def __init__(self,
state_dim,
action_dim,
gamma=0.99,
noise_std=0.02,
hidden_dim=64,
actor_lr=0.001,
critic_lr=0.001,
verbose=False):
self.gamma = gamma
# self.tau = 0.01
self.tau = 0.001
self.actor = Actor(state_dim,
noise_std=noise_std,
hidden_dim=hidden_dim)
self.actor_target = Actor(state_dim,
noise_std=noise_std,
hidden_dim=hidden_dim)
self.critic = Critic(state_dim,
action_dim,
hidden_dim=hidden_dim)
self.critic_target = Critic(state_dim,
action_dim,
hidden_dim=hidden_dim)
self.actor_target.load_state_dict(self.actor.state_dict())
self.critic_target.load_state_dict(self.critic.state_dict())
self.actor_optimizer = torch.optim.Adam(self.actor.parameters(),
lr=actor_lr)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(),
lr=critic_lr)
self.buffer = ReplayBuffer(max_size=1e5)
self.logging_period = 10 if verbose else 100
# --- ModelIO ---
self.modelio = ModelIO(model_path=Path(__file__).resolve().parent /
'models')
def __init__(self,
state_dim,
action_dim,
gamma=0.99,
hidden_dim=64,
actor_lr=0.001,
critic_lr=0.001,
K_epochs=5,
eps_clip=0.2,
entropy_coeff=0.02,
d2c=None,
verbose=False):
self.gamma = gamma
self.eps_clip = eps_clip
self.K_epochs = K_epochs
self.entropy_coeff = entropy_coeff
self.d2c = d2c
self.verbose = verbose
self.critic = Critic(state_dim, hidden_dim=hidden_dim).to(device)
self.actor = Actor(state_dim, action_dim,
hidden_dim=hidden_dim).to(device)
self.actor_old = Actor(state_dim, action_dim,
hidden_dim=hidden_dim).to(device)
self.actor_optimizer = torch.optim.Adam(self.actor.parameters(),
lr=actor_lr)
self.critic_optimizer = torch.optim.Adam(self.critic.parameters(),
lr=critic_lr)
self.actor_old.load_state_dict(self.actor.state_dict())
self.buffer = Buffer()
# --- ModelIO ---
self.modelio = ModelIO(model_path=Path(__file__).resolve().parent /
'models')
def __init__(self, state_dim, action_dim, gamma, d2c=None):
self._V = StateValueFunction(state_dim)
self._pi = Policy(state_dim, action_dim)
self.d2c = d2c # discrete to continuous actions
# self._V.cuda()
# self._pi.cuda()
self._gamma = gamma
self._loss_function = nn.MSELoss()
self._V_optimizer = optim.Adam(self._V.parameters(), lr=0.001)
self._pi_optimizer = optim.Adam(self._pi.parameters(), lr=0.0001)
self._action_dim = action_dim
# --- ModelIO ---
self._modelio = ModelIO(model_path=Path(__file__).resolve().parent /
'models')
self._baseline_model_name = 'ac_baseline.pt'
self._policy_model_name = 'ac_policy.pt'
def __init__(self, state_dim, action_dim, gamma, d2c=None):
self._q = Q(state_dim, action_dim)
self._q_target = Q(state_dim, action_dim)
# self._q.cuda()
# self._q_target.cuda()
self._gamma = gamma
self._loss_function = nn.MSELoss()
self._q_optimizer = optim.Adam(self._q.parameters(), lr=0.0001)
self._action_dim = action_dim
self._replay_buffer = ReplayBuffer(5000)
self._d2c = d2c
# --- ModelIO ---
self._modelio = ModelIO(model_path=Path(__file__).resolve().parent /
'models')
self._q_model_name = 'q.pt'
self._target_model_name = 'target.pt'
def __init__(self,
state_dim,
action_dim,
gamma=0.99,
hidden_dim=64,
policy_lr=0.001,
baseline_lr=0.001):
self._V = StateValueFunction(state_dim, hidden_dim=hidden_dim)
self._pi = Policy(state_dim, action_dim, hidden_dim=hidden_dim)
# self._V.cuda()
# self._pi.cuda()
self._gamma = gamma
self._loss_function = nn.MSELoss()
self._V_optimizer = optim.Adam(self._V.parameters(), lr=baseline_lr)
self._pi_optimizer = optim.Adam(self._pi.parameters(), lr=policy_lr)
self._action_dim = action_dim
# --- ModelIO ---
self._modelio = ModelIO(model_path=Path(__file__).resolve().parent /
'models')
