-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathexperience.py
149 lines (126 loc) · 6.26 KB
/
experience.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
from typing import Tuple
import torch
class ExperienceBatch:
def __init__(self,
observations,
actions,
prev_actions,
action_log_probs,
returns,
value_predictions,
advantage_targets,
masks,
recurrent_hidden_states):
num_steps, num_envs = actions.shape[:2]
self.observations = self._flatten(observations, num_steps, num_envs)
self.actions = self._flatten(actions, num_steps, num_envs)
self.prev_actions = self._flatten(prev_actions, num_steps, num_envs)
self.action_log_probs = self._flatten(action_log_probs, num_steps, num_envs)
self.returns = self._flatten(returns, num_steps, num_envs)
self.value_predictions = self._flatten(value_predictions, num_steps, num_envs)
self.advantage_targets = self._flatten(advantage_targets, num_steps, num_envs)
self.masks = self._flatten(masks, num_steps, num_envs)
self.recurrent_hidden_states = recurrent_hidden_states.view(num_envs, -1)
def action_eval_input(self):
return (self.observations,
self.recurrent_hidden_states,
self.masks,
self.prev_actions,
self.actions)
def _flatten(self, tensor, num_steps, num_envs):
return tensor.view(num_steps * num_envs, *tensor.shape[2:])
class ExperienceStorage:
def __init__(self,
num_steps: int,
num_envs: int,
observation_shape: Tuple,
recurrent_hidden_size: int,
device: torch.device):
self._num_steps = num_steps
self._num_envs = num_envs
self._step = 0
self._device = device
self.observations = torch.zeros(num_steps + 1,
num_envs,
*observation_shape,
dtype=torch.uint8).to(device)
self.actions = torch.zeros(num_steps, num_envs, 1, dtype=torch.long).to(device)
self.action_log_probs = torch.zeros(num_steps, num_envs, 1).to(device)
self.rewards = torch.zeros(num_steps, num_envs, 1).to(device)
self.value_predictions = torch.zeros(num_steps + 1, num_envs, 1).to(device)
self.returns = torch.zeros(num_steps + 1, num_envs, 1).to(device)
self.masks = torch.ones(num_steps + 1, num_envs, 1).to(device)
self.recurrent_hidden_states = torch.zeros(
num_steps + 1, num_envs, recurrent_hidden_size).to(device)
def insert(self,
observations,
actions,
action_log_probs,
rewards,
value_predictions,
masks,
recurrent_hidden_states):
self.observations[self._step + 1].copy_(observations)
self.actions[self._step].copy_(actions)
self.action_log_probs[self._step].copy_(action_log_probs)
self.rewards[self._step].copy_(rewards)
self.value_predictions[self._step].copy_(value_predictions)
self.masks[self._step + 1].copy_(masks)
self.recurrent_hidden_states[self._step + 1].copy_(recurrent_hidden_states)
self._step = (self._step + 1) % self._num_steps
def insert_initial_observations(self, observations):
self.observations[0].copy_(observations)
def get_actor_input(self, step):
states = self.observations[step]
rnn_hxs = self.recurrent_hidden_states[step]
masks = self.masks[step]
prev_actions = self.get_prev_actions(step)
return states, rnn_hxs, masks, prev_actions
def get_prev_actions(self, step, last_n=4):
prev_action_indices = [step - i for i in range(1, last_n + 1)]
prev_actions = self.actions[prev_action_indices, :].permute(1, 0, 2)
return prev_actions
def get_critic_input(self):
return self.get_actor_input(step=-1)
def compute_gae_returns(self, next_value, gamma, gae_lambda):
self.value_predictions[-1] = next_value
gae = 0
for step in reversed(range(self.rewards.size(0))):
delta = self.rewards[step] + \
gamma * self.value_predictions[step + 1] * self.masks[step + 1] - \
self.value_predictions[step]
gae = delta + gamma * gae_lambda * self.masks[step + 1] * gae
self.returns[step] = gae + self.value_predictions[step]
def after_update(self):
self.observations[0].copy_(self.observations[-1])
self.recurrent_hidden_states[0].copy_(self.recurrent_hidden_states[-1])
self.masks[0].copy_(self.masks[-1])
def compute_advantages(self, eps: float = 1e-5):
advantages = self.returns[:-1] - self.value_predictions[:-1]
norm_advantages = (advantages - advantages.mean()) / (advantages.std() + eps)
return norm_advantages
def batches(self, advantages: torch.tensor, minibatches: int):
"""Yield experience batches for recurrent policy training."""
assert (self._num_envs % minibatches) == 0
num_envs_per_batch = self._num_envs // minibatches
random_env_indices = torch.randperm(self._num_envs)
for start in range(0, self._num_envs, num_envs_per_batch):
end = start + num_envs_per_batch
indices = random_env_indices[start:end]
prev_actions_shape = (self._num_steps, num_envs_per_batch, 4, 1)
prev_actions = torch.zeros(*prev_actions_shape,
dtype=torch.long).to(self._device)
for step in range(self._num_steps):
actions = self.get_prev_actions(step)
prev_actions[step, :] = actions[indices]
yield ExperienceBatch(
observations=self.observations[:-1, indices],
actions=self.actions[:, indices],
prev_actions=prev_actions,
action_log_probs=self.action_log_probs[:, indices],
returns=self.returns[:-1, indices],
value_predictions=self.value_predictions[:-1, indices],
advantage_targets=advantages[:, indices],
masks=self.masks[:-1, indices],
recurrent_hidden_states=self.recurrent_hidden_states[:1, indices]
)