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 | from typing import Any
import warnings
from gymnasium import spaces
import numpy as np
import torch
from lit_sb3.buffers.base import BaseBuffer
from lit_sb3.core.types import ReplayBufferSamples
from lit_sb3.vec_env import VecNormalize
try:
import psutil
except ImportError:
psutil = None # ty:ignore[invalid-assignment]
class ReplayBuffer(BaseBuffer):
"""Replay buffer used in off-policy algorithms like SAC/TD3.
Args:
buffer_size: Max number of element in the buffer
observation_space: Observation space
action_space: Action space
device: PyTorch device
n_envs: Number of parallel environments
optimize_memory_usage: Enable a memory efficient variant
of the replay buffer which reduces by almost a factor two the memory used,
at a cost of more complexity.
See https://github.com/DLR-RM/stable-baselines3/issues/37#issuecomment-637501195
and https://github.com/DLR-RM/stable-baselines3/pull/28#issuecomment-637559274
Cannot be used in combination with handle_timeout_termination.
handle_timeout_termination: Handle timeout termination (due to timelimit)
separately and treat the task as infinite horizon task.
https://github.com/DLR-RM/stable-baselines3/issues/284
"""
observations: np.ndarray
next_observations: np.ndarray
actions: np.ndarray
rewards: np.ndarray
dones: np.ndarray
timeouts: np.ndarray
def __init__(
self,
buffer_size: int,
observation_space: spaces.Space,
action_space: spaces.Space,
device: torch.device | str = "auto",
n_envs: int = 1,
optimize_memory_usage: bool = False,
handle_timeout_termination: bool = True,
) -> None:
super().__init__(buffer_size, observation_space, action_space, device, n_envs=n_envs)
# Adjust buffer size
self.buffer_size = max(buffer_size // n_envs, 1)
# Check that the replay buffer can fit into the memory
if psutil is not None:
mem_available = psutil.virtual_memory().available
# there is a bug if both optimize_memory_usage and handle_timeout_termination are true
# see https://github.com/DLR-RM/stable-baselines3/issues/934
if optimize_memory_usage and handle_timeout_termination:
raise ValueError(
"ReplayBuffer does not support optimize_memory_usage = True "
"and handle_timeout_termination = True simultaneously."
)
self.optimize_memory_usage = optimize_memory_usage
self.observations = np.zeros((self.buffer_size, self.n_envs, *self.obs_shape), dtype=observation_space.dtype)
if not optimize_memory_usage:
# When optimizing memory, `observations` contains also the next observation
self.next_observations = np.zeros(
(self.buffer_size, self.n_envs, *self.obs_shape), dtype=observation_space.dtype
)
self.actions = np.zeros(
(self.buffer_size, self.n_envs, self.action_dim),
dtype=self._maybe_cast_dtype(action_space.dtype), # ty:ignore[invalid-argument-type]
)
self.rewards = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
self.dones = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
# Handle timeouts termination properly if needed
# see https://github.com/DLR-RM/stable-baselines3/issues/284
self.handle_timeout_termination = handle_timeout_termination
self.timeouts = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
if psutil is not None:
total_memory_usage: float = (
self.observations.nbytes + self.actions.nbytes + self.rewards.nbytes + self.dones.nbytes
)
if not optimize_memory_usage:
total_memory_usage += self.next_observations.nbytes
if total_memory_usage > mem_available:
# Convert to GB
total_memory_usage /= 1e9
mem_available /= 1e9
warnings.warn(
"This system does not have apparently enough memory to store the complete "
f"replay buffer {total_memory_usage:.2f}GB > {mem_available:.2f}GB",
stacklevel=2,
)
def add(
self,
obs: np.ndarray,
next_obs: np.ndarray,
action: np.ndarray,
reward: np.ndarray,
done: np.ndarray,
infos: list[dict[str, Any]],
) -> None:
# Reshape needed when using multiple envs with discrete observations
# as numpy cannot broadcast (n_discrete,) to (n_discrete, 1)
if isinstance(self.observation_space, spaces.Discrete):
obs = obs.reshape((self.n_envs, *self.obs_shape))
next_obs = next_obs.reshape((self.n_envs, *self.obs_shape))
# Reshape to handle multi-dim and discrete action spaces, see GH #970 #1392
action = action.reshape((self.n_envs, self.action_dim))
# Copy to avoid modification by reference
self.observations[self.pos] = np.array(obs)
if self.optimize_memory_usage:
self.observations[(self.pos + 1) % self.buffer_size] = np.array(next_obs)
else:
self.next_observations[self.pos] = np.array(next_obs)
self.actions[self.pos] = np.array(action)
self.rewards[self.pos] = np.array(reward)
self.dones[self.pos] = np.array(done)
if self.handle_timeout_termination:
self.timeouts[self.pos] = np.array([info.get("TimeLimit.truncated", False) for info in infos])
self.pos += 1
if self.pos == self.buffer_size:
self.full = True
self.pos = 0
def sample(self, batch_size: int, env: VecNormalize | None = None) -> ReplayBufferSamples:
"""Sample elements from the replay buffer.
Custom sampling when using memory efficient variant,
as we should not sample the element with index `self.pos`
See https://github.com/DLR-RM/stable-baselines3/pull/28#issuecomment-637559274.
Args:
batch_size: Number of element to sample
env: associated gym VecEnv
to normalize the observations/rewards when sampling
"""
if not self.optimize_memory_usage:
return super().sample(batch_size=batch_size, env=env) # ty:ignore[invalid-return-type]
# Do not sample the element with index `self.pos` as the transitions is invalid
# (we use only one array to store `obs` and `next_obs`)
if self.full:
batch_inds = (np.random.randint(1, self.buffer_size, size=batch_size) + self.pos) % self.buffer_size
else:
batch_inds = np.random.randint(0, self.pos, size=batch_size)
return self._get_samples(batch_inds, env=env)
def _get_samples(self, batch_inds: np.ndarray, env: VecNormalize | None = None) -> ReplayBufferSamples:
# Sample randomly the env idx
env_indices = np.random.randint(0, high=self.n_envs, size=(len(batch_inds),))
if self.optimize_memory_usage:
next_obs = self._normalize_obs(self.observations[(batch_inds + 1) % self.buffer_size, env_indices, :], env)
else:
next_obs = self._normalize_obs(self.next_observations[batch_inds, env_indices, :], env)
data = (
self._normalize_obs(self.observations[batch_inds, env_indices, :], env),
self.actions[batch_inds, env_indices, :],
next_obs,
# Only use dones that are not due to timeouts
# deactivated by default (timeouts is initialized as an array of False)
(self.dones[batch_inds, env_indices] * (1 - self.timeouts[batch_inds, env_indices])).reshape(-1, 1),
self._normalize_reward(self.rewards[batch_inds, env_indices].reshape(-1, 1), env),
)
return ReplayBufferSamples(*tuple(map(self.to_torch, data))) # ty:ignore[invalid-argument-type]
@staticmethod
def _maybe_cast_dtype(dtype: np.typing.DTypeLike) -> np.typing.DTypeLike:
"""Cast `np.float64` action datatype to `np.float32`,.
keep the others dtype unchanged.
See GH#1572 for more information.
Args:
dtype: The original action space dtype
Returns:
``np.float32`` if the dtype was float64,
the original dtype otherwise.
"""
if dtype == np.float64:
return np.float32
return dtype
|