forked from KuoHaoZeng/Visual_Reaction
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
executable file
·566 lines (485 loc) · 29.2 KB
/
main.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
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
import time, os, argparse, jsonlines, progressbar
import numpy as np
from threading import Thread, Barrier, Lock
from multiprocessing import Queue
import torch
import torch.nn.functional as F
import torch.optim as optim
from torch.nn.utils import clip_grad_norm_
from network.network import *
from utils.storage import *
from utils.event import *
from utils.utils import *
from utils.config import Config
from mpc.mpc import mpc_func
from data.data_mr import *
class trainer_Online_VA(Thread):
def __init__(self, t_id, cfg, log_queue, data_queue, memory_buffer, mode, explorer,
share_forecaster, share_actor, optimizer_f, optimizer_a, scheduler_f):
super(trainer_Online_VA, self).__init__()
### somethings
set_random_seed(t_id + cfg.framework.seed)
self.t_id = t_id
self.cfg = cfg
self.mode = mode
self.data_queue = data_queue
self.memory_buffer = memory_buffer
self.gpu_id = t_id % cfg.framework.num_gpu + cfg.thor.x_display_offset
self.train_iter = 0
self.num_traj = 0
self.total_traj = len(explorer.trajectory['scene'])
### logging
self.log_keys = ['ball_pos', 'ball_vel', 'ball_acc', 'agent_pos', 'agent_acc',
'agent_pos_est', 'agent_vel_est', 'agent_acc_est', 'agent_angle_y', 'agent_angle_xz',
'success', 'num_SimObj_hits', 'num_Floor_hits', 'num_Structure_hits', 'tidx',
'angle_x', 'angle_y', 'force', 'object_name', 'mass', 'initial_agnet_pos', 'launcher_pos',
'ball_pos_est', 'ball_vel_est', 'ball_acc_est']
self.log_queue = log_queue
if self.t_id == 0:
self.log_dir = "{}/{}_{}".format(cfg.base_dir, mode, cfg.log_dir)
self.log_writer = open(self.log_dir, 'wb')
if mode == "train":
widgets = ['Training phase [', progressbar.SimpleProgress(), '] [', progressbar.Percentage(), '] ',
progressbar.Bar(marker='█'), ' (', progressbar.Timer(), ' ', progressbar.ETA(), ') ', ]
else:
widgets = ['Evaluation phase [', progressbar.SimpleProgress(), '] [', progressbar.Percentage(), '] ',
progressbar.Bar(marker='█'), ' (', progressbar.Timer(), ' ', progressbar.ETA(), ') ', ]
if mode == "train":
max_value = self.total_traj * cfg.train.num_epoch
else:
max_value = self.total_traj
self.bar = progressbar.ProgressBar(max_value=max_value, widgets=widgets, term_width=100)
else:
self.log_dir, self.writer = None, None
### define loss functions
self.criterionL1 = nn.L1Loss()
### explorer
self.explorer = explorer
### forecaster
self.share_forecaster = share_forecaster
self.forecaster = forecaster_protocols[cfg.agent.protocol](1,
cfg.agent.ND,
cfg.agent.hidden_size).cuda(self.gpu_id)
if mode == "train":
self.forecaster.train()
else:
self.forecaster.eval()
self.optimizer_f, self.scheduler_f = optimizer_f, scheduler_f
### model
self.model = model_protocols[cfg.agent.protocol_model](1,
cfg.agent.ND,
cfg.agent.hidden_size,
cfg.thor.delta_time,
True,
self.gpu_id).cuda(self.gpu_id)
self.model.gpu_id = self.gpu_id
### actor
self.share_actor = share_actor
if not isinstance(self.cfg.agent.protocol_actor, type(None)):
self.actor = actor_protocols[cfg.agent.protocol_actor](1,
cfg.agent.ND,
cfg.agent.hidden_size,
cfg.agent.CD).cuda(self.gpu_id)
self.dist = torch.distributions.Normal
if mode == "train":
self.actor.train()
else:
self.actor.eval()
self.forecaster.eval()
else:
self.actor, self.dist = None, None
self.optimizer_a = optimizer_a
if not isinstance(self.cfg.checkpoint_file, type(None)) and len(os.listdir(self.cfg.checkpoint_dir)) > 0:
checkpoint = "{}/{:07d}".format(self.cfg.checkpoint_dir, self.cfg.checkpoint_file)
self.load_checkpoints(checkpoint)
def save_checkpoints(self):
dir = "{}/{:07d}".format(self.cfg.checkpoint_dir, self.explorer.iters * self.cfg.framework.num_thread)
if not os.path.isdir(dir):
os.makedirs(dir)
sd = {}
if not isinstance(self.forecaster, type(None)):
sd['forecaster'] = self.share_forecaster.module.state_dict()
if not isinstance(self.actor, type(None)):
sd['actor'] = self.share_actor.module.state_dict()
checkpoint_dir = "{}/model.pt".format(dir)
torch.save(sd, checkpoint_dir)
def load_checkpoints(self, checkpoint_dir):
sd = torch.load("{}/model.pt".format(checkpoint_dir), map_location=torch.device('cpu'))
self.train_iter = 0
if not isinstance(self.share_forecaster, type(None)):
if 'forecaster' in sd.keys():
self.share_forecaster.module.load_state_dict(sd['forecaster'])
if not isinstance(self.share_actor, type(None)):
if 'actor' in sd.keys():
self.share_actor.module.load_state_dict(sd['actor'])
def run(self):
self.explorer.start()
while not self.data_queue.empty():
### Sync with the shared model
if not isinstance(self.forecaster, type(None)):
self.forecaster.load_state_dict(self.share_forecaster.module.state_dict())
if not isinstance(self.actor, type(None)):
self.actor.load_state_dict(self.share_actor.module.state_dict())
if self.t_id == 0:
if self.mode == 'train':
if self.explorer.iters > 1:
if (self.explorer.iters * self.cfg.framework.num_thread) % self.cfg.train.save_iter == 0:
self.save_checkpoints()
if len(self.memory_buffer) < self.cfg.framework.buffer_size or self.mode != 'train':
storage = Storage()
### reset the initial velocity
self.model.reset_vel()
### Reset the environment
self.explorer.reset = True
## Exploration
while self.explorer.reset:
time.sleep(0.01)
if not self.explorer.good_to_go:
if self.data_queue.empty():
break
else:
continue
## first
event = self.explorer.event
storage, agent_pos, ball_pos = record(storage, event, [0, 0, 0])
agent_angle_y, agent_angle_xz = get_camera_pos_numpy(event)
storage = record_agnet(storage, agent_pos, [0, 0, 0], [0, 0, 0], storage.dict['frame'][-1],
agent_angle_y, agent_angle_xz, ball_pos, [0, 0, 0], [0, 0, 0])
storage = record_meta(storage, event, self.explorer)
## register what's the first time step the data is ready to used
storage.add_to('first', 3)
while len(storage) < self.cfg.thor.max_time:
## wait till the first three frames are collected
if len(storage) < 3:
event = self.explorer.controller.step(dict(action='Pass'))
tag = check_with_time_difference(event, storage.dict['event'][-1])
if tag == 'fail':
break
elif tag == 'continue':
continue
else:
pass
storage, agent_pos, ball_pos = record(storage, event, [0, 0, 0])
agent_angle_y, agent_angle_xz = get_camera_pos_numpy(event)
storage = record_agnet(storage, agent_pos, [0, 0, 0], [0, 0, 0], storage.dict['frame'][-1],
agent_angle_y, agent_angle_xz, ball_pos, storage.dict['ball_vel'][-1],
storage.dict['ball_acc'][-1])
## or to policy evaluation
else:
# Data processing
input_frames = storage.dict['frame'][-1]
# check if it's the first time policy evaluation
if len(storage) == 3:
# expand the first dimension for batch size
agent_pos_est = torch.Tensor(np.array(agent_pos)).cuda(self.gpu_id)
agent_vel_est = torch.Tensor(np.array([0, 0, 0])).cuda(self.gpu_id)
agent_acc_est = torch.Tensor(np.array([0, 0, 0])).cuda(self.gpu_id)
agent_pos_est = agent_pos_est.unsqueeze(0)
agent_vel_est = agent_vel_est.unsqueeze(0)
agent_acc_est = agent_acc_est.unsqueeze(0)
ball_pos_est = torch.Tensor(np.array(ball_pos)).cuda(self.gpu_id)
ball_vel_est = torch.Tensor(np.array(storage.dict['ball_vel'][-1])).cuda(self.gpu_id)
ball_acc_est = torch.Tensor(np.array(storage.dict['ball_acc'][-1])).cuda(self.gpu_id)
ball_pos_est = ball_pos_est.unsqueeze(0)
ball_vel_est = ball_vel_est.unsqueeze(0)
ball_acc_est = ball_acc_est.unsqueeze(0)
agent_angle_y = torch.Tensor([agent_angle_y]).cuda(self.gpu_id)
agent_angle_xz = torch.Tensor([agent_angle_xz]).cuda(self.gpu_id)
storage = record_agnet(storage,
agent_pos_est.squeeze(0).cpu().numpy(),
agent_vel_est.squeeze(0).cpu().numpy(),
agent_acc_est.squeeze(0).cpu().numpy(),
input_frames,
agent_angle_y.item(),
agent_angle_xz.item(),
ball_pos_est.squeeze(0).cpu().numpy(),
ball_vel_est.squeeze(0).cpu().numpy(),
ball_acc_est.squeeze(0).cpu().numpy())
## Planner, Input: first three images
# forward forecaster to obtain the anticipated trajectory
input_frames = torch.Tensor(input_frames).cuda(self.gpu_id)
ball_pos_est, ball_vel_est, ball_acc_est, _ = self.forecaster(input_frames,
agent_pos_est,
agent_vel_est,
agent_acc_est,
agent_angle_y,
agent_angle_xz)
# detach the graph
ball_pos_est, ball_vel_est, ball_acc_est = ball_pos_est.detach(), ball_vel_est.detach(), ball_acc_est.detach()
# obtain the predicted vel and acc by the highest prob.
p_pred = ball_pos_est.clone()
v_pred = ball_vel_est.clone()
a_pred = ball_acc_est.clone()
# anticipate trajectory
data_forecast = self.forecaster.forecast(p_pred, v_pred, a_pred, self.cfg.agent.Horizon)
horizon = self.cfg.agent.Horizon
# minus 0.44 to let agent fly below the target trajectory
data_forecast[:, 1] = data_forecast[:, 1] - self.cfg.agent.height_difference
# clone the dynamics model for MPC usage
model_to_mpc = model_protocols[self.cfg.agent.protocol_model](1,
self.cfg.agent.ND,
self.cfg.agent.hidden_size,
self.cfg.thor.delta_time,
True,
self.gpu_id).cuda(self.gpu_id)
model_to_mpc.vel = self.model.vel.clone()
# do MPC
# weight_horizon = [0.95**ele for ele in range(self.cfg.agent.Horizon)]
weight_horizon = None
dis, action = mpc_func(data_forecast,
agent_pos_est,
model_to_mpc,
self.cfg.agent.ND,
self.cfg.agent.CD,
horizon,
weight_horizon,
self.cfg.agent.NSample,
self.actor,
self.forecaster,
self.gpu_id,
self.cfg.agent.DScale)
# detach the graph
dis, action = dis.detach(), action.detach()
# forward again by the selected action
inp_s = torch.cat((agent_pos_est.unsqueeze(0), action.unsqueeze(1)), dim=1)
agent_pos_est = self.model(inp_s)
agent_vel_est = self.model.vel.detach().clone()
agent_acc_est = action.clone()
# detach the graph
agent_pos_est = agent_pos_est.detach()
# compute camera pose
ball_pos_est_for_camera = data_forecast[0, :].unsqueeze(0)
ball_pos_est_for_camera[:, 1] += self.cfg.agent.height_difference
agent_pos_est_for_camera = agent_pos_est.clone()
if self.cfg.agent.gt_camera_pose:
agent_angle_y, agent_angle_xz = get_camera_pos_numpy(event)
agent_angle_y = torch.Tensor([agent_angle_y]).cuda(self.gpu_id)
agent_angle_xz = torch.Tensor([agent_angle_xz]).cuda(self.gpu_id)
else:
agent_angle_y, agent_angle_xz = get_camera_pos(ball_pos_est_for_camera,
agent_pos_est_for_camera)
agent_angle_y = torch.round(
agent_angle_y.clone() * (1 / self.cfg.agent.PScale)) * self.cfg.agent.PScale
agent_angle_xz = torch.round(
agent_angle_xz.clone() * (1 / self.cfg.agent.PScale)) * self.cfg.agent.PScale
# send action commands to THOR
event = self.explorer.controller.step(dict(action='FlyTo',
x=action[0, 0].item(),
y=action[0, 1].item(),
z=action[0, 2].item(),
horizon=-agent_angle_y.item(),
rotation=agent_angle_xz.item()))
# check time difference
tag = check_with_time_difference(event, storage.dict['event'][-1])
while tag == 'continue':
event = self.explorer.controller.step(dict(action='Pass'))
tag = check_with_time_difference(event, storage.dict['event'][-1])
if tag == 'fail':
break
else:
pass
# record data
storage, agent_pos, ball_pos = record(storage,
event,
[action[0, 0].item(), action[0, 1].item(),
action[0, 2].item()])
# check if the trajectory ends
if event.metadata['objects'][0]['isCaught']:
break
elif get_distance(storage.dict['event'][-1], storage.dict['event'][-2]) < 0.00001:
if ((storage.dict['ball_vel'][-1] ** 2).sum() ** 0.5) < 0.00001:
if ((storage.dict['ball_acc'][-1] ** 2).sum() ** 0.5) < 0.00001:
break
if tag != 'fail':
lock.acquire()
start_time = storage.dict['first'][0]
log_meta = {}
storage.dict['terminal'][-1] = True
for k, v in storage.dict.items():
if k != 'event' and k != 'frames':
if self.mode == 'train':
if k == 'ball_pos':
self.memory_buffer.add_to(k, v[start_time - 1: -1])
self.memory_buffer.add_to('ball_pos_next', v[start_time:])
else:
self.memory_buffer.add_to(k, v[start_time:])
if k in self.log_keys:
log_meta[k] = v.tolist()
lock.release()
self.log_queue.put(log_meta)
if self.t_id == 0:
log_meta = []
while not self.log_queue.empty():
log_meta.append(self.log_queue.get())
if len(log_meta) > 0:
self.num_traj = write_log(self.log_writer, log_meta, self.num_traj, self.bar)
else:
self.data_queue.put(self.explorer.tidx)
### training
if len(self.memory_buffer) >= self.cfg.framework.batch_size and self.mode == 'train' and self.t_id == 0:
## get data
lock.acquire()
memory_data = get_memory_data(self.memory_buffer)
lock.release()
# get ball data
ball_pos = torch.Tensor(memory_data['ball_pos']).cuda(self.gpu_id)
ball_pos_next = torch.Tensor(memory_data['ball_pos_next']).cuda(self.gpu_id)
ball_vel = torch.Tensor(memory_data['ball_vel']).cuda(self.gpu_id)
ball_acc = torch.Tensor(memory_data['ball_acc']).cuda(self.gpu_id)
# get agent data
agent_pos_est = torch.Tensor(memory_data['agent_pos_est']).cuda(self.gpu_id)
agent_vel_est = torch.Tensor(memory_data['agent_vel_est']).cuda(self.gpu_id)
agent_acc_est = torch.Tensor(memory_data['agent_acc_est']).cuda(self.gpu_id)
agent_pos = torch.Tensor(memory_data['agent_pos']).cuda(self.gpu_id)
agent_vel = torch.Tensor(memory_data['agent_vel']).cuda(self.gpu_id)
agent_acc = torch.Tensor(memory_data['agent_acc']).cuda(self.gpu_id)
agent_angle_y = torch.Tensor(memory_data['agent_angle_y']).cuda(self.gpu_id)
agent_angle_xz = torch.Tensor(memory_data['agent_angle_xz']).cuda(self.gpu_id)
# get frames
input_frames = torch.Tensor(memory_data['input_frames']).cuda(self.gpu_id)
success = torch.Tensor(memory_data['success']).cuda(self.gpu_id)
terminal = torch.Tensor(memory_data['terminal']).cuda(self.gpu_id)
if self.cfg.agent.forecaster.train:
## Forecaster training
ball_pos_est, ball_vel_est, ball_acc_est, _ = self.share_forecaster(input_frames,
agent_pos_est,
agent_vel_est,
agent_acc_est,
agent_angle_y,
agent_angle_xz)
# get ground truth and compute sub-losses
gt_pos = ball_pos.clone()
loss_pos = self.criterionL1(ball_pos_est, gt_pos)
gt_vel = ball_vel.clone()
loss_vel = self.criterionL1(ball_vel_est, gt_vel)
gt_acc = ball_acc.clone()
loss_acc = self.criterionL1(ball_acc_est, gt_acc)
# compute loss
loss_forecaster = loss_pos + 0.1 * loss_vel + 0.1 * loss_acc
# backward and update
self.share_forecaster.zero_grad()
loss_forecaster.backward()
self.optimizer_f.step()
self.scheduler_f.step()
if self.cfg.agent.actor.train and not isinstance(self.actor, type(None)):
## actor training
ball_pos_next[:, 1] -= self.cfg.agent.height_difference
distance = ((ball_pos_next - agent_pos) ** 2).sum(1) ** 0.5
ball_pos_est, ball_vel_est, ball_acc_est, representation = self.share_forecaster(input_frames,
agent_pos_est,
agent_vel_est,
agent_acc_est,
agent_angle_y,
agent_angle_xz)
# anticipate trajectory
data_forecast = ball_pos_est + ball_vel_est
data_forecast = torch.cat((data_forecast, agent_pos_est, agent_vel_est), dim=1).clone().detach()
representation = representation.clone().detach()
# compute loss
v_est, action_est_mu, action_est_sigma = self.share_actor(data_forecast, representation)
m = self.dist(action_est_mu, action_est_sigma)
log_prob = m.log_prob(agent_acc)
index = list(range(self.cfg.framework.batch_size - 1))
index.reverse()
loss_v, loss_a = 0., 0.
R = success[-1] - 0.01 * distance[-1]
for ii in index:
if terminal[ii]:
R = success[ii] - 0.01 * distance[ii]
continue
v = v_est[ii]
R = self.cfg.agent.actor.gamma * R + success[ii] - 0.01 * distance[ii]
advangtage = R.clone().detach() - v.clone().detach()
loss_v = loss_v + 0.5 * (R.clone().detach() - v).pow(2)
loss_a = loss_a - (advangtage * log_prob[ii]).mean()
loss_a /= self.cfg.framework.batch_size
loss_v /= self.cfg.framework.batch_size
loss_actor = (loss_a + 0.5 * loss_v - self.cfg.agent.actor.beta * m.entropy().mean())
# backward and update
self.share_actor.zero_grad()
loss_actor.backward()
clip_grad_norm_(self.share_actor.parameters(), 0.5)
self.optimizer_a.step()
self.train_iter += 1
# wait all threads finish
if self.t_id == 0:
log_meta = []
while not self.log_queue.empty():
log_meta.append(self.log_queue.get())
if len(log_meta) > 0:
self.num_traj = write_log(self.log_writer, log_meta, self.num_traj, self.bar)
if self.mode == 'train':
self.save_checkpoints()
class master_Online():
def __init__(self, cfg, mode):
### allocate explorer
data_queue = Queue()
explorers = [drone_explorer(t_id, cfg, data_queue, mode) for t_id in range(cfg.framework.num_thread)]
device_ids = [g_id for g_id in range(cfg.thor.x_display_offset, cfg.thor.x_display_offset + cfg.framework.num_gpu)]
### construct a forecaster and, if necessary, its optimizer as well as scheduler
share_forecaster = forecaster_protocols[cfg.agent.protocol](
cfg.framework.batch_size // cfg.framework.num_gpu,
cfg.agent.ND,
cfg.agent.hidden_size)
share_forecaster = share_forecaster.cuda(cfg.thor.x_display_offset)
share_forecaster = nn.DataParallel(share_forecaster, device_ids=device_ids)
if mode == 'train':
share_forecaster.train()
share_forecaster.share_memory()
optimizer_f = optim.SGD(share_forecaster.parameters(), lr=cfg.train.lr_f)
scheduler_f = optim.lr_scheduler.MultiStepLR(optimizer_f,
milestones=cfg.train.lr_f_ms,
gamma=0.1)
else:
optimizer_f, scheduler_f = None, None
### construct an action sampler and, if necessary, its optimizer as well as scheduler
if not isinstance(cfg.agent.protocol_actor, type(None)):
share_actor = actor_protocols[cfg.agent.protocol_actor](
cfg.framework.batch_size // cfg.framework.num_gpu,
cfg.agent.ND,
cfg.agent.hidden_size,
cfg.agent.CD).cuda(cfg.thor.x_display_offset)
share_actor = nn.DataParallel(share_actor, device_ids=device_ids)
if cfg.agent.actor.train:
optimizer_a = optim.Adam(share_actor.parameters(), lr=cfg.train.lr_a)
else:
optimizer_a = None
else:
share_actor, optimizer_a = None, None
if isinstance(optimizer_a, type(None)):
memory_buffer = Memory_WF(cfg.framework.buffer_size, cfg.framework.batch_size)
else:
memory_buffer = Memory_Seq(cfg.framework.buffer_size, cfg.framework.batch_size)
### prepare training/validation/testing jobs
self.job = []
log_queue = Queue()
for t_id in range(cfg.framework.num_thread):
self.job.append(protocols[cfg.agent.protocol](t_id, cfg, log_queue, data_queue,
memory_buffer, mode, explorers[t_id],
share_forecaster, share_actor,
optimizer_f, optimizer_a, scheduler_f))
def run(self):
for job in self.job:
job.start()
for job in self.job:
job.join()
def get_configs():
parser = argparse.ArgumentParser(description="Online mode training")
parser.add_argument("--config", type=str, default="configs/pretrained_action_sampler_test.yaml")
args = parser.parse_args()
config = Config(args.config)
return config
def main(cfg):
if not os.path.isdir(cfg.base_dir):
os.makedirs(cfg.base_dir)
m = master_Online(cfg, cfg.mode)
m.run()
if __name__ == '__main__':
protocols = {'VA': trainer_Online_VA}
forecaster_protocols = {'VA': NetForecaster}
model_protocols = {'PM': DroneModel}
actor_protocols = {'RL': NetPolicy}
config = get_configs()
barrier = Barrier(config.framework.num_thread)
lock = Lock()
main(config)