communications

attention_comm

class xuance.torch.communications.attention_comm.TarMAC(*args: Any, **kwargs: Any)

Bases: IC3NetComm

attention_block(q, k, v)

forward(obs: torch.Tensor, msg_send: dict, alive_ally: dict, gate_control: dict = None, agent_key: str | None = None) → torch.Tensor

base_comm

class xuance.torch.communications.base_comm.BaseComm(*args: Any, **kwargs: Any)

Bases: Module

forward(hidden_features: torch.Tensor)

class xuance.torch.communications.base_comm.NoneComm(*args: Any, **kwargs: Any)

Bases: Module

forward(msg: torch.Tensor, **kwargs)

comm_net

class xuance.torch.communications.comm_net.CommNet(*args: Any, **kwargs: Any)

Bases: Module

create_mlp(input_shape: int, layers: list, out_shape: int, activation: torch.nn.Module, device: str | torch.device) → torch.nn.Sequential

forward(obs: torch.Tensor, msg_send: dict, alive_ally: dict) → torch.Tensor

message_encode(message: torch.Tensor) → torch.Tensor

obs_encode(observation)

emergent_comm

gnn_comm

class xuance.torch.communications.gnn_comm.DGNComm(*args: Any, **kwargs: Any)

Bases: CommNet

build_gcn(num_layers, num_heads, activation=torch.nn.ReLU)

create_adjacency_matrix(alive_ally)

forward(obs: torch.Tensor, msg_send: dict, alive_ally: dict) → torch.Tensor

gcn_block(x, adj_matrix)

class xuance.torch.communications.gnn_comm.GraphMultiHeadAttentionLayer(*args: Any, **kwargs: Any)

Bases: Module

forward(x, adj)

x: node feature matrix [batch_size, n_agents, input_dim] adj: adjacency matrix [batch_size, n_agents, n_agents]

reset_parameters()

ic3net_comm

class xuance.torch.communications.ic3net_comm.IC3NetComm(*args: Any, **kwargs: Any)

Bases: CommNet

forward(obs: torch.Tensor, msg_send: dict, alive_ally: dict, gate_control: dict = None)