def __init__(self, num_inputs, action_space, args):
self.gamma = args.gamma
self.tau = args.tau
self.alpha = args.alpha
self.policy_type = args.policy
self.target_update_interval = args.target_update_interval
self.automatic_entropy_tuning = args.automatic_entropy_tuning
self.device = torch.device("cuda" if args.cuda else "cpu")
self.critic = QNetwork(num_inputs, action_space.shape[0], args.hidden_size).to(device=self.device)
self.critic_optim = Adam(self.critic.parameters(), lr=args.lr)
self.critic_target = QNetwork(num_inputs, action_space.shape[0], args.hidden_size).to(self.device)
hard_update(self.critic_target, self.critic)
if self.policy_type == "Gaussian":
# Target Entropy = −dim(A) (e.g. , -6 for HalfCheetah-v2) as given in the paper
if self.automatic_entropy_tuning == True:
self.target_entropy = -torch.prod(torch.Tensor(action_space.shape).to(self.device)).item()
self.log_alpha = torch.zeros(1, requires_grad=True, device=self.device)
self.alpha_optim = Adam([self.log_alpha], lr=args.lr)
self.policy = GaussianPolicy(num_inputs, action_space.shape[0], args.hidden_size, action_space).to(self.device)
self.policy_optim = Adam(self.policy.parameters(), lr=args.lr)
else:
self.alpha = 0
self.automatic_entropy_tuning = False
self.policy = DeterministicPolicy(num_inputs, action_space.shape[0], args.hidden_size, action_space).to(self.device)
self.policy_optim = Adam(self.policy.parameters(), lr=args.lr)
def __init__(
self,
env,
train_tasks,
eval_tasks,
latent_dim,
nets,
policy_lr=3e-4,
qf_lr=1e-3,
vf_lr=1e-3,
alpha=1,
automatic_entropy_tuning=True,
lr=3e-4,
context_lr=3e-4,
kl_lambda=1.,
policy_mean_reg_weight=1e-3,
policy_std_reg_weight=1e-3,
policy_pre_activation_weight=0.,
optimizer_class=optim.Adam,
recurrent=False,
use_information_bottleneck=True,
sparse_rewards=False,
#soft_target_tau=1e-2,
soft_target_tau=0.005,
plotter=None,
render_eval_paths=False,
**kwargs):
super().__init__(env=env,
agent=nets[0],
train_tasks=train_tasks,
eval_tasks=eval_tasks,
**kwargs)
self.soft_target_tau = soft_target_tau
self.policy_mean_reg_weight = policy_mean_reg_weight
self.policy_std_reg_weight = policy_std_reg_weight
self.policy_pre_activation_weight = policy_pre_activation_weight
self.plotter = plotter
self.render_eval_paths = render_eval_paths
self.alpha = alpha
self.automatic_entropy_tuning = automatic_entropy_tuning
self.recurrent = recurrent
self.latent_dim = latent_dim
# self.qf_criterion = nn.MSELoss()
# self.vf_criterion = nn.MSELoss()
self.vib_criterion = nn.MSELoss()
self.l2_reg_criterion = nn.MSELoss()
self.kl_lambda = kl_lambda
self.use_information_bottleneck = use_information_bottleneck
self.sparse_rewards = sparse_rewards
# self.qf1, self.qf2, self.vf = nets[1:]
self.critic, self.critic_target = nets[1:] #q1,q2,target q1,target q2
# self.target_vf = self.vf.copy()
self.critic_optimizer = Adam(self.critic.parameters(), lr=lr)
hard_update(self.critic_target, self.critic)
self.policy_optimizer = Adam(self.agent.policy.parameters(),
lr=policy_lr)
self.target_entropy = -torch.prod(torch.Tensor(
env.action_space.shape)).to(
"cuda").item() # torch.prod(input) : 返回所有元素的乘积
self.log_alpha = torch.zeros(1, requires_grad=True, device="cuda")
self.alpha_optim = Adam([self.log_alpha], lr=lr)
# self.policy_optimizer = optimizer_class(
# self.agent.policy.parameters(),
# lr=policy_lr,
# )
# self.qf1_optimizer = optimizer_class(
# self.qf1.parameters(),
# lr=qf_lr,
# )
# self.qf2_optimizer = optimizer_class(
# self.qf2.parameters(),
# lr=qf_lr,
# )
# self.vf_optimizer = optimizer_class(
# self.vf.parameters(),
# lr=vf_lr,
# )
self.context_optimizer = optimizer_class(
self.agent.context_encoder.parameters(),
lr=context_lr,
)
def __init__(self,
action_space,
policy: str = "Gaussian",
gamma: float = 0.99,
tau: float = 0.005,
lr: float = 0.0003,
alpha: float = 0.2,
automatic_temperature_tuning: bool = False,
batch_size: int = 256,
hidden_size: int = 256,
target_update_interval: int = 1,
input_dim: int = 32):
self.gamma = gamma
self.tau = tau
self.alpha = alpha
self.lr = lr
self.policy_type = policy
self.target_update_interval = target_update_interval
self.automatic_temperature_tuning = automatic_temperature_tuning
self.input_dim = input_dim
self.hidden_size = hidden_size
self.bs = batch_size
self.critic = QNetwork(input_dim, action_space.shape[0],
hidden_size).to(DEVICE)
self.critic_optim = Adam(self.critic.parameters(), lr=self.lr)
self.critic_target = QNetwork(input_dim, action_space.shape[0],
hidden_size).to(DEVICE)
hard_update(self.critic_target, self.critic)
if self.policy_type == "Gaussian":
if self.automatic_temperature_tuning == True:
self.target_entropy = -torch.prod(
torch.Tensor(action_space.shape).to(DEVICE)).item()
self.log_alpha = torch.zeros(1,
requires_grad=True,
device=DEVICE)
self.alpha_optim = Adam([self.log_alpha], lr=self.lr)
self.policy = GaussianPolicy(input_dim, action_space.shape[0],
hidden_size).to(DEVICE)
self.policy_optim = Adam(self.policy.parameters(), lr=self.lr)
else:
self.alpha = 0
self.automatic_temperature_tuning = False
self.policy = DeterministicPolicy(input_dim, action_space.shape[0],
hidden_size).to(DEVICE)
self.policy_optim = Adam(self.policy.parameters(), lr=self.lr)
settings = (
f"INITIALIZING SAC ALGORITHM WITH {self.policy_type} POLICY"
f"\nRunning on: {DEVICE}"
f"\nSettings: Automatic Temperature tuning = {self.automatic_temperature_tuning}, Update Interval = {self.target_update_interval}"
f"\nParameters: Learning rate = {self.lr}, Batch Size = {self.bs} Gamma = {self.gamma}, Tau = {self.tau}, Alpha = {self.alpha}"
f"\nArchitecture: Input dimension = {self.input_dim}, Hidden layer dimension = {self.hidden_size}"
"\n--------------------------")
print(settings)