def setup_mixins(policy: Policy, obs_space: gym.spaces.Space,
action_space: gym.spaces.Space,
config: TrainerConfigDict) -> None:
ValueNetworkMixin.__init__(policy, obs_space, action_space, config)
# Set up a tf-var for the moving avg (do this here to make it work with
# eager mode); "c^2" in the paper.
policy._moving_average_sqd_adv_norm = get_variable(
100.0,
framework="tf",
tf_name="moving_average_of_advantage_norm",
trainable=False)
def setup_mixins(policy: Policy, obs_space: gym.spaces.Space,
action_space: gym.spaces.Space,
config: TrainerConfigDict) -> None:
# Setup Value branch of our NN.
ValueNetworkMixin.__init__(policy, obs_space, action_space, config)
# Not needed for pure BC.
if policy.config["beta"] != 0.0:
# Set up a torch-var for the squared moving avg. advantage norm.
policy._moving_average_sqd_adv_norm = torch.tensor(
[policy.config["moving_average_sqd_adv_norm_start"]],
dtype=torch.float32,
requires_grad=False).to(policy.device)
def setup_mixins(policy: Policy, obs_space: gym.spaces.Space,
action_space: gym.spaces.Space,
config: TrainerConfigDict) -> None:
# Setup Value branch of our NN.
ValueNetworkMixin.__init__(policy, obs_space, action_space, config)
# Not needed for pure BC.
if policy.config["beta"] != 0.0:
# Set up a tf-var for the moving avg (do this here to make it work
# with eager mode); "c^2" in the paper.
policy._moving_average_sqd_adv_norm = get_variable(
policy.config["moving_average_sqd_adv_norm_start"],
framework="tf",
tf_name="moving_average_of_advantage_norm",
trainable=False)
