NAS Interface#

(hivenas.core.nas.nas_interface)

The NAS Interface encapsulating the Evaluation Strategy and Search Space.

class NASInterface(space_config=None, eval_config=None)[source]#

An interface that combines the Search Space & Evaluation Strategy for the NAS Search Algorithm (ABC)

cls.eval_strategy#

NASEval instance used to instantiate and evaluate candidates

cls.search_space#

NASSearchSpace instance used to sample candidates and neighbors

evaluate(candidate)[source]#

Evaluates a given candidate architecture

Parameters: candidate (str) – string-encoded representation of the architecture to be evaluated
Returns: a dictionary containing all relevant results to be saved, including: fitness, number of training epochs conducted (in case of ACT), hashed file name, number of trainable parameters, and the last epoch’s momentum value if applicable
Return type: dict

fully_train_best_model(from_arch=True)[source]#

Fully-train best-performing model (relies on paths set in Params)

Parameters: from_arch (bool, optional) – determines whether to train model from scratch using the string representations of the architecture (from_arch = True) or load the saved model file and continue training (from_arch = False). Note: optimizer settings are typically not saved, therefore training continuation from a model’s file can result in a worse overall accuracy (read more…).
Returns: a dictionary containing all relevant results to be saved, including: fitness, number of training epochs conducted (not including any previous trainings), hashed file name, number of trainable parameters
Return type: dict
Raises: FileNotFoundError – raises an error if the results file or model h5 file (when applicable) do not exist

get_neighbor(orig_arch)[source]#

Returns a random architecture with 1 op diff to the given candidate

Parameters: orig_arch (str) – string-encoded representation of the candidate architecture
Returns: string-encoded representation of the neighbor architecture
Return type: str

property is_minimize#

Used by the optimization algorithm to determine whether this is a minimization or maximization problem

Returns: hard-coded False; the search algorithm is always maximizing accuracy
Return type: bool

momentum_eval(candidate, weights_filename, m_epochs)[source]#

Trains a given network for additional m_epochs

Parameters

candidate (str) – string-encoded representation of the candidate architecture
weights_filename (str) – the SHA1-hashed unique string ID for the given architecture
m_epochs (int) – the additional momentum epochs the candidate should be trained for

Returns

final fitness value (accuracy) after training continuation

Return type

dict

Samples new random candidate architecture from the search space

Returns: string-encoded representation of the sampled candidate architecture
Return type: str

train_custom_arch(arch_str)[source]#

Fully trains a given string-encoded architecture (primarily used for debugging)

Parameters: arch_str (str) – string-encoded representation of the architecture
Returns: a dictionary containing all relevant results to be saved, including: fitness, number of training epochs conducted (not including any previous trainings), hashed file name, number of trainable parameters
Return type: dict