def _build_optimizer(self):
"""Creates the optimizer"""
optimizer = getters.get_optimizer(
self.optimizer_config.module,
learning_rate=self.optimizer_config.learning_rate,
decay_type=self.optimizer_config.decay_type,
decay_steps=self.optimizer_config.decay_steps,
decay_rate=self.optimizer_config.decay_rate,
start_decay_at=self.optimizer_config.start_decay_at,
stop_decay_at=self.optimizer_config.stop_decay_at,
min_learning_rate=self.optimizer_config.min_learning_rate,
staircase=self.optimizer_config.staircase,
**self.optimizer_config.params)
# Optionally wrap with SyncReplicasOptimizer
if self.optimizer_config.sync_replicas > 0:
optimizer = tf.train.SyncReplicasOptimizer(
opt=optimizer,
replicas_to_aggregate=self.optimizer_config.
sync_replicas_to_aggregate,
total_num_replicas=self.optimizer_config.sync_replicas)
# This is really ugly, but we need to do this to make the optimizer
# accessible outside of the model.
configs.SYNC_REPLICAS_OPTIMIZER = optimizer
return optimizer
def _build_optimizer(self):
"""Creates the optimizer"""
optimizer = getters.get_optimizer(
self.optimizer_config.module,
learning_rate=self.optimizer_config.learning_rate,
decay_type=self.optimizer_config.decay_type,
decay_steps=self.optimizer_config.decay_steps,
decay_rate=self.optimizer_config.decay_rate,
start_decay_at=self.optimizer_config.start_decay_at,
stop_decay_at=self.optimizer_config.stop_decay_at,
min_learning_rate=self.optimizer_config.min_learning_rate,
staircase=self.optimizer_config.staircase,
**self.optimizer_config.params)
# Optionally wrap with SyncReplicasOptimizer
if self.optimizer_config.sync_replicas > 0:
optimizer = tf.train.SyncReplicasOptimizer(
opt=optimizer,
replicas_to_aggregate=self.optimizer_config.sync_replicas_to_aggregate,
total_num_replicas=self.optimizer_config.sync_replicas)
# This is really ugly, but we need to do this to make the optimizer
# accessible outside of the model.
configs.SYNC_REPLICAS_OPTIMIZER = optimizer
return optimizer
def _build_optimizer(self):
"""Creates the optimizer"""
optimizer = getters.get_optimizer(self.optimizer.IDENTIFIER,
**self.optimizer.to_dict())
# TODO: use the _SyncReplicasOptimizerHook
# # Optionally wrap with SyncReplicasOptimizer
# if self.optimizer.sync_replicas > 0:
# optimizer = tf.train.SyncReplicasOptimizer(
# opt=optimizer,
# replicas_to_aggregate=self.optimizer.sync_replicas_to_aggregate,
# total_num_replicas=self.optimizer.sync_replicas)
# # This is really ugly, but we need to do this to make the optimizer
# # accessible outside of the model.
# configs.SYNC_REPLICAS_OPTIMIZER = optimizer
return optimizer