NAS-Bench-ASR

Bases: Graph

Interface to the tabular benchmark for nas-bench-asr architectures.

This class extends the Graph class to provide a structure for ASR neural network architectures. It includes methods to create macro graphs, cells blocks, cells, and query methods for searching optimal architectures.

Attributes:

Name	Type	Description
QUERYABLE	bool	Whether the architecture can be queried.
OPTIMIZER_SCOPE	list of str	List of the names of cell stages.

Note

Currently, building a NASLib object for nas-bench-asr architectures is not supported.

__init__()

Initialize the NasBenchASRSearchSpace object.

Set the properties to default values, which will be used to create the neural network architecture.

encode(encoding_type=EncodingType.ADJACENCY_ONE_HOT)

Encode the architecture based on the specified encoding type.

Parameters:

Name	Type	Description	Default
encoding_type	EncodingType	The type of encoding to use. Defaults to EncodingType.ADJACENCY_ONE_HOT.	EncodingType.ADJACENCY_ONE_HOT

Returns:

Name	Type	Description
object		The encoded architecture.

get_compact()

Get the compact representation of the architecture.

Returns:

Type	Description
	The compact representation of the architecture.

Raises:

Type	Description
AssertionError	If the compact representation is not set.

get_hash()

Get the hash of the architecture based on its compact representation.

Returns:

Type	Description
	The hash of the architectures.

get_max_epochs()

Get the maximum number of epochs for training.

Returns:

Name	Type	Description
int		The maximum number of epochs.

get_nbhd(dataset_api=None)

Get all neighbors of the current architecture.

Parameters:

Name	Type	Description	Default
dataset_api	optional	The dataset API instance for neighbor fetching. Defaults to None.	None

Returns:

Name	Type	Description
list		List of all neighbor architectures.

get_type()

Get the type of the search space.

Returns:

Name	Type	Description
str		The type of the search space, in this case, 'asr'.

mutate(parent, mutation_rate=1, dataset_api=None)

Mutate the architecture.

Parameters:

Name	Type	Description	Default
parent	NasBenchASRSearchSpace	The parent architecture.	required
mutation_rate	int	The rate of mutation. Defaults to 1.	1
dataset_api	DatasetAPI	The dataset API instance for the mutation. Defaults to None.	None

Returns:

Name	Type	Description
None		The architecture is mutated in-place.

Note

This will mutate the cell in one of two ways: change an edge; change an op.

Todo: mutate by adding/removing nodes. Todo: mutate the list of hidden nodes. Todo: edges between initial hidden nodes are not mutated.

query(metric=None, dataset=None, path=None, epoch=-1, full_lc=False, dataset_api=None)

Query results from the nas-bench-asr benchmark.

Parameters:

Name	Type	Description	Default
metric	Metric	The performance metric to query for.	None
dataset	str	The dataset to query on.	None
path	str	The file path to save the results.	None
epoch	int	The epoch number at which to query the performance metric.	-1
full_lc	bool	Whether to return the full learning curve.	False
dataset_api	dict	The dataset API to use for querying.	None

Returns:

Type	Description
	float or list: The value(s) of the queried metric.

sample_random_architecture(dataset_api)

Sample a random architecture based on the dataset API.

Parameters:

Name	Type	Description	Default
dataset_api		The dataset API instance for the architecture sampling.	required

Returns:

Type	Description
	The compact representation of the sampled architecture.

set_compact(compact)

Set the compact representation of the architecture.

Parameters:

Name	Type	Description	Default
compact		The new compact representation of the architecture.	required