Evaluation Strategy#

(hivenas.core.nas.evaluation_strategy)

The Evaluation Strategy phase of the NAS framework.

class NASEval(config)[source]#

Bases: object

Responsible for instantiating and evaluating candidate architectures

config#

the predefined operational parameters pertaining to evaluation (defined in evaluation_strategy_config)

Type

dict

datagen#

the data generator used to train all candidates (instantiated once to preserve resources)

Type

tensorflow.keras.preprocessing.image.ImageDataGenerator

model#

the candidate model to be evaluated (overwritten and deleted with every evaluation process to prevent leaks)

Type

Model

evaluate(arch)[source]#

Evaluates the candidate architecture given a string-encoded representation of the model

Parameters

arch (list) – a list of architecture operations ([str]), encoded by NASSearchSpace

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(model_file=None, arch=None)[source]#

Loads and continues training of a partially-trained model using either a string-encoded architecture (instantiates the model and trains from scratch) or a model h5 file

Parameters

  • model_file (str, optional) – a previously saved h5 model file (continues training)

  • arch (list, optional) – a list of architecture operations ([str]), encoded by NASSearchSpace

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

get_weights_filename(arch)[source]#

Hashes the architecture op-list into a unique filename using SHA1

Parameters

arch (list) – a list of architecture operations ([str]), encoded by NASSearchSpace

Returns

SHA1-hashed unique string ID for the given architecture

Return type

str

instantiate_network(arch)[source]#

Instantiates a Keras network given an architecture op list

Parameters

arch (list) – a list of architecture operations ([str]), encoded by NASSearchSpace

momentum_training(weights_file, m_epochs)[source]#

Loads and continues training of a partially-trained model

Parameters

  • weights_file (str) – the previously saved h5 model file (continues training)

  • m_epochs (int) – number of momentum epochs to continue training for

Returns

final fitness value (accuracy) after training continuation

Return type

dict