Random Search

Bases: MetaOptimizer

RandomSearch is a class that implements the Random Search algorithm for Neural Architecture Search (NAS). It is derived from the MetaOptimizer class. The random search is done by randomly sampling 'k' architectures and training them for 'n' epochs, then selecting the best architecture. In the DARTS paper, 'k' equals 24 and 'n' equals 100 for the CIFAR-10 dataset.

Attributes:

Name	Type	Description
using_step_function	bool	Flag indicating the absence of a step function for this optimizer.
performance_metric	Metric	The performance metric for evaluating the architectures.
dataset	str	The dataset to be used for evaluation.
fidelity	int	The number of epochs for each sampled architecture's training.
device	torch.device	The device to be used for computations, either CUDA or CPU.
sampled_archs	list	A list to store the sampled architectures.
history	torch.nn.ModuleList	A list to store the history of architectures.

__init__(config)

Initializes the RandomSearch class with configuration settings.

Parameters:

Name	Type	Description	Default
config	CfgNode	Configuration settings for the search process.	required

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

Adapts the search space for the random search.

Parameters:

Name	Type	Description	Default
search_space	Graph	The search space to be adapted.	required
scope	str	The scope for the search. Defaults to None.	None
dataset_api	dict	API for the dataset. Defaults to None.	None

get_checkpointables()

Gets the models that can be checkpointed.

Returns:

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

get_final_architecture()

Gets the final (best) architecture from the search.

Returns:

Name	Type	Description
Graph		The best architecture found during the search.

get_op_optimizer()

Gets the optimizer for the operations. This method is not implemented in this class and raises an error when called.

Raises:

Type	Description
NotImplementedError	Always, because this method is not implemented in this class.

new_epoch(epoch)

Starts a new epoch in the search process, sampling a new architecture to train.

Parameters:

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

test_statistics()

Reports the test statistics.

Returns:

Name	Type	Description
float		The raw performance metric for the best architecture.

train_statistics(report_incumbent=True)

Reports the statistics after training.

Parameters:

Name	Type	Description	Default
report_incumbent	bool	Whether to report the incumbent or the most recent architecture. Defaults to True.	True

Returns:

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