-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathbuffers.py
184 lines (156 loc) · 8.2 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
import numpy as np
from utils import log_trajectory_statistics
class ReplayBuffer(object):
"""Basic replay buffer."""
def __init__(self, buffer_size, initial_data={}):
self.buffer_size = buffer_size
if initial_data == {}:
self.N = -1
else:
self._initial_setup(initial_data)
def _initial_setup(self, initial_data={}):
self.obs = initial_data['obs'].astype('float32')
self.nobs = initial_data['nobs'].astype('float32')
self.act = initial_data['act'].astype('float32')
self.rew = initial_data['rew'].astype('float32')
self.don = initial_data['don']
self.N = initial_data['n']
def add(self, other_data):
"""Add collected data from Sampler."""
if self.N == -1:
self._initial_setup(other_data)
else:
self.N += other_data['n']
offset_index = int(np.amax(np.array([self.N - self.buffer_size, 0])))
self.N -= offset_index
self.obs = np.concatenate((self.obs[offset_index:],
other_data['obs'].astype('float32')), axis=0)
self.nobs = np.concatenate((self.nobs[offset_index:],
other_data['nobs'].astype('float32')), axis=0)
self.act = np.concatenate((self.act[offset_index:],
other_data['act'].astype('float32')), axis=0)
self.rew = np.concatenate((self.rew[offset_index:],
other_data['rew'].astype('float32')), axis=0)
self.don = np.concatenate((self.don[offset_index:],
other_data['don']), axis=0)
def gather_indices(self, indices):
out_dict = {}
out_dict['obs'], out_dict['act'], out_dict['nobs'] = (self.obs[indices],
self.act[indices],
self.nobs[indices])
out_dict['rew'], out_dict['don'] = self.rew[indices], self.don[indices]
return out_dict
def get_random_batch(self, batch_size):
"""Get random batch of data."""
indices = np.random.randint(self.N, size=batch_size)
return self.gather_indices(indices)
def get_stats_previous_timesteps(self, num_timesteps):
"""Get basic statistics of latest collected data."""
if num_timesteps > self.N:
print('Not enough samples in the buffer')
return
latest_obs = self.obs[-num_timesteps - 1:]
latest_rew = self.rew[-num_timesteps - 1:]
latest_don = self.don[-num_timesteps - 1:]
trajectory_startpoints = np.where(latest_don)[0] + 1
number_startpoints = trajectory_startpoints.shape[0]
trajectory_rewards = []
for i in range(number_startpoints - 1):
trajectory_rewards.append(np.sum(
latest_rew[trajectory_startpoints[i]:trajectory_startpoints[i + 1]]))
if len(trajectory_rewards) == 0:
print('Latest number of completed trajectories - {}'.format(
len(trajectory_rewards)))
else:
log_trajectory_statistics(trajectory_rewards)
class VisualReplayBuffer(ReplayBuffer):
"""Replay buffer with added support for visual observations."""
def __init__(self, buffer_size, initial_data={}):
super(VisualReplayBuffer, self).__init__(buffer_size, initial_data)
def _initial_setup(self, initial_data={}):
super(VisualReplayBuffer, self)._initial_setup(initial_data)
self.ims = initial_data['ims'].astype(np.uint8)
def add(self, other_data):
if self.N == -1:
self._initial_setup(other_data)
else:
self.N += other_data['n']
offset_index = int(np.amax(np.array([self.N - self.buffer_size, 0])))
self.N -= offset_index
self.obs = np.concatenate((self.obs[offset_index:],
other_data['obs'].astype('float32')), axis=0)
self.nobs = np.concatenate((self.nobs[offset_index:],
other_data['nobs'].astype('float32')), axis=0)
self.act = np.concatenate((self.act[offset_index:],
other_data['act'].astype('float32')), axis=0)
self.rew = np.concatenate((self.rew[offset_index:],
other_data['rew'].astype('float32')), axis=0)
self.don = np.concatenate((self.don[offset_index:],
other_data['don']), axis=0)
self.ims = np.concatenate((self.ims[offset_index:],
other_data['ims'].astype(np.uint8)), axis=0)
def gather_indices(self, indices):
"""Get random batch of data."""
out_dict = super(VisualReplayBuffer, self).gather_indices(indices)
out_dict['ims'] = ((self.ims[indices].astype('float32') + 0.5) / 256)
return out_dict
class LearnerAgentReplayBuffer(VisualReplayBuffer):
"""Replay buffer computing calculating the pseudo-rewards from a discriminator."""
def __init__(self, gail, buffer_size, reward_noise=True,
initial_data={}):
super(LearnerAgentReplayBuffer, self).__init__(buffer_size, initial_data)
self._pre = gail._pre
self._disc = gail._disc
self._rn = reward_noise
def get_random_batch(self, batch_size, re_eval_rw=True):
"""Get random batch of data.
Parameters
----------
batch_size : Batch size of experience to collect.
re_eval_rw : Compute pseudo-rewards for batch, default is True.
"""
out = super(LearnerAgentReplayBuffer, self).get_random_batch(
batch_size)
if re_eval_rw:
if self._rn:
out['pre'] = self._pre(out['ims'])
else:
out['pre'], _ = self._pre.get_distribution_info(out['ims'])
out['rew'] = self._disc.get_reward(out['pre'])
return out
class DemonstrationsReplayBuffer(object):
"""Replay buffer efficiently storing priorly collected visual observations."""
def __init__(self, initial_data):
self.ims = initial_data['ims'][:, 0, :, :, :].astype(np.uint8)
self.N = self.ims.shape[0]
self.ids = initial_data['ids']
self.past_frames = initial_data['ims'].shape[1]
self.idx_shifts = np.expand_dims(np.arange(self.past_frames), axis=0)
self.pad_image = (np.zeros_like(self.ims[0]).astype('float32') + 0.5) / 256
_, self.first_indices = np.unique(self.ids, return_index=True)
self.first_ims = initial_data['ims'][self.first_indices, 1, :, :, :].astype(np.uint8)
retrieval_indices = np.arange(self.first_indices.shape[0])
self.padded_retrieval_list = np.zeros([self.N]).astype('int') + 1000000
self.padded_retrieval_list[self.first_indices] = retrieval_indices
def gather_indices(self, indices):
all_indices = np.expand_dims(indices, axis=-1) - self.idx_shifts
images = (self.ims[all_indices].astype('float32') + 0.5) / 256
start_indices_mask = np.isin(all_indices[:, :-1], self.first_indices)
start_indices_x, start_indices_y = np.where(start_indices_mask)
first_indices_y = start_indices_y + 1
trajectories_start_indices = all_indices[start_indices_x, start_indices_y]
images[start_indices_x, first_indices_y] = (self.first_ims[
self.padded_retrieval_list[trajectories_start_indices]].astype(
'float32') + 0.5) / 256
pad_indices_mask = start_indices_mask[:, :-1]
for i in range(self.past_frames - 3):
pad_indices_mask[:, i + 1] = np.logical_or(pad_indices_mask[:, i + 1],
pad_indices_mask[:, i])
pad_indices_x, pad_indices_y = np.where(pad_indices_mask)
pad_indices_y += 2
images[pad_indices_x, pad_indices_y] = self.pad_image
return {'ims': images}
def get_random_batch(self, batch_size):
"""Get random batch of data."""
indices = np.random.randint(self.N, size=batch_size)
return self.gather_indices(indices)