from typing import Union, List import torch import torch.nn as nn import torch.nn.functional as F from functools import reduce from ding.utils import list_split, MODEL_REGISTRY from ding.torch_utils.network.nn_module import fc_block, MLP from ding.torch_utils.network.transformer import ScaledDotProductAttention from .q_learning import DRQN from ding.model.template.qmix import Mixer class MixerStar(nn.Module): """ Overview: Mixer network for Q_star in WQMIX(https://arxiv.org/abs/2006.10800), which mix up the independent q_value of \ each agent to a total q_value and is diffrent from the QMIX's mixer network, \ here the mixing network is a feedforward network with 3 hidden layers of 256 dim. \ This Q_star mixing network is not constrained to be monotonic by using non-negative weights and \ having the state and agent_q be inputs, as opposed to having hypernetworks take the state as input \ and generate the weights in QMIX. Interface: ``__init__``, ``forward``. """ def __init__(self, agent_num: int, state_dim: int, mixing_embed_dim: int) -> None: """ Overview: Initialize the mixer network of Q_star in WQMIX. Arguments: - agent_num (:obj:`int`): The number of agent, e.g., 8. - state_dim(:obj:`int`): The dimension of global observation state, e.g., 16. - mixing_embed_dim (:obj:`int`): The dimension of mixing state emdedding, e.g., 128. """ super(MixerStar, self).__init__() self.agent_num = agent_num self.state_dim = state_dim self.embed_dim = mixing_embed_dim self.input_dim = self.agent_num + self.state_dim # shape N+A non_lin = nn.ReLU() self.net = nn.Sequential( nn.Linear(self.input_dim, self.embed_dim), non_lin, nn.Linear(self.embed_dim, self.embed_dim), non_lin, nn.Linear(self.embed_dim, self.embed_dim), non_lin, nn.Linear(self.embed_dim, 1) ) # V(s) instead of a bias for the last layers self.V = nn.Sequential(nn.Linear(self.state_dim, self.embed_dim), non_lin, nn.Linear(self.embed_dim, 1)) def forward(self, agent_qs: torch.FloatTensor, states: torch.FloatTensor) -> torch.FloatTensor: """ Overview: Forward computation graph of the mixer network for Q_star in WQMIX. This mixer network for \ is a feed-forward network that takes the state and the appropriate actions' utilities as input. Arguments: - agent_qs (:obj:`torch.FloatTensor`): The independent q_value of each agent. - states (:obj:`torch.FloatTensor`): The emdedding vector of global state. Returns: - q_tot (:obj:`torch.FloatTensor`): The total mixed q_value. Shapes: - agent_qs (:obj:`torch.FloatTensor`): :math:`(T,B, N)`, where T is timestep, \ B is batch size, A is agent_num, N is obs_shape. - states (:obj:`torch.FloatTensor`): :math:`(T, B, M)`, where M is global_obs_shape. - q_tot (:obj:`torch.FloatTensor`): :math:`(T, B, )`. """ # in below annotations about the shape of the variables, T is timestep, # B is batch_size A is agent_num, N is obs_shape, for example, # in 3s5z, we can set T=10, B=32, A=8, N=216 bs = agent_qs.shape[:-1] # (T*B, A) states = states.reshape(-1, self.state_dim) # T*B, N), agent_qs = agent_qs.reshape(-1, self.agent_num) # (T, B, A) -> (T*B, A) inputs = torch.cat([states, agent_qs], dim=1) # (T*B, N) (T*B, A)-> (T*B, N+A) advs = self.net(inputs) # (T*B, 1) vs = self.V(states) # (T*B, 1) y = advs + vs q_tot = y.view(*bs) # (T*B, 1) -> (T, B) return q_tot @MODEL_REGISTRY.register('wqmix') class WQMix(nn.Module): """ Overview: WQMIX (https://arxiv.org/abs/2006.10800) network, There are two components: \ 1) Q_tot, which is same as QMIX network and composed of agent Q network and mixer network. \ 2) An unrestricted joint action Q_star, which is composed of agent Q network and mixer_star network. \ The QMIX paper mentions that all agents share local Q network parameters, so only one Q network is initialized \ in Q_tot or Q_star. Interface: ``__init__``, ``forward``. """ def __init__( self, agent_num: int, obs_shape: int, global_obs_shape: int, action_shape: int, hidden_size_list: list, lstm_type: str = 'gru', dueling: bool = False ) -> None: """ Overview: Initialize WQMIX neural network according to arguments, i.e. agent Q network and mixer, \ Q_star network and mixer_star. Arguments: - agent_num (:obj:`int`): The number of agent, such as 8. - obs_shape (:obj:`int`): The dimension of each agent's observation state, such as 8. - global_obs_shape (:obj:`int`): The dimension of global observation state, such as 8. - action_shape (:obj:`int`): The dimension of action shape, such as 6. - hidden_size_list (:obj:`list`): The list of hidden size for ``q_network``, \ the last element must match mixer's ``mixing_embed_dim``. - lstm_type (:obj:`str`): The type of RNN module in ``q_network``, now support \ ['normal', 'pytorch', 'gru'], default to gru. - dueling (:obj:`bool`): Whether choose ``DuelingHead`` (True) or ``DiscreteHead (False)``, \ default to False. """ super(WQMix, self).__init__() self._act = nn.ReLU() self._q_network = DRQN(obs_shape, action_shape, hidden_size_list, lstm_type=lstm_type, dueling=dueling) self._q_network_star = DRQN(obs_shape, action_shape, hidden_size_list, lstm_type=lstm_type, dueling=dueling) embedding_size = hidden_size_list[-1] self._mixer = Mixer(agent_num, global_obs_shape, mixing_embed_dim=embedding_size) self._mixer_star = MixerStar( agent_num, global_obs_shape, mixing_embed_dim=256 ) # the mixing network of Q_star is a feedforward network with 3 hidden layers of 256 dim self._global_state_encoder = nn.Identity() # nn.Sequential() def forward(self, data: dict, single_step: bool = True, q_star: bool = False) -> dict: """ Overview: Forward computation graph of qmix network. Input dict including time series observation and \ related data to predict total q_value and each agent q_value. Determine whether to calculate \ Q_tot or Q_star based on the ``q_star`` parameter. Arguments: - data (:obj:`dict`): Input data dict with keys ['obs', 'prev_state', 'action']. - agent_state (:obj:`torch.Tensor`): Time series local observation data of each agents. - global_state (:obj:`torch.Tensor`): Time series global observation data. - prev_state (:obj:`list`): Previous rnn state for ``q_network`` or ``_q_network_star``. - action (:obj:`torch.Tensor` or None): If action is None, use argmax q_value index as action to\ calculate ``agent_q_act``. - single_step (:obj:`bool`): Whether single_step forward, if so, add timestep dim before forward and\ remove it after forward. - Q_star (:obj:`bool`): Whether Q_star network forward. If True, using the Q_star network, where the\ agent networks have the same architecture as Q network but do not share parameters and the mixing\ network is a feedforward network with 3 hidden layers of 256 dim; if False, using the Q network,\ same as the Q network in Qmix paper. Returns: - ret (:obj:`dict`): Output data dict with keys [``total_q``, ``logit``, ``next_state``]. - total_q (:obj:`torch.Tensor`): Total q_value, which is the result of mixer network. - agent_q (:obj:`torch.Tensor`): Each agent q_value. - next_state (:obj:`list`): Next rnn state. Shapes: - agent_state (:obj:`torch.Tensor`): :math:`(T, B, A, N)`, where T is timestep, B is batch_size\ A is agent_num, N is obs_shape. - global_state (:obj:`torch.Tensor`): :math:`(T, B, M)`, where M is global_obs_shape. - prev_state (:obj:`list`): math:`(T, B, A)`, a list of length B, and each element is a list of length A. - action (:obj:`torch.Tensor`): :math:`(T, B, A)`. - total_q (:obj:`torch.Tensor`): :math:`(T, B)`. - agent_q (:obj:`torch.Tensor`): :math:`(T, B, A, P)`, where P is action_shape. - next_state (:obj:`list`): math:`(T, B, A)`, a list of length B, and each element is a list of length A. """ if q_star: # forward using Q_star network agent_state, global_state, prev_state = data['obs']['agent_state'], data['obs']['global_state'], data[ 'prev_state'] action = data.get('action', None) if single_step: agent_state, global_state = agent_state.unsqueeze(0), global_state.unsqueeze(0) T, B, A = agent_state.shape[:3] assert len(prev_state) == B and all( [len(p) == A for p in prev_state] ), '{}-{}-{}-{}'.format([type(p) for p in prev_state], B, A, len(prev_state[0])) prev_state = reduce(lambda x, y: x + y, prev_state) agent_state = agent_state.reshape(T, -1, *agent_state.shape[3:]) output = self._q_network_star( { 'obs': agent_state, 'prev_state': prev_state, 'enable_fast_timestep': True } ) # here is the forward pass of the agent networks of Q_star agent_q, next_state = output['logit'], output['next_state'] next_state, _ = list_split(next_state, step=A) agent_q = agent_q.reshape(T, B, A, -1) if action is None: # For target forward process if len(data['obs']['action_mask'].shape) == 3: action_mask = data['obs']['action_mask'].unsqueeze(0) else: action_mask = data['obs']['action_mask'] agent_q[action_mask == 0.0] = -9999999 action = agent_q.argmax(dim=-1) agent_q_act = torch.gather(agent_q, dim=-1, index=action.unsqueeze(-1)) agent_q_act = agent_q_act.squeeze(-1) # T, B, A global_state_embedding = self._global_state_encoder(global_state) total_q = self._mixer_star( agent_q_act, global_state_embedding ) # here is the forward pass of the mixer networks of Q_star if single_step: total_q, agent_q = total_q.squeeze(0), agent_q.squeeze(0) return { 'total_q': total_q, 'logit': agent_q, 'next_state': next_state, 'action_mask': data['obs']['action_mask'] } else: # forward using Q network agent_state, global_state, prev_state = data['obs']['agent_state'], data['obs']['global_state'], data[ 'prev_state'] action = data.get('action', None) if single_step: agent_state, global_state = agent_state.unsqueeze(0), global_state.unsqueeze(0) T, B, A = agent_state.shape[:3] assert len(prev_state) == B and all( [len(p) == A for p in prev_state] ), '{}-{}-{}-{}'.format([type(p) for p in prev_state], B, A, len(prev_state[0])) prev_state = reduce(lambda x, y: x + y, prev_state) agent_state = agent_state.reshape(T, -1, *agent_state.shape[3:]) output = self._q_network( { 'obs': agent_state, 'prev_state': prev_state, 'enable_fast_timestep': True } ) # here is the forward pass of the agent networks of Q agent_q, next_state = output['logit'], output['next_state'] next_state, _ = list_split(next_state, step=A) agent_q = agent_q.reshape(T, B, A, -1) if action is None: # For target forward process if len(data['obs']['action_mask'].shape) == 3: action_mask = data['obs']['action_mask'].unsqueeze(0) else: action_mask = data['obs']['action_mask'] agent_q[action_mask == 0.0] = -9999999 action = agent_q.argmax(dim=-1) agent_q_act = torch.gather(agent_q, dim=-1, index=action.unsqueeze(-1)) agent_q_act = agent_q_act.squeeze(-1) # T, B, A global_state_embedding = self._global_state_encoder(global_state) total_q = self._mixer( agent_q_act, global_state_embedding ) # here is the forward pass of the mixer networks of Q if single_step: total_q, agent_q = total_q.squeeze(0), agent_q.squeeze(0) return { 'total_q': total_q, 'logit': agent_q, 'next_state': next_state, 'action_mask': data['obs']['action_mask'] }