def update_targets(self, tau=1.0):
"""Performs a soft update of the target network parameters.
For each weight w_s in the actor/critic networks, and its corresponding
weight w_t in the target actor/critic networks, a soft update is:
w_t = (1- tau) x w_t + tau x ws
Args:
tau: A float scalar in [0, 1]
Returns:
An operation that performs a soft update of the target network parameters.
Raises:
ValueError: If `tau` is not in [0, 1].
"""
if tau < 0 or tau > 1:
raise ValueError('Input `tau` should be in [0, 1].')
update_actor = utils.soft_variables_update(
slim.get_trainable_variables(
utils.join_scope(self._scope, self.ACTOR_NET_SCOPE)),
slim.get_trainable_variables(
utils.join_scope(self._scope, self.TARGET_ACTOR_NET_SCOPE)),
tau)
# NOTE: This updates both critic networks.
update_critic = utils.soft_variables_update(
slim.get_trainable_variables(
utils.join_scope(self._scope, self.CRITIC_NET_SCOPE)),
slim.get_trainable_variables(
utils.join_scope(self._scope, self.TARGET_CRITIC_NET_SCOPE)),
tau)
return tf.group(update_actor, update_critic, name='update_targets')
def update_targets(self, tau=1.0):
"""Performs a soft update of the target network parameters.
For each weight w_s in the actor/critic networks, and its corresponding
weight w_t in the target actor/critic networks, a soft update is:
w_t = (1- tau) x w_t + tau x ws
Args:
tau: A float scalar in [0, 1]
Returns:
An operation that performs a soft update of the target network parameters.
Raises:
ValueError: If `tau` is not in [0, 1].
"""
if tau < 0 or tau > 1:
raise ValueError('Input `tau` should be in [0, 1].')
update_actor = utils.soft_variables_update(
slim.get_trainable_variables(
utils.join_scope(self._scope, self.ACTOR_NET_SCOPE)),
slim.get_trainable_variables(
utils.join_scope(self._scope, self.TARGET_ACTOR_NET_SCOPE)),
tau)
# NOTE: This updates both critic networks.
update_critic = utils.soft_variables_update(
slim.get_trainable_variables(
utils.join_scope(self._scope, self.CRITIC_NET_SCOPE)),
slim.get_trainable_variables(
utils.join_scope(self._scope, self.TARGET_CRITIC_NET_SCOPE)),
tau)
return tf.group(update_actor, update_critic, name='update_targets')