-
Notifications
You must be signed in to change notification settings - Fork 831
/
envs.py
executable file
·254 lines (202 loc) · 7.81 KB
/
envs.py
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
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
import os
import gym
import numpy as np
import torch
from gym.spaces.box import Box
from stable_baselines3.common.monitor import Monitor
from stable_baselines3.common.atari_wrappers import (
NoopResetEnv, MaxAndSkipEnv, EpisodicLifeEnv, FireResetEnv, WarpFrame, ClipRewardEnv)
from stable_baselines3.common.vec_env import VecEnvWrapper, DummyVecEnv, SubprocVecEnv
from stable_baselines3.common.vec_env.vec_normalize import \
VecNormalize as VecNormalize_
try:
import dm_control2gym
except ImportError:
pass
try:
import roboschool
except ImportError:
pass
try:
import pybullet_envs
except ImportError:
pass
def make_env(env_id, seed, rank, log_dir, allow_early_resets):
def _thunk():
if env_id.startswith("dm"):
_, domain, task = env_id.split('.')
env = dm_control2gym.make(domain_name=domain, task_name=task)
else:
env = gym.make(env_id)
is_atari = hasattr(gym.envs, 'atari') and isinstance(
env.unwrapped, gym.envs.atari.atari_env.AtariEnv)
if is_atari:
env = NoopResetEnv(env, noop_max=30)
env = MaxAndSkipEnv(env, skip=4)
env.seed(seed + rank)
if str(env.__class__.__name__).find('TimeLimit') >= 0:
env = TimeLimitMask(env)
if log_dir is not None:
env = Monitor(
env,
os.path.join(log_dir, str(rank)),
allow_early_resets=allow_early_resets)
if is_atari:
if len(env.observation_space.shape) == 3:
env = EpisodicLifeEnv(env)
if "FIRE" in env.unwrapped.get_action_meanings():
env = FireResetEnv(env)
env = WarpFrame(env, width=84, height=84)
env = ClipRewardEnv(env)
elif len(env.observation_space.shape) == 3:
raise NotImplementedError(
"CNN models work only for atari,\n"
"please use a custom wrapper for a custom pixel input env.\n"
"See wrap_deepmind for an example.")
# If the input has shape (W,H,3), wrap for PyTorch convolutions
obs_shape = env.observation_space.shape
if len(obs_shape) == 3 and obs_shape[2] in [1, 3]:
env = TransposeImage(env, op=[2, 0, 1])
return env
return _thunk
def make_vec_envs(env_name,
seed,
num_processes,
gamma,
log_dir,
device,
allow_early_resets,
num_frame_stack=None):
envs = [
make_env(env_name, seed, i, log_dir, allow_early_resets)
for i in range(num_processes)
]
if len(envs) > 1:
envs = SubprocVecEnv(envs)
else:
envs = DummyVecEnv(envs)
if len(envs.observation_space.shape) == 1:
if gamma is None:
envs = VecNormalize(envs, norm_reward=False)
else:
envs = VecNormalize(envs, gamma=gamma)
envs = VecPyTorch(envs, device)
if num_frame_stack is not None:
envs = VecPyTorchFrameStack(envs, num_frame_stack, device)
elif len(envs.observation_space.shape) == 3:
envs = VecPyTorchFrameStack(envs, 4, device)
return envs
# Checks whether done was caused my timit limits or not
class TimeLimitMask(gym.Wrapper):
def step(self, action):
obs, rew, done, info = self.env.step(action)
if done and self.env._max_episode_steps == self.env._elapsed_steps:
info['bad_transition'] = True
return obs, rew, done, info
def reset(self, **kwargs):
return self.env.reset(**kwargs)
# Can be used to test recurrent policies for Reacher-v2
class MaskGoal(gym.ObservationWrapper):
def observation(self, observation):
if self.env._elapsed_steps > 0:
observation[-2:] = 0
return observation
class TransposeObs(gym.ObservationWrapper):
def __init__(self, env=None):
"""
Transpose observation space (base class)
"""
super(TransposeObs, self).__init__(env)
class TransposeImage(TransposeObs):
def __init__(self, env=None, op=[2, 0, 1]):
"""
Transpose observation space for images
"""
super(TransposeImage, self).__init__(env)
assert len(op) == 3, "Error: Operation, " + str(op) + ", must be dim3"
self.op = op
obs_shape = self.observation_space.shape
self.observation_space = Box(
self.observation_space.low[0, 0, 0],
self.observation_space.high[0, 0, 0], [
obs_shape[self.op[0]], obs_shape[self.op[1]],
obs_shape[self.op[2]]
],
dtype=self.observation_space.dtype)
def observation(self, ob):
return ob.transpose(self.op[0], self.op[1], self.op[2])
class VecPyTorch(VecEnvWrapper):
def __init__(self, venv, device):
"""Return only every `skip`-th frame"""
super(VecPyTorch, self).__init__(venv)
self.device = device
# TODO: Fix data types
def reset(self):
obs = self.venv.reset()
obs = torch.from_numpy(obs).float().to(self.device)
return obs
def step_async(self, actions):
if isinstance(actions, torch.LongTensor):
# Squeeze the dimension for discrete actions
actions = actions.squeeze(1)
actions = actions.cpu().numpy()
self.venv.step_async(actions)
def step_wait(self):
obs, reward, done, info = self.venv.step_wait()
obs = torch.from_numpy(obs).float().to(self.device)
reward = torch.from_numpy(reward).unsqueeze(dim=1).float()
return obs, reward, done, info
class VecNormalize(VecNormalize_):
def __init__(self, *args, **kwargs):
super(VecNormalize, self).__init__(*args, **kwargs)
self.training = True
def _obfilt(self, obs, update=True):
if self.ob_rms:
if self.training and update:
self.ob_rms.update(obs)
obs = np.clip((obs - self.ob_rms.mean) /
np.sqrt(self.ob_rms.var + self.epsilon),
-self.clipob, self.clipob)
return obs
else:
return obs
def train(self):
self.training = True
def eval(self):
self.training = False
# Derived from
# https://github.com/openai/baselines/blob/master/baselines/common/vec_env/vec_frame_stack.py
class VecPyTorchFrameStack(VecEnvWrapper):
def __init__(self, venv, nstack, device=None):
self.venv = venv
self.nstack = nstack
wos = venv.observation_space # wrapped ob space
self.shape_dim0 = wos.shape[0]
low = np.repeat(wos.low, self.nstack, axis=0)
high = np.repeat(wos.high, self.nstack, axis=0)
if device is None:
device = torch.device('cpu')
self.stacked_obs = torch.zeros((venv.num_envs, ) +
low.shape).to(device)
observation_space = gym.spaces.Box(
low=low, high=high, dtype=venv.observation_space.dtype)
VecEnvWrapper.__init__(self, venv, observation_space=observation_space)
def step_wait(self):
obs, rews, news, infos = self.venv.step_wait()
self.stacked_obs[:, :-self.shape_dim0] = \
self.stacked_obs[:, self.shape_dim0:].clone()
for (i, new) in enumerate(news):
if new:
self.stacked_obs[i] = 0
self.stacked_obs[:, -self.shape_dim0:] = obs
return self.stacked_obs, rews, news, infos
def reset(self):
obs = self.venv.reset()
if torch.backends.cudnn.deterministic:
self.stacked_obs = torch.zeros(self.stacked_obs.shape)
else:
self.stacked_obs.zero_()
self.stacked_obs[:, -self.shape_dim0:] = obs
return self.stacked_obs
def close(self):
self.venv.close()