Source code for xuance.torch.policies.coordination_graph

import torch
import numpy as np
from xuance.common import Optional, Union
from xuance.torch import nn, Tensor
try:
    import torch_scatter
except ImportError:
    print("The module torch_scatter is not installed.")


[docs] class DCG_utility(nn.Module): """ The utility module for deep coordination graph. Args: dim_input (int): The dimension of input for the utility module. dim_hidden (int): The dimension of hidden layer for the utility module. dim_output (int): The dimension of output for the utility module. device (Optional[Union[str, int, torch.device]]): The device for running the model, default is None. """ def __init__(self, dim_input: int, dim_hidden: int, dim_output: int, device: Optional[Union[str, int, torch.device]] = None): super(DCG_utility, self).__init__() self.dim_input = dim_input self.dim_hidden = dim_hidden self.dim_output = dim_output self.device = device '''All utilities share the same parameters''' self.output = nn.Sequential(nn.Linear(self.dim_input, self.dim_hidden), nn.ReLU(), nn.Linear(self.dim_hidden, self.dim_output)).to(device)
[docs] def forward(self, hidden_states_n: Tensor): """ Calculate the utility values for multiple agents. Args: hidden_states_n (Tensor): The hidden states for the representations of n agents. Returns: The utility values for multiple agents. """ return self.output(hidden_states_n)
[docs] class DCG_payoff(DCG_utility): """ The payoff module for deep coordination graph. Args: dim_input (int): The dimension of input for the payoff module. dim_hidden (int): The dimension of hidden layer for the payoff module. dim_act (int): The dimension of actions. low_rank_payoff (int): The low rank payoff. payoff_rank (int): The rank of payoff. device (Optional[Union[str, int, torch.device]]): The device for running the model, default is None. """ def __init__(self, dim_input: int, dim_hidden: int, dim_act: int, low_rank_payoff: int, payoff_rank: int, device: Optional[Union[str, int, torch.device]] = None): self.dim_act = dim_act self.low_rank_payoff = low_rank_payoff self.payoff_rank = payoff_rank dim_payoff_out = 2 * self.payoff_rank * self.dim_act if self.low_rank_payoff else self.dim_act ** 2 super(DCG_payoff, self).__init__(dim_input, dim_hidden, dim_payoff_out, device)
[docs] def forward(self, hidden_states_n: Tensor, edges_from: Tensor=None, edges_to: Tensor=None): """ Calculate the payoff values for the graph constructed by multiple agents. Args: hidden_states_n: The hidden states for the representations of n agents. edges_from: The edges from others to self, default is None. edges_to: The edges from self to others, default is None. Returns: Mean of payoff values for edge_from and edge_to. """ input_payoff = torch.stack([torch.cat([hidden_states_n[:, edges_from], hidden_states_n[:, edges_to]], dim=-1), torch.cat([hidden_states_n[:, edges_to], hidden_states_n[:, edges_from]], dim=-1)], dim=0) payoffs = self.output(input_payoff) dim = payoffs.shape[0:-1] if self.low_rank_payoff: payoffs = payoffs.view(np.prod(dim) * self.payoff_rank, 2, self.dim_act) payoffs = torch.matmul(payoffs[:, 0, :].unsqueeze(dim=-1), payoffs[:, 1, :].unsqueeze( dim=-2)) # (dim_act * 1) * (1 * dim_act) -> (dim_act * dim_act) payoffs = payoffs.view(list(dim) + [self.payoff_rank, self.dim_act, self.dim_act]).sum(dim=-3) else: payoffs = payoffs.view(list(dim) + [self.dim_act, self.dim_act]) payoffs[1] = payoffs[1].transpose(dim0=-1, dim1=-2).clone() # f_ij(a_i, a_j) <-> f_ji(a_j, a_i) return payoffs.mean(dim=0) # f^E_{ij} = (f_ij(a_i, a_j) + f_ji(a_j, a_i)) / 2
[docs] class Coordination_Graph(object): """ Construct a deep coordination graph. Args: n_vertexes (int): The number of vertexes in the graph. graph_type (str): The type of graph, default is "FULL". """ def __init__(self, n_vertexes: int, graph_type: str = "FULL", device: Optional[Union[str, int, torch.device]] = None): self.n_vertexes = n_vertexes self.device = device self.edges = [] if graph_type == "CYCLE": self.edges = [(i, i + 1) for i in range(self.n_vertexes - 1)] + [(self.n_vertexes - 1, 0)] elif graph_type == "LINE": self.edges = [(i, i + 1) for i in range(self.n_vertexes - 1)] elif graph_type == "STAR": self.edges = [(0, i + 1) for i in range(self.n_vertexes - 1)] elif graph_type == "VDN": pass elif graph_type == "FULL": self.edges = [[(j, i + j + 1) for i in range(self.n_vertexes - j - 1)] for j in range(self.n_vertexes - 1)] self.edges = [e for l in self.edges for e in l] else: raise AttributeError("There is no graph type named {}!".format(graph_type)) self.n_edges = len(self.edges) self.edges_from = None self.edges_to = None
[docs] def set_coordination_graph(self): """ Reset the coordination graph. """ self.edges_from = torch.zeros(self.n_edges).long().to(self.device) self.edges_to = torch.zeros(self.n_edges).long().to(self.device) for i, edge in enumerate(self.edges): self.edges_from[i] = edge[0] self.edges_to[i] = edge[1] self.edges_n_in = torch_scatter.scatter_add(src=self.edges_to.new_ones(len(self.edges_to)), index=self.edges_to, dim=0, dim_size=self.n_vertexes) \ + torch_scatter.scatter_add(src=self.edges_to.new_ones(len(self.edges_to)), index=self.edges_from, dim=0, dim_size=self.n_vertexes) self.edges_n_in = self.edges_n_in.float() return