Spaces:
Sleeping
Sleeping
| from typing import Any, List, Union, Optional | |
| import time | |
| import gym | |
| import copy | |
| import numpy as np | |
| from easydict import EasyDict | |
| from ding.envs import BaseEnv, BaseEnvTimestep | |
| from ding.torch_utils import to_ndarray, to_list | |
| from ding.utils import ENV_REGISTRY | |
| from ding.envs import ObsPlusPrevActRewWrapper | |
| class CartPoleEnv(BaseEnv): | |
| def __init__(self, cfg: dict = {}) -> None: | |
| self._cfg = cfg | |
| self._init_flag = False | |
| self._replay_path = None | |
| 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) -> np.ndarray: | |
| if not self._init_flag: | |
| self._env = gym.make('CartPole-v0') | |
| if self._replay_path is not None: | |
| self._env = gym.wrappers.RecordVideo( | |
| self._env, | |
| video_folder=self._replay_path, | |
| episode_trigger=lambda episode_id: True, | |
| name_prefix='rl-video-{}'.format(id(self)) | |
| ) | |
| 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._env.seed(self._seed + np_seed) | |
| self._action_space.seed(self._seed + np_seed) | |
| elif hasattr(self, '_seed'): | |
| self._env.seed(self._seed) | |
| self._action_space.seed(self._seed) | |
| self._observation_space = self._env.observation_space | |
| self._eval_episode_return = 0 | |
| obs = self._env.reset() | |
| obs = to_ndarray(obs) | |
| return obs | |
| def close(self) -> None: | |
| if self._init_flag: | |
| self._env.close() | |
| self._init_flag = False | |
| def seed(self, seed: int, dynamic_seed: bool = True) -> None: | |
| self._seed = seed | |
| self._dynamic_seed = dynamic_seed | |
| np.random.seed(self._seed) | |
| def step(self, action: Union[int, np.ndarray]) -> BaseEnvTimestep: | |
| if isinstance(action, np.ndarray) and action.shape == (1, ): | |
| action = action.squeeze() # 0-dim array | |
| obs, rew, done, info = self._env.step(action) | |
| self._eval_episode_return += rew | |
| if done: | |
| info['eval_episode_return'] = self._eval_episode_return | |
| obs = to_ndarray(obs) | |
| rew = to_ndarray([rew]).astype(np.float32) # wrapped to be transfered to a array with shape (1,) | |
| return BaseEnvTimestep(obs, rew, done, info) | |
| def enable_save_replay(self, replay_path: Optional[str] = None) -> None: | |
| if replay_path is None: | |
| replay_path = './video' | |
| self._replay_path = replay_path | |
| def random_action(self) -> np.ndarray: | |
| random_action = self.action_space.sample() | |
| random_action = to_ndarray([random_action], dtype=np.int64) | |
| return random_action | |
| def observation_space(self) -> gym.spaces.Space: | |
| return self._observation_space | |
| def action_space(self) -> gym.spaces.Space: | |
| return self._action_space | |
| def reward_space(self) -> gym.spaces.Space: | |
| return self._reward_space | |
| def __repr__(self) -> str: | |
| return "DI-engine CartPole Env" | |