File size: 5,838 Bytes
b18ddcc
 
 
 
 
 
 
 
 
 
 
 
1e1c086
b18ddcc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1e1c086
b18ddcc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import numpy as np
import torch
import torch.nn as nn

from collections import defaultdict
from dataclasses import dataclass, asdict
from torch.optim import Adam
from torch.utils.tensorboard.writer import SummaryWriter
from typing import Optional, Sequence, TypeVar

from rl_algo_impls.shared.algorithm import Algorithm
from rl_algo_impls.shared.callbacks.callback import Callback
from rl_algo_impls.shared.gae import compute_rtg_and_advantage_from_trajectories
from rl_algo_impls.shared.trajectory import Trajectory, TrajectoryAccumulator
from rl_algo_impls.vpg.policy import VPGActorCritic
from rl_algo_impls.wrappers.vectorable_wrapper import VecEnv


@dataclass
class TrainEpochStats:
    pi_loss: float
    entropy_loss: float
    v_loss: float
    envs_with_done: int = 0
    episodes_done: int = 0

    def write_to_tensorboard(self, tb_writer: SummaryWriter, global_step: int) -> None:
        for name, value in asdict(self).items():
            tb_writer.add_scalar(f"losses/{name}", value, global_step=global_step)


class VPGTrajectoryAccumulator(TrajectoryAccumulator):
    def __init__(self, num_envs: int) -> None:
        super().__init__(num_envs, trajectory_class=Trajectory)
        self.completed_per_env: defaultdict[int, int] = defaultdict(int)

    def on_done(self, env_idx: int, trajectory: Trajectory) -> None:
        self.completed_per_env[env_idx] += 1


VanillaPolicyGradientSelf = TypeVar(
    "VanillaPolicyGradientSelf", bound="VanillaPolicyGradient"
)


class VanillaPolicyGradient(Algorithm):
    def __init__(
        self,
        policy: VPGActorCritic,
        env: VecEnv,
        device: torch.device,
        tb_writer: SummaryWriter,
        gamma: float = 0.99,
        pi_lr: float = 3e-4,
        val_lr: float = 1e-3,
        train_v_iters: int = 80,
        gae_lambda: float = 0.97,
        max_grad_norm: float = 10.0,
        n_steps: int = 4_000,
        sde_sample_freq: int = -1,
        ent_coef: float = 0.0,
    ) -> None:
        super().__init__(policy, env, device, tb_writer)
        self.policy = policy

        self.gamma = gamma
        self.gae_lambda = gae_lambda
        self.pi_optim = Adam(self.policy.pi.parameters(), lr=pi_lr)
        self.val_optim = Adam(self.policy.v.parameters(), lr=val_lr)
        self.max_grad_norm = max_grad_norm

        self.n_steps = n_steps
        self.train_v_iters = train_v_iters
        self.sde_sample_freq = sde_sample_freq

        self.ent_coef = ent_coef

    def learn(
        self: VanillaPolicyGradientSelf,
        total_timesteps: int,
        callback: Optional[Callback] = None,
    ) -> VanillaPolicyGradientSelf:
        timesteps_elapsed = 0
        epoch_cnt = 0
        while timesteps_elapsed < total_timesteps:
            epoch_cnt += 1
            accumulator = self._collect_trajectories()
            epoch_stats = self.train(accumulator.all_trajectories)
            epoch_stats.envs_with_done = len(accumulator.completed_per_env)
            epoch_stats.episodes_done = sum(accumulator.completed_per_env.values())
            epoch_steps = accumulator.n_timesteps()
            timesteps_elapsed += epoch_steps
            epoch_stats.write_to_tensorboard(
                self.tb_writer, global_step=timesteps_elapsed
            )
            print(
                " | ".join(
                    [
                        f"Epoch: {epoch_cnt}",
                        f"Pi Loss: {round(epoch_stats.pi_loss, 2)}",
                        f"Epoch Loss: {round(epoch_stats.entropy_loss, 2)}",
                        f"V Loss: {round(epoch_stats.v_loss, 2)}",
                        f"Total Steps: {timesteps_elapsed}",
                    ]
                )
            )
            if callback:
                callback.on_step(timesteps_elapsed=epoch_steps)
        return self

    def train(self, trajectories: Sequence[Trajectory]) -> TrainEpochStats:
        self.policy.train()
        obs = torch.as_tensor(
            np.concatenate([np.array(t.obs) for t in trajectories]), device=self.device
        )
        act = torch.as_tensor(
            np.concatenate([np.array(t.act) for t in trajectories]), device=self.device
        )
        rtg, adv = compute_rtg_and_advantage_from_trajectories(
            trajectories, self.policy, self.gamma, self.gae_lambda, self.device
        )

        _, logp, entropy = self.policy.pi(obs, act)
        pi_loss = -(logp * adv).mean()
        entropy_loss = entropy.mean()

        actor_loss = pi_loss - self.ent_coef * entropy_loss

        self.pi_optim.zero_grad()
        actor_loss.backward()
        nn.utils.clip_grad_norm_(self.policy.pi.parameters(), self.max_grad_norm)
        self.pi_optim.step()

        v_loss = 0
        for _ in range(self.train_v_iters):
            v = self.policy.v(obs)
            v_loss = ((v - rtg) ** 2).mean()

            self.val_optim.zero_grad()
            v_loss.backward()
            nn.utils.clip_grad_norm_(self.policy.v.parameters(), self.max_grad_norm)
            self.val_optim.step()

        return TrainEpochStats(
            pi_loss.item(),
            entropy_loss.item(),
            v_loss.item(),  # type: ignore
        )

    def _collect_trajectories(self) -> VPGTrajectoryAccumulator:
        self.policy.eval()
        obs = self.env.reset()
        accumulator = VPGTrajectoryAccumulator(self.env.num_envs)
        self.policy.reset_noise()
        for i in range(self.n_steps):
            if self.sde_sample_freq > 0 and i > 0 and i % self.sde_sample_freq == 0:
                self.policy.reset_noise()
            action, value, _, clamped_action = self.policy.step(obs)
            next_obs, reward, done, _ = self.env.step(clamped_action)
            accumulator.step(obs, action, next_obs, reward, done, value)
            obs = next_obs
        return accumulator