NAS-Bench-201

Bases: Graph

Represents the NASBench201 search space.

This class provides methods for querying and manipulating architectures within the search space, including methods for mutation and random sampling.

Attributes:

Name	Type	Description
num_classes	int	Number of classes for classification tasks.
in_channels	int	Number of input channels.
max_epoch	int	Maximum number of epochs for training.
space_name	str	The name of the search space.
labeled_archs	list	A list of labeled architectures.
instantiate_model	bool	Boolean indicating whether to instantiate the model during initialization.
sample_without_replacement	bool	Boolean indicating whether to sample architectures without replacement.
channels	list	Number of channels at different stages of the architecture.
op_indices	list	Indices of the operations.
OPTIMIZER_SCOPE	list	A list of the stages in the architecture, useful for scoping during optimization.
QUERYABLE	bool	A boolean indicating whether the search space is queryable or not.

__init__(n_classes=10, in_channels=3)

Constructor method.

This initializes the NasBench201SearchSpace object with provided number of classes and input channels.

Parameters:

Name	Type	Description	Default
n_classes	int	The number of classes for the classification task. Defaults to 10.	10
in_channels	int	The number of input channels. Defaults to 3.	3

encode(encoding_type=EncodingType.ADJACENCY_ONE_HOT)

Encodes the current architecture based on a given encoding type.

Parameters:

Name	Type	Description	Default
encoding_type	EncodingType	The encoding type for the architecture.	EncodingType.ADJACENCY_ONE_HOT

Returns:

Name	Type	Description
Any	Union[List, np.ndarray, dict]	The encoded architecture. The return type depends on the chosen encoding type.

Raises:

Type	Description
NotImplementedError	If the given encoding type is not yet supported as an architecture encoding for nb201.

forward_before_global_avg_pool(x)

Performs a forward pass until the global average pooling layer and returns the outputs.

Parameters:

Name	Type	Description	Default
x	torch.Tensor	The input tensor.	required

Returns:

Name	Type	Description
list	list	List of outputs from the forward pass.

get_arch_iterator(dataset_api=None)

Returns an iterator for all possible architectures in the search space. The iterator is a product of the number of operations for each edge in the graph.

Parameters:

Name	Type	Description	Default
dataset_api	optional	The dataset api. Defaults to None.	None

Returns:

Name	Type	Description
Iterator	Iterator	An iterator over all possible architectures.

get_hash()

Gets a hash representation of the architecture. The hash is a tuple of the operation indices.

Returns:

Name	Type	Description
tuple	tuple	The hash of the architecture.

get_loss_fn()

Returns the loss function to be used for this architecture.

Returns:

Name	Type	Description
Callable	Callable	A callable object (cross entropy loss function) that can be used as a loss function.

get_nbhd(dataset_api=None)

Returns all neighbors of the architecture.

Parameters:

Name	Type	Description	Default
dataset_api	dict	The api containing nasbench201 data. Defaults to None.	None

Returns:

Name	Type	Description
list	list	List of neighbor models.

get_op_indices()

Gets the operation indices of the architecture. If they are not defined yet, it will convert the naslib object to operation indices and save them.

Returns:

Name	Type	Description
list	list	The operation indices of the architecture.

get_type()

Returns the type of the search space.

Returns:

Name	Type	Description
str	str	The type of the search space, "nasbench201" in this case.

mutate(parent, dataset_api=None)

Mutates one operation from the parent operation indices and sets them as the operation indices of the current object.

Parameters:

Name	Type	Description	Default
parent	Graph	The parent Graph object from which to mutate.	required
dataset_api	dict	The api containing nasbench201 data. Defaults to None.	None

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

Query results from the nasbench201 database based on the specified metric and dataset.

Parameters:

Name	Type	Description	Default
metric	Metric	The performance metric to query for.	required
dataset	str	The dataset to query for.	required
path	str	The path to the nasbench201 database. Defaults to None.	None
epoch	int	The training epoch to query for. Defaults to -1, which means the last epoch.	-1
full_lc	bool	If True, returns the full learning curve. Defaults to False.	False
dataset_api	dict	The api containing nasbench201 data. Defaults to None.	None

Raises:

Type	Description
NotImplementedError	If the metric is Metric.ALL or the dataset_api is not provided.

Returns:

Name	Type	Description
float	float	The queried result.

sample_random_architecture(dataset_api=None, load_labeled=False)

Samples a random architecture and sets it as the current architecture.

Parameters:

Name	Type	Description	Default
dataset_api	dict	The api containing nasbench201 data. Defaults to None.	None
load_labeled	bool	If True, a random labeled architecture is sampled instead. Defaults to False.	False

sample_random_labeled_architecture()

Samples a random labeled architecture and sets it as the current architecture.

set_op_indices(op_indices)

Sets the operation indices for the current architecture. If the model should be instantiated, it will convert the operation indices to a naslib object.

Parameters:

Name	Type	Description	Default
op_indices	list	List of operation indices to set.	required

set_spec(op_indices, dataset_api=None)

Sets the specifications of the architecture.

Parameters:

Name	Type	Description	Default
op_indices	list	List of operation indices to set.	required
dataset_api	optional	The dataset api. Defaults to None.	None