-
Notifications
You must be signed in to change notification settings - Fork 0
/
replay_buffer.py
223 lines (184 loc) · 7.12 KB
/
replay_buffer.py
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
import numpy as np
from typing import Deque, Dict, List, Tuple
from segment_tree import MinSegmentTree, SumSegmentTree
from collections import deque
import random
class ReplayBuffer:
"""A simple numpy replay buffer."""
def __init__(
self,
obs_dim: Tuple[int, ...],
size: int,
batch_size: int = 32,
n_step: int = 1,
gamma: float = 0.99
):
self.obs_buf = np.zeros([size, *obs_dim], dtype=np.float32)
self.next_obs_buf = np.zeros([size, *obs_dim], dtype=np.float32)
self.acts_buf = np.zeros([size], dtype=np.float32)
self.rews_buf = np.zeros([size], dtype=np.float32)
self.done_buf = np.zeros(size, dtype=np.float32)
self.max_size, self.batch_size = size, batch_size
self.ptr, self.size, = 0, 0
# for N-step Learning
self.n_step_buffer = deque(maxlen=n_step)
self.n_step = n_step
self.gamma = gamma
def store(
self,
obs: np.ndarray,
act: np.ndarray,
rew: float,
next_obs: np.ndarray,
done: bool,
) -> Tuple[np.ndarray, np.ndarray, float, np.ndarray, bool]:
transition = (obs, act, rew, next_obs, done)
self.n_step_buffer.append(transition)
# single step transition is not ready
if len(self.n_step_buffer) < self.n_step:
return ()
# make a n-step transition
rew, next_obs, done = self._get_n_step_info(
self.n_step_buffer, self.gamma
)
obs, act = self.n_step_buffer[0][:2]
self.obs_buf[self.ptr] = obs
self.next_obs_buf[self.ptr] = next_obs
self.acts_buf[self.ptr] = act
self.rews_buf[self.ptr] = rew
self.done_buf[self.ptr] = done
self.ptr = (self.ptr + 1) % self.max_size
self.size = min(self.size + 1, self.max_size)
return self.n_step_buffer[0]
def sample_batch(self) -> Dict[str, np.ndarray]:
idxs = np.random.choice(self.size, size=self.batch_size, replace=False)
return dict(
obs=self.obs_buf[idxs],
next_obs=self.next_obs_buf[idxs],
acts=self.acts_buf[idxs],
rews=self.rews_buf[idxs],
done=self.done_buf[idxs],
# for N-step Learning
indices=idxs,
)
def sample_batch_from_idxs(
self, idxs: np.ndarray
) -> Dict[str, np.ndarray]:
# for N-step Learning
return dict(
obs=self.obs_buf[idxs],
next_obs=self.next_obs_buf[idxs],
acts=self.acts_buf[idxs],
rews=self.rews_buf[idxs],
done=self.done_buf[idxs],
)
def _get_n_step_info(
self, n_step_buffer: Deque, gamma: float
) -> Tuple[np.int64, np.ndarray, bool]:
"""Return n step rew, next_obs, and done."""
# info of the last transition
rew, next_obs, done = n_step_buffer[-1][-3:]
for transition in reversed(list(n_step_buffer)[:-1]):
r, n_o, d = transition[-3:]
rew = r + gamma * rew * (1 - d)
next_obs, done = (n_o, d) if d else (next_obs, done)
return rew, next_obs, done
def __len__(self) -> int:
return self.size
class PrioritizedReplayBuffer(ReplayBuffer):
"""Prioritized Replay buffer.
Attributes:
max_priority (float): max priority
tree_ptr (int): next index of tree
alpha (float): alpha parameter for prioritized replay buffer
sum_tree (SumSegmentTree): sum tree for prior
min_tree (MinSegmentTree): min tree for min prior to get max weight
"""
def __init__(
self,
obs_dim: tuple[int, ...],
size: int,
batch_size: int = 32,
alpha: float = 0.6,
n_step: int = 1,
gamma: float = 0.99,
):
"""Initialization."""
assert alpha >= 0
super(PrioritizedReplayBuffer, self).__init__(
obs_dim, size, batch_size, n_step, gamma
)
self.max_priority, self.tree_ptr = 1.0, 0
self.alpha = alpha
# capacity must be positive and a power of 2.
tree_capacity = 1
while tree_capacity < self.max_size:
tree_capacity *= 2
self.sum_tree = SumSegmentTree(tree_capacity)
self.min_tree = MinSegmentTree(tree_capacity)
def store(
self,
obs: np.ndarray,
act: int,
rew: float,
next_obs: np.ndarray,
done: bool,
) -> Tuple[np.ndarray, np.ndarray, float, np.ndarray, bool]:
"""Store experience and priority."""
transition = super().store(obs, act, rew, next_obs, done)
if transition:
self.sum_tree[self.tree_ptr] = self.max_priority ** self.alpha
self.min_tree[self.tree_ptr] = self.max_priority ** self.alpha
self.tree_ptr = (self.tree_ptr + 1) % self.max_size
return transition
def sample_batch(self, beta: float = 0.4) -> Dict[str, np.ndarray]:
"""Sample a batch of experiences."""
assert len(self) >= self.batch_size
assert beta > 0
indices = self._sample_proportional()
obs = self.obs_buf[indices]
next_obs = self.next_obs_buf[indices]
acts = self.acts_buf[indices]
rews = self.rews_buf[indices]
done = self.done_buf[indices]
weights = np.array([self._calculate_weight(i, beta) for i in indices])
return dict(
obs=obs,
next_obs=next_obs,
acts=acts,
rews=rews,
done=done,
weights=weights,
indices=indices,
)
def update_priorities(self, indices: List[int], priorities: np.ndarray):
"""Update priorities of sampled transitions."""
assert len(indices) == len(priorities)
for idx, priority in zip(indices, priorities):
assert priority > 0
assert 0 <= idx < len(self)
self.sum_tree[idx] = priority ** self.alpha
self.min_tree[idx] = priority ** self.alpha
self.max_priority = max(self.max_priority, priority)
def _sample_proportional(self) -> List[int]:
"""Sample indices based on proportions."""
indices = []
p_total = self.sum_tree.sum(0, len(self) - 1)
segment = p_total / self.batch_size
for i in range(self.batch_size):
a = segment * i
b = segment * (i + 1)
upperbound = random.uniform(a, b)
idx = self.sum_tree.retrieve(upperbound)
indices.append(idx)
return indices
def _calculate_weight(self, idx: int, beta: float):
"""Calculate the weight of the experience at idx."""
# get max weight
p_min = self.min_tree.min() / self.sum_tree.sum()
max_weight = (p_min * len(self)) ** (-beta)
# calculate weights
p_sample = self.sum_tree[idx] / self.sum_tree.sum()
weight = (p_sample * len(self)) ** (-beta)
weight = weight / max_weight
return weight