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 | from collections.abc import Generator
from typing import NamedTuple
from gymnasium import spaces
import numpy as np
import torch
from lit_sb3.buffers.dict_rollout import DictRolloutBuffer
from lit_sb3.core.types import RNNStates, TensorDict
from lit_sb3.utils.sequence import create_sequencers
from lit_sb3.vec_env import VecNormalize
class RecurrentDictRolloutBufferSamples(NamedTuple):
observations: TensorDict
actions: torch.Tensor
old_values: torch.Tensor
old_log_prob: torch.Tensor
advantages: torch.Tensor
returns: torch.Tensor
lstm_states: RNNStates
episode_starts: torch.Tensor
mask: torch.Tensor
class RecurrentDictRolloutBuffer(DictRolloutBuffer):
"""Dict Rollout buffer used in on-policy algorithms like A2C/PPO.
Extends the RecurrentRolloutBuffer to use dictionary observations.
Args:
buffer_size: Max number of element in the buffer
observation_space: Observation space
action_space: Action space
hidden_state_shape: Shape of the buffer that will collect lstm states
device: PyTorch device
gae_lambda: Factor for trade-off of bias vs variance for Generalized Advantage Estimator
Equivalent to classic advantage when set to 1.
gamma: Discount factor
n_envs: Number of parallel environments
"""
def __init__(
self,
buffer_size: int,
observation_space: spaces.Space,
action_space: spaces.Space,
hidden_state_shape: tuple[int, int, int, int],
device: torch.device | str = "auto",
gae_lambda: float = 1,
gamma: float = 0.99,
n_envs: int = 1,
) -> None:
self.hidden_state_shape = hidden_state_shape
self.seq_start_indices, self.seq_end_indices = None, None
super().__init__(buffer_size, observation_space, action_space, device, gae_lambda, gamma, n_envs=n_envs) # ty:ignore[invalid-argument-type]
def reset(self) -> None:
super().reset()
self.hidden_states_pi = np.zeros(self.hidden_state_shape, dtype=np.float32)
self.cell_states_pi = np.zeros(self.hidden_state_shape, dtype=np.float32)
self.hidden_states_vf = np.zeros(self.hidden_state_shape, dtype=np.float32)
self.cell_states_vf = np.zeros(self.hidden_state_shape, dtype=np.float32)
def add(self, *args, lstm_states: RNNStates, **kwargs) -> None: # ty:ignore[invalid-method-override]
"""Args.
hidden_states: LSTM cell and hidden state.
"""
self.hidden_states_pi[self.pos] = np.array(lstm_states.pi[0].cpu().numpy())
self.cell_states_pi[self.pos] = np.array(lstm_states.pi[1].cpu().numpy())
self.hidden_states_vf[self.pos] = np.array(lstm_states.vf[0].cpu().numpy())
self.cell_states_vf[self.pos] = np.array(lstm_states.vf[1].cpu().numpy())
super().add(*args, **kwargs)
def get(self, batch_size: int | None = None) -> Generator[RecurrentDictRolloutBufferSamples, None, None]: # ty:ignore[invalid-method-override]
assert self.full, "Rollout buffer must be full before sampling from it"
# Prepare the data
if not self.generator_ready:
# hidden_state_shape = (self.n_steps, lstm.num_layers, self.n_envs, lstm.hidden_size)
# swap first to (self.n_steps, self.n_envs, lstm.num_layers, lstm.hidden_size)
for tensor in ["hidden_states_pi", "cell_states_pi", "hidden_states_vf", "cell_states_vf"]:
self.__dict__[tensor] = self.__dict__[tensor].swapaxes(1, 2)
for key, obs in self.observations.items():
self.observations[key] = self.swap_and_flatten(obs)
for tensor in [
"actions",
"values",
"log_probs",
"advantages",
"returns",
"hidden_states_pi",
"cell_states_pi",
"hidden_states_vf",
"cell_states_vf",
"episode_starts",
]:
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
# Trick to shuffle a bit: keep the sequence order
# but split the indices in two
split_index = np.random.randint(self.buffer_size * self.n_envs)
indices = np.arange(self.buffer_size * self.n_envs)
indices = np.concatenate((indices[split_index:], indices[:split_index]))
env_change = np.zeros(self.buffer_size * self.n_envs).reshape(self.buffer_size, self.n_envs)
# Flag first timestep as change of environment
env_change[0, :] = 1.0
env_change = self.swap_and_flatten(env_change)
start_idx = 0
while start_idx < self.buffer_size * self.n_envs:
batch_inds = indices[start_idx : start_idx + batch_size]
yield self._get_samples(batch_inds, env_change)
start_idx += batch_size
def _get_samples(
self,
batch_inds: np.ndarray,
env_change: np.ndarray,
env: VecNormalize | None = None,
) -> RecurrentDictRolloutBufferSamples: # ty:ignore[invalid-method-override]
# Retrieve sequence starts and utility function
self.seq_start_indices, self.pad, self.pad_and_flatten = create_sequencers(
self.episode_starts[batch_inds], env_change[batch_inds], self.device
)
n_seq = len(self.seq_start_indices)
max_length = self.pad(self.actions[batch_inds]).shape[1]
padded_batch_size = n_seq * max_length
# We retrieve the lstm hidden states that will allow
# to properly initialize the LSTM at the beginning of each sequence
lstm_states_pi = (
# (n_envs * n_steps, n_layers, dim) -> (n_layers, n_seq, dim)
self.hidden_states_pi[batch_inds][self.seq_start_indices].swapaxes(0, 1),
self.cell_states_pi[batch_inds][self.seq_start_indices].swapaxes(0, 1),
)
lstm_states_vf = (
# (n_envs * n_steps, n_layers, dim) -> (n_layers, n_seq, dim)
self.hidden_states_vf[batch_inds][self.seq_start_indices].swapaxes(0, 1),
self.cell_states_vf[batch_inds][self.seq_start_indices].swapaxes(0, 1),
)
lstm_states_pi = (self.to_torch(lstm_states_pi[0]).contiguous(), self.to_torch(lstm_states_pi[1]).contiguous())
lstm_states_vf = (self.to_torch(lstm_states_vf[0]).contiguous(), self.to_torch(lstm_states_vf[1]).contiguous())
observations = {key: self.pad(obs[batch_inds]) for (key, obs) in self.observations.items()}
observations = {
key: obs.reshape((padded_batch_size,) + self.obs_shape[key]) for (key, obs) in observations.items()
}
return RecurrentDictRolloutBufferSamples(
observations=observations,
actions=self.pad(self.actions[batch_inds]).reshape((padded_batch_size, *self.actions.shape[1:])),
old_values=self.pad_and_flatten(self.values[batch_inds]),
old_log_prob=self.pad_and_flatten(self.log_probs[batch_inds]),
advantages=self.pad_and_flatten(self.advantages[batch_inds]),
returns=self.pad_and_flatten(self.returns[batch_inds]),
lstm_states=RNNStates(lstm_states_pi, lstm_states_vf),
episode_starts=self.pad_and_flatten(self.episode_starts[batch_inds]),
mask=self.pad_and_flatten(np.ones_like(self.returns[batch_inds])),
)
|