def __init__(self,
state_shape,
action_size,
hiddens,
layer_fn,
activation_fn=nn.ReLU,
norm_fn=None,
bias=True,
out_activation=nn.Tanh):
super().__init__()
# hack to prevent cycle imports
from catalyst.modules.modules import name2nn
layer_fn = name2nn(layer_fn)
activation_fn = name2nn(activation_fn)
norm_fn = name2nn(norm_fn)
out_activation = name2nn(out_activation)
state_size = reduce(lambda x, y: x * y, state_shape)
self.feature_net = SequentialNet(hiddens=[state_size] + hiddens,
layer_fn=layer_fn,
activation_fn=activation_fn,
norm_fn=norm_fn,
bias=bias)
self.policy_net = SequentialNet(hiddens=[hiddens[-1], action_size],
layer_fn=nn.Linear,
activation_fn=out_activation,
norm_fn=None,
bias=True)
inner_init = create_optimal_inner_init(nonlinearity=activation_fn)
self.feature_net.apply(inner_init)
self.policy_net.apply(out_init)
def __init__(self,
state_shape,
action_size,
hiddens,
layer_fn,
concat_at=1,
n_atoms=1,
activation_fn=nn.ReLU,
norm_fn=None,
bias=True,
out_activation=None):
super().__init__()
# hack to prevent cycle imports
from catalyst.modules.modules import name2nn
layer_fn = name2nn(layer_fn)
activation_fn = name2nn(activation_fn)
norm_fn = name2nn(norm_fn)
out_activation = name2nn(out_activation)
self.n_atoms = n_atoms
state_size = reduce(lambda x, y: x * y, state_shape)
if concat_at > 0:
hiddens_ = [state_size] + hiddens[0:concat_at]
self.observation_net = SequentialNet(hiddens=hiddens_,
layer_fn=layer_fn,
activation_fn=activation_fn,
norm_fn=norm_fn,
bias=bias)
hiddens_ = \
[hiddens[concat_at - 1] + action_size] + hiddens[concat_at:]
self.feature_net = SequentialNet(hiddens=hiddens_,
layer_fn=layer_fn,
activation_fn=activation_fn,
norm_fn=norm_fn,
bias=bias)
else:
self.observation_net = None
hiddens_ = [state_size + action_size] + hiddens
self.feature_net = SequentialNet(hiddens=hiddens_,
layer_fn=layer_fn,
activation_fn=activation_fn,
norm_fn=norm_fn,
bias=bias)
self.value_net = SequentialNet(hiddens=[hiddens[-1], n_atoms],
layer_fn=nn.Linear,
activation_fn=out_activation,
norm_fn=None,
bias=True)
inner_init = create_optimal_inner_init(nonlinearity=activation_fn)
if self.observation_net is not None:
self.observation_net.apply(inner_init)
self.feature_net.apply(inner_init)
self.value_net.apply(out_init)
def __init__(self,
state_shape,
action_size,
hiddens,
layer_fn,
activation_fn=nn.ReLU,
norm_fn=None,
bias=True,
out_activation=nn.Sigmoid):
super().__init__()
# hack to prevent cycle imports
from catalyst.modules.modules import name2nn
self.n_action = action_size
layer_fn = name2nn(layer_fn)
activation_fn = name2nn(activation_fn)
norm_fn = name2nn(norm_fn)
out_activation = name2nn(out_activation)
state_size = reduce(lambda x, y: x * y, state_shape)
self.feature_net = SequentialNet(hiddens=[state_size] + hiddens,
layer_fn=layer_fn,
activation_fn=activation_fn,
norm_fn=norm_fn,
bias=bias)
self.embedding_net = SequentialNet(
hiddens=[hiddens[-1], action_size * 2],
layer_fn=layer_fn,
activation_fn=None,
norm_fn=norm_fn,
bias=bias)
self.coupling1 = CouplingLayer(action_size=action_size,
layer_fn=layer_fn,
activation_fn=activation_fn,
norm_fn=None,
bias=bias,
parity="odd")
self.coupling2 = CouplingLayer(action_size=action_size,
layer_fn=layer_fn,
activation_fn=activation_fn,
norm_fn=None,
bias=bias,
parity="even")
self.squasher = SquashingLayer(out_activation)
inner_init = create_optimal_inner_init(nonlinearity=activation_fn)
self.feature_net.apply(inner_init)
self.embedding_net.apply(inner_init)
def __init__(self,
state_shape,
action_size,
hiddens,
layer_fn,
activation_fn=nn.ReLU,
norm_fn=None,
bias=True,
out_activation=nn.Tanh):
super().__init__()
# hack to prevent cycle imports
from catalyst.modules.modules import name2nn
layer_fn = name2nn(layer_fn)
activation_fn = name2nn(activation_fn)
norm_fn = name2nn(norm_fn)
out_activation = name2nn(out_activation)
state_size = state_shape[-1]
self.feature_net = SequentialNet(hiddens=[state_size] + hiddens,
layer_fn=layer_fn,
activation_fn=activation_fn,
norm_fn=norm_fn,
bias=bias)
self.attn = nn.Sequential(
nn.Conv1d(in_channels=hiddens[-1],
out_channels=1,
kernel_size=1,
bias=True), nn.Softmax(dim=1))
self.feature_net2 = SequentialNet(
hiddens=[hiddens[-1] * 4, hiddens[-1]],
layer_fn=layer_fn,
activation_fn=activation_fn,
norm_fn=norm_fn,
bias=bias)
self.policy_net = SequentialNet(hiddens=[hiddens[-1], action_size],
layer_fn=nn.Linear,
activation_fn=out_activation,
norm_fn=None,
bias=True)
inner_init = create_optimal_inner_init(nonlinearity=activation_fn)
self.feature_net.apply(inner_init)
self.attn.apply(out_init)
self.feature_net2.apply(inner_init)
self.policy_net.apply(out_init)
def __init__(self,
action_size,
layer_fn,
activation_fn=nn.ReLU,
norm_fn=None,
bias=True,
parity="odd"):
""" Conditional affine coupling layer used in Real NVP Bijector.
Original paper: https://arxiv.org/abs/1605.08803
Adaptation to RL: https://arxiv.org/abs/1804.02808
Important notes
---------------
1. State embeddings are supposed to have size (action_size * 2).
2. Scale and translation networks used in the Real NVP Bijector
both have one hidden layer of (action_size) (activation_fn) units.
3. Parity ("odd" or "even") determines which part of the input
is being copied and which is being transformed.
"""
super().__init__()
# hack to prevent cycle imports
from catalyst.modules.modules import name2nn
layer_fn = name2nn(layer_fn)
activation_fn = name2nn(activation_fn)
norm_fn = name2nn(norm_fn)
self.parity = parity
if self.parity == "odd":
self.copy_size = action_size // 2
else:
self.copy_size = action_size - action_size // 2
self.scale_prenet = SequentialNet(
hiddens=[action_size * 2 + self.copy_size, action_size],
layer_fn=layer_fn,
activation_fn=activation_fn,
norm_fn=None,
bias=bias)
self.scale_net = SequentialNet(
hiddens=[action_size, action_size - self.copy_size],
layer_fn=layer_fn,
activation_fn=None,
norm_fn=None,
bias=True)
self.translation_prenet = SequentialNet(
hiddens=[action_size * 2 + self.copy_size, action_size],
layer_fn=layer_fn,
activation_fn=activation_fn,
norm_fn=None,
bias=bias)
self.translation_net = SequentialNet(
hiddens=[action_size, action_size - self.copy_size],
layer_fn=layer_fn,
activation_fn=None,
norm_fn=None,
bias=True)
inner_init = create_optimal_inner_init(nonlinearity=activation_fn)
self.scale_prenet.apply(inner_init)
self.scale_net.apply(out_init)
self.translation_prenet.apply(inner_init)
self.translation_net.apply(out_init)