def __init__(self, input_channels=4, activation_str='relu'):
super().__init__()
gain = nn.init.calculate_gain(activation_str)
activation = activations_dict[activation_str]
# (input_channels, 42, 42) -> (32, 4, 4)
self.conv = nn.Sequential(
init(nn.Conv2d(input_channels, 32, kernel_size=3, stride=2), gain=gain), activation(),
init(nn.Conv2d(32, 32, kernel_size=3, stride=2), gain=gain), activation(),
init(nn.Conv2d(32, 32, kernel_size=3, stride=2), gain=gain), activation()
)
def __init__(self, input_size, hidden_size):
super().__init__()
gain = nn.init.calculate_gain('leaky_relu', 0.2)
self.net = nn.Sequential(
init(nn.Linear(input_size, hidden_size), gain=gain),
nn.LeakyReLU(0.2),
init(nn.Linear(hidden_size, hidden_size), gain=gain),
nn.LeakyReLU(0.2),
init(nn.Linear(hidden_size, hidden_size), gain=gain),
nn.LeakyReLU(0.2),
init(nn.Linear(hidden_size, input_size), gain=0.1))
def __init__(
self,
input_size, action_size,
critic_size=1,
detach_actor=False, detach_critic=False
):
super().__init__()
self.actor = init(nn.Linear(input_size, action_size), gain=0.01)
self.critic = init(nn.Linear(input_size, critic_size))
self.detach_actor = detach_actor
self.detach_critic = detach_critic
def __init__(self, input_channels=4, activation_str='relu'):
super().__init__()
gain = nn.init.calculate_gain(activation_str)
activation = activations_dict[activation_str]
# (input_channels, 96, 96) -> (256, 1, 1)
self.conv = nn.Sequential(
init(nn.Conv2d(input_channels, 8, kernel_size=4, stride=2), gain=gain), activation(),
init(nn.Conv2d(8, 16, kernel_size=3, stride=2), gain=gain), activation(),
init(nn.Conv2d(16, 32, kernel_size=3, stride=2), gain=gain), activation(),
init(nn.Conv2d(32, 64, kernel_size=3, stride=2), gain=gain), activation(),
init(nn.Conv2d(64, 128, kernel_size=3, stride=1), gain=gain), activation(),
init(nn.Conv2d(128, 256, kernel_size=3, stride=1), gain=gain), activation()
)