-
-
Notifications
You must be signed in to change notification settings - Fork 57
/
multiproc_vec.py
264 lines (221 loc) · 8.47 KB
/
multiproc_vec.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
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
import copy
import multiprocessing as mp
import time
import traceback
import gymnasium.vector
import numpy as np
from gymnasium.vector.utils import (
concatenate,
create_empty_array,
create_shared_memory,
iterate,
read_from_shared_memory,
write_to_shared_memory,
)
from .utils.shared_array import SharedArray
def compress_info(infos):
non_empty_infs = [(i, info) for i, info in enumerate(infos) if info]
return non_empty_infs
def decompress_info(num_envs, idx_starts, comp_infos):
all_info = [{}] * num_envs
for idx_start, comp_infos in zip(idx_starts, comp_infos):
for i, info in comp_infos:
all_info[idx_start + i] = info
return all_info
def write_observations(vec_env, env_start_idx, shared_obs, obs):
obs = list(iterate(vec_env.observation_space, obs))
for i in range(vec_env.num_envs):
write_to_shared_memory(
vec_env.observation_space,
env_start_idx + i,
obs[i],
shared_obs,
)
def numpy_deepcopy(buf):
if isinstance(buf, dict):
return {name: numpy_deepcopy(v) for name, v in buf.items()}
elif isinstance(buf, tuple):
return tuple(numpy_deepcopy(v) for v in buf)
elif isinstance(buf, np.ndarray):
return buf.copy()
else:
raise ValueError("numpy_deepcopy ")
def async_loop(
vec_env_constr, inpt_p, pipe, shared_obs, shared_rews, shared_terms, shared_truncs
):
inpt_p.close()
try:
vec_env = vec_env_constr()
pipe.send(vec_env.num_envs)
env_start_idx = pipe.recv()
env_end_idx = env_start_idx + vec_env.num_envs
while True:
instr = pipe.recv()
comp_infos = []
if instr == "close":
vec_env.close()
elif isinstance(instr, tuple):
name, data = instr
if name == "reset":
observations, infos = vec_env.reset(seed=data[0], options=data[1])
comp_infos = compress_info(infos)
write_observations(vec_env, env_start_idx, shared_obs, observations)
shared_terms.np_arr[env_start_idx:env_end_idx] = False
shared_truncs.np_arr[env_start_idx:env_end_idx] = False
shared_rews.np_arr[env_start_idx:env_end_idx] = 0.0
elif name == "step":
actions = data
actions = concatenate(
vec_env.action_space,
actions,
create_empty_array(vec_env.action_space, n=len(actions)),
)
observations, rewards, terms, truncs, infos = vec_env.step(actions)
write_observations(vec_env, env_start_idx, shared_obs, observations)
shared_terms.np_arr[env_start_idx:env_end_idx] = terms
shared_truncs.np_arr[env_start_idx:env_end_idx] = truncs
shared_rews.np_arr[env_start_idx:env_end_idx] = rewards
comp_infos = compress_info(infos)
elif name == "env_is_wrapped":
comp_infos = vec_env.env_is_wrapped(data)
else:
raise AssertionError("bad tuple instruction name: " + name)
elif instr == "render":
render_result = vec_env.render()
if vec_env.render_mode == "rgb_array":
comp_infos = render_result
elif instr == "terminate":
return
else:
raise AssertionError("bad instruction: " + instr)
pipe.send(comp_infos)
except BaseException as e:
tb = traceback.format_exc()
pipe.send((e, tb))
class ProcConcatVec(gymnasium.vector.VectorEnv):
def __init__(
self, vec_env_constrs, observation_space, action_space, tot_num_envs, metadata
):
self.observation_space = observation_space
self.action_space = action_space
self.num_envs = num_envs = tot_num_envs
self.metadata = metadata
self.shared_obs = create_shared_memory(self.observation_space, n=self.num_envs)
self.shared_act = create_shared_memory(self.action_space, n=self.num_envs)
self.shared_rews = SharedArray((num_envs,), dtype=np.float32)
self.shared_terms = SharedArray((num_envs,), dtype=np.uint8)
self.shared_truncs = SharedArray((num_envs,), dtype=np.uint8)
self.observations_buffers = read_from_shared_memory(
self.observation_space, self.shared_obs, n=self.num_envs
)
self.graceful_shutdown_timeout = 10
pipes = []
procs = []
for constr in vec_env_constrs:
inpt, outpt = mp.Pipe()
constr = gymnasium.vector.async_vector_env.CloudpickleWrapper(constr)
proc = mp.Process(
target=async_loop,
args=(
constr,
inpt,
outpt,
self.shared_obs,
self.shared_rews,
self.shared_terms,
self.shared_truncs,
),
)
proc.start()
outpt.close()
pipes.append(inpt)
procs.append(proc)
self.pipes = pipes
self.procs = procs
num_envs = 0
env_nums = self._receive_info()
idx_starts = []
for pipe, cnum_env in zip(self.pipes, env_nums):
cur_env_idx = num_envs
num_envs += cnum_env
pipe.send(cur_env_idx)
idx_starts.append(cur_env_idx)
idx_starts.append(num_envs)
assert num_envs == tot_num_envs
self.idx_starts = idx_starts
def reset(self, seed=None, options=None):
for i, pipe in enumerate(self.pipes):
if seed is not None:
pipe.send(("reset", (seed + i, options)))
else:
pipe.send(("reset", (seed, options)))
info = self._receive_info()
return numpy_deepcopy(self.observations_buffers), copy.deepcopy(info)
def step_async(self, actions):
actions = list(iterate(self.action_space, actions))
for i, pipe in enumerate(self.pipes):
start, end = self.idx_starts[i : i + 2]
pipe.send(("step", actions[start:end]))
def _receive_info(self):
all_data = []
for cin in self.pipes:
data = cin.recv()
if isinstance(data, tuple):
e, tb = data
print(tb)
raise e
all_data.append(data)
return all_data
def step_wait(self):
compressed_infos = self._receive_info()
infos = decompress_info(self.num_envs, self.idx_starts, compressed_infos)
rewards = self.shared_rews.np_arr
terms = self.shared_terms.np_arr
truncs = self.shared_truncs.np_arr
return (
numpy_deepcopy(self.observations_buffers),
rewards.copy(),
terms.astype(bool).copy(),
truncs.astype(bool).copy(),
copy.deepcopy(infos),
)
def step(self, actions):
self.step_async(actions)
return self.step_wait()
def __del__(self):
self.close()
def render(self):
self.pipes[0].send("render")
render_result = self.pipes[0].recv()
if isinstance(render_result, tuple):
e, tb = render_result
print(tb)
raise e
return render_result
def close(self):
try:
for pipe, proc in zip(self.pipes, self.procs):
if proc.is_alive():
pipe.send(("close", None))
except OSError:
pass
else:
deadline = (
None
if self.graceful_shutdown_timeout is None
else time.monotonic() + self.graceful_shutdown_timeout
)
for proc in self.procs:
timeout = None if deadline is None else deadline - time.monotonic()
if timeout is not None and timeout <= 0:
break
proc.join(timeout)
for pipe, proc in zip(self.pipes, self.procs):
if proc.is_alive():
proc.kill()
pipe.close()
def env_is_wrapped(self, wrapper_class, indices=None):
for i, pipe in enumerate(self.pipes):
pipe.send(("env_is_wrapped", wrapper_class))
results = self._receive_info()
return sum(results, [])