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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345 | from copy import deepcopy
import inspect
import pathlib
from pathlib import Path
import pickle
from typing import Any, Optional, Union
from gymnasium import spaces
import numpy as np
from lit_sb3.core.preprocessing import is_image_space
from lit_sb3.core.running_mean_std import RunningMeanStd
from lit_sb3.utils.spaces import check_shape_equal
from lit_sb3.vec_env.base_vec_env import VecEnv, VecEnvStepReturn, VecEnvWrapper
class VecNormalize(VecEnvWrapper):
"""A moving average, normalizing wrapper for vectorized environment.
has support for saving/loading moving average,.
Args:
venv: the vectorized environment to wrap
training: Whether to update or not the moving average
norm_obs: Whether to normalize observation or not (default: True)
norm_reward: Whether to normalize rewards or not (default: True)
clip_obs: Max absolute value for observation
clip_reward: Max value absolute for discounted reward
gamma: discount factor
epsilon: To avoid division by zero
norm_obs_keys: Which keys from observation dict to normalize.
If not specified, all keys will be normalized.
"""
obs_spaces: dict[str, spaces.Space]
old_obs: np.ndarray | dict[str, np.ndarray]
def __init__(
self,
venv: VecEnv,
training: bool = True,
norm_obs: bool = True,
norm_reward: bool = True,
clip_obs: float = 10.0,
clip_reward: float = 10.0,
gamma: float = 0.99,
epsilon: float = 1e-8,
norm_obs_keys: list[str] | None = None,
) -> None:
VecEnvWrapper.__init__(self, venv)
self.norm_obs = norm_obs
self.norm_obs_keys = norm_obs_keys
# Check observation spaces
if self.norm_obs:
# Note: mypy doesn't take into account the sanity checks, which lead to several type: ignore...
self._sanity_checks()
if isinstance(self.observation_space, spaces.Dict):
self.obs_spaces = self.observation_space.spaces
self.obs_rms = {key: RunningMeanStd(shape=self.obs_spaces[key].shape) for key in self.norm_obs_keys} # type: ignore[arg-type, union-attr]
# Update observation space when using image
# See explanation below and GH #1214
for key in self.obs_rms:
if is_image_space(self.obs_spaces[key]):
self.observation_space.spaces[key] = spaces.Box(
low=-clip_obs,
high=clip_obs,
shape=self.obs_spaces[key].shape,
dtype=np.float32,
)
else:
self.obs_rms = RunningMeanStd(shape=self.observation_space.shape) # type: ignore[assignment, arg-type]
# Update observation space when using image
# See GH #1214
# This is to raise proper error when
# VecNormalize is used with an image-like input and
# normalize_images=True.
# For correctness, we should also update the bounds
# in other cases but this will cause backward-incompatible change
# and break already saved policies.
if is_image_space(self.observation_space):
self.observation_space = spaces.Box(
low=-clip_obs,
high=clip_obs,
shape=self.observation_space.shape,
dtype=np.float32,
)
self.ret_rms = RunningMeanStd(shape=())
self.clip_obs = clip_obs
self.clip_reward = clip_reward
# Returns: discounted rewards
self.returns = np.zeros(self.num_envs)
self.gamma = gamma
self.epsilon = epsilon
self.training = training
self.norm_obs = norm_obs
self.norm_reward = norm_reward
self.old_reward = np.array([])
def _sanity_checks(self) -> None:
"""Check the observations that are going to be normalized are of the correct type (spaces.Box)."""
if isinstance(self.observation_space, spaces.Dict):
# By default, we normalize all keys
if self.norm_obs_keys is None:
self.norm_obs_keys = list(self.observation_space.spaces.keys())
# Check that all keys are of type Box
for obs_key in self.norm_obs_keys:
if not isinstance(self.observation_space.spaces[obs_key], spaces.Box):
raise TypeError(
f"VecNormalize only supports `gym.spaces.Box` observation spaces but {obs_key} "
f"is of type {self.observation_space.spaces[obs_key]}. "
"You should probably explicitly pass the observation keys "
" that should be normalized via the `norm_obs_keys` parameter."
)
elif isinstance(self.observation_space, spaces.Box):
if self.norm_obs_keys is not None:
raise ValueError("`norm_obs_keys` param is applicable only with `gym.spaces.Dict` observation spaces")
else:
raise TypeError(
"VecNormalize only supports `gym.spaces.Box` and `gym.spaces.Dict` observation spaces, "
f"not {self.observation_space}"
)
def __getstate__(self) -> dict[str, Any]:
"""Gets state for pickling.
Excludes self.venv, as in general VecEnv's may not be pickleable.
"""
state = self.__dict__.copy()
# these attributes are not pickleable
del state["venv"]
del state["class_attributes"]
# these attributes depend on the above and so we would prefer not to pickle
del state["returns"]
return state
def __setstate__(self, state: dict[str, Any]) -> None:
"""Restores pickled state.
User must call set_venv() after unpickling before using.
Args:
state:
state: The state to restore.
"""
# Backward compatibility
if "norm_obs_keys" not in state and isinstance(state["observation_space"], spaces.Dict):
state["norm_obs_keys"] = list(state["observation_space"].spaces.keys())
self.__dict__.update(state)
assert "venv" not in state
self.venv = None
def set_venv(self, venv: VecEnv) -> None:
"""Sets the vector environment to wrap to venv.
Also sets attributes derived from this such as `num_env`.
Args:
venv:
venv: The venv parameter.
"""
if self.venv is not None:
raise ValueError("Trying to set venv of already initialized VecNormalize wrapper.")
self.venv = venv
self.num_envs = venv.num_envs
self.class_attributes = dict(inspect.getmembers(self.__class__))
self.render_mode = venv.render_mode
# Check that the observation_space shape match
check_shape_equal(self.observation_space, venv.observation_space)
self.returns = np.zeros(self.num_envs)
def step_wait(self) -> VecEnvStepReturn:
"""Apply sequence of actions to sequence of environments.
actions -> (observations, rewards, dones).
where ``dones`` is a boolean vector indicating whether each element is new.
"""
obs, rewards, dones, infos = self.venv.step_wait() # ty:ignore[unresolved-attribute]
assert isinstance(obs, (np.ndarray, dict)) # for mypy
self.old_obs = obs
self.old_reward = rewards
if self.training and self.norm_obs:
if isinstance(obs, dict) and isinstance(self.obs_rms, dict):
for key in self.obs_rms:
self.obs_rms[key].update(obs[key]) # ty:ignore[invalid-argument-type, unresolved-attribute]
else:
self.obs_rms.update(obs) # ty:ignore[invalid-argument-type, no-matching-overload]
obs = self.normalize_obs(obs)
if self.training:
self._update_reward(rewards)
rewards = self.normalize_reward(rewards)
# Normalize the terminal observations
for idx, done in enumerate(dones):
if not done:
continue
if "terminal_observation" in infos[idx]:
infos[idx]["terminal_observation"] = self.normalize_obs(infos[idx]["terminal_observation"])
self.returns[dones] = 0
return obs, rewards, dones, infos
def _update_reward(self, reward: np.ndarray) -> None:
"""Update reward normalization statistics."""
self.returns = self.returns * self.gamma + reward
self.ret_rms.update(self.returns)
def _normalize_obs(self, obs: np.ndarray, obs_rms: RunningMeanStd) -> np.ndarray:
"""Helper to normalize observation.
Args:
obs:
obs_rms: associated statistics
obs: The current observation.
Returns:
normalized observation
"""
return np.clip((obs - obs_rms.mean) / np.sqrt(obs_rms.var + self.epsilon), -self.clip_obs, self.clip_obs)
def _unnormalize_obs(self, obs: np.ndarray, obs_rms: RunningMeanStd) -> np.ndarray:
"""Helper to unnormalize observation.
Args:
obs:
obs_rms: associated statistics
obs: The current observation.
Returns:
unnormalized observation
"""
return (obs * np.sqrt(obs_rms.var + self.epsilon)) + obs_rms.mean
def normalize_obs(self, obs: np.ndarray | dict[str, np.ndarray]) -> np.ndarray | dict[str, np.ndarray]:
"""Normalize observations using this VecNormalize's observations statistics.
Calling this method does not update statistics.
"""
# Avoid modifying by reference the original object
obs_ = deepcopy(obs)
if self.norm_obs:
if isinstance(obs, dict) and isinstance(self.obs_rms, dict):
assert self.norm_obs_keys is not None
# Only normalize the specified keys
for key in self.norm_obs_keys:
obs_[key] = self._normalize_obs(obs[key], self.obs_rms[key]).astype(np.float32) # type: ignore[call-overload]
else:
assert isinstance(self.obs_rms, RunningMeanStd)
obs_ = self._normalize_obs(obs, self.obs_rms).astype(np.float32) # ty:ignore[invalid-argument-type]
return obs_
def normalize_reward(self, reward: np.ndarray) -> np.ndarray:
"""Normalize rewards using this VecNormalize's rewards statistics.
Calling this method does not update statistics.
"""
if self.norm_reward:
reward = np.clip(reward / np.sqrt(self.ret_rms.var + self.epsilon), -self.clip_reward, self.clip_reward)
# Note: we cast to float32 as it correspond to Python default float type
# This cast is needed because `RunningMeanStd` keeps stats in float64
return reward.astype(np.float32)
def unnormalize_obs(self, obs: np.ndarray | dict[str, np.ndarray]) -> np.ndarray | dict[str, np.ndarray]:
# Avoid modifying by reference the original object
obs_ = deepcopy(obs)
if self.norm_obs:
if isinstance(obs, dict) and isinstance(self.obs_rms, dict):
assert self.norm_obs_keys is not None
for key in self.norm_obs_keys:
obs_[key] = self._unnormalize_obs(obs[key], self.obs_rms[key]) # type: ignore[call-overload]
else:
assert isinstance(self.obs_rms, RunningMeanStd)
obs_ = self._unnormalize_obs(obs, self.obs_rms) # ty:ignore[invalid-argument-type]
return obs_
def unnormalize_reward(self, reward: np.ndarray) -> np.ndarray:
if self.norm_reward:
return reward * np.sqrt(self.ret_rms.var + self.epsilon)
return reward
def get_original_obs(self) -> np.ndarray | dict[str, np.ndarray]:
"""Returns an unnormalized version of the observations from the most recent.
step or reset.
"""
return deepcopy(self.old_obs)
def get_original_reward(self) -> np.ndarray:
"""Returns an unnormalized version of the rewards from the most recent step."""
return self.old_reward.copy()
def reset(self) -> np.ndarray | dict[str, np.ndarray]:
"""Reset all environments.
Returns:
first observation of the episode
"""
obs = self.venv.reset() # ty:ignore[unresolved-attribute]
assert isinstance(obs, (np.ndarray, dict))
self.old_obs = obs
self.returns = np.zeros(self.num_envs)
if self.training and self.norm_obs:
if isinstance(obs, dict) and isinstance(self.obs_rms, dict):
for key in self.obs_rms:
self.obs_rms[key].update(obs[key]) # ty:ignore[invalid-argument-type, unresolved-attribute]
else:
assert isinstance(self.obs_rms, RunningMeanStd)
self.obs_rms.update(obs) # ty:ignore[invalid-argument-type]
return self.normalize_obs(obs)
@staticmethod
def load(load_path: str | Path, venv: VecEnv) -> "VecNormalize":
"""Loads a saved VecNormalize object.
Args:
load_path: the path to load from.
venv: the VecEnv to wrap.
"""
with Path(load_path).open("rb") as file_handler:
vec_normalize = pickle.load(file_handler) # noqa: S301
vec_normalize.set_venv(venv)
return vec_normalize
def save(self, save_path: str | Path) -> None:
"""Save current VecNormalize object with.
all running statistics and settings (e.g. clip_obs).
Args:
save_path: The path to save to
"""
with Path(save_path).open("wb") as file_handler:
pickle.dump(self, file_handler)
|