-
Notifications
You must be signed in to change notification settings - Fork 160
/
dagger_training.py
executable file
·203 lines (190 loc) · 9.68 KB
/
dagger_training.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
#!/usr/bin/env python3
import argparse
import os
import sys
import time
import numpy as np
import rospy
import shutil
from PlannerLearning import PlannerLearning
from std_msgs.msg import Bool, Empty
from common import setup_sim, place_quad_at_start, MessageHandler
import json
import random
from config.settings import create_settings
MAX_TIME_EXP = 500 # in second, if it takes more the process needs to be killed
class Trainer():
def __init__(self, settings):
rospy.init_node('iterative_learning_node', anonymous=False)
self.settings = settings
np.random.seed(self.settings.random_seed)
self.expert_done = False
self.label_sub = rospy.Subscriber("/hummingbird/labelling_completed", Bool,
self.callback_expert, queue_size=1) # Expert is done, decide what to do.
self.msg_handler = MessageHandler()
def callback_expert(self, data):
# Will receive a true bool when expert is done
print("Expert done with data labelling")
self.expert_done = data.data
def start_experiment(self, rollout_idx):
self.msg_handler.publish_reset()
place_quad_at_start(self.msg_handler)
print("Doing experiment {}".format(rollout_idx))
# Save point_cloud
if self.settings.execute_nw_predictions:
self.msg_handler.publish_save_pc()
else:
# We use expert to collect data
print("Using expert to collect data")
msg = Bool()
msg.data = True
self.learner.expert_pub.publish(msg)
def perform_training(self):
self.learner = PlannerLearning.PlanLearning(
self.settings, mode="iterative")
rollout_idx = 0
# Wipe out expert dir to avoid problems
removable_rollout_folders = os.listdir(self.settings.expert_folder)
if len(removable_rollout_folders) > 0:
removable_rollout_folders = [os.path.join(self.settings.expert_folder, d) \
for d in removable_rollout_folders]
removable_rollout_folders = [d for d in removable_rollout_folders if os.path.isdir(d)]
for d in removable_rollout_folders:
string = "rm -rf {}".format(d)
os.system(string)
while rollout_idx < self.settings.max_rollouts:
if len(os.listdir(self.settings.expert_folder)) > 0:
rollout_dir = os.path.join(self.settings.expert_folder,
sorted(os.listdir(self.settings.expert_folder))[-1])
rm_string = "rm -rf {}".format(rollout_dir)
os.system(rm_string)
self.learner.maneuver_complete = False # Just to be sure
self.expert_done = False # Re-init to be sure
spacing = random.choice(self.settings.tree_spacings)
# whatever is not in the environment will be ignored
self.msg_handler.publish_tree_spacing(spacing)
self.msg_handler.publish_obj_spacing(spacing)
unity_start_pos = setup_sim(self.msg_handler, config=self.settings)
self.start_experiment(rollout_idx)
start = time.time()
exp_failed = False
while not self.learner.maneuver_complete:
time.sleep(0.1)
duration = time.time() - start
if duration > MAX_TIME_EXP:
exp_failed = True
self.learner.publish_stop_recording_msg()
break
if (exp_failed or self.learner.exp_failed):
print("Current experiment failed, will not save data")
if len(os.listdir(self.settings.expert_folder)) > 0:
rollout_dir = os.path.join(self.settings.expert_folder,
sorted(os.listdir(self.settings.expert_folder))[-1])
rm_string = "rm -rf {}".format(rollout_dir)
os.system(rm_string)
else: # Experiment Worked: label it and save it
# final logging if experiment worked
metrics_experiment = self.learner.experiment_report()
for name, value in metrics_experiment.items():
print("{} is {:.3f}".format(name, value))
rollout_idx += 1
# Wait for expert to be done labelling (block gazebo meanwhile)
os.system("rosservice call /gazebo/pause_physics")
# Send message to get expert running
self.learner.run_mppi_expert()
while not self.expert_done:
time.sleep(1)
# Mv data to train folder the labelled data
rollout_dir = os.path.join(self.settings.expert_folder,
sorted(os.listdir(self.settings.expert_folder))[-1])
move_string = "mv {} {}".format(
rollout_dir, self.settings.train_dir)
os.system(move_string)
if rollout_idx % self.settings.train_every_n_rollouts == 0:
self.learner.train()
if rollout_idx % self.settings.increase_net_usage_every_n_rollouts == 0:
self.settings.fallback_radius_expert = \
np.minimum(
self.settings.fallback_radius_expert + 0.5, 50.0)
print("Setting threshold to {}".format(
self.settings.fallback_radius_expert))
os.system("rosservice call /gazebo/unpause_physics")
def perform_testing(self):
self.learner = PlannerLearning.PlanLearning(
self.settings, mode="testing")
tree_spacings = self.settings.tree_spacings
removable_rollout_folders = os.listdir(self.settings.expert_folder)
if len(removable_rollout_folders) > 0:
removable_rollout_folders = [os.path.join(self.settings.expert_folder, d) \
for d in removable_rollout_folders]
removable_rollout_folders = [d for d in removable_rollout_folders if os.path.isdir(d)]
for d in removable_rollout_folders:
string = "rm -rf {}".format(d)
os.system(string)
for spacing in tree_spacings:
self.msg_handler.publish_tree_spacing(spacing)
self.msg_handler.publish_obj_spacing(spacing)
exp_log_dir = os.path.join(self.settings.log_dir, "tree_{}_obj_{}".format(spacing,spacing))
os.makedirs(exp_log_dir)
# Start Experiment
rollout_idx = 0
report_buffer = []
while rollout_idx < self.settings.max_rollouts:
self.learner.maneuver_complete = False # Just to be sure
unity_start_pos = setup_sim(self.msg_handler, config=self.settings)
self.start_experiment(rollout_idx)
output_file_buffer = os.path.join(exp_log_dir,
"experiment_metrics.json")
start = time.time()
exp_failed = False
self.expert_done = False # Re-init to be sure
while not self.learner.maneuver_complete:
time.sleep(0.1)
duration = time.time() - start
if duration > MAX_TIME_EXP:
print("Current experiment failed. Will try again")
exp_failed = True
break
if ((not exp_failed) and (self.learner.planner_succed)):
# final logging
metrics_experiment = self.learner.experiment_report()
report_buffer.append(metrics_experiment)
print("------- {} Rollout ------------".format(rollout_idx+1))
for name, value in metrics_experiment.items():
print("{} is {:.3f}".format(name, value))
print("-------------------------------")
rollout_idx += 1
rollout_dir = os.path.join(self.settings.expert_folder,
sorted(os.listdir(self.settings.expert_folder))[-1])
# Wait one second to stop recording
time.sleep(1)
if self.settings.verbose:
# Mv data record to log folder
move_string = "mv {} {}".format(
rollout_dir, exp_log_dir)
os.system(move_string)
else:
print("Rollout dir is {}".format(rollout_dir))
shutil.rmtree(rollout_dir)
# Save latest version of report buffer
with open(output_file_buffer, 'w') as fout:
json.dump(report_buffer, fout)
else:
# Wait one second to stop recording
time.sleep(1)
# remove folder
rollout_dir = os.path.join(self.settings.expert_folder,
sorted(os.listdir(self.settings.expert_folder))[-1])
rm_string = "rm -rf {}".format(rollout_dir)
os.system(rm_string)
def main():
parser = argparse.ArgumentParser(description='Train Planning network.')
parser.add_argument('--settings_file',
help='Path to settings yaml', required=True)
args = parser.parse_args()
settings_filepath = args.settings_file
settings = create_settings(settings_filepath, mode='dagger')
trainer = Trainer(settings)
trainer.perform_training()
if __name__ == "__main__":
main()