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 | from typing import Any
import warnings
from gymnasium import spaces
import numpy as np
import torch
from lit_sb3.buffers.replay import ReplayBuffer
from lit_sb3.core.types import DictReplayBufferSamples
from lit_sb3.vec_env import VecNormalize
try:
import psutil
except ImportError:
psutil = None # ty:ignore[invalid-assignment]
class DictReplayBuffer(ReplayBuffer):
"""Dict Replay buffer used in off-policy algorithms like SAC/TD3.
Extends the ReplayBuffer to use dictionary observations.
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
Disabled for now (see https://github.com/DLR-RM/stable-baselines3/pull/243#discussion_r531535702)
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
"""
observation_space: spaces.Dict
obs_shape: dict[str, tuple[int, ...]]
observations: dict[str, np.ndarray]
next_observations: dict[str, np.ndarray]
def __init__(
self,
buffer_size: int,
observation_space: spaces.Dict,
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(ReplayBuffer, self).__init__(buffer_size, observation_space, action_space, device, n_envs=n_envs)
assert isinstance(self.obs_shape, dict), "DictReplayBuffer must be used with Dict obs space only"
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
assert not optimize_memory_usage, "DictReplayBuffer does not support optimize_memory_usage"
# disabling as this adds quite a bit of complexity
# https://github.com/DLR-RM/stable-baselines3/pull/243#discussion_r531535702
self.optimize_memory_usage = optimize_memory_usage
self.observations = {
key: np.zeros((self.buffer_size, self.n_envs, *_obs_shape), dtype=observation_space[key].dtype)
for key, _obs_shape in self.obs_shape.items()
}
self.next_observations = {
key: np.zeros((self.buffer_size, self.n_envs, *_obs_shape), dtype=observation_space[key].dtype)
for key, _obs_shape in self.obs_shape.items()
}
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:
obs_nbytes = 0
for _, obs in self.observations.items():
obs_nbytes += obs.nbytes
total_memory_usage: float = obs_nbytes + self.actions.nbytes + self.rewards.nbytes + self.dones.nbytes
if not optimize_memory_usage:
next_obs_nbytes = 0
for _, obs in self.observations.items():
next_obs_nbytes += obs.nbytes
total_memory_usage += next_obs_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( # type: ignore[override]
self,
obs: dict[str, np.ndarray],
next_obs: dict[str, np.ndarray],
action: np.ndarray,
reward: np.ndarray,
done: np.ndarray,
infos: list[dict[str, Any]],
) -> None:
# Copy to avoid modification by reference
for key in self.observations:
# 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[key], spaces.Discrete):
obs[key] = obs[key].reshape((self.n_envs,) + self.obs_shape[key])
self.observations[key][self.pos] = np.array(obs[key])
for key in self.next_observations:
if isinstance(self.observation_space.spaces[key], spaces.Discrete):
next_obs[key] = next_obs[key].reshape((self.n_envs,) + self.obs_shape[key])
self.next_observations[key][self.pos] = np.array(next_obs[key])
# Reshape to handle multi-dim and discrete action spaces, see GH #970 #1392
action = action.reshape((self.n_envs, self.action_dim))
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( # type: ignore[override]
self,
batch_size: int,
env: VecNormalize | None = None,
) -> DictReplayBufferSamples:
"""Sample elements from the replay buffer.
Args:
batch_size: Number of element to sample
env: associated gym VecEnv
to normalize the observations/rewards when sampling
"""
return super(ReplayBuffer, self).sample(batch_size=batch_size, env=env) # ty:ignore[invalid-return-type]
def _get_samples( # type: ignore[override]
self,
batch_inds: np.ndarray,
env: VecNormalize | None = None,
) -> DictReplayBufferSamples:
# Sample randomly the env idx
env_indices = np.random.randint(0, high=self.n_envs, size=(len(batch_inds),))
# Normalize if needed and remove extra dimension (we are using only one env for now)
obs_ = self._normalize_obs(
{key: obs[batch_inds, env_indices, :] for key, obs in self.observations.items()}, env
)
next_obs_ = self._normalize_obs(
{key: obs[batch_inds, env_indices, :] for key, obs in self.next_observations.items()}, env
)
assert isinstance(obs_, dict)
assert isinstance(next_obs_, dict)
# Convert to torch tensor
observations = {key: self.to_torch(obs) for key, obs in obs_.items()} # ty:ignore[invalid-argument-type]
next_observations = {key: self.to_torch(obs) for key, obs in next_obs_.items()} # ty:ignore[invalid-argument-type]
return DictReplayBufferSamples(
observations=observations, # ty:ignore[invalid-argument-type]
actions=self.to_torch(self.actions[batch_inds, env_indices]),
next_observations=next_observations, # ty:ignore[invalid-argument-type]
# Only use dones that are not due to timeouts
# deactivated by default (timeouts is initialized as an array of False)
dones=self.to_torch(
self.dones[batch_inds, env_indices] * (1 - self.timeouts[batch_inds, env_indices])
).reshape(-1, 1),
rewards=self.to_torch(self._normalize_reward(self.rewards[batch_inds, env_indices].reshape(-1, 1), env)),
)
|