relayrl_framework 0.4.52

A system-oriented, distributed reinforcement learning framework using a Rust backend with Python interfaces.
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

import sys, os
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..', '..')))
from _common._algorithms.BaseAlgorithm import AlgorithmAbstract

import numpy as np
import torch
from torch.optim import Adam

from .kernel import PolicyWithoutBaseline, PolicyWithBaseline
from .replay_buffer import ReplayBuffer

from utils.logger import EpochLogger, setup_logger_kwargs
from relayrl_framework import RelayRLTrajectory, ConfigLoader


class REINFORCE(AlgorithmAbstract):
    def __init__(self, env_dir: str, config_path: str, obs_dim: int, act_dim: int, buf_size: int):
        super().__init__()

        """Import and load RelayRL/config.json REINFORCE algorithm configurations and applies them to
        the current instance.
        
        Loads defaults if config.json is unavailable or key error thrown.
        """
        config_loader = ConfigLoader(algorithm_name='REINFORCE', config_path=config_path)
        hyperparams = config_loader.get_algorithm_params()['REINFORCE']
        self.save_model_path = config_loader.get_server_model_path()

        # Hyperparameters
        self._discrete = hyperparams['discrete']
        self._with_vf_baseline = hyperparams['with_vf_baseline']
        self._seed = hyperparams['seed']
        self._traj_per_epoch = hyperparams['traj_per_epoch']
        self._gamma = hyperparams['gamma']
        self._lam = hyperparams['lam']
        self._pi_lr = hyperparams['pi_lr']
        self._vf_lr = hyperparams['vf_lr']
        self._train_vf_iters = hyperparams['train_vf_iters']

        seed = self._seed + 10000 * os.getpid()
        torch.manual_seed(seed)
        np.random.seed(seed)

        self._replay_buffer = ReplayBuffer(obs_dim, act_dim, buf_size, self._gamma, self._lam, self._with_vf_baseline)
        if self._with_vf_baseline:
            self._model = PolicyWithBaseline(obs_dim, act_dim, self._discrete, [128, 128], torch.nn.ReLU)
            self._vf_optimizer = Adam(self._model.baseline.parameters(), lr=self._vf_lr)
        else:
            self._model = PolicyWithoutBaseline(obs_dim, act_dim, self._discrete, [128, 128])
        self._pi_optimizer = Adam(self._model.parameters(), lr=self._pi_lr)

        # set up logger
        log_data_dir = os.path.join(env_dir, './logs/')
        logger_kwargs = setup_logger_kwargs(
            "relayrl-reinforce-vf-info" if self._with_vf_baseline else "relayrl-reinforce-info", seed=seed, data_dir=log_data_dir)
        self.logger = EpochLogger(**logger_kwargs)
        self.logger.save_config(locals())

        self.traj = 0
        self.epoch = 0
        
        print("[REINFORCE Algorithm] Initialized")

    def save(self) -> None:
        self._model.eval()
        model_script = torch.jit.script(self._model)
        torch.jit.save(model_script, self.save_model_path)
        self._model.train()

    def receive_trajectory(self, trajectory: RelayRLTrajectory) -> bool:
        self.traj += 1
        ep_ret, ep_len = 0, 0

        for r4a in trajectory.get_actions():
            ep_ret += r4a.get_rew()
            ep_len += 1
            
            data = r4a.get_data()
            if not r4a.get_done():
                if self._with_vf_baseline:
                    self._replay_buffer.store(r4a.get_obs(), r4a.get_act(), r4a.get_mask(), r4a.get_rew(), data['v'], data['logp_a'])
                    self.logger.store(VVals=data['v'])
                else:
                    self._replay_buffer.store(r4a.get_obs(), r4a.get_act(), r4a.get_mask(), r4a.get_rew(), None, data['logp_a'])
            else:
                self._replay_buffer.finish_path(r4a.get_rew())
                self.logger.store(EpRet=ep_ret, EpLen=ep_len)

        if self.traj > 0 and self.traj % self._traj_per_epoch == 0:
            self.epoch += 1
            self.train_model()
            self.log_epoch()
            return True

        return False

    def train_model(self) -> None:
        data = self._replay_buffer.get()

        pi_l_old, pi_info_old = self.compute_loss_pi(data)
        pi_l_old = pi_l_old.item()
        if self._with_vf_baseline:
            v_l_old = self.compute_loss_vf(data).item()

        self._pi_optimizer.zero_grad()
        loss_pi, pi_info = self.compute_loss_pi(data)
        loss_pi.backward()
        self._pi_optimizer.step()

        if self._with_vf_baseline:
            for i in range(self._train_vf_iters):
                self._vf_optimizer.zero_grad()
                loss_v = self.compute_loss_vf(data)
                loss_v.backward()
                self._vf_optimizer.step()

        kl, ent = pi_info['kl'], pi_info['ent']
        delta_loss_pi = loss_pi.item() - pi_l_old
        delta_loss_v = 0 if not self._with_vf_baseline else loss_v.item() - v_l_old

        self.logger.store(LossPi=loss_pi.item(), DeltaLossPi=delta_loss_pi,
                          KL=kl, Entropy=ent)

        if self._with_vf_baseline:
            self.logger.store(LossV=loss_v.item(), DeltaLossV=delta_loss_v)

    def log_epoch(self) -> None:
        self.logger.log_tabular('Epoch', self.epoch)
        self.logger.log_tabular('EpRet', with_min_and_max=True)
        self.logger.log_tabular('EpLen', average_only=True)
        self.logger.log_tabular('LossPi', average_only=True)
        self.logger.log_tabular('DeltaLossPi', average_only=True)
        if self._with_vf_baseline:
            self.logger.log_tabular('VVals', with_min_and_max=True)
            self.logger.log_tabular('LossV', average_only=True)
            self.logger.log_tabular('DeltaLossV', average_only=True)
        self.logger.log_tabular('KL', average_only=True)
        self.logger.log_tabular('Entropy', average_only=True)
        self.logger.dump_tabular()

    def compute_loss_pi(self, data):
        obs, mask, act, adv, logp_old = data['obs'], data['mask'], data['act'], data['adv'], data['logp']
        pi_probs, pi_logits, logp = self._model.policy.forward(obs, mask, act)
        loss_pi = -(logp * adv).mean()

        approx_kl = (logp_old - logp).mean().item()
        
        # Entropy calculation
        min_real = torch.finfo(pi_logits.dtype).min
        logits = torch.clamp(pi_logits, min=min_real)
        p_log_p = logits * pi_probs
        entropy = -p_log_p.sum(-1).mean().item()

        pi_info = dict(kl=approx_kl, ent=entropy)

        return loss_pi, pi_info

    def compute_loss_vf(self, data):
        obs, mask, ret = data['obs'], data['mask'], data['ret']
        return ((self._model.baseline.forward(obs, mask) - ret) ** 2).mean()