Merge pull request #12 from opendilab/gl-dev

update version 0.3.0

CHANGELOG

@@ -1,3 +1,17 @@
+## v0.3.0 (2022.1.7)
+- Add MetaDrive support (env, entry, doc)
+- Add simulator selection when import
+- Rename all common modules without 'carla'
+- Add MetaDrive basic env training
+- Add MetaDrive Macro env training and test
+- Add all env registry
+- Update DriveEnvWrapper
+- Update logo
+- Update doc, add metadrive guide and tutorial
+- Reformat simple rl, all policies now have common entry(train eval test)
+- Fix LBC planner bug
+- Delete config helper
+
 ## v0.2.2 (2021.12.3)
 - Add fail count and retry for benchmark collector
 - Add LBC Image training phase 0 & 1

README.md

@@ -1,8 +1,8 @@
 # DI-drive
 
-<img src="./docs/figs/DI-drive.png" width="200" alt="icon"/>
+<img src="./docs/figs/di-drive_logo.png" width="200" alt="icon"/>
 
-Updated on 2021.12.3 DI-drive-v0.2.2 (beta)
+Updated on 2022.1.7 DI-drive-v0.3.0 (beta)
 
 DI-drive - Decision Intelligence Platform for Autonomous Driving simulation.
 
@@ -24,16 +24,18 @@ and most importantly, to **put these all together!**
 
 **DI-drive** uses [DI-engine](https://github.com/opendilab/DI-engine), a Reinforcement Learning
 platform to build most of the running modules and demos. **DI-drive** currently supports [Carla](http://carla.org),
-an open-source Autonomous Drining simulator to operate driving simualtion.
+an open-source Autonomous Drining simulator to operate driving simualtion, and [MetaDrive](https://decisionforce.github.io/metadrive/),
+a diverse driving scenarios for Generalizable Reinforcement Learning. Users can specify any of them to run in global config under `core`.
 
 ## Installation
 
-**DI-drive** needs to run [Carla](http://carla.org) server for simulation. Besides, The client needs to have the following modules installed:
+**DI-drive** needs to have the following modules installed:
 
 - Pytorch
 - DI-engine
-- Carla Python API
 
+[MetaDrive](https://decisionforce.github.io/metadrive/) can be easily installed via `pip`.
+If [Carla](http://carla.org) server is used for simulation, users need to install 'Carla Python API' in addition.
 Please refer to the [documentation](https://opendilab.github.io/DI-drive/) for details about installation and user guide of **DI-drive**.
 We provide IL and RL tutorials, and full guidance for quick run existing policy for beginners.
 
@@ -52,6 +54,7 @@ Please refer to [FAQ](https://opendilab.github.io/DI-drive/faq/index.html) for f
 - BeV Speed RL
 - [Implicit Affordance](https://arxiv.org/abs/1911.10868)
 - [Latent DRL](https://arxiv.org/abs/2001.08726)
+- MetaDrive Macro RL
 
 ## DI-drive Casezoo
 
@@ -71,10 +74,12 @@ DI-engine released under the Apache 2.0 license.
 
 ## Citation
 
+```latex
 @misc{didrive,
     title={{DI-drive: OpenDILab} Decision Intelligence platform for Autonomous Driving simulation},
     author={DI-drive Contributors},
     publisher = {GitHub},
-    howpublished = {\url{`https://github.com/opendilab/DI-drive`}},
+    howpublished = {\url{https://github.com/opendilab/DI-drive}},
     year={2021},
 }
+```

core/__init__.py

@@ -1,6 +1,19 @@
 __TITLE__ = "DI-drive"
-__VERSION__ = "0.2.2"
+__VERSION__ = "0.3.0"
 __DESCRIPTION__ = "Decision AI Auto-Driving Platform"
 __AUTHOR__ = "OpenDILab Contributors"
 __AUTHOR_EMAIL__ = "[email protected]"
 __version__ = __VERSION__
+
+SIMULATORS = ['carla', 'metadrive']
+
+# if 'carla' in SIMULATORS:
+#     try:
+#         import carla
+#     except:
+#         raise ImportError("Import carla failed! Please install carla Python API first.")
+# if 'metadrive' in SIMULATORS:
+#     try:
+#         import metadrive
+#     except:
+#         raise ImportError("Import metadrive failed! Please install metadrive simulator first.")

core/data/base_collector.py

@@ -3,7 +3,7 @@
 from typing import Any, Dict
 from easydict import EasyDict
 
-from core.utils.others.config_helper import deep_merge_dicts
+from ding.utils.default_helper import deep_merge_dicts
 from ding.utils import EasyTimer
 
 

core/data/lbc_dataset.py

@@ -60,6 +60,8 @@ def __init__(
         self._lmdb_list = []
         self._idx_list = []
 
+        n_episodes = 0
+
         for item in epi_folder:
             lmdb_file = lmdb.open(os.path.join(root_dir, item, 'measurements.lmdb')).begin(write=False)
             max_len = int(lmdb_file.get('len'.encode())) - self._gap * self._n_step
@@ -68,12 +70,22 @@ def __init__(
             ]
             png_files.sort()
             for i in range(max_len):
+                if self._max_frames and len(self) >= self._max_frames:
+                    break
+
                 png_file = png_files[i]
                 index = int(png_file.split('_')[1].split('.')[0])
                 self._img_list.append(os.path.join(root_dir, item, png_file))
                 self._idx_list.append(index)
                 self._lmdb_list.append(lmdb_file)
 
+            n_episodes += 1
+
+            if self._max_frames and len(self) >= self._max_frames:
+                break
+
+        print('%s: %d frames, %d episodes.' % (root_dir, len(self), n_episodes))
+
     def __len__(self):
         return len(self._img_list)
 
@@ -142,7 +154,7 @@ def __getitem__(self, index) -> Any:
             locations.append([pixel_x, pixel_y])
             orientations.append([ori_dx, ori_dy])
 
-        #birdview = self.bird_view_transform(birdview)
+        birdview = self.bird_view_transform(birdview)
 
         # Create mask
         output_size = self._crop_size // self._down_ratio
@@ -205,6 +217,8 @@ def __init__(
         self._lmdb_list = []
         self._idx_list = []
 
+        count = 0
+
         for item in epi_folder:
             lmdb_file = lmdb.open(os.path.join(root_dir, item, 'measurements.lmdb')).begin(write=False)
             max_len = int(lmdb_file.get('len'.encode())) - self._gap * self._n_step
@@ -219,6 +233,10 @@ def __init__(
                 self._idx_list.append(index)
                 self._lmdb_list.append(lmdb_file)
 
+            count += max_len
+
+        print('Finished loading %s. Length: %d' % (root_dir, count))
+
     def __len__(self):
         return len(self._img_list)
 
@@ -298,7 +316,9 @@ def __getitem__(self, index) -> Any:
             #     import pdb; pdb.set_trace()
             rgb_images = torch.stack([self.rgb_transform(img) for img in rgb_images])
 
-        # birdview = self.bird_view_transform(birdview)
+        birdview = self.bird_view_transform(birdview)
+
+        self._batch_read_number += 1
 
         return {
             'rgb': rgb_images,

core/envs/__init__.py

@@ -2,7 +2,32 @@
 Copyright 2021 OpenDILab. All Rights Reserved:
 Description:
 '''
-from .base_carla_env import BaseCarlaEnv
-from .carla_env_wrapper import CarlaEnvWrapper, BenchmarkEnvWrapper
-from .simple_carla_env import SimpleCarlaEnv
-from .scenario_carla_env import ScenarioCarlaEnv
+from gym.envs.registration import register, registry
+from core import SIMULATORS
+from .base_drive_env import BaseDriveEnv
+from .drive_env_wrapper import DriveEnvWrapper, BenchmarkEnvWrapper
+
+envs = []
+env_map = {}
+
+if 'carla' in SIMULATORS:
+    from .simple_carla_env import SimpleCarlaEnv
+    from .scenario_carla_env import ScenarioCarlaEnv
+    env_map.update({
+        "SimpleCarla-v1": 'core.envs.simple_carla_env.SimpleCarlaEnv',
+        "ScenarioCarla-v1": 'core.envs.scenario_carla_env.ScenarioCarlaEnv'
+    })
+
+if 'metadrive' in SIMULATORS:
+    from .md_macro_env import MetaDriveMacroEnv
+    env_map.update({
+        "Macro-v1": 'core.envs.md_macro_env:MetaDriveMacroEnv',
+    })
+
+for k, v in env_map.items():
+    if k not in registry.env_specs:
+        envs.append(k)
+        register(id=k, entry_point=v)
+
+if len(envs) > 0:
+    print("[ENV] Register environments: {}.".format(envs))

core/envs/base_carla_env.py → core/envs/base_drive_env.py

@@ -10,16 +10,16 @@
 import gym
 from gym import utils
 
-from core.utils.others.config_helper import deep_merge_dicts
+from ding.utils.default_helper import deep_merge_dicts
 
 
-class BaseCarlaEnv(gym.Env, utils.EzPickle):
+class BaseDriveEnv(gym.Env, utils.EzPickle):
     """
     Base class for environments. It is inherited from `gym.Env` and uses the same interfaces.
-    All Carla Env class is supposed to inherit from this class.
+    All Drive Env class is supposed to inherit from this class.
 
     Note:
-        To run Reinforcement Learning on DI-engine platform, the environment should be wrapped with `CarlaEnvWrapper`.
+        To run Reinforcement Learning on DI-engine platform, the environment should be wrapped with `DingEnvWrapper`.
 
     :Arguments:
         - cfg (Dict): Config Dict.

core/envs/carla_env_wrapper.py → core/envs/drive_env_wrapper.py

@@ -7,28 +7,28 @@
 
 from core.data.benchmark import ALL_SUITES
 from core.eval.carla_benchmark_evaluator import get_suites_list, read_pose_txt, get_benchmark_dir
-from .base_carla_env import BaseCarlaEnv
-from core.utils.others.config_helper import deep_merge_dicts
+from .base_drive_env import BaseDriveEnv
+from ding.utils.default_helper import deep_merge_dicts
 from ding.envs.env.base_env import BaseEnvTimestep, BaseEnvInfo
 from ding.envs.common.env_element import EnvElementInfo
 from ding.torch_utils.data_helper import to_ndarray
 
 
-class CarlaEnvWrapper(gym.Wrapper):
+class DriveEnvWrapper(gym.Wrapper):
     """
     Environment wrapper to make ``gym.Env`` align with DI-engine definitions, so as to use utilities in DI-engine.
     It changes ``step``, ``reset`` and ``info`` method of ``gym.Env``, while others are straightly delivered.
 
     :Arguments:
-        - env (BaseCarlaEnv): The environment to be wrapped.
+        - env (BaseDriveEnv): The environment to be wrapped.
         - cfg (Dict): Config dict.
 
     :Interfaces: reset, step, info, render, seed, close
     """
 
     config = dict()
 
-    def __init__(self, env: BaseCarlaEnv, cfg: Dict = None, **kwargs) -> None:
+    def __init__(self, env: BaseDriveEnv, cfg: Dict = None, **kwargs) -> None:
         if cfg is None:
             self._cfg = self.__class__.default_config()
         elif 'cfg_type' not in cfg:
@@ -47,7 +47,9 @@ def reset(self, *args, **kwargs) -> Any:
             Any: Observations from environment
         """
         obs = self.env.reset(*args, **kwargs)
-        obs = to_ndarray(obs)
+        obs = to_ndarray(obs, dtype=np.float32)
+        if isinstance(obs, np.ndarray) and len(obs.shape) == 3:
+            obs = obs.transpose((2, 0, 1))
         self._final_eval_reward = 0.0
         return obs
 
@@ -68,12 +70,19 @@ def step(self, action: Any = None) -> BaseEnvTimestep:
 
         obs, rew, done, info = self.env.step(action)
         self._final_eval_reward += rew
-        obs = to_ndarray(obs)
+        obs = to_ndarray(obs, dtype=np.float32)
+        if isinstance(obs, np.ndarray) and len(obs.shape) == 3:
+            obs = obs.transpose((2, 0, 1))
         rew = to_ndarray([rew], dtype=np.float32)
         if done:
             info['final_eval_reward'] = self._final_eval_reward
         return BaseEnvTimestep(obs, rew, done, info)
 
+    def seed(self, seed: int, dynamic_seed: bool = True) -> None:
+        self._seed = seed
+        self._dynamic_seed = dynamic_seed
+        np.random.seed(self._seed)
+
     def info(self) -> BaseEnvInfo:
         """
         Interface of ``info`` method to suit DI-engine format env.
@@ -83,9 +92,25 @@ def info(self) -> BaseEnvInfo:
         :Returns:
             BaseEnvInfo: Env information instance defined in DI-engine.
         """
-        obs_space = self.env.observation_space
-        act_space = self.env.action_space
-        return BaseEnvInfo(agent_num=1, obs_space=obs_space, act_space=act_space, use_wrappers=None)
+        obs_space = EnvElementInfo(shape=self.env.observation_space, value={'min': 0., 'max': 1., 'dtype': np.float32})
+        act_space = EnvElementInfo(
+            shape=self.env.action_space,
+            value={
+                'min': np.float32("-inf"),
+                'max': np.float32("inf"),
+                'dtype': np.float32
+            },
+        )
+        rew_space = EnvElementInfo(
+            shape=1,
+            value={
+                'min': np.float32("-inf"),
+                'max': np.float32("inf")
+            },
+        )
+        return BaseEnvInfo(
+            agent_num=1, obs_space=obs_space, act_space=act_space, rew_space=rew_space, use_wrappers=None
+        )
 
     def enable_save_replay(self, replay_path: Optional[str] = None) -> None:
         if replay_path is None:
@@ -106,15 +131,15 @@ def render(self):
         self.env.render()
 
 
-class BenchmarkEnvWrapper(CarlaEnvWrapper):
+class BenchmarkEnvWrapper(DriveEnvWrapper):
     """
     Environment Wrapper for Carla Benchmark suite evaluations. It wraps an environment with Benchmark
     suite so that the env will always run with a benchmark suite setting. It has 2 mode to get reset
     params in a suite: 'random' will randomly get reset param, 'order' will get all reset params in
     order.
 
     :Arguments:
-        - env (BaseCarlaEnv): The environment to be wrapped.
+        - env (BaseDriveEnv): The environment to be wrapped.
         - cfg (Dict): Config dict.
     """
 
@@ -124,7 +149,7 @@ class BenchmarkEnvWrapper(CarlaEnvWrapper):
         mode='random',
     )
 
-    def __init__(self, env: BaseCarlaEnv, cfg: Dict, **kwargs) -> None:
+    def __init__(self, env: BaseDriveEnv, cfg: Dict, **kwargs) -> None:
         super().__init__(env, cfg=cfg, **kwargs)
         suite = self._cfg.suite
         benchmark_dir = self._cfg.benchmark_dir
@@ -208,9 +233,9 @@ def step(self, action: Dict) -> Any:
 
 
 # TODO: complete scenario env wrapper
-class ScenarioEnvWrapper(CarlaEnvWrapper):
+class ScenarioEnvWrapper(DriveEnvWrapper):
 
     config = dict()
 
-    def __init__(self, env: BaseCarlaEnv, cfg: Dict, **kwargs) -> None:
+    def __init__(self, env: BaseDriveEnv, cfg: Dict, **kwargs) -> None:
         super().__init__(env, cfg=cfg, **kwargs)