-cost proxies.

__init__(config, zc_api=None)

Constructor of Npenas optimizer.

Parameters:

Name	Type	Description	Default
config	obj	Configuration object containing setup parameters.	required
zc_api	obj	Zero cost API object to query zero-cost proxies.	None

adapt_search_space(search_space, scope=None, dataset_api=None)

Adapts the search space for optimization.

Parameters:

Name	Type	Description	Default
search_space	obj	The search space object that contains all possible architectures.	required
scope	str	Defines the scope of the search space for the optimizer.	None
dataset_api	obj	API for querying dataset specific metrics.	None

Raises:

Type	Description
AssertionError	If search_space is not queryable.

get_arch_as_string(arch)

Returns a string representation of the given architecture.

Parameters:

Name	Type	Description	Default
arch	obj	The architecture to convert to string.	required

Returns:

Name	Type	Description
str		The string representation of the architecture.

get_checkpointables()

Returns the models that should be checkpointed.

Returns:

Name	Type	Description
dict		A dictionary with 'model' as the key and the history of models as the value.

get_final_architecture()

Returns the final/best architecture from the history.

Returns:

Name	Type	Description
obj		The final/best architecture.

get_model_size()

Returns the size of the model.

Returns:

Name	Type	Description
float		The size of the model in MB.

get_op_optimizer()

Method to get the optimizer of the operation. Not implemented in this class.

Raises:

Type	Description
NotImplementedError	Always, as this method is not implemented.

get_zero_cost_predictors()

Creates and returns a dictionary of zero cost predictors based on the zc_names configuration.

Returns:

Name	Type	Description
dict		A dictionary where keys are zero cost predictor names and values are the predictor objects.

new_epoch(epoch)

Conducts one epoch of the optimization process.

Parameters:

Name	Type	Description	Default
epoch	int	The current epoch number.	required

query_zc_scores(arch)

Queries and returns the zero-cost scores for a given architecture.

Parameters:

Name	Type	Description	Default
arch	obj	The architecture to query.	required

Returns:

Name	Type	Description
dict		A dictionary where keys are zero cost predictor names and values are the corresponding scores.

test_statistics()

Returns the test statistics of the final architecture.

Returns:

Name	Type	Description
int		The raw test metric of the final architecture.

train_statistics(report_incumbent=True)

Returns the training statistics of the best architecture or the last trained one.

Parameters:

Name	Type	Description	Default
report_incumbent	bool	Whether to return the statistics of the best architecture. Defaults to True.	True

Returns:

Name	Type	Description
tuple		A tuple containing the training accuracy, validation accuracy, test accuracy, and training time of the architecture.