File size: 12,508 Bytes
079c32c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 |
from typing import Dict, Union
import torch
import torch.nn as nn
from functools import reduce
from ding.torch_utils import one_hot, MLP
from ding.utils import squeeze, list_split, MODEL_REGISTRY, SequenceType
from .q_learning import DRQN
class COMAActorNetwork(nn.Module):
"""
Overview:
Decentralized actor network in COMA algorithm.
Interface:
``__init__``, ``forward``
"""
def __init__(
self,
obs_shape: int,
action_shape: int,
hidden_size_list: SequenceType = [128, 128, 64],
):
"""
Overview:
Initialize COMA actor network
Arguments:
- obs_shape (:obj:`int`): the dimension of each agent's observation state
- action_shape (:obj:`int`): the dimension of action shape
- hidden_size_list (:obj:`list`): the list of hidden size, default to [128, 128, 64]
"""
super(COMAActorNetwork, self).__init__()
self.main = DRQN(obs_shape, action_shape, hidden_size_list)
def forward(self, inputs: Dict) -> Dict:
"""
Overview:
The forward computation graph of COMA actor network
Arguments:
- inputs (:obj:`dict`): input data dict with keys ['obs', 'prev_state']
- agent_state (:obj:`torch.Tensor`): each agent local state(obs)
- action_mask (:obj:`torch.Tensor`): the masked action
- prev_state (:obj:`torch.Tensor`): the previous hidden state
Returns:
- output (:obj:`dict`): output data dict with keys ['logit', 'next_state', 'action_mask']
ArgumentsKeys:
- necessary: ``obs`` { ``agent_state``, ``action_mask`` }, ``prev_state``
ReturnsKeys:
- necessary: ``logit``, ``next_state``, ``action_mask``
Examples:
>>> T, B, A, N = 4, 8, 3, 32
>>> embedding_dim = 64
>>> action_dim = 6
>>> data = torch.randn(T, B, A, N)
>>> model = COMAActorNetwork((N, ), action_dim, [128, embedding_dim])
>>> prev_state = [[None for _ in range(A)] for _ in range(B)]
>>> for t in range(T):
>>> inputs = {'obs': {'agent_state': data[t], 'action_mask': None}, 'prev_state': prev_state}
>>> outputs = model(inputs)
>>> logit, prev_state = outputs['logit'], outputs['next_state']
"""
agent_state = inputs['obs']['agent_state']
prev_state = inputs['prev_state']
if len(agent_state.shape) == 3: # B, A, N
agent_state = agent_state.unsqueeze(0)
unsqueeze_flag = True
else:
unsqueeze_flag = False
T, B, A = agent_state.shape[:3]
agent_state = agent_state.reshape(T, -1, *agent_state.shape[3:])
prev_state = reduce(lambda x, y: x + y, prev_state)
output = self.main({'obs': agent_state, 'prev_state': prev_state, 'enable_fast_timestep': True})
logit, next_state = output['logit'], output['next_state']
next_state, _ = list_split(next_state, step=A)
logit = logit.reshape(T, B, A, -1)
if unsqueeze_flag:
logit = logit.squeeze(0)
return {'logit': logit, 'next_state': next_state, 'action_mask': inputs['obs']['action_mask']}
class COMACriticNetwork(nn.Module):
"""
Overview:
Centralized critic network in COMA algorithm.
Interface:
``__init__``, ``forward``
"""
def __init__(
self,
input_size: int,
action_shape: int,
hidden_size: int = 128,
):
"""
Overview:
initialize COMA critic network
Arguments:
- input_size (:obj:`int`): the size of input global observation
- action_shape (:obj:`int`): the dimension of action shape
- hidden_size_list (:obj:`list`): the list of hidden size, default to 128
Returns:
- output (:obj:`dict`): output data dict with keys ['q_value']
Shapes:
- obs (:obj:`dict`): ``agent_state``: :math:`(T, B, A, N, D)`, ``action_mask``: :math:`(T, B, A, N, A)`
- prev_state (:obj:`list`): :math:`[[[h, c] for _ in range(A)] for _ in range(B)]`
- logit (:obj:`torch.Tensor`): :math:`(T, B, A, N, A)`
- next_state (:obj:`list`): :math:`[[[h, c] for _ in range(A)] for _ in range(B)]`
- action_mask (:obj:`torch.Tensor`): :math:`(T, B, A, N, A)`
"""
super(COMACriticNetwork, self).__init__()
self.action_shape = action_shape
self.act = nn.ReLU()
self.mlp = nn.Sequential(
MLP(input_size, hidden_size, hidden_size, 2, activation=self.act), nn.Linear(hidden_size, action_shape)
)
def forward(self, data: Dict) -> Dict:
"""
Overview:
forward computation graph of qmix network
Arguments:
- data (:obj:`dict`): input data dict with keys ['obs', 'prev_state', 'action']
- agent_state (:obj:`torch.Tensor`): each agent local state(obs)
- global_state (:obj:`torch.Tensor`): global state(obs)
- action (:obj:`torch.Tensor`): the masked action
ArgumentsKeys:
- necessary: ``obs`` { ``agent_state``, ``global_state`` }, ``action``, ``prev_state``
ReturnsKeys:
- necessary: ``q_value``
Examples:
>>> agent_num, bs, T = 4, 3, 8
>>> obs_dim, global_obs_dim, action_dim = 32, 32 * 4, 9
>>> coma_model = COMACriticNetwork(
>>> obs_dim - action_dim + global_obs_dim + 2 * action_dim * agent_num, action_dim)
>>> data = {
>>> 'obs': {
>>> 'agent_state': torch.randn(T, bs, agent_num, obs_dim),
>>> 'global_state': torch.randn(T, bs, global_obs_dim),
>>> },
>>> 'action': torch.randint(0, action_dim, size=(T, bs, agent_num)),
>>> }
>>> output = coma_model(data)
"""
x = self._preprocess_data(data)
q = self.mlp(x)
return {'q_value': q}
def _preprocess_data(self, data: Dict) -> torch.Tensor:
"""
Overview:
preprocess data to make it can be used by MLP net
Arguments:
- data (:obj:`dict`): input data dict with keys ['obs', 'prev_state', 'action']
- agent_state (:obj:`torch.Tensor`): each agent local state(obs)
- global_state (:obj:`torch.Tensor`): global state(obs)
- action (:obj:`torch.Tensor`): the masked action
ArgumentsKeys:
- necessary: ``obs`` { ``agent_state``, ``global_state``} , ``action``, ``prev_state``
Return:
- x (:obj:`torch.Tensor`): the data can be used by MLP net, including \
``global_state``, ``agent_state``, ``last_action``, ``action``, ``agent_id``
"""
t_size, batch_size, agent_num = data['obs']['agent_state'].shape[:3]
agent_state_ori, global_state = data['obs']['agent_state'], data['obs']['global_state']
# splite obs, last_action and agent_id
agent_state = agent_state_ori[..., :-self.action_shape - agent_num]
last_action = agent_state_ori[..., -self.action_shape - agent_num:-agent_num]
last_action = last_action.reshape(t_size, batch_size, 1, -1).repeat(1, 1, agent_num, 1)
agent_id = agent_state_ori[..., -agent_num:]
action = one_hot(data['action'], self.action_shape) # T, B, A，N
action = action.reshape(t_size, batch_size, -1, agent_num * self.action_shape).repeat(1, 1, agent_num, 1)
action_mask = (1 - torch.eye(agent_num).to(action.device))
action_mask = action_mask.view(-1, 1).repeat(1, self.action_shape).view(agent_num, -1) # A, A*N
action = (action_mask.unsqueeze(0).unsqueeze(0)) * action # T, B, A, A*N
global_state = global_state.unsqueeze(2).repeat(1, 1, agent_num, 1)
x = torch.cat([global_state, agent_state, last_action, action, agent_id], -1)
return x
@MODEL_REGISTRY.register('coma')
class COMA(nn.Module):
"""
Overview:
The network of COMA algorithm, which is QAC-type actor-critic.
Interface:
``__init__``, ``forward``
Properties:
- mode (:obj:`list`): The list of forward mode, including ``compute_actor`` and ``compute_critic``
"""
mode = ['compute_actor', 'compute_critic']
def __init__(
self, agent_num: int, obs_shape: Dict, action_shape: Union[int, SequenceType],
actor_hidden_size_list: SequenceType
) -> None:
"""
Overview:
initialize COMA network
Arguments:
- agent_num (:obj:`int`): the number of agent
- obs_shape (:obj:`Dict`): the observation information, including agent_state and \
global_state
- action_shape (:obj:`Union[int, SequenceType]`): the dimension of action shape
- actor_hidden_size_list (:obj:`SequenceType`): the list of hidden size
"""
super(COMA, self).__init__()
action_shape = squeeze(action_shape)
actor_input_size = squeeze(obs_shape['agent_state'])
critic_input_size = squeeze(obs_shape['agent_state']) + squeeze(obs_shape['global_state']) + \
agent_num * action_shape + (agent_num - 1) * action_shape
critic_hidden_size = actor_hidden_size_list[-1]
self.actor = COMAActorNetwork(actor_input_size, action_shape, actor_hidden_size_list)
self.critic = COMACriticNetwork(critic_input_size, action_shape, critic_hidden_size)
def forward(self, inputs: Dict, mode: str) -> Dict:
"""
Overview:
forward computation graph of COMA network
Arguments:
- inputs (:obj:`dict`): input data dict with keys ['obs', 'prev_state', 'action']
- agent_state (:obj:`torch.Tensor`): each agent local state(obs)
- global_state (:obj:`torch.Tensor`): global state(obs)
- action (:obj:`torch.Tensor`): the masked action
ArgumentsKeys:
- necessary: ``obs`` { ``agent_state``, ``global_state``, ``action_mask`` }, ``action``, ``prev_state``
ReturnsKeys:
- necessary:
- compute_critic: ``q_value``
- compute_actor: ``logit``, ``next_state``, ``action_mask``
Shapes:
- obs (:obj:`dict`): ``agent_state``: :math:`(T, B, A, N, D)`, ``action_mask``: :math:`(T, B, A, N, A)`
- prev_state (:obj:`list`): :math:`[[[h, c] for _ in range(A)] for _ in range(B)]`
- logit (:obj:`torch.Tensor`): :math:`(T, B, A, N, A)`
- next_state (:obj:`list`): :math:`[[[h, c] for _ in range(A)] for _ in range(B)]`
- action_mask (:obj:`torch.Tensor`): :math:`(T, B, A, N, A)`
- q_value (:obj:`torch.Tensor`): :math:`(T, B, A, N, A)`
Examples:
>>> agent_num, bs, T = 4, 3, 8
>>> agent_num, bs, T = 4, 3, 8
>>> obs_dim, global_obs_dim, action_dim = 32, 32 * 4, 9
>>> coma_model = COMA(
>>> agent_num=agent_num,
>>> obs_shape=dict(agent_state=(obs_dim, ), global_state=(global_obs_dim, )),
>>> action_shape=action_dim,
>>> actor_hidden_size_list=[128, 64],
>>> )
>>> prev_state = [[None for _ in range(agent_num)] for _ in range(bs)]
>>> data = {
>>> 'obs': {
>>> 'agent_state': torch.randn(T, bs, agent_num, obs_dim),
>>> 'action_mask': None,
>>> },
>>> 'prev_state': prev_state,
>>> }
>>> output = coma_model(data, mode='compute_actor')
>>> data= {
>>> 'obs': {
>>> 'agent_state': torch.randn(T, bs, agent_num, obs_dim),
>>> 'global_state': torch.randn(T, bs, global_obs_dim),
>>> },
>>> 'action': torch.randint(0, action_dim, size=(T, bs, agent_num)),
>>> }
>>> output = coma_model(data, mode='compute_critic')
"""
assert mode in self.mode, "not support forward mode: {}/{}".format(mode, self.mode)
if mode == 'compute_actor':
return self.actor(inputs)
elif mode == 'compute_critic':
return self.critic(inputs)
|