def __init__(self, observation_space, action_space, config):
config = dict(
ray.rllib.algorithms.marwil.marwil.MARWILConfig().to_dict(),
**config)
TorchPolicyV2.__init__(
self,
observation_space,
action_space,
config,
max_seq_len=config["model"]["max_seq_len"],
)
ValueNetworkMixin.__init__(self, config)
PostprocessAdvantages.__init__(self)
# Not needed for pure BC.
if config["beta"] != 0.0:
# Set up a torch-var for the squared moving avg. advantage norm.
self._moving_average_sqd_adv_norm = torch.tensor(
[config["moving_average_sqd_adv_norm_start"]],
dtype=torch.float32,
requires_grad=False,
).to(self.device)
# TODO: Don't require users to call this manually.
self._initialize_loss_from_dummy_batch()
def __init__(self, observation_space, action_space, config):
config = dict(ray.rllib.agents.ppo.appo.DEFAULT_CONFIG, **config)
# Although this is a no-op, we call __init__ here to make it clear
# that base.__init__ will use the make_model() call.
VTraceOptimizer.__init__(self)
LearningRateSchedule.__init__(self, config["lr"], config["lr_schedule"])
TorchPolicyV2.__init__(
self,
observation_space,
action_space,
config,
max_seq_len=config["model"]["max_seq_len"],
)
EntropyCoeffSchedule.__init__(
self, config["entropy_coeff"], config["entropy_coeff_schedule"]
)
ValueNetworkMixin.__init__(self, config)
KLCoeffMixin.__init__(self, config)
# TODO: Don't require users to call this manually.
self._initialize_loss_from_dummy_batch()
# Initiate TargetNetwork ops after loss initialization.
TargetNetworkMixin.__init__(self)
def __init__(
self,
observation_space: gym.spaces.Space,
action_space: gym.spaces.Space,
config: AlgorithmConfigDict,
):
config = dict(ray.rllib.algorithms.ddpg.ddpg.DDPGConfig().to_dict(), **config)
# Create global step for counting the number of update operations.
self.global_step = 0
# Validate action space for DDPG
validate_spaces(self, observation_space, action_space)
TorchPolicyV2.__init__(
self,
observation_space,
action_space,
config,
max_seq_len=config["model"]["max_seq_len"],
)
ComputeTDErrorMixin.__init__(self)
# TODO: Don't require users to call this manually.
self._initialize_loss_from_dummy_batch()
TargetNetworkMixin.__init__(self)
def __init__(self, observation_space, action_space, config):
config = dict(ray.rllib.algorithms.pg.PGConfig().to_dict(), **config)
TorchPolicyV2.__init__(
self,
observation_space,
action_space,
config,
max_seq_len=config["model"]["max_seq_len"],
)
# TODO: Don't require users to call this manually.
self._initialize_loss_from_dummy_batch()
def __init__(self, observation_space, action_space, config):
config = dict(ray.rllib.algorithms.maml.maml.DEFAULT_CONFIG, **config)
validate_config(config)
TorchPolicyV2.__init__(
self,
observation_space,
action_space,
config,
max_seq_len=config["model"]["max_seq_len"],
)
KLCoeffMixin.__init__(self, config)
ValueNetworkMixin.__init__(self, config)
# TODO: Don't require users to call this manually.
self._initialize_loss_from_dummy_batch()
def __init__(
self,
observation_space: gym.spaces.Space,
action_space: gym.spaces.Space,
config: TrainerConfigDict,
):
self.target_model = None # assign it in self.make_model
self._is_action_discrete = isinstance(action_space, gym.spaces.Discrete)
TorchPolicyV2.__init__(
self,
observation_space,
action_space,
config,
max_seq_len=config["model"]["max_seq_len"],
)
"""
def __init__(self, observation_space, action_space, config):
config = dict(ray.rllib.agents.a3c.a3c.A3CConfig().to_dict(), **config)
TorchPolicyV2.__init__(
self,
observation_space,
action_space,
config,
max_seq_len=config["model"]["max_seq_len"],
)
ValueNetworkMixin.__init__(self, config)
LearningRateSchedule.__init__(self, config["lr"],
config["lr_schedule"])
EntropyCoeffSchedule.__init__(self, config["entropy_coeff"],
config["entropy_coeff_schedule"])
# TODO: Don't require users to call this manually.
self._initialize_loss_from_dummy_batch()
def __init__(self, observation_space, action_space, config):
config = dict(ray.rllib.agents.impala.impala.DEFAULT_CONFIG, **config)
VTraceOptimizer.__init__(self)
# Need to initialize learning rate variable before calling
# TorchPolicyV2.__init__.
LearningRateSchedule.__init__(self, config["lr"],
config["lr_schedule"])
EntropyCoeffSchedule.__init__(self, config["entropy_coeff"],
config["entropy_coeff_schedule"])
TorchPolicyV2.__init__(
self,
observation_space,
action_space,
config,
max_seq_len=config["model"]["max_seq_len"],
)
# TODO: Don't require users to call this manually.
self._initialize_loss_from_dummy_batch()
def __init__(self, observation_space, action_space, config):
config = dict(ray.rllib.algorithms.ppo.ppo.PPOConfig().to_dict(),
**config)
# TODO: Move into Policy API, if needed at all here. Why not move this into
# `PPOConfig`?.
validate_config(config)
TorchPolicyV2.__init__(
self,
observation_space,
action_space,
config,
max_seq_len=config["model"]["max_seq_len"],
)
ValueNetworkMixin.__init__(self, config)
LearningRateSchedule.__init__(self, config["lr"],
config["lr_schedule"])
EntropyCoeffSchedule.__init__(self, config["entropy_coeff"],
config["entropy_coeff_schedule"])
KLCoeffMixin.__init__(self, config)
# TODO: Don't require users to call this manually.
self._initialize_loss_from_dummy_batch()
def set_weights(self: TorchPolicyV2, weights):
# Makes sure that whenever we restore weights for this policy's
# model, we sync the target network (from the main model)
# at the same time.
TorchPolicyV2.set_weights(self, weights)
self.update_target()