marl-value-factorisation/src/learners/mappo_learner.py at main · leyancode/marl-value-factorisation · GitHub

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
"""MAPPO learner with a centralised state-value critic."""

import copy

import torch
from torch.optim import Adam

from components.episode_buffer import EpisodeBatch
from modules.critics import REGISTRY as critic_registry
from utils.rl_utils import build_td_lambda_targets


class MAPPOLearner:
    def __init__(self, mac, scheme, logger, args):
        self.args = args
        self.n_agents = args.n_agents
        self.mac = mac
        self.logger = logger

        self.agent_params = list(mac.parameters())
        self.agent_optimiser = Adam(self.agent_params, lr=args.lr)
        self.critic = critic_registry[args.critic_type](scheme, args)
        self.target_critic = copy.deepcopy(self.critic)
        self.critic_params = list(self.critic.parameters())
        self.critic_optimiser = Adam(self.critic_params, lr=args.critic_lr)

        self.last_target_update_step = 0
        self.critic_training_steps = 0
        self.log_stats_t = -args.learner_log_interval - 1

    def _policy_outputs(self, batch):
        self.mac.init_hidden(batch.batch_size)
        return torch.stack(
            [self.mac.forward(batch, t=t) for t in range(batch.max_seq_length - 1)],
            dim=1,
        )

    def train(self, batch: EpisodeBatch, t_env: int, episode_num: int):
        del episode_num
        rewards = batch["reward"][:, :-1].squeeze(-1)
        actions = batch["actions"][:, :-1]
        mask = batch["filled"][:, :-1].float().squeeze(-1)
        terminated = batch["terminated"][:, :-1].float().squeeze(-1)
        mask[:, 1:] *= 1 - terminated[:, :-1]
        agent_mask = mask.unsqueeze(-1).expand(-1, -1, self.n_agents)

        old_policy = self._policy_outputs(batch)
        old_log_prob = torch.log(
            torch.gather(old_policy, dim=3, index=actions).squeeze(3) + 1e-10
        )

        states = batch["state"]
        target_values = self.target_critic(states).squeeze(-1)
        targets = build_td_lambda_targets(
            rewards,
            terminated,
            mask,
            target_values,
            1,
            self.args.gamma,
            self.args.td_lambda,
        )
        values = self.critic(states)[:, :-1].squeeze(-1)
        td_error = targets.detach() - values
        critic_loss = (td_error.square() * mask).sum() / mask.sum()

        self.critic_optimiser.zero_grad()
        critic_loss.backward()
        critic_grad_norm = torch.nn.utils.clip_grad_norm_(
            self.critic_params, self.args.grad_norm_clip
        )
        self.critic_optimiser.step()

        advantages = td_error.detach().unsqueeze(-1)
        for _ in range(self.args.ppo_epoch):
            policy = self._policy_outputs(batch)
            log_prob = torch.log(
                torch.gather(policy, dim=3, index=actions).squeeze(3) + 1e-10
            )
            ratio = torch.exp(log_prob - old_log_prob.detach())
            unclipped = ratio * advantages
            clipped = torch.clamp(
                ratio,
                1 - self.args.clip_param,
                1 + self.args.clip_param,
            ) * advantages
            entropy = -(policy * torch.log(policy + 1e-10)).sum(-1)
            policy_loss = -(
                (torch.minimum(unclipped, clipped) + self.args.entropy_coef * entropy)
                * agent_mask
            ).sum() / agent_mask.sum()

            self.agent_optimiser.zero_grad()
            policy_loss.backward()
            agent_grad_norm = torch.nn.utils.clip_grad_norm_(
                self.agent_params, self.args.grad_norm_clip
            )
            self.agent_optimiser.step()

        self.critic_training_steps += 1
        if (
            self.critic_training_steps - self.last_target_update_step
            >= self.args.target_update_interval
        ):
            self.target_critic.load_state_dict(self.critic.state_dict())
            self.last_target_update_step = self.critic_training_steps

        if t_env - self.log_stats_t >= self.args.learner_log_interval:
            self.logger.log_stat("critic_loss", critic_loss.item(), t_env)
            self.logger.log_stat("critic_grad_norm", critic_grad_norm.item(), t_env)
            self.logger.log_stat("policy_loss", policy_loss.item(), t_env)
            self.logger.log_stat("agent_grad_norm", agent_grad_norm.item(), t_env)
            advantage_mean = (advantages * agent_mask).sum().item() / agent_mask.sum().item()
            self.logger.log_stat("advantage_mean", advantage_mean, t_env)
            self.log_stats_t = t_env

    def cuda(self):
        self.mac.cuda()
        self.critic.cuda()
        self.target_critic.cuda()

    def save_models(self, path):
        self.mac.save_models(path)
        torch.save(self.critic.state_dict(), f"{path}/critic.th")
        torch.save(self.agent_optimiser.state_dict(), f"{path}/agent_opt.th")
        torch.save(self.critic_optimiser.state_dict(), f"{path}/critic_opt.th")

    def load_models(self, path):
        self.mac.load_models(path)
        self.critic.load_state_dict(torch.load(f"{path}/critic.th", map_location="cpu"))
        self.target_critic.load_state_dict(self.critic.state_dict())
        self.agent_optimiser.load_state_dict(
            torch.load(f"{path}/agent_opt.th", map_location="cpu")
        )
        self.critic_optimiser.load_state_dict(
            torch.load(f"{path}/critic_opt.th", map_location="cpu")
        )