-
Notifications
You must be signed in to change notification settings - Fork 1
/
rl_main_by_data.py
202 lines (176 loc) · 9.16 KB
/
rl_main_by_data.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
# -*- utf-8 -*-
from rl_env import TrajComp, TrajSam
import time
from rl_brain import AGENT_Sel, AGENT_Sam
import warnings
import data_utils as F
import random
random.seed(0)
warnings.filterwarnings("ignore")
globaltime = str(time.time())
def run_dqn():
start_test = time.time()
SampleList_val = []
for sam_i in range(int(sam_size_v)):
observation_oub = env_rl_sam_v.starting(level_start)
while True:
action_oub = RL_OUB.online_act(observation_oub)
if action_oub == 8 or env_rl_sam_v.level_cur == level_end:
break
observation_oub = env_rl_sam_v.get_agent_oub_state(action_oub)
SampleList_val.append(env_rl_sam_v.cubeID)
print('done sampling')
env_rl_sam_v.get_ID2Grids(SampleList_val)
print('init the whole database for training with sample list size', len(SampleList_val))
env_rl_sel_v = TrajComp(traj_path, val_start, val_start + validDB_amount, K, LEN, val_DB_DISTRI, env_rl_sam_v.ID2Grids, 'V')
buffer_valid_size = int(ratio*env_rl_sel_v.DB_size)+2*validDB_amount
for trajID in range(validDB_amount):
steps = len(env_rl_sel_v.ori_traj_set[trajID])
s = (trajID, 0)
e = (trajID, steps-1)
env_rl_sel_v.read(s, e)
env_rl_sel_v.add_buffer(s, e)
print('DB size {}, simDB size {}, intial size {} for validation'.format(env_rl_sel_v.DB_size, buffer_valid_size, len(env_rl_sel_v.F_ward)))
epoch_i_c = 0
while len(env_rl_sel_v.F_ward) < buffer_valid_size:
if epoch_i_c % 100 == 0:
print('time', len(env_rl_sel_v.F_ward), '/', buffer_valid_size)
observation_inb = env_rl_sel_v.get_agent_inb_state(SampleList_val[epoch_i_c])
action_inb = RL_INB.online_act(observation_inb)
env_rl_sel_v.update_inb(action_inb)
epoch_i_c += 1
print('test time', time.time()-start_test)
F1 = env_rl_sel_v.submit_query(Rtree_ref_val, test_query)
return F1
def Training():
batch_size = 32
check_res = -1
query_count = 50
for epoch in range(12):
train_start = time.time()
env_rl_sel = TrajComp(traj_path, epoch*trainDB_amount, (epoch+1)*trainDB_amount, K, LEN)
tra_set = env_rl_sel.ori_traj_set
tra_DB_DISTRI, _, tra_DB_DISTRI_trajID = F.get_distribution_feature_data(tra_set)
_, query_part1, _ = F.get_query_workload_data(tra_DB_DISTRI)
train_query = query_part1
Rtree_ref_tra, _ = F.build_Rtree(tra_set, traj_path+'tra_tree_'+str(epoch*trainDB_amount)+'_'+str((epoch+1)*trainDB_amount))
print('bounds', Rtree_ref_tra.idx.get_bounds())
Rtree_ref_tra.idx.close()
Rtree_ref_tra = F.obtain_Rtree(traj_path+'tra_tree_'+str(epoch*trainDB_amount)+'_'+str((epoch+1)*trainDB_amount))
print('re-check bounds', Rtree_ref_tra.idx.get_bounds())
buffer_train_size = int(ratio*env_rl_sel.DB_size)+2*trainDB_amount
for epoch_i in range(5):
Transition_oub, SampleList_tra = [], []
sam_size = int(env_rl_sel.DB_size*ratio)
env_rl_sam = TrajSam(tra_set, level_start, 'data')
print('done env_rl_sam init')
for sam_i in range(int(sam_size*1.2)):
obs_oub = []
act_oub = []
transition_oub = []
observation_oub = env_rl_sam.starting(level_start)
obs_oub.append(observation_oub)
while True:
action_oub = RL_OUB.act(observation_oub)
#print('action_oub', action_oub)
act_oub.append(action_oub)
if action_oub == 8 or env_rl_sam.level_cur == level_end:
break
observation_oub = env_rl_sam.get_agent_oub_state(action_oub)
obs_oub.append(observation_oub)
if len(obs_oub) >= 2:
transition_oub.append([obs_oub[-2], act_oub[-1], obs_oub[-1], False])
if len(transition_oub) > 0:
transition_oub[-1][-1] = True
Transition_oub.append(transition_oub)
SampleList_tra.append(env_rl_sam.cubeID)
env_rl_sam.get_ID2Grids(SampleList_tra)
print('init the whole database for training with sample list size', len(SampleList_tra))
env_rl_sel = TrajComp(traj_path, epoch*trainDB_amount, (epoch+1)*trainDB_amount, K, LEN, tra_DB_DISTRI, env_rl_sam.ID2Grids, 'T') #softcopy
for trajID in range(trainDB_amount):
steps = len(env_rl_sel.ori_traj_set[trajID])
s = (trajID, 0)
e = (trajID, steps-1)
env_rl_sel.read(s, e)
env_rl_sel.add_buffer(s, e)
print('DB size {}, simDB size {}, intial size {} for training'.format(env_rl_sel.DB_size, buffer_train_size, len(env_rl_sel.F_ward)))
obs_inb = []
act_inb = []
transition_inb = []
transition_oub_tmp = []
epoch_i_c = 0
while len(env_rl_sel.F_ward) < buffer_train_size:
transition_oub_tmp.append(Transition_oub[epoch_i_c])
observation_inb = env_rl_sel.get_agent_inb_state(SampleList_tra[epoch_i_c])
obs_inb.append(observation_inb)
action_inb = RL_INB.act(observation_inb)
env_rl_sel.update_inb(action_inb)
if len(env_rl_sel.F_ward) == buffer_train_size:
done = True
else:
done = False
if len(obs_inb) >= 2:
transition_inb.append([obs_inb[-2], act_inb[-1], obs_inb[-1], done])
act_inb.append(action_inb)
# training
if epoch_i_c % query_count == 0:
print(epoch, 'training epoch...', epoch_i, 'at', epoch_i_c)
reward = env_rl_sel.submit_query(Rtree_ref_tra, train_query)
print('training query performance reward {} on epoch_i_c {}'.format(reward, epoch_i_c))
for tr in transition_inb:
RL_INB.remember(tr[0], tr[1], reward, tr[2], tr[3])
for Tr in transition_oub_tmp:
for tr in Tr:
if len(tr) > 0:
RL_OUB.remember(tr[0], tr[1], reward, tr[2], tr[3])
transition_inb.clear()
transition_oub_tmp.clear()
if epoch_i_c % query_count == 0 or done:
if len(RL_INB.memory) > batch_size:
RL_INB.replay(batch_size)
if len(RL_OUB.memory) > batch_size:
RL_OUB.replay(batch_size)
epoch_i_c += 1
RL_INB.update_target_model()
RL_OUB.update_target_model()
# validation
print('validation...')
res1 = run_dqn()
RL_INB.save(model_path+'save/'+globaltime+'_data_RL_INB_F1_' + str(res1)+'.h5')
RL_OUB.save(model_path+'save/'+globaltime+'_data_RL_OUB_F1_' + str(res1)+'.h5')
print('epoch {} epoch_i {} Save model at scanning trajID {} with query performance {}'.format(epoch, epoch_i, trajID, res1))
if res1 >= check_res:
check_res = res1
print('==>current best model with query performance {}'.format(check_res))
if __name__ == "__main__":
traj_path = './TrajData/Geolife_out/'
model_path = './'
trainDB_amount = 500
validDB_amount = 1000
K = 2
LEN = 2
ratio = 0.001
a_size_inb, s_size_inb = K, K*LEN
level_start, level_end = 9, 12
#reinforcement_learning
RL_INB = AGENT_Sel(s_size_inb, a_size_inb)
RL_OUB = AGENT_Sam(2*8, 9)
start = time.time()
val_start = 6000
env_rl_sel_v = TrajComp(traj_path, val_start, val_start + validDB_amount, K, LEN)
sam_size_v = int(env_rl_sel_v.DB_size*ratio)
val_set = env_rl_sel_v.ori_traj_set
env_rl_sam_v = TrajSam(val_set, level_start, 'data')
#only one time (too slow) and dump then,*can comment the lines after*
Rtree_ref_val, _ = F.build_Rtree(val_set, traj_path+'val_tree_'+str(val_start)+'_'+str(val_start+validDB_amount))
print('bounds', Rtree_ref_val.idx.get_bounds())
Rtree_ref_val.idx.close()
#only one time (too slow) and dump then,*can comment the lines after*
Rtree_ref_val = F.obtain_Rtree(traj_path+'val_tree_'+str(val_start)+'_'+str(val_start+validDB_amount))
print('re-check bounds', Rtree_ref_val.idx.get_bounds())
val_DB_DISTRI, _, val_DB_DISTRI_trajID = F.get_distribution_feature_data(val_set)
_, query_part1, query_part2 = F.get_query_workload_data(val_DB_DISTRI)
test_query = query_part1 + query_part2
print('Training...')
Training()
print("Training elapsed time = %s", float(time.time() - start))