def __init__(
self,
*,
input_size: int,
filters: Tuple[Tuple[int]] = (
(1024, 5, 2),
(128, 5, 2),
(64, 6, 2),
(32, 6, 2),
),
initializer="default",
bias_init=0,
activation_fn: str = "relu",
output_shape: Tuple[int] = (3, 64, 64)
):
"""Initializes a TransposedConv2DStack instance.
Args:
input_size: The size of the 1D input vector, from which to
generate the image distribution.
filters (Tuple[Tuple[int]]): Tuple of filter setups (1 for each
ConvTranspose2D layer): [in_channels, kernel, stride].
initializer (Union[str]):
bias_init: The initial bias values to use.
activation_fn: Activation function descriptor (str).
output_shape (Tuple[int]): Shape of the final output image.
"""
super().__init__()
self.activation = get_activation_fn(activation_fn, framework="torch")
self.output_shape = output_shape
initializer = get_initializer(initializer, framework="torch")
in_channels = filters[0][0]
self.layers = [
# Map from 1D-input vector to correct initial size for the
# Conv2DTransposed stack.
nn.Linear(input_size, in_channels),
# Reshape from the incoming 1D vector (input_size) to 1x1 image
# format (channels first).
Reshape([-1, in_channels, 1, 1]),
]
for i, (_, kernel, stride) in enumerate(filters):
out_channels = (
filters[i + 1][0] if i < len(filters) - 1 else output_shape[0]
)
conv_transp = nn.ConvTranspose2d(in_channels, out_channels, kernel, stride)
# Apply initializer.
initializer(conv_transp.weight)
nn.init.constant_(conv_transp.bias, bias_init)
self.layers.append(conv_transp)
# Apply activation function, if provided and if not last layer.
if self.activation is not None and i < len(filters) - 1:
self.layers.append(self.activation())
# num-outputs == num-inputs for next layer.
in_channels = out_channels
self._model = nn.Sequential(*self.layers)
def __init__(
self,
policy: Policy,
gamma: float,
model: ModelConfigDict = None,
n_iters: int = 160,
lr: float = 1e-3,
delta: float = 1e-4,
clip_grad_norm: float = 100.0,
batch_size: int = 32,
) -> None:
"""
Args:
policy: Policy to evaluate.
gamma: Discount factor of the environment.
# The ModelConfigDict for self.q_model
model = {
"fcnet_hiddens": [8, 8],
"fcnet_activation": "relu",
"vf_share_layers": True,
},
# Maximum number of training iterations to run on the batch
n_iters = 160,
# Learning rate for Q-function optimizer
lr = 1e-3,
# Early stopping if the mean loss < delta
delta = 1e-4,
# Clip gradients to this maximum value
clip_grad_norm = 100.0,
# Minibatch size for training Q-function
batch_size = 32,
"""
self.policy = policy
self.gamma = gamma
self.observation_space = policy.observation_space
self.action_space = policy.action_space
if model is None:
model = {
"fcnet_hiddens": [8, 8],
"fcnet_activation": "relu",
"vf_share_layers": True,
}
self.device = self.policy.device
self.q_model: TorchModelV2 = ModelCatalog.get_model_v2(
self.observation_space,
self.action_space,
self.action_space.n,
model,
framework="torch",
name="TorchQModel",
).to(self.device)
self.n_iters = n_iters
self.lr = lr
self.delta = delta
self.clip_grad_norm = clip_grad_norm
self.batch_size = batch_size
self.optimizer = torch.optim.Adam(self.q_model.variables(), self.lr)
initializer = get_initializer("xavier_uniform", framework="torch")
def f(m):
if isinstance(m, nn.Linear):
initializer(m.weight)
self.initializer = f
def __init__(
self,
policy: Policy,
gamma: float,
model: ModelConfigDict = None,
n_iters: int = 1,
lr: float = 1e-3,
delta: float = 1e-4,
clip_grad_norm: float = 100.0,
minibatch_size: int = None,
tau: float = 1.0,
) -> None:
"""
Args:
policy: Policy to evaluate.
gamma: Discount factor of the environment.
model: The ModelConfigDict for self.q_model, defaults to:
{
"fcnet_hiddens": [8, 8],
"fcnet_activation": "relu",
"vf_share_layers": True,
},
n_iters: Number of gradient steps to run on batch, defaults to 1
lr: Learning rate for Q-model optimizer
delta: Early stopping threshold if the mean loss < delta
clip_grad_norm: Clip gradients to this maximum value
minibatch_size: Minibatch size for training Q-function;
if None, train on the whole batch
tau: Polyak averaging factor for target Q-function
"""
self.policy = policy
assert isinstance(
policy.action_space, Discrete
), f"{self.__class__.__name__} only supports discrete action spaces!"
self.gamma = gamma
self.observation_space = policy.observation_space
self.action_space = policy.action_space
if model is None:
model = {
"fcnet_hiddens": [8, 8],
"fcnet_activation": "relu",
"vf_share_layers": True,
}
self.device = self.policy.device
self.q_model: TorchModelV2 = ModelCatalog.get_model_v2(
self.observation_space,
self.action_space,
self.action_space.n,
model,
framework="torch",
name="TorchQModel",
).to(self.device)
self.target_q_model: TorchModelV2 = ModelCatalog.get_model_v2(
self.observation_space,
self.action_space,
self.action_space.n,
model,
framework="torch",
name="TargetTorchQModel",
).to(self.device)
self.n_iters = n_iters
self.lr = lr
self.delta = delta
self.clip_grad_norm = clip_grad_norm
self.minibatch_size = minibatch_size
self.tau = tau
self.optimizer = torch.optim.Adam(self.q_model.variables(), self.lr)
initializer = get_initializer("xavier_uniform", framework="torch")
# Hard update target
self.update_target(tau=1.0)
def f(m):
if isinstance(m, nn.Linear):
initializer(m.weight)
self.initializer = f