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 | from collections.abc import Generator
from gymnasium import spaces
import numpy as np
import torch
from lit_sb3.buffers.rollout import RolloutBuffer
from lit_sb3.core.types import DictRolloutBufferSamples
from lit_sb3.vec_env import VecNormalize
class DictRolloutBuffer(RolloutBuffer):
"""Dict Rollout buffer used in on-policy algorithms like A2C/PPO.
Extends the RolloutBuffer to use dictionary observations.
It corresponds to ``buffer_size`` transitions collected
using the current policy.
This experience will be discarded after the policy update.
In order to use PPO objective, we also store the current value of each state
and the log probability of each taken action.
The term rollout here refers to the model-free notion and should not
be used with the concept of rollout used in model-based RL or planning.
Hence, it is only involved in policy and value function training but not action selection.
Args:
buffer_size: Max number of element in the buffer
observation_space: Observation space
action_space: Action space
device: PyTorch device
gae_lambda: Factor for trade-off of bias vs variance for Generalized Advantage Estimator
Equivalent to Monte-Carlo advantage estimate when set to 1.
gamma: Discount factor
n_envs: Number of parallel environments
"""
observation_space: spaces.Dict
obs_shape: dict[str, tuple[int, ...]]
observations: dict[str, np.ndarray]
def __init__(
self,
buffer_size: int,
observation_space: spaces.Dict,
action_space: spaces.Space,
device: torch.device | str = "auto",
gae_lambda: float = 1,
gamma: float = 0.99,
n_envs: int = 1,
) -> None:
super(RolloutBuffer, self).__init__(buffer_size, observation_space, action_space, device, n_envs=n_envs)
assert isinstance(self.obs_shape, dict), "DictRolloutBuffer must be used with Dict obs space only"
self.gae_lambda = gae_lambda
self.gamma = gamma
self.generator_ready = False
self.reset()
def reset(self) -> None:
self.observations = {}
for key, obs_input_shape in self.obs_shape.items():
self.observations[key] = np.zeros(
(self.buffer_size, self.n_envs, *obs_input_shape), dtype=self.observation_space[key].dtype
)
self.actions = np.zeros((self.buffer_size, self.n_envs, self.action_dim), dtype=self.action_space.dtype)
self.rewards = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
self.returns = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
self.episode_starts = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
self.values = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
self.log_probs = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
self.advantages = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
self.generator_ready = False
super(RolloutBuffer, self).reset()
def add( # type: ignore[override]
self,
obs: dict[str, np.ndarray],
action: np.ndarray,
reward: np.ndarray,
episode_start: np.ndarray,
value: torch.Tensor,
log_prob: torch.Tensor,
) -> None:
"""Args.
obs: Observation
action: Action
reward:
episode_start: Start of episode signal.
value: estimated value of the current state
following the current policy.
log_prob: log probability of the action
following the current policy.
"""
if len(log_prob.shape) == 0:
# Reshape 0-d tensor to avoid error
log_prob = log_prob.reshape(-1, 1)
for key in self.observations:
obs_ = np.array(obs[key])
# 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_ = obs_.reshape((self.n_envs,) + self.obs_shape[key])
self.observations[key][self.pos] = obs_
# 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.episode_starts[self.pos] = np.array(episode_start)
self.values[self.pos] = value.clone().cpu().numpy().flatten()
self.log_probs[self.pos] = log_prob.clone().cpu().numpy()
self.pos += 1
if self.pos == self.buffer_size:
self.full = True
def get( # type: ignore[override]
self,
batch_size: int | None = None,
) -> Generator[DictRolloutBufferSamples, None, None]:
assert self.full, ""
indices = np.random.permutation(self.buffer_size * self.n_envs)
# Prepare the data
if not self.generator_ready:
for key, obs in self.observations.items():
self.observations[key] = self.swap_and_flatten(obs)
_tensor_names = ["actions", "values", "log_probs", "advantages", "returns"]
for tensor in _tensor_names:
self.__dict__[tensor] = self.swap_and_flatten(self.__dict__[tensor])
self.generator_ready = True
# Return everything, don't create minibatches
if batch_size is None:
batch_size = self.buffer_size * self.n_envs
start_idx = 0
while start_idx < self.buffer_size * self.n_envs:
yield self._get_samples(indices[start_idx : start_idx + batch_size])
start_idx += batch_size
def _get_samples( # type: ignore[override]
self,
batch_inds: np.ndarray,
env: VecNormalize | None = None,
) -> DictRolloutBufferSamples:
return DictRolloutBufferSamples(
observations={key: self.to_torch(obs[batch_inds]) for (key, obs) in self.observations.items()},
# Cast to float32 (backward compatible), this would lead to RuntimeError for MultiBinary space
actions=self.to_torch(self.actions[batch_inds].astype(np.float32, copy=False)),
old_values=self.to_torch(self.values[batch_inds].flatten()),
old_log_prob=self.to_torch(self.log_probs[batch_inds].flatten()),
advantages=self.to_torch(self.advantages[batch_inds].flatten()),
returns=self.to_torch(self.returns[batch_inds].flatten()),
)
|