-
Notifications
You must be signed in to change notification settings - Fork 0
/
server.py
351 lines (295 loc) · 13.1 KB
/
server.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
from http.server import BaseHTTPRequestHandler, HTTPServer
import logging
from time import gmtime, strftime, time
from datetime import datetime
import json
import traceback
import numpy as np
import argparse
import scipy.stats
# https://gist.github.com/mdonkers/63e115cc0c79b4f6b8b3a6b797e485c7
class ParaManager:
def __init__(self, args):
self.repeat = args.repeat
self.path = args.path
self.stop = args.stop
self.p_timeout = 3600 * args.timeout
self.n_pop = 30
self.v_max = 0.1
self.f_w = 0.5
self.f_c1 = 2
self.f_c2 = 2
# type, min, max
self.paras = [
(float, 0, 10, 'Predict Future'), # weights
(float, 0, 10, 'Extract Reward'),
(float, 0, 10, 'BYOL'),
(float, 0, 10, 'AutoEncoder'),
(float, 0, 10, 'Rotation CLS'),
]
self.n_para = len(self.paras)
# particle_pos
self.p_pos = np.random.rand(self.n_pop, self.n_para)
self.p_vel = (np.random.rand(self.n_pop, self.n_para) - 0.5) * self.v_max * 2
# initialization
self.p_pos[0] = 0
n_weights = len(self.paras)
self.p_pos[1:n_weights+1] = np.eye(n_weights, n_weights) * 0.1
# personal best
self.p_best = self.p_pos.copy()
self.p_best_val = np.zeros(self.n_pop, dtype=np.float32)
# global best
self.g_best = self.p_pos[0]
self.g_best_val = 0
# particle status
self.p_status = np.zeros(self.n_pop, dtype=np.uint8)
self.p_iter = np.zeros(self.n_pop, dtype=np.uint8) # iteration/generation
self.p_idx = 0
# task assgined time
self.p_time = np.ones(self.n_pop) * time()
# load previous tasks
self.load()
print("Particle status:")
print(self.p_status[:10])
print("*" * 8)
print("Particle pos:")
print(self.p_pos[:10])
print("*" * 8)
print("Particle vel:")
print(self.p_vel[:10])
def load(self, path=None):
try:
if path is None:
path = self.path
with open(path, 'rb') as f:
tmp = np.load(f)
assert tmp.shape == self.p_pos.shape
self.p_pos = tmp
self.p_vel = np.load(f)
self.p_best = np.load(f)
self.p_best_val = np.load(f)
self.g_best = np.load(f)
self.g_best_val = float(np.load(f))
self.p_status = np.load(f)
self.p_iter = np.load(f)
self.p_idx = int(np.load(f))
self.p_time = np.load(f)
logging.info(f'Load PSO states from {path}')
except ValueError:
logging.warning("Failed to load some values")
except FileNotFoundError:
pass
except AssertionError:
logging.warning(f'Dim from {path} does not match, will not load file.')
def save(self, path=None):
if path is None:
path = self.path
with open(path, 'wb') as f:
np.save(f, self.p_pos)
np.save(f, self.p_vel)
np.save(f, self.p_best)
np.save(f, self.p_best_val)
np.save(f, self.g_best)
np.save(f, self.g_best_val)
np.save(f, self.p_status)
np.save(f, self.p_iter)
np.save(f, self.p_idx)
np.save(f, self.p_time)
logging.info('Save PSO states to %s', path)
def pos2data(self, pos):
data = []
for v, para in zip(pos, self.paras):
t, v_min, v_max, des = para
v = np.clip(v, 0, 1)
val = v * (v_max - v_min) + v_min
if t == int:
val = int(val)
data.append(val)
return data
def check_timeout(self):
t_now = time()
for i in range(self.n_pop):
if self.p_status[i] > 0 and t_now - self.p_time[i] > self.p_timeout:
self.p_status[i] = 0
logging.warning(f"Particle {i} time out! {t_now} > {self.p_time[i]}")
def get_status(self, per_particle=False):
self.check_timeout()
def unix2str(ts):
return datetime.utcfromtimestamp(ts).strftime('%Y-%m-%d %H:%M:%S')
def clip_float(l):
return [round(i, 3) for i in l]
data = {
'global_best_val': round(self.g_best_val, 3),
'global_best': clip_float(self.pos2data(self.g_best)),
'running_jobs': int(sum(self.p_status)),
'status': [int(v) for v in self.p_status],
'task_index': self.p_idx,
'time': unix2str(time())
}
if per_particle:
data['particles'] = [{
'idx': i,
'assigned_time': unix2str(self.p_time[i]),
'iter': int(self.p_iter[i]),
'best_val': round(float(self.p_best_val[i]), 3),
'best': clip_float(self.pos2data(self.p_best[i])),
'pos': clip_float(self.p_pos[i]),
'vel': clip_float(self.p_vel[i])
} for i in range(self.n_pop)]
# print(data)
return json.dumps(data)
def get_eval(self):
data = self.pos2data(self.g_best)
p_data = {
'id': self.p_idx,
'tid': self.p_idx,
'pos': [float(i) for i in self.g_best],
'seed': list(range(self.repeat)),
'data': data
}
return json.dumps(p_data)
def get_next(self):
self.check_timeout()
if self.stop:
return json.dumps({'stop': True})
self.p_idx += 1
# Better way to find next task
min_iter = (1 - self.p_status) * (self.p_iter + 1)
min_iter[min_iter == 0] = np.iinfo(np.uint8).max
p_idx = int(np.argmin(min_iter))
self.p_status[p_idx] = 1
self.p_time[p_idx] = time()
data = self.pos2data(self.p_pos[p_idx])
p_data = {
'id': p_idx,
'tid': self.p_idx,
'pos': [float(i) for i in self.p_pos[p_idx]],
'seed': list(range(self.repeat)),
'data': data
}
return json.dumps(p_data)
def post_para(self, data_str):
try:
p = json.loads(data_str)
logging.info(f"Parse {data_str}")
# particle id
p_idx = p["id"]
# validate
if "pos" in p:
for i, value in enumerate(p["pos"]):
if abs(value - self.p_pos[p_idx][i]) > 1e-3:
logging.warning(f"Particle {p_idx} value doesn't match, submission abandoned.")
return 0
self.p_status[p_idx] = 0
if p["status"] > 0:
# task canceled
print(f"Particle {p_idx} canceled")
logging.info(f"Particle {p_idx} canceled")
return 0
results_dist = np.array(p["results"])
# ref: https://github.com/google-research/rliable/blob/master/rliable/metrics.py
p_val = scipy.stats.trim_mean(results_dist, proportiontocut=0.25, axis=None)
if p_val > self.p_best_val[p_idx]:
self.p_best_val[p_idx] = p_val
self.p_best[p_idx] = self.p_pos[p_idx].copy()
logging.info(f"Particle {p_idx}'s local best is update to {p_val:.3f}")
if p_val > self.g_best_val:
self.g_best_val = p_val
self.g_best = self.p_pos[p_idx].copy()
print(f"Global best is updated to {p_val:.3f}")
logging.info(f"Global best is updated to {p_val:.3f}")
# print(p_idx, 'vel_before', self.p_vel[p_idx])
# update velocity
self.p_vel[p_idx] = self.p_vel[p_idx] * self.f_w + \
self.f_c1 * np.random.rand(self.n_para) * (self.p_best[p_idx] - self.p_pos[p_idx]) + \
self.f_c2 * np.random.rand(self.n_para) * (self.g_best - self.p_pos[p_idx])
# print(p_idx, 'vel_after', self.p_vel[p_idx])
# clip velocity
self.p_vel[p_idx] = np.clip(self.p_vel[p_idx], -self.v_max, self.v_max)
# print(p_idx, 'pos_before', self.p_pos[p_idx])
# update position
self.p_pos[p_idx] += self.p_vel[p_idx]
# clip position
# self.p_pos[p_idx] = np.clip(self.p_pos[p_idx], 0, 1)
# print(p_idx, 'pos_after', self.p_pos[p_idx])
self.p_iter[p_idx] += 1
self.save()
return 0
except:
traceback.print_exc()
logging.warning(f"Fail to parse {data_str}")
return 1
def make_handler(para_manager):
class ParaRequestHandler(BaseHTTPRequestHandler):
def __init__(self, *args, **kwargs):
self.para = para_manager
# https://stackoverflow.com/questions/21631799/how-can-i-pass-parameters-to-a-requesthandler
# """One thing to bear in mind is that BaseHTTPRequestHandler actually runs the handler functions like do_GET inside its __init__ method, so you have to do your initialization before calling super().__init__, contrary to more typical best-practice-by-default""" -- mtraceur
super(ParaRequestHandler, self).__init__(*args, **kwargs)
def _set_response(self, code=200):
self.send_response(code)
self.send_header('Content-type', 'text/html')
self.end_headers()
def do_GET(self):
# logging.info("GET request,\nPath: %s\nHeaders:\n%s\n", str(self.path), str(self.headers))
if self.path == '/get_task':
data = self.para.get_next()
logging.info("GET request from %s:%s, send task: %s", *self.client_address, data)
self._set_response()
self.wfile.write(data.encode('utf-8'))
elif self.path == '/get_eval':
data = self.para.get_eval()
logging.info("GET request from %s:%s, send task: %s", *self.client_address, data)
self._set_response()
self.wfile.write(data.encode('utf-8'))
elif self.path =='/get_status':
data = self.para.get_status()
self._set_response()
self.wfile.write(data.encode('utf-8'))
elif self.path =='/get_full_status':
data = self.para.get_status(True)
self._set_response()
self.wfile.write(data.encode('utf-8'))
else:
self._set_response(404)
def do_POST(self):
if self.path == '/submit_result':
content_length = int(self.headers['Content-Length']) # <--- Gets the size of data
post_data = self.rfile.read(content_length) # <--- Gets the data itself
# logging.info("POST request,\nPath: %s\nHeaders:\n%s\n\nBody:\n%s\n",
# str(self.path), str(self.headers), post_data.decode('utf-8'))
data_str = post_data.decode('utf-8')
logging.info("POST request from %s:%s, get data: %s", *self.client_address, data_str)
if not self.para.post_para(data_str):
self._set_response()
self.wfile.write("accepted".encode('utf-8'))
else:
self._set_response(400)
else:
self._set_response(404)
return ParaRequestHandler
def run():
parser = argparse.ArgumentParser(description='Search server')
parser.add_argument('--port', type=int, help='port to listen (default: 8888)', default=8888)
parser.add_argument('--repeat', type=int, help='test one set of parameters multiple time. (default: 5)', default=5)
parser.add_argument('--timeout', type=int, help='particle busy state reset timeout (unit: hour, default: 12)', default=12)
parser.add_argument('--path', type=str, help='where to save search result file (default: server/evol_rl.npy)', default='server/evol_rl.npy')
parser.add_argument('--log', type=str, help='where to save log file (default: server/evol_rl.log)', default='server/evol_rl.log')
parser.add_argument('--stop', type=bool, help='stop the client after the current task is completed (default: False)', default=False)
args = parser.parse_args()
logging.basicConfig(filename=args.log, filemode='a', level=logging.INFO, format='%(asctime)s:%(levelname)s:%(message)s', force=True)
para_manager = ParaManager(args)
handler = make_handler(para_manager)
server_address = ('', args.port)
httpd = HTTPServer(server_address, handler)
logging.info('Starting httpd...')
try:
httpd.serve_forever()
except KeyboardInterrupt:
pass
logging.info('Stopping httpd...')
httpd.server_close()
para_manager.save()
logging.shutdown()
if __name__ == '__main__':
run()