def __init__(self, experiment_name: str, search_config: AgingEvoConfig,
training_config: TrainingConfig, bound_config: BoundConfig):
self.log = logging.getLogger(
name=f"AgingEvoSearch [{experiment_name}]")
self.config = search_config
self.trainer = ModelTrainer(training_config)
self.root_dir = Path(search_config.checkpoint_dir)
self.root_dir.mkdir(parents=True, exist_ok=True)
self.experiment_name = experiment_name
if training_config.pruning and not training_config.pruning.structured:
self.log.warning(
"For unstructured pruning, we can only meaningfully use the model "
"size resource metric.")
bound_config.peak_mem_bound = None
bound_config.mac_bound = None
self.pruning = training_config.pruning
# We establish an order of objective in the feature vector, all functions must ensure the order is the same
self.constraint_bounds = [
bound_config.error_bound, bound_config.peak_mem_bound,
bound_config.model_size_bound, bound_config.mac_bound
]
self.history: List[EvaluatedPoint] = []
self.population: List[EvaluatedPoint] = []
self.population_size = search_config.population_size
self.initial_population_size = search_config.initial_population_size or self.population_size
self.rounds = search_config.rounds
self.sample_size = search_config.sample_size
num_gpus = len(tf.config.experimental.list_physical_devices("GPU"))
self.max_parallel_evaluations = search_config.max_parallel_evaluations or num_gpus
def __init__(self,
experiment_name: str,
search_config: BayesOptConfig,
training_config: TrainingConfig,
bound_config: BoundConfig):
assert search_config.starting_points >= 1
self.log = logging.getLogger(name=f"BayesOpt [{experiment_name}]")
self.config = search_config
self.trainer = ModelTrainer(training_config)
self.root_dir = Path(search_config.checkpoint_dir)
self.root_dir.mkdir(parents=True, exist_ok=True)
self.experiment_name = experiment_name
if training_config.pruning and not training_config.pruning.structured:
self.log.warning("For unstructured pruning, we can only use the model size resource metric.")
bound_config.peak_mem_bound = None
bound_config.mac_bound = None
# We establish an order of objective in the feature vector, all functions must ensure the order is the same
self.constraint_bounds = [bound_config.error_bound, bound_config.peak_mem_bound,
bound_config.model_size_bound, bound_config.mac_bound]