class CaltechExperimentRep(CaltechExperiment):
"""
Performs experiments over Caltech-101 dataset using the proposed representative-selection algorithm
"""
optimizer_name = const.TR_REP
general_config = None
trainer = None
def _prepare_trainer(self):
tester = Tester(self.neural_net, self.data_input, self.input_tensor, self.output_tensor)
self.trainer = RepresentativesTrainer(self.general_config, self.neural_net, self.data_input,
self.input_tensor, self.output_tensor,
tester=tester, checkpoint=self.ckp_path)
def _prepare_config(self, str_optimizer: str, train_mode: TrainMode):
self.general_config = CRILConfig(train_mode, 0.0001, self.summary_interval, self.ckp_interval,
config_name=str_optimizer, model_name=self.dataset_name,
n_candidates=20, memory_size=1, buffer_size=1)
# Creates configuration for 5 mega-batches
if train_mode == TrainMode.INCREMENTAL or train_mode == TrainMode.ACUMULATIVE:
for i in range(5):
train_conf = MegabatchConfig(60, batch_size=128)
self.general_config.add_train_conf(train_conf)
else:
raise OptionNotSupportedError("The requested Experiment class: {} doesn't support the requested training"
" mode: {}".format(self.__class__, train_mode))
def _prepare_scenarios(self, base_config):
scenarios = None
scenarios = self._add_scenario(scenarios, base_config, 'Test with 1% of data stored as representatives')
scenario = copy.copy(base_config)
scenario.memory_size = 10
scenarios = self._add_scenario(scenarios, scenario, 'Test with 10% of data stored as representatives')
return scenarios
def _prepare_config(self, str_optimizer: str, train_mode: TrainMode):
self.general_config = CRILConfig(train_mode, 0.0001, self.summary_interval, self.ckp_interval,
config_name=str_optimizer, model_name=self.dataset_name,
n_candidates=20, memory_size=1, buffer_size=1)
# Creates configuration for 5 mega-batches
if train_mode == TrainMode.INCREMENTAL or train_mode == TrainMode.ACUMULATIVE:
for i in range(5):
train_conf = MegabatchConfig(60, batch_size=128)
self.general_config.add_train_conf(train_conf)
else:
raise OptionNotSupportedError("The requested Experiment class: {} doesn't support the requested training"
" mode: {}".format(self.__class__, train_mode))