def reward(state, dataset_handler, log_path="./workspace"):
"""Perform the training of the network, given (state, dataset) pair."""
try:
train_features, train_labels = dataset_handler.current_train_set()
val_features, val_labels = dataset_handler.current_validation_set()
hash_state = compute_str_hash(state_to_string(state))
nas_trainer = EarlyStopNASTrainer(
encoded_network=state,
input_shape=infer_data_shape(train_features),
n_classes=infer_n_classes(train_labels),
batch_size=256,
log_path="{lp}/trainer-{h}".format(lp=log_path, h=hash_state),
mu=0.5,
rho=0.5,
variable_scope="cnn-{h}".format(h=hash_state))
train_features = normalize_dataset(dataset=train_features,
baseline=255)
train_labels = train_labels.astype(np.int32)
nas_trainer.train(
train_data=train_features,
train_labels=train_labels,
train_input_fn="default",
n_epochs=12 # As specified by BlockQNN
)
val_features = normalize_dataset(dataset=val_features,
baseline=255)
val_labels = val_labels.astype(np.int32)
res = nas_trainer.evaluate(eval_data=val_features,
eval_labels=val_labels,
eval_input_fn="default")
accuracy = res['accuracy']
# Compute the refined reward as defined
reward = accuracy*100 - nas_trainer.weighted_log_density - \
nas_trainer.weighted_log_flops
return reward
except Exception as ex: # pylint: disable=broad-except
# TODO: Make sure exceptions are printed correctly.
print("Reward computation failed with exception:", ex)
return 0.
def test_train(self):
"""Test the Default Training procedure."""
tf.reset_default_graph()
if os.path.isdir(self.training_dir):
shutil.rmtree(self.training_dir)
nas_trainer = DefaultNASTrainer(
encoded_network=self.net_nsc,
input_shape=infer_data_shape(self.train_data),
n_classes=infer_n_classes(self.train_labels),
batch_size=self.batch_size,
log_path=self.training_dir,
variable_scope="cnn")
nas_trainer.train(train_data=self.train_data,
train_labels=self.train_labels,
train_input_fn="default")
self.assertTrue(os.path.isdir(self.training_dir))
def test_evaluate(self):
"""Test the Default Training procedure."""
tf.reset_default_graph()
if os.path.isdir(self.training_dir):
shutil.rmtree(self.training_dir)
nas_trainer = EarlyStopNASTrainer(
encoded_network=self.net_nsc,
input_shape=infer_data_shape(self.train_data),
n_classes=infer_n_classes(self.train_labels),
batch_size=self.batch_size,
log_path=self.training_dir,
mu=0.5,
rho=0.5,
variable_scope="cnn")
nas_trainer.train(train_data=self.train_data,
train_labels=self.train_labels,
train_input_fn="default")
res = nas_trainer.evaluate(eval_data=self.eval_data,
eval_labels=self.eval_labels,
eval_input_fn="default")
self.assertTrue(os.path.isdir(self.training_dir))
self.assertTrue("accuracy" in list(res.keys()))
self.assertTrue(nas_trainer.density is not None)
self.assertTrue(nas_trainer.density != 0.)
self.assertTrue(nas_trainer.weighted_log_density is not None)
self.assertTrue(nas_trainer.weighted_log_density != 0.)
self.assertTrue(nas_trainer.flops is not None)
self.assertTrue(nas_trainer.flops != 0.)
self.assertTrue(nas_trainer.weighted_log_flops is not None)
self.assertTrue(nas_trainer.weighted_log_flops != 0.)