-
Notifications
You must be signed in to change notification settings - Fork 0
/
farm.py
368 lines (299 loc) · 10 KB
/
farm.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
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
# farm.py
# a single instance of a farm.
# a farm should consist of a pool of instances
# and expose those instances as one giant callable class
import multiprocessing, time, random, threading
from multiprocessing import Process, Pipe, Queue
import traceback
from create_env import create_env
# for hacking seeds
from test_multi import set_all_seeds
#import numpy as np
ncpu = multiprocessing.cpu_count()
# separate process that holds a separate RunEnv instance.
# This has to be done since RunEnv() in the same process result in interleaved running of simulations.
def standalone_headless_isolated(pq, cq, plock):
# locking to prevent mixed-up printing.
plock.acquire()
print('starting headless...',pq,cq)
try:
# create a training environment
env = create_env(train=True, render=False)
except Exception as ex:
print('error on start of standalone')
traceback.print_exc()
plock.release()
return
else:
plock.release()
def report(env):
# a way to report errors ( since you can't just throw them over a pipe )
# env should be a string
print('(standalone) got error!!!')
cq.put(('error',env))
#def floatify(np):
# return [float(np[i]) for i in range(len(np))]
try:
while True:
msg = pq.get()
# messages should be tuples,
# msg[0] should be string
if msg[0] == 'reset':
# only play bad episodes
#seeds = [4, 7, 8, 10, 15, 20, 24, 25, 26, 30, 31, 39, 44, 45, 50, 56, 57, 58, 59]
#seed = random.choice(seeds)
#set_all_seeds(seed)
#print("Picked seed", seed)
o = env.reset(project=False)
#o = env.reset() # for gym env
#o = floatify(o) # obs is np in gym env
# send as tuple
cq.put((o,))
elif msg[0] == 'step':
o,r,d,i = env.step(msg[1], project=False)
#o,r,d,i = env.step(msg[1]) # for gym env
#o = floatify(o)
cq.put((o,r,d,i))
else:
cq.close()
pq.close()
del env
break
except Exception as ex:
traceback.print_exc()
report(str(ex))
return # end process
# global process lock
plock = multiprocessing.Lock()
# global thread lock
tlock = threading.Lock()
# global id issurance
eid = int(random.random()*100000)
def get_eid():
global eid,tlock
tlock.acquire()
i = eid
eid+=1
tlock.release()
return i
# class that manages the interprocess communication and expose itself as a RunEnv.
# reinforced: this class should be long-running. it should reload the process on errors.
class ei: # Environment Instance
def __init__(self):
self.occupied = False # is this instance occupied by a remote client
self.id = get_eid() # what is the id of this environment
self.pretty('instance creating')
self.newproc()
import threading as th
self.lock = th.Lock()
def timer_update(self):
self.last_interaction = time.time()
def is_occupied(self):
if self.occupied == False:
return False
else:
if time.time() - self.last_interaction > 20*60:
# if no interaction for more than X minutes
self.pretty('no interaction for too long, self-releasing now. applying for a new id.')
self.id = get_eid() # apply for a new id.
self.occupied == False
self.pretty('self-released.')
return False
else:
return True
def occupy(self):
self.lock.acquire()
if self.is_occupied() == False:
self.occupied = True
self.id = get_eid()
self.lock.release()
return True # on success
else:
self.lock.release()
return False # failed
def release(self):
self.lock.acquire()
self.occupied = False
self.id = get_eid()
self.lock.release()
# create a new RunEnv in a new process.
def newproc(self):
global plock
self.timer_update()
self.pq, self.cq = Queue(1), Queue(1) # two queue needed
self.p = Process(
target = standalone_headless_isolated,
args=(self.pq, self.cq, plock)
)
self.p.daemon = True
self.p.start()
self.reset_count = 0 # how many times has this instance been reset
self.step_count = 0
self.timer_update()
return
# send x to the process
def send(self,x):
return self.pq.put(x)
# receive from the process.
def recv(self):
# receive and detect if we got any errors
r = self.cq.get()
#print(r)
# isinstance is dangerous, commented out
# if isinstance(r,tuple):
if r[0] == 'error':
# read the exception string
e = r[1]
self.pretty('got exception')
self.pretty(e)
raise Exception(e)
return r
def reset(self, project=False):
self.timer_update()
if not self.is_alive():
# if our process is dead for some reason
self.pretty('process found dead on reset(). reloading.')
self.kill()
self.newproc()
if self.reset_count>50 or self.step_count>10000: # if reset for more than X times
self.pretty('Environment has been reset too many times. Reloading...')
self.kill()
self.newproc()
self.reset_count += 1
self.send(('reset',))
r = self.recv()
self.timer_update()
# return first of tuple
return r[0]
def step(self, actions, project=False):
self.timer_update()
self.send(('step',actions,))
r = self.recv()
self.timer_update()
self.step_count+=1
return r
def kill(self):
if not self.is_alive():
self.pretty('process already dead, no need for kill.')
else:
self.send(('exit',))
self.pretty('waiting for join()...')
while 1:
self.p.join(timeout=5)
if not self.is_alive():
break
else:
self.pretty('process is not joining after 5s, still waiting...')
self.pretty('process joined.')
def __del__(self):
self.pretty('__del__')
self.kill()
self.pretty('__del__ accomplished.')
def is_alive(self):
return self.p.is_alive()
# pretty printing
def pretty(self,s):
print(('(ei) {} ').format(self.id)+str(s))
# class that other classes acquires and releases EIs from.
class eipool: # Environment Instance Pool
def pretty(self,s):
print(('(eipool) ')+str(s))
def __init__(self,n=1):
import threading as th
self.pretty('starting '+str(n)+' instance(s)...')
self.pool = [ei() for i in range(n)]
self.lock = th.Lock()
def acq_env(self):
self.lock.acquire()
for env in self.pool:
if env.occupy() == True: # successfully occupied an environment
self.lock.release()
return env # return the envinstance
self.lock.release()
return False # no available ei
def rel_env(self,ei):
self.lock.acquire()
for env in self.pool:
if env == ei:
env.release() # freed
self.lock.release()
def get_env_by_id(self,id):
for e in self.pool:
if e.id == id:
return e
return False
def __del__(self):
for e in self.pool:
del e
# farm
# interface with eipool via eids.
# ! this class is a singleton. must be made thread-safe.
import traceback
class farm:
def pretty(self,s):
print(('(farm) ')+str(s))
def __init__(self):
# on init, create a pool
# self.renew()
import threading as th
self.lock = th.Lock()
def acq(self,n=None):
self.renew_if_needed(n)
result = self.eip.acq_env() # thread-safe
if result == False:
ret = False
else:
self.pretty('new env '+str(result.id))
ret = result.id
return ret
def rel(self,id):
e = self.eip.get_env_by_id(id)
#if e == False:
# self.pretty(str(id)+' not found on rel(), might already be released')
#else:
if e != False:
self.eip.rel_env(e)
#self.pretty('rel '+str(id))
def step(self,id,actions):
e = self.eip.get_env_by_id(id)
if e == False:
self.pretty(str(id)+' not found on step(), might already be released')
return False
try:
ordi = e.step(actions)
return ordi
except Exception as e:
traceback.print_exc()
raise e
def reset(self,id):
e = self.eip.get_env_by_id(id)
if e == False:
self.pretty(str(id)+' not found on reset(), might already be released')
return False
try:
oo = e.reset()
return oo
except Exception as e:
traceback.print_exc()
raise e
def renew_if_needed(self,n=None):
self.lock.acquire()
if not hasattr(self,'eip'):
self.pretty('renew because no eipool present')
self._new(n)
self.lock.release()
def forcerenew(self,n=None):
self.lock.acquire()
self.pretty('forced pool renew')
if hasattr(self,'eip'): # if eip exists
del self.eip
self._new(n)
self.lock.release()
def _new(self,n=None):
self.eip = eipool(ncpu if n is None else n)
# expose the farm via Pyro4
def main():
from pyro_helper import pyro_expose
pyro_expose(farm,20099,'farm')
if __name__ == '__main__':
main()