/
maddpg_learner.py
203 lines (168 loc) · 8.99 KB
/
maddpg_learner.py
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
# code heavily adapted from https://github.com/oxwhirl/facmac/
import copy
from components.episode_buffer import EpisodeBatch
from modules.critics.maddpg import MADDPGCritic
import torch as th
from torch.optim import RMSprop, Adam
from controllers.maddpg_controller import gumbel_softmax
from modules.critics import REGISTRY as critic_registry
from components.standarize_stream import RunningMeanStd
class MADDPGLearner:
def __init__(self, mac, scheme, logger, args):
self.args = args
self.n_agents = args.n_agents
self.n_actions = args.n_actions
self.logger = logger
self.mac = mac
self.target_mac = copy.deepcopy(self.mac)
self.agent_params = list(mac.parameters())
self.critic = critic_registry[args.critic_type](scheme, args)
self.target_critic = copy.deepcopy(self.critic)
self.critic_params = list(self.critic.parameters())
self.agent_optimiser = Adam(params=self.agent_params, lr=self.args.lr)
self.critic_optimiser = Adam(params=self.critic_params, lr=self.args.lr)
self.log_stats_t = -self.args.learner_log_interval - 1
self.last_target_update_episode = 0
device = "cuda" if args.use_cuda else "cpu"
if self.args.standardise_returns:
self.ret_ms = RunningMeanStd(shape=(self.n_agents,), device=device)
if self.args.standardise_rewards:
self.rew_ms = RunningMeanStd(shape=(1,), device=device)
def train(self, batch: EpisodeBatch, t_env: int, episode_num: int):
# Get the relevant quantities
rewards = batch["reward"][:, :-1]
actions = batch["actions_onehot"]
terminated = batch["terminated"][:, :-1].float()
rewards = rewards.unsqueeze(2).expand(-1, -1, self.n_agents, -1)
terminated = terminated.unsqueeze(2).expand(-1, -1, self.n_agents, -1)
mask = 1 - terminated
batch_size = batch.batch_size
if self.args.standardise_rewards:
self.rew_ms.update(rewards)
rewards = (rewards - self.rew_ms.mean) / th.sqrt(self.rew_ms.var)
# Train the critic
inputs = self._build_inputs(batch)
actions = actions.view(batch_size, -1, 1, self.n_agents * self.n_actions).expand(-1, -1, self.n_agents, -1)
q_taken = self.critic(inputs[:, :-1], actions[:, :-1].detach())
q_taken = q_taken.view(batch_size, -1, 1)
# Use the target actor and target critic network to compute the target q
self.target_mac.init_hidden(batch.batch_size)
target_actions = []
for t in range(1, batch.max_seq_length):
agent_target_outs = self.target_mac.target_actions(batch, t)
target_actions.append(agent_target_outs)
target_actions = th.stack(target_actions, dim=1) # Concat over time
target_actions = target_actions.view(batch_size, -1, 1, self.n_agents * self.n_actions).expand(-1, -1, self.n_agents, -1)
target_vals = self.target_critic(inputs[:, 1:], target_actions.detach())
target_vals = target_vals.view(batch_size, -1, 1)
if self.args.standardise_returns:
target_vals = target_vals * th.sqrt(self.ret_ms.var) + self.ret_ms.mean
targets = rewards.reshape(-1, 1) + self.args.gamma * (1 - terminated.reshape(-1, 1)) * target_vals.reshape(-1, 1).detach()
if self.args.standardise_returns:
self.ret_ms.update(targets)
targets = (targets - self.ret_ms.mean) / th.sqrt(self.ret_ms.var)
td_error = (q_taken.view(-1, 1) - targets.detach())
masked_td_error = td_error * mask.reshape(-1, 1)
loss = (masked_td_error ** 2).mean()
self.critic_optimiser.zero_grad()
loss.backward()
critic_grad_norm = th.nn.utils.clip_grad_norm_(self.critic_params, self.args.grad_norm_clip)
self.critic_optimiser.step()
# Train the actor
self.mac.init_hidden(batch_size)
pis = []
actions = []
for t in range(batch.max_seq_length-1):
pi = self.mac.forward(batch, t=t).view(batch_size, 1, self.n_agents, -1)
actions.append(gumbel_softmax(pi, hard=True))
pis.append(pi)
actions = th.cat(actions, dim=1)
actions = actions.view(batch_size, -1, 1, self.n_agents * self.n_actions).expand(-1, -1, self.n_agents, -1)
new_actions = []
for i in range(self.n_agents):
temp_action = th.split(actions[:, :, i, :], self.n_actions, dim=2)
actions_i = []
for j in range(self.n_agents):
if i == j:
actions_i.append(temp_action[j])
else:
actions_i.append(temp_action[j].detach())
actions_i = th.cat(actions_i, dim=-1)
new_actions.append(actions_i.unsqueeze(2))
new_actions = th.cat(new_actions, dim=2)
pis = th.cat(pis, dim=1)
pis[pis==-1e10] = 0
pis = pis.reshape(-1, 1)
q = self.critic(inputs[:, :-1], new_actions)
q = q.reshape(-1, 1)
mask = mask.reshape(-1, 1)
# Compute the actor loss
pg_loss = -(q * mask).mean() + self.args.reg * (pis ** 2).mean()
# Optimise agents
self.agent_optimiser.zero_grad()
pg_loss.backward()
agent_grad_norm = th.nn.utils.clip_grad_norm_(self.agent_params, self.args.grad_norm_clip)
self.agent_optimiser.step()
if self.args.target_update_interval_or_tau > 1 and (episode_num - self.last_target_update_episode) / self.args.target_update_interval_or_tau >= 1.0:
self._update_targets_hard()
self.last_target_update_episode = episode_num
elif self.args.target_update_interval_or_tau <= 1.0:
self._update_targets_soft(self.args.target_update_interval_or_tau)
if t_env - self.log_stats_t >= self.args.learner_log_interval:
self.logger.log_stat("critic_loss", loss.item(), t_env)
self.logger.log_stat("critic_grad_norm", critic_grad_norm.item(), t_env)
self.logger.log_stat("agent_grad_norm", agent_grad_norm.item(), t_env)
mask_elems = mask.sum().item()
self.logger.log_stat("td_error_abs", masked_td_error.abs().sum().item() / mask_elems, t_env)
self.logger.log_stat("q_taken_mean", (q_taken).sum().item() / mask_elems, t_env)
self.logger.log_stat("target_mean", targets.sum().item() / mask_elems, t_env)
self.logger.log_stat("pg_loss", pg_loss.item(), t_env)
self.logger.log_stat("agent_grad_norm", agent_grad_norm, t_env)
self.log_stats_t = t_env
def _build_inputs(self, batch, t=None):
bs = batch.batch_size
max_t = batch.max_seq_length if t is None else 1
ts = slice(None) if t is None else slice(t, t + 1)
inputs = []
inputs.append(batch["state"][:, ts].unsqueeze(2).expand(-1, -1, self.n_agents, -1))
if self.args.obs_individual_obs:
inputs.append(batch["obs"][:, ts])
# last actions
if self.args.obs_last_action:
if t == 0:
inputs.append(th.zeros_like(batch["actions_onehot"][:, 0:1]))
elif isinstance(t, int):
inputs.append(batch["actions_onehot"][:, slice(t - 1, t)])
else:
last_actions = th.cat([th.zeros_like(batch["actions_onehot"][:, 0:1]), batch["actions_onehot"][:, :-1]],
dim=1)
# last_actions = last_actions.view(bs, max_t, 1, -1).repeat(1, 1, self.n_agents, 1)
inputs.append(last_actions)
if self.args.obs_agent_id:
inputs.append(th.eye(self.n_agents, device=batch.device).unsqueeze(0).unsqueeze(0).expand(bs, max_t, -1, -1))
inputs = th.cat(inputs, dim=-1)
return inputs
def _update_targets_hard(self):
self.target_mac.load_state(self.mac)
self.target_critic.load_state_dict(self.critic.state_dict())
def _update_targets_soft(self, tau):
for target_param, param in zip(self.target_mac.parameters(), self.mac.parameters()):
target_param.data.copy_(target_param.data * (1.0 - tau) + param.data * tau)
for target_param, param in zip(self.target_critic.parameters(), self.critic.parameters()):
target_param.data.copy_(target_param.data * (1.0 - tau) + param.data * tau)
def cuda(self):
self.mac.cuda()
self.target_mac.cuda()
self.critic.cuda()
self.target_critic.cuda()
def save_models(self, path):
self.mac.save_models(path)
th.save(self.critic.state_dict(), "{}/critic.th".format(path))
th.save(self.agent_optimiser.state_dict(), "{}/agent_opt.th".format(path))
th.save(self.critic_optimiser.state_dict(), "{}/critic_opt.th".format(path))
def load_models(self, path):
self.mac.load_models(path)
# Not quite right but I don't want to save target networks
self.target_mac.load_models(path)
self.agent_optimiser.load_state_dict(
th.load("{}/agent_opt.th".format(path), map_location=lambda storage, loc: storage))