def make_ddpg_optimizers(policy: Policy, config: TrainerConfigDict) -> None:
# Create separate optimizers for actor & critic losses.
if policy.config["framework"] in ["tf2", "tfe"]:
policy._actor_optimizer = tf.keras.optimizers.Adam(
learning_rate=config["actor_lr"])
policy._critic_optimizer = tf.keras.optimizers.Adam(
learning_rate=config["critic_lr"])
else:
policy._actor_optimizer = tf1.train.AdamOptimizer(
learning_rate=config["actor_lr"])
policy._critic_optimizer = tf1.train.AdamOptimizer(
learning_rate=config["critic_lr"])
# TODO: (sven) make this function return both optimizers and
# TFPolicy handle optimizers vs loss terms correctly (like torch).
return None
def make_ddpg_optimizers(policy: Policy,
config: AlgorithmConfigDict) -> Tuple[LocalOptimizer]:
"""Create separate optimizers for actor & critic losses."""
# Set epsilons to match tf.keras.optimizers.Adam's epsilon default.
policy._actor_optimizer = torch.optim.Adam(
params=policy.model.policy_variables(),
lr=config["actor_lr"],
eps=1e-7)
policy._critic_optimizer = torch.optim.Adam(
params=policy.model.q_variables(), lr=config["critic_lr"], eps=1e-7)
# Return them in the same order as the respective loss terms are returned.
return policy._actor_optimizer, policy._critic_optimizer
