def start_drone(self):
# Get the arguments for this subcommand.
# Create the dagger object and train it.
#pdb.set_trace()
if (self.iteration > 1):
self.dag = dagger.DAgger(self.learning)
if (os.path.isfile('../data/coef.txt')):
print('Load training')
self.dag.load_coef()
else:
print('Start training')
self.dag.train()
self.dag.save_coef()
# Feature extraction parameters.
window_size = (10, 5)
overlap = 0.25
cmd_history_feats = 7 # the approximate number of cmd history features
cmd_history_length = 10 # keep a running list of the last 10 cmds
nav_history_feats = 7 # the approximate number of nav history features
nav_history_length = 10 # keep a running list of the last 10 nav data
# Create the drone object.
self.debug_queue.put({'MSG': ':: Initializing parrot.', 'PRIORITY': 1})
self.debugger.debug()
self.drone = parrot.Parrot()
self.feature_queue = Queue.Queue(maxsize=1)
init_image = self.drone.get_image()
self.feature_extractor = feature_extractor.FeatureExtractor(
self.feature_queue, init_image, window_size, overlap,
cmd_history_feats, cmd_history_length, nav_history_feats,
nav_history_length)
def train(self, args):
""" Starts training. Make sure you have annotated the images first.
"""
self.iterations = args.iterations
self.learning = args.learning
# Create the dagger object.
self.dag = dagger.DAgger(self.learning)
self.dag.aggregate(self.iterations)
def execute(self, args):
# Get the arguments for this subcommand.
self.address = args.address
self.learning = args.learning
self.iteration = args.iteration
self.trajectory = args.trajectory
# Create the dagger object and train it.
pdb.set_trace()
self.dag = dagger.DAgger(self.learning)
self.dag.train()
# Feature extraction parameters.
window_size = (10, 5)
overlap = 0.25
cmd_history_feats = 7 # the approximate number of cmd history features
cmd_history_length = 10 # keep a running list of the last 10 cmds
nav_history_feats = 7 # the approximate number of nav history features
nav_history_length = 10 # keep a running list of the last 10 nav data
self.debug_queue.put({
'MSG': 'Parrot AR 2 Flying Tool :: Execution Mode',
'PRIORITY': 1
})
self.debug_queue.put({
'MSG': ':: GUI flag set to %s.' % str(self.gui),
'PRIORITY': 1
})
self.debug_queue.put({
'MSG': ':: Verbosity set to %d.' % self.verbosity,
'PRIORITY': 1
})
self.debug_queue.put({
'MSG':
':: Accessing controller server at: %s.' % self.address[0],
'PRIORITY':
1
})
self.debug_queue.put({
'MSG':
':: Accessing navigation data server at: %s.' % self.address[1],
'PRIORITY':
1
})
self.debug_queue.put({
'MSG':
':: Accessing camera stream server at: tcp://192.168.1.1:5555.',
'PRIORITY': 1
})
# Create the drone object.
self.debug_queue.put({'MSG': ':: Initializing parrot.', 'PRIORITY': 1})
self.debugger.debug()
self.drone = parrot.Parrot(self.debug_queue, self.error_queue,
self.address, self.learning, self.iteration,
self.trajectory)
self.feature_queue = Queue.Queue(maxsize=1)
init_image = self.drone.get_image()
self.feature_extractor = feature_extractor.FeatureExtractor(
self.feature_queue, init_image, window_size, overlap,
cmd_history_feats, cmd_history_length, nav_history_feats,
nav_history_length)
directory = './data/%s/%s/' % (self.iteration, self.trajectory)
# Start training.
if self.iteration == 1:
self.debug_queue.put({
'MSG':
'Since this is the first iteration, all input will come from the expert.',
'PRIORITY': 1
})
self.debug_queue.put({'MSG': 'Waiting to take off...', 'PRIORITY': 1})
self.debugger.debug()
# Start when the drone takes off.
while True:
cmd = self.drone.get_cmd()
self.drone.send_cmd(cmd)
if cmd is not None:
if cmd['T']:
break
self.debug_queue.put({
'MSG': 'Starting training for iteration %s, trajectory %s.',
'PRIORITY': 1
})
self.debugger.debug()
features_filename = directory + 'features.data'
cmd_filename = directory + 'drone_cmds.data'
# Loop until the drone has landed.
self.time_step = 1
feature_flag = False
while True:
# Land to avoid a crash.
emergency_cmd = self.drone.get_cmd()
if emergency_cmd is not None:
if emergency_cmd['L']:
self.drone.send_cmd(self.drone.remote.land())
break
image_filename = directory + '%s.jpg' % self.time_step
expert_cmd = self.drone.get_cmd()
expert_cmd['X'] = expert_cmd['X'] * 0.15
if self.iteration == 1:
expert_cmd['Y'] = 0.02
if expert_cmd is not None and not feature_flag:
image = self.drone.get_image()
navdata = self.drone.get_navdata()
self.feature_extractor.extract(image)
self.feature_extractor.update(cmd, navdata)
feature_flag = True
try:
features = self.feature_queue.get(block=False)
# Save the features and command.
self.save_image(image, image_filename)
self.save_features(features, features_filename)
self.save_cmd(expert_cmd, cmd_filename)
self.time_step += 1
feature_flag = False
except Queue.Empty:
pass
self.drone.send_cmd(expert_cmd)
else:
if not feature_flag:
image = self.drone.get_image()
navdata = self.drone.get_navdata()
self.feature_extractor.extract(image)
feature_flag = True
try:
features = self.feature_queue.get(block=False)
# Get the command associated with this state.
blah = np.array([0])
blah.shape = (1, 1)
features = np.hstack((blah, features))
x = self.dag.test(features, self.iteration)
cmd = self.drone.default_cmd
cmd['Y'] = 0.02
cmd['X'] = x[0, 0] * 0.15
print(x)
self.feature_extractor.update(cmd, navdata)
# Save the features and command.
self.save_image(image, image_filename)
self.save_features(features, features_filename)
self.save_cmd(cmd, cmd_filename)
self.time_step += 1
self.drone.send_cmd(cmd)
feature_flag = False
except Queue.Empty:
pass