def __init__(self,
action_dim: int,
history_length: int = 4,
fc_layers: Tuple = (256, )):
super(DuelingDQNValueCNN, self).__init__()
self.action_dim = action_dim
self.conv, output_size = cnn((history_length, 16, 32))
self.advantage = mlp([output_size] + list(fc_layers) + [action_dim])
self.value = mlp([output_size] + list(fc_layers) + [1])
def __init__(
self,
action_dim: int,
history_length: int = 4,
fc_layers: Tuple = (128, ),
noisy_layers: Tuple = (128, 128),
):
super(NoisyDQNValueCNN, self).__init__()
self.conv, output_size = cnn((history_length, 16, 32))
self.model = noisy_mlp([output_size] + list(fc_layers),
list(noisy_layers) + [action_dim])
def __init__(self,
action_dim,
history_length=4,
val_type="Qs",
fc_layers=(256, )):
super(CNNValue, self).__init__()
self.action_dim = action_dim
self.conv, output_size = cnn((history_length, 16, 32))
self.fc = _get_val_model(mlp, val_type, output_size, fc_layers,
action_dim)
def __init__(self,
framestack: int,
action_dim: spaces.Space,
fc_layers: Tuple = (256, ),
val_type: str = "V",
discrete: bool = True,
*args,
**kwargs):
super(CNNActorCritic, self).__init__()
self.feature, output_size = cnn((framestack, 16, 32))
self.actor = MlpPolicy(output_size, action_dim, fc_layers, discrete,
**kwargs)
self.critic = MlpValue(output_size, action_dim, val_type, fc_layers)
def __init__(self,
framestack: int,
action_dim: int,
hidden: Tuple = (32, 32),
discrete: bool = True,
*args,
**kwargs):
super(CNNPolicy, self).__init__(framestack, action_dim, hidden,
discrete, **kwargs)
self.action_dim = action_dim
self.conv, output_size = cnn((framestack, 16, 32))
self.fc = mlp([output_size] + list(hidden) + [action_dim],
sac=self.sac)
def __init__(
self,
action_dim: int,
num_atoms: int,
history_length: int = 4,
fc_layers: Tuple = (128, 128),
noisy_layers: Tuple = (128, 512),
):
super(CategoricalDQNValueCNN, self).__init__()
self.action_dim = action_dim
self.num_atoms = num_atoms
self.conv, output_size = cnn((history_length, 16, 32))
self.model = noisy_mlp(
[output_size] + list(fc_layers),
list(noisy_layers) + [self.action_dim * self.num_atoms],
)
def __init__(
self,
framestack: int,
action_dim: int,
val_type: str = "Qs",
fc_layers: Tuple = (256, ),
**kwargs,
):
super(CNNValue, self).__init__()
self.action_dim = action_dim
activation = kwargs["activation"] if "activation" in kwargs else "relu"
self.conv, output_size = cnn((framestack, 16, 32),
activation=activation)
self.fc = _get_val_model(mlp, val_type, output_size, fc_layers,
action_dim)
def __init__(self, input_dim, action_dim, hidden_dims=[128, 64]):
super(Discriminator, self).__init__()
self.conv, self.conv_output_size = cnn()
self.fc1 = nn.Linear(self.conv_output_size, hidden_dims[0])
self.fc2 = nn.Linear(hidden_dims[0] + action_dim, hidden_dims[1])
self.fc3 = nn.Linear(hidden_dims[1], 1)
