def __init__(self, num_inputs, num_outputs):
super(Bernoulli, self).__init__()
init_ = lambda m: init(m, nn.init.orthogonal_, lambda x: nn.init.
constant_(x, 0))
self.linear = init_(nn.Linear(num_inputs, num_outputs))
def __init__(self, num_inputs, num_outputs):
super(DiagGaussian, self).__init__()
init_ = lambda m: init(m, nn.init.orthogonal_, lambda x: nn.init.
constant_(x, 0))
self.fc_mean = init_(nn.Linear(num_inputs, num_outputs))
self.logstd = AddBias(torch.zeros(num_outputs))
def __init__(self, num_inputs, recurrent=False, hidden_size=512):
super(CNNBase, self).__init__(recurrent, hidden_size, hidden_size)
init_ = lambda m: init(m, nn.init.orthogonal_, lambda x: nn.init.
constant_(x, 0), nn.init.calculate_gain('relu'))
self.main = nn.Sequential(
init_(nn.Conv2d(num_inputs, 32, 8, stride=4)), nn.ReLU(),
init_(nn.Conv2d(32, 64, 4, stride=2)), nn.ReLU(),
init_(nn.Conv2d(64, 32, 3, stride=1)), nn.ReLU(), Flatten(),
init_(nn.Linear(32 * 7 * 7, hidden_size)), nn.ReLU())
init_ = lambda m: init(m, nn.init.orthogonal_, lambda x: nn.init.
constant_(x, 0))
self.critic_linear = init_(nn.Linear(hidden_size, 1))
self.train()
def __init__(self, num_inputs, num_outputs):
super(Categorical, self).__init__()
init_ = lambda m: init(
m,
nn.init.orthogonal_,
lambda x: nn.init.constant_(x, 0),
gain=0.01)
self.linear = init_(nn.Linear(num_inputs, num_outputs))
def __init__(self, obs_shape, action_space):
super(Discriminator, self).__init__()
# atari game
if (action_space.__class__.__name__ == "Discrete") and (len(obs_shape)
== 3):
action_dim = action_space.n
state_dim = obs_shape[0]
relu_init_ = lambda m: init(m, nn.init.orthogonal_, lambda x: nn.
init.constant_(x, 0),
nn.init.calculate_gain('relu'))
init_ = lambda m: init(m, nn.init.orthogonal_, lambda x: nn.init.
constant_(x, 0))
self.main = nn.Sequential(
relu_init_(nn.Conv2d(state_dim, 32, 8, stride=4)), nn.ReLU(),
relu_init_(nn.Conv2d(32, 64, 4, stride=2)), nn.ReLU(),
relu_init_(nn.Conv2d(64, 32, 3, stride=1)), nn.ReLU(),
Flatten(), relu_init_(nn.Linear(32 * 7 * 7, 512)), nn.ReLU(),
init_(nn.Linear(512, action_dim)))
# mujoco
elif (action_space.__class__.__name__ == "Box") and (len(obs_shape)
== 1):
action_dim = action_space.shape[0]
state_dim = obs_shape[0]
init_ = lambda m: init(m, nn.init.orthogonal_, lambda x: nn.init.
constant_(x, 0),
nn.init.calculate_gain('relu'))
self.main = nn.Sequential(
init_(nn.Linear(state_dim + action_dim, 64)), nn.ReLU(),
init_(nn.Linear(64, 64)), nn.ReLU(), init_(nn.Linear(64, 1)))
# elif action_space.__class__.__name__ == "MultiBinary":
# action_dim = action_space.shape[0]
else:
raise NotImplementedError
self.train()
def __init__(self, obs_shape, action_space):
super(Critic, self).__init__()
if (action_space.__class__.__name__ == "Box") and (len(obs_shape)
== 1):
action_dim = action_space.shape[0]
state_dim = obs_shape[0]
# init_ = lambda m: init(m, nn.init.orthogonal_, lambda x: nn.init.
# constant_(x, 0), nn.init.calculate_gain('tanh'))
init_ = lambda m: init(m, nn.init.orthogonal_, lambda x: nn.init.
constant_(x, 0), np.sqrt(2))
self.main = nn.Sequential(
init_(nn.Linear(state_dim + action_dim, 64)), nn.Tanh(),
init_(nn.Linear(64, 64)), nn.Tanh(), init_(nn.Linear(64, 1)))
else:
raise NotImplementedError
self.train()
def __init__(self, obs_shape, action_space):
super(Actor, self).__init__()
if (action_space.__class__.__name__ == "Box") and (len(obs_shape)
== 1):
action_dim = action_space.shape[0]
state_dim = obs_shape[0]
# init_ = lambda m: init(m, nn.init.orthogonal_, lambda x: nn.init.
# constant_(x, 0), nn.init.calculate_gain('tanh'))
init_ = lambda m: init(m, nn.init.orthogonal_, lambda x: nn.init.
constant_(x, 0), np.sqrt(2))
self.base = nn.Sequential(init_(nn.Linear(state_dim, 64)),
nn.Tanh(), init_(nn.Linear(64, 64)),
nn.Tanh())
num_outputs = action_space.shape[0]
self.dist = DiagGaussian(64, num_outputs)
else:
raise NotImplementedError
def __init__(self, num_inputs, recurrent=False, hidden_size=64):
super(MLPBase, self).__init__(recurrent, num_inputs, hidden_size)
if recurrent:
num_inputs = hidden_size
init_ = lambda m: init(m, nn.init.orthogonal_, lambda x: nn.init.
constant_(x, 0), np.sqrt(2))
self.actor = nn.Sequential(init_(nn.Linear(num_inputs, hidden_size)),
nn.Tanh(),
init_(nn.Linear(hidden_size, hidden_size)),
nn.Tanh())
self.critic = nn.Sequential(init_(nn.Linear(num_inputs, hidden_size)),
nn.Tanh(),
init_(nn.Linear(hidden_size, hidden_size)),
nn.Tanh())
self.critic_linear = init_(nn.Linear(hidden_size, 1))
self.train()
