File size: 7,557 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
import copy
from datetime import datetime
from typing import Union, Optional, Dict

import gymnasium as gym
import numpy as np
from ding.envs import BaseEnv, BaseEnvTimestep
from ding.envs import ObsPlusPrevActRewWrapper
from ding.torch_utils import to_ndarray
from ding.utils import ENV_REGISTRY
from easydict import EasyDict


@ENV_REGISTRY.register('cartpole_lightzero')
class CartPoleEnv(BaseEnv):
    """
    LightZero version of the classic CartPole environment. This class includes methods for resetting, closing, and
    stepping through the environment, as well as seeding for reproducibility, saving replay videos, and generating random
    actions. It also includes properties for accessing the observation space, action space, and reward space of the
    environment.
    """

    config = dict(
        # env_name (str): The name of the environment.
        env_name="CartPole-v0",
        # replay_path (str): The path to save the replay video. If None, the replay will not be saved.
        # Only effective when env_manager.type is 'base'.
        replay_path=None,
    )

    @classmethod
    def default_config(cls: type) -> EasyDict:
        cfg = EasyDict(copy.deepcopy(cls.config))
        cfg.cfg_type = cls.__name__ + 'Dict'
        return cfg

    def __init__(self, cfg: dict = {}) -> None:
        """
        Initialize the environment with a configuration dictionary. Sets up spaces for observations, actions, and rewards.
        """
        self._cfg = cfg
        self._init_flag = False
        self._continuous = False
        self._replay_path = cfg.replay_path
        self._observation_space = gym.spaces.Box(
            low=np.array([-4.8, float("-inf"), -0.42, float("-inf")]),
            high=np.array([4.8, float("inf"), 0.42, float("inf")]),
            shape=(4,),
            dtype=np.float32
        )
        self._action_space = gym.spaces.Discrete(2)
        self._action_space.seed(0)  # default seed
        self._reward_space = gym.spaces.Box(low=0.0, high=1.0, shape=(1,), dtype=np.float32)

    def reset(self) -> Dict[str, np.ndarray]:
        """
        Reset the environment. If it hasn't been initialized yet, this method also handles that. It also handles seeding
        if necessary. Returns the first observation.
        """
        if not self._init_flag:
            self._env = gym.make('CartPole-v0', render_mode="rgb_array")
            if self._replay_path is not None:
                timestamp = datetime.now().strftime("%Y%m%d%H%M%S")
                video_name = f'{self._env.spec.id}-video-{timestamp}'
                self._env = gym.wrappers.RecordVideo(
                    self._env,
                    video_folder=self._replay_path,
                    episode_trigger=lambda episode_id: True,
                    name_prefix=video_name
                )
            if hasattr(self._cfg, 'obs_plus_prev_action_reward') and self._cfg.obs_plus_prev_action_reward:
                self._env = ObsPlusPrevActRewWrapper(self._env)
            self._init_flag = True
        if hasattr(self, '_seed') and hasattr(self, '_dynamic_seed') and self._dynamic_seed:
            np_seed = 100 * np.random.randint(1, 1000)
            self._seed = self._seed + np_seed
            self._action_space.seed(self._seed)
            obs, _ = self._env.reset(seed=self._seed)
        elif hasattr(self, '_seed'):
            self._action_space.seed(self._seed)
            obs, _ = self._env.reset(seed=self._seed)
        else:
            obs, _ = self._env.reset()
        self._observation_space = self._env.observation_space
        self._eval_episode_return = 0
        obs = to_ndarray(obs)

        action_mask = np.ones(self.action_space.n, 'int8')
        obs = {'observation': obs, 'action_mask': action_mask, 'to_play': -1}

        return obs

    def step(self, action: Union[int, np.ndarray]) -> BaseEnvTimestep:
        """
        Overview:
            Perform a step in the environment using the provided action, and return the next state of the environment.
            The next state is encapsulated in a BaseEnvTimestep object, which includes the new observation, reward,
            done flag, and info dictionary.
        Arguments:
            - action (:obj:`Union[int, np.ndarray]`): The action to be performed in the environment. If the action is
              a 1-dimensional numpy array, it is squeezed to a 0-dimension array.
        Returns:
            - timestep (:obj:`BaseEnvTimestep`): An object containing the new observation, reward, done flag,
              and info dictionary.
        .. note::
            - The cumulative reward (`_eval_episode_return`) is updated with the reward obtained in this step.
            - If the episode ends (done is True), the total reward for the episode is stored in the info dictionary
              under the key 'eval_episode_return'.
            - An action mask is created with ones, which represents the availability of each action in the action space.
            - Observations are returned in a dictionary format containing 'observation', 'action_mask', and 'to_play'.
        """
        if isinstance(action, np.ndarray) and action.shape == (1,):
            action = action.squeeze()  # 0-dim array

        obs, rew, terminated, truncated, info = self._env.step(action)
        done = terminated or truncated

        self._eval_episode_return += rew
        if done:
            info['eval_episode_return'] = self._eval_episode_return

        action_mask = np.ones(self.action_space.n, 'int8')
        obs = {'observation': obs, 'action_mask': action_mask, 'to_play': -1}

        return BaseEnvTimestep(obs, rew, done, info)

    def close(self) -> None:
        """
        Close the environment, and set the initialization flag to False.
        """
        if self._init_flag:
            self._env.close()
        self._init_flag = False

    def seed(self, seed: int, dynamic_seed: bool = True) -> None:
        """
        Set the seed for the environment's random number generator. Can handle both static and dynamic seeding.
        """
        self._seed = seed
        self._dynamic_seed = dynamic_seed
        np.random.seed(self._seed)

    def enable_save_replay(self, replay_path: Optional[str] = None) -> None:
        """
        Enable the saving of replay videos. If no replay path is given, a default is used.
        """
        if replay_path is None:
            replay_path = './video'
        self._replay_path = replay_path

    def random_action(self) -> np.ndarray:
        """
         Generate a random action using the action space's sample method. Returns a numpy array containing the action.
         """
        random_action = self.action_space.sample()
        random_action = to_ndarray([random_action], dtype=np.int64)
        return random_action

    @property
    def observation_space(self) -> gym.spaces.Space:
        """
        Property to access the observation space of the environment.
        """
        return self._observation_space

    @property
    def action_space(self) -> gym.spaces.Space:
        """
        Property to access the action space of the environment.
        """
        return self._action_space

    @property
    def reward_space(self) -> gym.spaces.Space:
        """
        Property to access the reward space of the environment.
        """
        return self._reward_space

    def __repr__(self) -> str:
        """
        String representation of the environment.
        """
        return "LightZero CartPole Env"