-
Notifications
You must be signed in to change notification settings - Fork 2
/
gym_wrapper.py
203 lines (182 loc) · 7.83 KB
/
gym_wrapper.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
import gym
import numpy as np
import threading
class FakeMultiThread(threading.Thread):
def __init__(self, func, args=()):
super().__init__()
self.func = func
self.args = args
def run(self):
self.result = self.func(*self.args)
def get_result(self):
try:
return self.result
except Exception:
return None
class gym_envs(object):
def __init__(self, gym_env_name, n, render_mode='first'):
'''
Input:
gym_env_name: gym training environment id, i.e. CartPole-v0
n: environment number
render_mode: mode of rendering, optional: first, last, all, random_[num] -> i.e. random_2, [list] -> i.e. [0, 2, 4]
'''
self.n = n # environments number
self.envs = [gym.make(gym_env_name) for _ in range(self.n)]
# process observation
self.obs_space = self.envs[0].observation_space
if isinstance(self.obs_space, gym.spaces.box.Box):
self.obs_high = self.obs_space.high
self.obs_low = self.obs_space.low
self.obs_type = 'visual' if len(self.obs_space.shape) == 3 else 'vector'
self.reward_threshold = self.envs[0].env.spec.reward_threshold # reward threshold refer to solved
# process action
self.action_space = self.envs[0].action_space
if isinstance(self.action_space, gym.spaces.box.Box):
self.action_type = 'continuous'
self.action_high = self.action_space.high
self.action_low = self.action_space.low
elif isinstance(self.action_space, gym.spaces.tuple.Tuple):
self.action_type = 'Tuple(Discrete)'
else:
self.action_type = 'discrete'
self.action_mu, self.action_sigma = self._get_action_normalize_factor()
self._get_render_index(render_mode)
def _get_render_index(self, render_mode):
'''
get render windows list, i.e. [0, 1] when there are 4 training enviornment.
'''
assert isinstance(render_mode, (list, str)), 'render_mode must have type of str or list.'
if isinstance(render_mode, list):
assert all([isinstance(i, int) for i in render_mode]), 'items in render list must have type of int'
assert min(index) >= 0, 'index must larger than zero'
assert max(index) <= self.n, 'render index cannot larger than environment number.'
self.render_index = render_mode
elif isinstance(render_mode, str):
if render_mode == 'first':
self.render_index = [0]
elif render_mode == 'last':
self.render_index = [-1]
elif render_mode == 'all':
self.render_index = [i for i in range(self.n)]
else:
a, b = render_mode.split('_')
if a == 'random' and 0 < int(b) <= self.n:
import random
self.render_index = random.sample([i for i in range(self.n)], int(b))
else:
raise Exception('render_mode must be first, last, all, [list] or random_[num]')
def render(self):
'''
render game windows.
'''
[self.envs[i].render() for i in self.render_index]
def close(self):
'''
close all environments.
'''
[env.close() for env in self.envs]
def sample_action(self):
'''
generate ramdom actions for all training environment.
'''
return np.array([env.action_space.sample() for env in self.envs])
def reset(self):
self.dones_index = []
threadpool = []
for i in range(self.n):
th = FakeMultiThread(self.envs[i].reset, args=())
threadpool.append(th)
for th in threadpool:
th.start()
for th in threadpool:
threading.Thread.join(th)
obs = np.array([threadpool[i].get_result() for i in range(self.n)])
obs = self._maybe_one_hot(obs)
return obs
# if self.obs_type == 'visual':
# return np.array([threadpool[i].get_result()[np.newaxis, :] for i in range(self.n)])
# else:
# return np.array([threadpool[i].get_result() for i in range(self.n)])
def step(self, actions, scale=True):
if scale == True:
actions = self.action_sigma * actions + self.action_mu
if self.action_type == 'discrete':
actions = actions.reshape(-1,)
elif self.action_type == 'Tuple(Discrete)':
actions = actions.reshape(self.n, -1).tolist()
threadpool = []
for i in range(self.n):
th = FakeMultiThread(self.envs[i].step, args=(actions[i], ))
threadpool.append(th)
for th in threadpool:
th.start()
for th in threadpool:
threading.Thread.join(th)
results = [threadpool[i].get_result() for i in range(self.n)]
# if self.obs_type == 'visual':
# results = [
# [threadpool[i].get_result()[0][np.newaxis, :], *threadpool[i].get_result()[1:]]
# for i in range(self.n)]
# else:
# results = [threadpool[i].get_result() for i in range(self.n)]
obs, reward, done, info = [np.array(e) for e in zip(*results)]
obs = self._maybe_one_hot(obs)
self.dones_index = np.where(done)[0]
return obs, reward, done, info
def partial_reset(self):
threadpool = []
for i in self.dones_index:
th = FakeMultiThread(self.envs[i].reset, args=())
threadpool.append(th)
for th in threadpool:
th.start()
for th in threadpool:
threading.Thread.join(th)
obs = np.array([threadpool[i].get_result() for i in range(self.dones_index.shape[0])])
obs = self._maybe_one_hot(obs, is_partial=True)
return obs
# if self.obs_type == 'visual':
# return np.array([threadpool[i].get_result()[np.newaxis, :] for i in range(self.dones_index.shape[0])])
# else:
# return np.array([threadpool[i].get_result() for i in range(self.dones_index.shape[0])])
def _get_action_normalize_factor(self):
'''
get action mu and sigma. mu: action bias. sigma: action scale
input:
self.action_low: [-2, -3],
self.action_high: [2, 6]
return:
mu: [0, 1.5],
sigma: [2, 4.5]
'''
if self.action_type == 'continuous':
return (self.action_high + self.action_low) / 2, (self.action_high - self.action_low) / 2
else:
return 0, 1
def _maybe_one_hot(self, obs, is_partial=False):
"""
Change discrete observation from list(int) to list(one_hot) format.
for example:
action: [[1, 0], [2, 1]]
observation space: [3, 4]
environment number: 2
then, output: [[0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0.]]
"""
obs_number = len(self.dones_index) if is_partial else self.n
if hasattr(self.obs_space, 'n'):
obs = obs.reshape(obs_number, -1)
if isinstance(self.obs_space.n, (int, np.int32)):
dim = [int(self.obs_space.n)]
else:
dim = list(self.obs_space.n) # 在CliffWalking-v0环境其类型为numpy.int32
multiplication_factor = dim[1:] + [1]
n = np.array(dim).prod()
ints = obs.dot(multiplication_factor)
x = np.zeros([obs.shape[0], n])
for i, j in enumerate(ints):
x[i, j] = 1
return x
else:
return obs