-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathbuffers.py
235 lines (209 loc) · 10.1 KB
/
buffers.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
224
225
226
227
228
229
230
231
232
233
234
235
import numpy as np
class FastReplayBuffer(object):
"""Replay buffer model."""
def __init__(self, buffer_size, initial_data={}, visual_data=True,
stacked_frames=3, ims_channels=3):
self.buffer_size = buffer_size
self.visual_data = visual_data
if initial_data == {}:
self.index = -1
else:
self._initial_setup(initial_data)
self.stacked_frames = stacked_frames
self.ims_channels = ims_channels
self.full = False
def gather_indices(self, indices):
return {'obs': self.obs[indices], 'nobs': self.nobs[indices], 'act': self.act[indices],
'rew': self.rew[indices], 'don': self.don[indices]}
def _initial_setup(self, initial_data={}):
self.index = 0
self.obs_shape = initial_data['obs'].shape[1:]
self.act_shape = initial_data['act'].shape[1:]
if len(self.obs_shape) == 0:
raise NotImplementedError
else:
self.obs = np.zeros([self.buffer_size, *self.obs_shape], dtype=np.float32)
self.nobs = np.zeros([self.buffer_size, *self.obs_shape], dtype=np.float32)
self.act = np.zeros([self.buffer_size, *self.act_shape], dtype=np.float32)
self.rew = np.zeros([self.buffer_size], dtype=np.float32)
self.don = np.zeros([self.buffer_size], dtype=np.float32)
self.first = np.zeros([self.buffer_size], dtype=np.bool_)
if self.visual_data:
self.ims_shape = initial_data['ims'].shape[1:]
self.ims = np.zeros([self.buffer_size, *self.ims_shape], dtype=np.uint8)
def add_data_point(self, index, data):
np.copyto(self.obs[index], data['obs'])
np.copyto(self.nobs[index], data['nobs'])
np.copyto(self.act[index], data['act'])
np.copyto(self.rew[index], data['rew'])
np.copyto(self.don[index], data['don'])
np.copyto(self.first[index], data['first'])
if self.visual_data:
np.copyto(self.ims[index], data['ims'])
def add_data_batch(self, indexes, data):
self.obs[indexes] = data['obs']
self.nobs[indexes] = data['nobs']
self.act[indexes] = data['act']
self.rew[indexes] = data['rew']
self.don[indexes] = data['don']
self.first[indexes] = data['first']
if self.visual_data:
self.ims[indexes] = data['ims']
def add_data_batch_indexes(self, indexes, data_indexes, data):
self.obs[indexes] = data['obs'][data_indexes]
self.nobs[indexes] = data['nobs'][data_indexes]
self.act[indexes] = data['act'][data_indexes]
self.rew[indexes] = data['rew'][data_indexes]
self.don[indexes] = data['don'][data_indexes]
self.first[indexes] = data['first'][data_indexes]
if self.visual_data:
self.ims[indexes] = data['ims'][data_indexes]
def add_frame(self, frame):
if self.index == -1:
self._initial_setup(frame)
self.add_data_point(index=self.index, data=frame)
self.index += 1
if self.index == self.buffer_size:
self.index = 0
self.full = True
self.first[self.index] = True
def add(self, other_data):
if self.index == -1:
self._initial_setup(other_data)
end_index = self.index + other_data['n']
if end_index > self.buffer_size:
distance_to_end = self.buffer_size - self.index
self.add_data_batch_indexes(indexes=np.arange(start=self.index, stop=self.buffer_size),
data_indexes=np.arange(distance_to_end), data=other_data)
remainder_index = end_index - self.buffer_size
self.add_data_batch_indexes(indexes=np.arange(remainder_index),
data_indexes=np.arange(start=distance_to_end, stop=other_data['n']),
data=other_data)
self.index = remainder_index
self.full = True
else:
self.add_data_batch(indexes=np.arange(start=self.index, stop=end_index), data=other_data)
self.index = end_index
if self.index == self.buffer_size:
self.index = 0
self.full = True
self.first[self.index] = True
def split_ims(self, ims):
nims = ims[..., :self.ims_channels * self.stacked_frames]
ims = ims[..., -self.ims_channels * self.stacked_frames:]
return nims, ims
def get_random_batch(self, batch_size):
if self.full:
indices = np.random.randint(self.buffer_size, size=batch_size)
else:
indices = np.random.randint(self.index, size=batch_size)
return self.gather_indices(indices)
def gather_n_steps_indices_slow(self, indices, n):
n_samples = indices.shape[0]
obs = self.obs[indices]
nobses = np.empty((n_samples, n, *self.obs_shape))
acts = np.empty((n_samples, n, *self.act_shape))
rews = np.empty((n_samples, n))
dones = np.empty((n_samples, n))
mask = np.empty((n_samples, n))
if self.visual_data:
imses = np.empty((n_samples, n, *self.ims_shape))
for i in range(n):
if i == 0:
mask[:, 0] = 1 - self.first[indices]
else:
mask[:, i] = mask[:, i - 1] * (1 - self.first[indices + i])
nobses[:, i] = self.nobs[indices + i]
acts[:, i] = self.act[indices + i]
rews[:, i] = self.rew[indices + i]
dones[:, i] = self.don[indices + i]
if self.visual_data:
imses[:, i] = self.ims[indices + i]
imses, nimses = self.split_ims(imses)
ims = imses[0]
return {'obs': obs, 'nobses': nobses, 'acts': acts, 'rews': rews,
'dones': dones, 'ims': ims, 'nimses': nimses, 'mask': mask}
def gather_n_steps_indices(self, indices, n):
n_samples = indices.shape[0]
gather_ranges = np.stack([np.arange(indices[i], indices[i] + n)
for i in range(n_samples)], axis=0) % self.buffer_size
obs = self.obs[indices]
nobses = self.nobs[gather_ranges]
acts = self.act[gather_ranges]
rews = self.rew[gather_ranges]
dones = self.don[gather_ranges]
mask = 1 - self.first[gather_ranges]
mask[0] = 1
for i in range(n - 2):
mask[:, i + 2] = mask[:, i + 1] * mask[:, i + 2]
if self.visual_data:
imses = self.ims[gather_ranges]
imses, nimses = self.split_ims(imses)
ims = imses[:, 0]
return {'obs': obs, 'nobses': nobses, 'acts': acts, 'rews': rews,
'dones': dones, 'ims': ims, 'nimses': nimses, 'mask': mask}
return {'obs': obs, 'nobses': nobses, 'acts': acts, 'rews': rews,
'dones': dones, 'mask': mask}
def gather_n_steps_actions(self, indices, n):
n_samples = indices.shape[0]
gather_ranges = np.stack([np.arange(indices[i], indices[i] + n)
for i in range(n_samples)], axis=0)
acts = self.act[gather_ranges]
mask = 1 - self.first[gather_ranges]
return {'acts': acts, 'mask': mask}
def get_n_steps_random_batch(self, batch_size, n):
if self.full:
indices = np.random.randint(self.buffer_size, size=batch_size)
else:
indices = np.random.randint(self.index, size=batch_size)
return self.gather_n_steps_indices(indices, n)
def get_n_steps_random_actions_batch(self, batch_size, n):
if self.full:
indices = np.random.randint(self.buffer_size, size=batch_size)
else:
indices = np.random.randint(self.index, size=batch_size)
return self.gather_n_steps_actions(indices, n)
def gather_n_steps_indices_for_representations(self, indices, n):
assert self.visual_data
n_samples = indices.shape[0]
gather_ranges = np.stack([np.arange(indices[i], indices[i] + n)
for i in range(n_samples)], axis=0)
obs = self.obs[indices]
nobs = self.nobs[indices]
acts = self.act[gather_ranges]
act = acts[:, 0]
rew = self.rew[indices]
done = self.don[indices]
mask = 1 - self.first[gather_ranges]
for i in range(n - 1):
mask[:, i + 1] = mask[:, i] * mask[:, i + 1]
imses = self.ims[gather_ranges]
imses, nimses = self.split_ims(imses)
ims = imses[:, 0]
nims = nimses[:, 0]
return {'obs': obs, 'nobs': nobs, 'act': act, 'acts': acts, 'rew': rew,
'done': done, 'ims': ims, 'nims': nims, 'nimses': nimses, 'mask': mask}
class FiGARReplayBuffer(FastReplayBuffer):
"""Replay buffer model, storing FiGAR data."""
def __init__(self, buffer_size, initial_data={}):
super(FiGARReplayBuffer, self).__init__(buffer_size=buffer_size,
initial_data=initial_data,
visual_data=False)
def gather_indices(self, indices):
out_dict = super(FiGARReplayBuffer, self).gather_indices(indices=indices)
out_dict['reps'] = self.reps[indices]
return out_dict
def _initial_setup(self, initial_data={}):
super(FiGARReplayBuffer, self)._initial_setup(initial_data=initial_data)
self.reps = np.zeros([self.buffer_size], dtype=np.int32)
def add_data_point(self, index, data):
super(FiGARReplayBuffer, self).add_data_point(index=index, data=data)
np.copyto(self.reps[index], data['reps'])
def add_data_batch(self, indexes, data):
super(FiGARReplayBuffer, self).add_data_batch(indexes=indexes, data=data)
self.reps[indexes] = data['reps']
def add_data_batch_indexes(self, indexes, data_indexes, data):
super(FiGARReplayBuffer, self).add_data_batch_indexes(indexes=indexes,
data_indexes=data_indexes,
data=data)
self.reps[indexes] = data['reps'][data_indexes]