-
Notifications
You must be signed in to change notification settings - Fork 4
/
vec_env.py
183 lines (166 loc) · 6.72 KB
/
vec_env.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
import numpy as np
from multiprocessing import Process, Pipe
import scipy.misc
import os
# Processes Doom screen image to produce cropped and resized image.
def process_frame(frame):
s = frame[10:-10,30:-30]
s = scipy.misc.imresize(s,[84,84])
s = np.reshape(s,[np.prod(s.shape)]) / 255.0
return s
def worker(remote, parent_remote, env_fn_wrapper):
parent_remote.close()
env = env_fn_wrapper.x
# prev_agent_health = 0
# prev_agent_ammo = 0
log_file = None
if env_fn_wrapper.log_file != "":
log_file = open(env_fn_wrapper.log_file, 'w')
get_bin = lambda x, n: format(x, 'b').zfill(n)
total_reward = 0.0
episode_reward = 0.0
episode_cnt = 0.0
total_episode_cnt = 0
total_kills = 0.0
episode_kills = 0.0
game_variables = None
while True:
cmd, data = remote.recv()
if data is None:
import random
data = random.randint(0, 2**env.get_available_buttons_size() - 1)
action = [True if i == '1' else False for i in get_bin(data, env.get_available_buttons_size())]
if cmd == 'step':
reward = env.make_action(action)
if not env.is_episode_finished():
ob = process_frame(env.get_state().screen_buffer)
game_variables = env.get_state().game_variables
episode_kills = game_variables[2]
# agent_health = env.get_state().game_variables[0]
# agent_ammo = env.get_state().game_variables[1]
# if prev_agent_health > agent_health: # we add a penalty if the agent is hit
# reward = reward - 0 # having a penalty doesn't seem to help
# if prev_agent_ammo > agent_ammo: # we add a penalty if the agent fires
# reward = reward - 0
# prev_agent_health = agent_health
# prev_agent_ammo = agent_ammo
reward = reward / 100.0 # normalizing the reward
episode_reward += reward
done = env.is_episode_finished()
if done:
total_kills += episode_kills
env.new_episode()
ob = process_frame(env.get_state().screen_buffer)
total_reward += episode_reward
episode_cnt += 1
total_episode_cnt += 1
episode_reward = 0.0
remote.send((ob, reward, done, 0.0))
elif cmd == 'gv':
remote.send(game_variables)
elif cmd == 'log':
if log_file is None:
continue
if episode_cnt == 0.0:
continue
avg_reward = round(total_reward / episode_cnt, 5)
log_file.write(str(total_episode_cnt) + ', ' + str(avg_reward) + '\n')
log_file.flush()
total_reward = 0.0
episode_cnt = 0.0
total_kills = 0.0
elif cmd == 'reset':
env.new_episode()
ob = process_frame(env.get_state().screen_buffer)
remote.send(ob)
elif cmd == 'reset_task':
print ('reset_task: Not implemented')
raise NotImplementedError
elif cmd == 'close':
print ('Terminating doom environment')
if log_file is not None:
log_file.close()
remote.close()
break
elif cmd == 'get_spaces':
remote.send((2**env.get_available_buttons_size(), (1, 84, 84)))
else:
raise NotImplementedError
class CloudpickleWrapper(object):
"""
Uses cloudpickle to serialize contents (otherwise multiprocessing tries to use pickle)
"""
def __init__(self, x, log_file=""):
self.x = x
self.log_file = log_file
def __getstate__(self):
import cloudpickle
return cloudpickle.dumps(self.x)
def __setstate__(self, ob):
import pickle
self.x = pickle.loads(ob)
class VecEnv():
def __init__(self, env_fns, logging=False, log_dir='/tmp/vizdoom/'):
"""
envs: list of vizdoom game environments to run in subprocesses
"""
self.closed = False
nenvs = len(env_fns)
self.remotes, self.work_remotes = zip(*[Pipe() for _ in range(nenvs)])
self.log_files = [os.path.join(log_dir, 'worker_' + str(i) + '.log') for i in range(nenvs)]
if logging is False:
self.ps = [Process(target=worker, args=(work_remote, remote, CloudpickleWrapper(env_fn)))
for (work_remote, remote, env_fn) in zip(self.work_remotes, self.remotes, env_fns)]
else:
self.ps = [Process(target=worker, args=(work_remote, remote, CloudpickleWrapper(env_fn, log_file)))
for (work_remote, remote, env_fn, log_file) in zip(self.work_remotes, self.remotes, env_fns, self.log_files)]
for p in self.ps:
p.daemon = True # if the main process crashes, we should not cause things to hang
p.start()
for remote in self.work_remotes:
remote.close()
self.remotes[0].send(('get_spaces', None))
self.action_space_shape, self.observation_space_shape = self.remotes[0].recv()
def step(self, actions):
cumul_rewards = None
cumul_dones = None
for _ in range(4):
for remote, action in zip(self.remotes, actions):
remote.send(('step', action))
results = [remote.recv() for remote in self.remotes]
obs, rews, dones, infos = zip(*results)
if cumul_rewards is None:
cumul_rewards = np.stack(rews)
else:
cumul_rewards += np.stack(rews)
if cumul_dones is None:
cumul_dones = np.stack(dones)
else:
cumul_dones |= np.stack(dones)
return np.stack(obs), cumul_rewards, cumul_dones, infos
def get_game_variables(self, id):
self.remotes[id].send(['gv', None])
return self.remotes[id].recv()
def log(self):
for remote in self.remotes:
remote.send(('log', None))
return
def reset(self):
for remote in self.remotes:
remote.send(('reset', None))
return np.stack([remote.recv() for remote in self.remotes])
def reset_task(self):
for remote in self.remotes:
remote.send(('reset_task', None))
return np.stack([remote.recv() for remote in self.remotes])
def close(self):
if self.closed:
return
for remote in self.remotes:
remote.send(('close', None))
for p in self.ps:
p.join()
self.closed = True
@property
def num_envs(self):
return len(self.remotes)