-
Notifications
You must be signed in to change notification settings - Fork 62
/
Copy pathstorage.py
203 lines (172 loc) · 8.43 KB
/
storage.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
# The following code is largely borrowed from:
# https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail/blob/master/a2c_ppo_acktr/storage.py
from collections import namedtuple
import numpy as np
import torch
from torch.utils.data.sampler import BatchSampler, SubsetRandomSampler
def _flatten_helper(T, N, _tensor):
return _tensor.view(T * N, *_tensor.size()[2:])
class RolloutStorage(object):
def __init__(self, num_steps, num_processes, obs_shape, action_space,
rec_state_size):
if action_space.__class__.__name__ == 'Discrete':
self.n_actions = 1
action_type = torch.long
else:
self.n_actions = action_space.shape[0]
action_type = torch.float32
self.obs = torch.zeros(num_steps + 1, num_processes, *obs_shape)
self.rec_states = torch.zeros(num_steps + 1, num_processes,
rec_state_size)
self.rewards = torch.zeros(num_steps, num_processes)
self.value_preds = torch.zeros(num_steps + 1, num_processes)
self.returns = torch.zeros(num_steps + 1, num_processes)
self.action_log_probs = torch.zeros(num_steps, num_processes)
self.actions = torch.zeros((num_steps, num_processes, self.n_actions),
dtype=action_type)
self.masks = torch.ones(num_steps + 1, num_processes)
self.num_steps = num_steps
self.step = 0
self.has_extras = False
self.extras_size = None
def to(self, device):
self.obs = self.obs.to(device)
self.rec_states = self.rec_states.to(device)
self.rewards = self.rewards.to(device)
self.value_preds = self.value_preds.to(device)
self.returns = self.returns.to(device)
self.action_log_probs = self.action_log_probs.to(device)
self.actions = self.actions.to(device)
self.masks = self.masks.to(device)
if self.has_extras:
self.extras = self.extras.to(device)
return self
def insert(self, obs, rec_states, actions, action_log_probs, value_preds,
rewards, masks):
self.obs[self.step + 1].copy_(obs)
self.rec_states[self.step + 1].copy_(rec_states)
self.actions[self.step].copy_(actions.view(-1, self.n_actions))
self.action_log_probs[self.step].copy_(action_log_probs)
self.value_preds[self.step].copy_(value_preds)
self.rewards[self.step].copy_(rewards)
self.masks[self.step + 1].copy_(masks)
self.step = (self.step + 1) % self.num_steps
def after_update(self):
self.obs[0].copy_(self.obs[-1])
self.rec_states[0].copy_(self.rec_states[-1])
self.masks[0].copy_(self.masks[-1])
if self.has_extras:
self.extras[0].copy_(self.extras[-1])
def compute_returns(self, next_value, use_gae, gamma, tau):
if use_gae:
self.value_preds[-1] = next_value
gae = 0
for step in reversed(range(self.rewards.size(0))):
delta = self.rewards[step] + gamma \
* self.value_preds[step + 1] * self.masks[step + 1] \
- self.value_preds[step]
gae = delta + gamma * tau * self.masks[step + 1] * gae
self.returns[step] = gae + self.value_preds[step]
else:
self.returns[-1] = next_value
for step in reversed(range(self.rewards.size(0))):
self.returns[step] = self.returns[step + 1] * gamma \
* self.masks[step + 1] + self.rewards[step]
def feed_forward_generator(self, advantages, num_mini_batch):
num_steps, num_processes = self.rewards.size()[0:2]
batch_size = num_processes * num_steps
mini_batch_size = batch_size // num_mini_batch
assert batch_size >= num_mini_batch, (
"PPO requires the number of processes ({}) "
"* number of steps ({}) = {} "
"to be greater than or equal to "
"the number of PPO mini batches ({})."
"".format(num_processes, num_steps, num_processes * num_steps,
num_mini_batch))
sampler = BatchSampler(SubsetRandomSampler(range(batch_size)),
mini_batch_size, drop_last=False)
for indices in sampler:
yield {
'obs': self.obs[:-1].view(-1, *self.obs.size()[2:])[indices],
'rec_states': self.rec_states[:-1].view(
-1, self.rec_states.size(-1))[indices],
'actions': self.actions.view(-1, self.n_actions)[indices],
'value_preds': self.value_preds[:-1].view(-1)[indices],
'returns': self.returns[:-1].view(-1)[indices],
'masks': self.masks[:-1].view(-1)[indices],
'old_action_log_probs': self.action_log_probs.view(-1)[indices],
'adv_targ': advantages.view(-1)[indices],
'extras': self.extras[:-1].view(
-1, self.extras_size)[indices]
if self.has_extras else None,
}
def recurrent_generator(self, advantages, num_mini_batch):
num_processes = self.rewards.size(1)
assert num_processes >= num_mini_batch, (
"PPO requires the number of processes ({}) "
"to be greater than or equal to the number of "
"PPO mini batches ({}).".format(num_processes, num_mini_batch))
num_envs_per_batch = num_processes // num_mini_batch
perm = torch.randperm(num_processes)
T, N = self.num_steps, num_envs_per_batch
for start_ind in range(0, num_processes, num_envs_per_batch):
obs = []
rec_states = []
actions = []
value_preds = []
returns = []
masks = []
old_action_log_probs = []
adv_targ = []
if self.has_extras:
extras = []
for offset in range(num_envs_per_batch):
ind = perm[start_ind + offset]
obs.append(self.obs[:-1, ind])
rec_states.append(self.rec_states[0:1, ind])
actions.append(self.actions[:, ind])
value_preds.append(self.value_preds[:-1, ind])
returns.append(self.returns[:-1, ind])
masks.append(self.masks[:-1, ind])
old_action_log_probs.append(self.action_log_probs[:, ind])
adv_targ.append(advantages[:, ind])
if self.has_extras:
extras.append(self.extras[:-1, ind])
# These are all tensors of size (T, N, ...)
obs = torch.stack(obs, 1)
actions = torch.stack(actions, 1)
value_preds = torch.stack(value_preds, 1)
returns = torch.stack(returns, 1)
masks = torch.stack(masks, 1)
old_action_log_probs = torch.stack(old_action_log_probs, 1)
adv_targ = torch.stack(adv_targ, 1)
if self.has_extras:
extras = torch.stack(extras, 1)
yield {
'obs': _flatten_helper(T, N, obs),
'actions': _flatten_helper(T, N, actions),
'value_preds': _flatten_helper(T, N, value_preds),
'returns': _flatten_helper(T, N, returns),
'masks': _flatten_helper(T, N, masks),
'old_action_log_probs': _flatten_helper(
T, N, old_action_log_probs),
'adv_targ': _flatten_helper(T, N, adv_targ),
'extras': _flatten_helper(
T, N, extras) if self.has_extras else None,
'rec_states': torch.stack(rec_states, 1).view(N, -1),
}
class GlobalRolloutStorage(RolloutStorage):
def __init__(self, num_steps, num_processes, obs_shape, action_space,
rec_state_size, extras_size):
super(GlobalRolloutStorage, self).__init__(
num_steps, num_processes, obs_shape, action_space, rec_state_size)
self.extras = torch.zeros((num_steps + 1, num_processes, extras_size),
dtype=torch.long)
self.has_extras = True
self.extras_size = extras_size
def insert(self, obs, rec_states, actions, action_log_probs, value_preds,
rewards, masks, extras):
self.extras[self.step + 1].copy_(extras)
super(GlobalRolloutStorage, self).insert(
obs, rec_states, actions,
action_log_probs, value_preds, rewards, masks)