class DistributedAgent():
def __init__(self, parameters):
required_parameters = [
'data_dir', 'max_epoch_runtime_sec', 'replay_memory_size',
'batch_size', 'min_epsilon', 'per_iter_epsilon_reduction',
'experiment_name', 'train_conv_layers'
]
for required_parameter in required_parameters:
if required_parameter not in parameters:
raise ValueError('Missing required parameter {0}'.format(
required_parameter))
parameters['role_type'] = 'agent'
print('Starting time: {0}'.format(datetime.datetime.utcnow()),
file=sys.stderr)
self.__model_buffer = None
self.__model = "sample_model.json"
self.__airsim_started = False
self.__data_dir = parameters['data_dir']
self.__per_iter_epsilon_reduction = float(
parameters['per_iter_epsilon_reduction'])
self.__min_epsilon = float(parameters['min_epsilon'])
self.__max_epoch_runtime_sec = float(
parameters['max_epoch_runtime_sec'])
self.__replay_memory_size = int(parameters['replay_memory_size'])
self.__batch_size = int(parameters['batch_size'])
self.__experiment_name = parameters['experiment_name']
self.__train_conv_layers = bool(
(parameters['train_conv_layers'].lower().strip() == 'true'))
self.__epsilon = 1
self.__num_batches_run = 0
self.__last_checkpoint_batch_count = 0
if 'batch_update_frequency' in parameters:
self.__batch_update_frequency = int(
parameters['batch_update_frequency'])
if 'weights_path' in parameters:
self.__weights_path = parameters['weights_path']
else:
self.__weights_path = None
if 'airsim_path' in parameters:
self.__airsim_path = parameters['airsim_path']
else:
self.__airsim_path = None
self.__local_run = 'local_run' in parameters
self.__car_client = airsim.CarClient()
self.__car_controls = airsim.CarControls()
self.__minibatch_dir = os.path.join(self.__data_dir, 'minibatches')
self.__output_model_dir = os.path.join(self.__data_dir, 'models')
self.__make_dir_if_not_exist(self.__minibatch_dir)
self.__make_dir_if_not_exist(self.__output_model_dir)
self.__last_model_file = ''
self.__possible_ip_addresses = []
self.__trainer_ip_address = None
self.__experiences = {}
self.__init_road_points()
self.__init_reward_points()
# Starts the agent
def start(self):
self.__run_function()
# The function that will be run during training.
# It will initialize the connection to the trainer, start AirSim, and continuously run training iterations.
def __run_function(self):
print('Starting run function')
# Once the trainer is online, it will write its IP to a file in (data_dir)\trainer_ip\trainer_ip.txt
# Wait for that file to exist
if not self.__local_run:
while True:
trainer_ip_dir = os.path.join(
os.path.join(self.__data_dir, 'trainer_ip'),
self.__experiment_name)
print('Checking {0}...'.format(trainer_ip_dir))
if os.path.isdir(trainer_ip_dir):
with open(os.path.join(trainer_ip_dir, 'trainer_ip.txt'),
'r') as f:
self.__possible_ip_addresses.append(f.read().replace(
'\n', ''))
break
time.sleep(5)
# We now have the IP address for the trainer. Attempt to ping the trainer.
ping_idx = -1
while True:
ping_idx += 1
try:
print('\tPinging {0}...'.format(
self.__possible_ip_addresses[ping_idx % len(
self.__possible_ip_addresses)]))
response = requests.get('http://{0}:80/ping'.format(
self.__possible_ip_addresses[ping_idx % len(
self.__possible_ip_addresses)])).json()
if response['message'] != 'pong':
raise ValueError(
'Received unexpected message: {0}'.format(
response))
print('Success!')
self.__trainer_ip_address = self.__possible_ip_addresses[
ping_idx % len(self.__possible_ip_addresses)]
break
except Exception as e:
print('Could not get response. Message is {0}'.format(e))
if (ping_idx % len(self.__possible_ip_addresses) == 0):
print('Waiting 5 seconds and trying again...')
time.sleep(5)
print('Getting model from the trainer')
sys.stdout.flush()
self.__model = RlModel(self.__weights_path,
self.__train_conv_layers)
self.__get_latest_model()
else:
print('Run is local. Skipping connection to trainer.')
self.__model = RlModel(self.__weights_path,
self.__train_conv_layers)
self.__connect_to_airsim()
while True:
try:
self.__run_airsim_epoch(True)
percent_full = 100.0 * len(
self.__experiences['actions']) / self.__replay_memory_size
if (percent_full >= 100.0):
break
except msgpackrpc.error.TimeoutError:
self.__connect_to_airsim()
if not self.__local_run:
self.__get_latest_model()
while True:
try:
if (self.__model is not None):
experiences, frame_count = self.__run_airsim_epoch(False)
# If we didn't immediately crash, train on the gathered experiences
if (frame_count > 0):
sampled_experiences = self.__sample_experiences(
experiences, frame_count, True)
self.__num_batches_run += frame_count
# If we successfully sampled, train on the collected minibatches and send the gradients to the trainer node
if (len(sampled_experiences) > 0):
print('Publishing AirSim Epoch.')
self.__publish_batch_and_update_model(
sampled_experiences, frame_count)
except msgpackrpc.error.TimeoutError:
print('Lost connection to AirSim. Attempting to reconnect.')
self.__connect_to_airsim()
# Connects to the AirSim Exe.
def __connect_to_airsim(self):
attempt_count = 0
while True:
try:
print('Attempting to connect to AirSim (attempt {0})'.format(
attempt_count))
self.__car_client = airsim.CarClient()
self.__car_client.confirmConnection()
self.__car_client.enableApiControl(True)
self.__car_controls = airsim.CarControls()
return
except:
print('Failed to connect.')
attempt_count += 1
if (attempt_count % 10 == 0):
print(
'10 consecutive failures to connect. Attempting to start AirSim on my own.'
)
if self.__local_run:
os.system('START "" powershell.exe {0}'.format(
os.path.join(
self.__airsim_path,
'AD_Cookbook_Start_AirSim.ps1 neighborhood -windowed'
)))
else:
os.system(
'START "" powershell.exe D:\\AD_Cookbook_AirSim\\Scripts\\DistributedRL\\restart_airsim_if_agent.ps1'
)
time.sleep(10)
# Appends a sample to a ring buffer.
# If the appended example takes the size of the buffer over buffer_size, the example at the front will be removed.
def __append_to_ring_buffer(self, item, buffer, buffer_size):
if (len(buffer) >= buffer_size):
buffer = buffer[1:]
buffer.append(item)
return buffer
# Runs an interation of data generation from AirSim.
# Data will be saved in the replay memory.
def __run_airsim_epoch(self, always_random):
print('Running AirSim epoch.')
# Pick a random starting point on the roads
starting_points, starting_direction = self.__get_next_starting_point()
# Initialize the state buffer.
# For now, save 4 images at 0.01 second intervals.
state_buffer_len = 4
state_buffer = []
wait_delta_sec = 0.01
self.__car_controls.steering = 0
self.__car_controls.throttle = 0
self.__car_controls.brake = 1
self.__car_client.setCarControls(self.__car_controls)
time.sleep(2)
# While the car is rolling, start initializing the state buffer
stop_run_time = datetime.datetime.now() + datetime.timedelta(seconds=2)
while (datetime.datetime.now() < stop_run_time):
time.sleep(wait_delta_sec)
state_buffer = self.__append_to_ring_buffer(
self.__get_image(), state_buffer, state_buffer_len)
done = False
actions = [] #records the state we go to
pre_states = []
post_states = []
rewards = []
predicted_rewards = []
car_state = self.__car_client.getCarState()
start_time = datetime.datetime.utcnow()
end_time = start_time + datetime.timedelta(
seconds=self.__max_epoch_runtime_sec)
num_random = 0
far_off = False
# Main data collection loop
while not done:
collision_info = self.__car_client.simGetCollisionInfo()
utc_now = datetime.datetime.utcnow()
# Check for terminal conditions:
# 1) Car has collided
# 2) Car is stopped
# 3) The run has been running for longer than max_epoch_runtime_sec.
# 4) The car has run off the road
if (collision_info.has_collided or car_state.speed < 2
or far_off): # or utc_now > end_time or far_off):
self.__car_client.reset()
sys.stderr.flush()
else:
# The Agent should occasionally pick random action instead of best action
do_greedy = np.random.random_sample()
pre_state = copy.deepcopy(state_buffer)
if (do_greedy < self.__epsilon or always_random):
num_random += 1
next_state = self.__model.get_random_state()
predicted_reward = 0
else:
next_state, predicted_reward = self.__model.predict_state(
pre_state)
print('Model predicts {0}'.format(next_state))
# Convert the selected state to a control signal
next_control_signals = self.__model.state_to_control_signals(
next_state, self.__car_client.getCarState())
# Take the action
self.__car_controls.steering = next_control_signals[0]
self.__car_controls.throttle = next_control_signals[1]
self.__car_controls.brake = next_control_signals[2]
self.__car_client.setCarControls(self.__car_controls)
# Wait for a short period of time to see outcome
time.sleep(wait_delta_sec)
# Observe outcome and compute reward from action
state_buffer = self.__append_to_ring_buffer(
self.__get_image(), state_buffer, state_buffer_len)
car_state = self.__car_client.getCarState()
collision_info = self.__car_client.simGetCollisionInfo()
reward, far_off = self.__compute_reward(
collision_info, car_state)
# Add the experience to the set of examples from this iteration
pre_states.append(pre_state)
post_states.append(state_buffer)
rewards.append(reward)
predicted_rewards.append(predicted_reward)
actions.append(next_state)
# Only the last state is a terminal state.
is_not_terminal = [1 for i in range(0, len(actions) - 1, 1)]
is_not_terminal.append(0)
# Add all of the states from this iteration to the replay memory
self.__add_to_replay_memory('pre_states', pre_states)
self.__add_to_replay_memory('post_states', post_states)
self.__add_to_replay_memory('actions', actions)
self.__add_to_replay_memory('rewards', rewards)
self.__add_to_replay_memory('predicted_rewards', predicted_rewards)
self.__add_to_replay_memory('is_not_terminal', is_not_terminal)
# If we are in the main loop, reduce the epsilon parameter so that the model will be called more often
if not always_random:
self.__epsilon -= self.__per_iter_epsilon_reduction
self.__epsilon = max(self.__epsilon, self.__min_epsilon)
return self.__experiences, len(actions)
# Adds a set of examples to the replay memory
def __add_to_replay_memory(self, field_name, data):
if field_name not in self.__experiences:
self.__experiences[field_name] = data
else:
self.__experiences[field_name] += data
start_index = max(
0,
len(self.__experiences[field_name]) -
self.__replay_memory_size)
self.__experiences[field_name] = self.__experiences[field_name][
start_index:]
# Sample experiences from the replay memory
def __sample_experiences(self, experiences, frame_count, sample_randomly):
sampled_experiences = {}
sampled_experiences['pre_states'] = []
sampled_experiences['post_states'] = []
sampled_experiences['actions'] = []
sampled_experiences['rewards'] = []
sampled_experiences['predicted_rewards'] = []
sampled_experiences['is_not_terminal'] = []
# Compute the surprise factor, which is the difference between the predicted an the actual Q value for each state.
# We can use that to weight examples so that we are more likely to train on examples that the model got wrong.
suprise_factor = np.abs(
np.array(experiences['rewards'], dtype=np.dtype(float)) -
np.array(experiences['predicted_rewards'], dtype=np.dtype(float)))
suprise_factor_normalizer = np.sum(suprise_factor)
suprise_factor /= float(suprise_factor_normalizer)
# Generate one minibatch for each frame of the run
for _ in range(0, frame_count, 1):
if sample_randomly:
idx_set = set(
np.random.choice(list(range(0, suprise_factor.shape[0],
1)),
size=(self.__batch_size),
replace=False))
else:
idx_set = set(
np.random.choice(list(range(0, suprise_factor.shape[0],
1)),
size=(self.__batch_size),
replace=False,
p=suprise_factor))
sampled_experiences['pre_states'] += [
experiences['pre_states'][i] for i in idx_set
]
sampled_experiences['post_states'] += [
experiences['post_states'][i] for i in idx_set
]
sampled_experiences['actions'] += [
experiences['actions'][i] for i in idx_set
]
sampled_experiences['rewards'] += [
experiences['rewards'][i] for i in idx_set
]
sampled_experiences['predicted_rewards'] += [
experiences['predicted_rewards'][i] for i in idx_set
]
sampled_experiences['is_not_terminal'] += [
experiences['is_not_terminal'][i] for i in idx_set
]
return sampled_experiences
# Train the model on minibatches and post to the trainer node.
def __publish_batch_and_update_model(self, batches, batches_count):
# Train and get the gradients
print(
'Publishing epoch data and getting latest model from parameter server...'
)
gradients = self.__model.get_gradient_update_from_batches(batches)
# Post the data to the trainer node
if not self.__local_run:
post_data = {}
post_data['gradients'] = gradients
post_data['batch_count'] = batches_count
response = requests.post('http://{0}:80/gradient_update'.format(
self.__trainer_ip_address),
json=post_data)
print('Response:')
print(response)
new_model_parameters = response.json()
# Update the existing model with the new parameters
self.__model.from_packet(new_model_parameters)
#If the trainer sends us a epsilon, allow it to override our local value
if ('epsilon' in new_model_parameters):
new_epsilon = float(new_model_parameters['epsilon'])
print(
'Overriding local epsilon with {0}, which was sent from trainer'
.format(new_epsilon))
self.__epsilon = new_epsilon
else:
if (self.__num_batches_run > self.__batch_update_frequency +
self.__last_checkpoint_batch_count):
self.__model.update_critic()
checkpoint = {}
checkpoint['model'] = self.__model.to_packet(get_target=True)
checkpoint['batch_count'] = batches_count
checkpoint_str = json.dumps(checkpoint)
checkpoint_dir = os.path.join(
os.path.join(self.__data_dir, 'checkpoint'),
self.__experiment_name)
if not os.path.isdir(checkpoint_dir):
try:
os.makedirs(checkpoint_dir)
except OSError as e:
if e.errno != errno.EEXIST:
raise
file_name = os.path.join(
checkpoint_dir, '{0}.json'.format(self.__num_batches_run))
with open(file_name, 'w') as f:
print('Checkpointing to {0}'.format(file_name))
f.write(checkpoint_str)
self.__last_checkpoint_batch_count = self.__num_batches_run
# Gets the latest model from the trainer node
def __get_latest_model(self):
print('Getting latest model from parameter server...')
response = requests.get('http://{0}:80/latest'.format(
self.__trainer_ip_address)).json()
self.__model.from_packet(response)
# Gets an image from AirSim
def __get_image(self):
image_response = self.__car_client.simGetImages(
[ImageRequest(0, AirSimImageType.Scene, False, False)])[0]
image1d = np.fromstring(image_response.image_data_uint8,
dtype=np.uint8)
image_rgba = image1d.reshape(108, 256, 4)
image_resized = image_rgba[49:108, 0:255, 0:3].astype(float)
return image_resized
# Computes the reward functinon based on the car position.
def __compute_reward(self, collision_info, car_state):
#Define some constant parameters for the reward function
THRESH_DIST = 20 # The maximum distance from the center of the road to compute the reward function
DISTANCE_DECAY_RATE = 1.2 # The rate at which the reward decays for the distance function
CENTER_SPEED_MULTIPLIER = 2.0 # The ratio at which we prefer the distance reward to the speed reward
# If the car has collided, the reward is always zero
if (collision_info.has_collided):
#return 0.0, True
return -5.0, True
# If the car is stopped, the reward is always zero
speed = car_state.speed
if (speed < 2):
return 0.0, True
#Get the car position
position_key = bytes('position', encoding='utf8')
x_val_key = bytes('x_val', encoding='utf8')
y_val_key = bytes('y_val', encoding='utf8')
#car_point = np.array([car_state.kinematics_true[position_key][x_val_key], car_state.kinematics_true[position_key][y_val_key], 0])
car_point = np.array([
car_state.kinematics_estimated.position.x_val,
car_state.kinematics_estimated.position.y_val, 0
])
# Distance component is exponential distance to nearest line
distance = 999
#Compute the distance to the nearest center line
for line in self.__reward_points:
local_distance = 0
length_squared = ((line[0][0] - line[1][0])**2) + (
(line[0][1] - line[1][1])**2)
if (length_squared != 0):
t = max(
0,
min(
1,
np.dot(car_point - line[0], line[1] - line[0]) /
length_squared))
proj = line[0] + (t * (line[1] - line[0]))
local_distance = np.linalg.norm(proj - car_point)
distance = min(local_distance, distance)
distance_reward = math.exp(-(distance * DISTANCE_DECAY_RATE))
if (distance_reward < THRESH_DIST):
distance_reward *= 10
return distance_reward, distance > THRESH_DIST
# Initializes the points used for determining the starting point of the vehicle
def __init_road_points(self):
self.__road_points = []
car_start_coords = [12961.722656, 6660.329102, 0]
with open(
os.path.join(os.path.join(self.__data_dir, 'data'),
'road_lines.txt'), 'r') as f:
for line in f:
points = line.split('\t')
first_point = np.array(
[float(p) for p in points[0].split(',')] + [0])
second_point = np.array(
[float(p) for p in points[1].split(',')] + [0])
self.__road_points.append(tuple((first_point, second_point)))
# Points in road_points.txt are in unreal coordinates
# But car start coordinates are not the same as unreal coordinates
for point_pair in self.__road_points:
for point in point_pair:
point[0] -= car_start_coords[0]
point[1] -= car_start_coords[1]
point[0] /= 100
point[1] /= 100
# Initializes the points used for determining the optimal position of the vehicle during the reward function
def __init_reward_points(self):
self.__reward_points = []
with open(
os.path.join(os.path.join(self.__data_dir, 'data'),
'reward_points.txt'), 'r') as f:
for line in f:
point_values = line.split('\t')
first_point = np.array(
[float(point_values[0]),
float(point_values[1]), 0])
second_point = np.array(
[float(point_values[2]),
float(point_values[3]), 0])
self.__reward_points.append(tuple((first_point, second_point)))
# Randomly selects a starting point on the road
# Used for initializing an iteration of data generation from AirSim
def __get_next_starting_point(self):
car_state = self.__car_client.getCarState()
random_line_index = np.random.randint(0, high=len(self.__road_points))
random_interp = (np.random.random_sample() * 0.4) + 0.3
random_direction_interp = np.random.random_sample()
random_line = self.__road_points[random_line_index]
random_start_point = list(random_line[0])
random_start_point[0] += (random_line[1][0] -
random_line[0][0]) * random_interp
random_start_point[1] += (random_line[1][1] -
random_line[0][1]) * random_interp
# Compute the direction that the vehicle will face
# Vertical line
if (np.isclose(random_line[0][1], random_line[1][1])):
if (random_direction_interp > 0.5):
random_direction = (0, 0, 0)
else:
random_direction = (0, 0, math.pi)
# Horizontal line
elif (np.isclose(random_line[0][0], random_line[1][0])):
if (random_direction_interp > 0.5):
random_direction = (0, 0, math.pi / 2)
else:
random_direction = (0, 0, -1.0 * math.pi / 2)
# The z coordinate is always zero
random_start_point[2] = -0
return (random_start_point, random_direction)
# A helper function to make a directory if it does not exist
def __make_dir_if_not_exist(self, directory):
if not (os.path.exists(directory)):
try:
os.makedirs(directory)
except OSError as e:
if e.errno != errno.EEXIST:
raise
class DistributedAgent():
def __init__(self, parameters):
required_parameters = ['data_dir', 'max_epoch_runtime_sec', 'replay_memory_size', 'batch_size', 'experiment_name', 'train_conv_layers', 'airsim_path', 'airsim_simulation_name', 'coordinator_address']
for required_parameter in required_parameters:
if required_parameter not in parameters:
raise ValueError('Missing required parameter {0}'.format(required_parameter))
print('Starting time: {0}'.format(datetime.datetime.utcnow()), file=sys.stderr)
self.__data_dir = parameters['data_dir']
self.__max_epoch_runtime_sec = float(parameters['max_epoch_runtime_sec'])
self.__replay_memory_size = int(parameters['replay_memory_size'])
self.__batch_size = int(parameters['batch_size'])
self.__experiment_name = parameters['experiment_name']
self.__weights_path = parameters['weights_path'] if 'weights_path' in parameters else None
self.__train_conv_layers = bool((parameters['train_conv_layers'].lower().strip() == 'true'))
self.__epsilon = 1
self.__experiences = {}
self.__possible_ip_addresses = [parameters['coordinator_address']]
self.__trainer_ip_address = None
self.__model = RlModel(parameters['weights_path'] if 'weights_path' in parameters else None, self.__train_conv_layers)
self.__car_connector = AirSimCarConnector(parameters['airsim_path'], parameters['airsim_simulation_name'], self.__data_dir)
self.__rewarder = CollisionRewarder(self.__data_dir)
# Starts the agent
def start(self):
self.__run_function()
# The function that will be run during training.
# It will initialize the connection to the trainer, start AirSim, and continuously run training iterations.
def __run_function(self):
def train():
try:
self.__experiences = {}
self.__model = RlModel(self.__weights_path, self.__train_conv_layers)
self.__fill_replay_memory()
self.__get_latest_model()
while True:
self.__train_model()
except Exception as e:
print(e, flush=True)
print('Error during training - reinitialization', flush=True)
return
self.__ping_coordinator()
self.__get_latest_model()
self.__car_connector.connect()
while True:
train()
# We have the IP address for the trainer. Attempt to ping the trainer.
def __ping_coordinator(self):
ping_idx = -1
while True:
ping_idx += 1
print('Attempting to ping trainer...', flush=True)
try:
print('\tPinging {0}...'.format(self.__possible_ip_addresses[ping_idx % len(self.__possible_ip_addresses)]), flush=True)
response = requests.get('http://{0}/ping'.format(self.__possible_ip_addresses[ping_idx % len(self.__possible_ip_addresses)])).json()
if response['message'] != 'pong':
raise ValueError('Received unexpected message: {0}'.format(response))
print('Success!', flush=True)
self.__trainer_ip_address = self.__possible_ip_addresses[ping_idx % len(self.__possible_ip_addresses)]
return
except Exception as e:
print('Could not get response. Message is {0}'.format(e), flush=True)
if (ping_idx % len(self.__possible_ip_addresses) == 0):
print('Waiting 5 seconds and trying again...', flush=True)
time.sleep(5)
# Fill the replay memory by driving randomly.
def __fill_replay_memory(self):
print('Filling replay memory...', flush=True)
while True:
print('Running Airsim Epoch.', flush=True)
try:
self.__run_training_epoch(True)
percent_full = 100.0 * len(self.__experiences['actions']) / self.__replay_memory_size
print('Replay memory now contains {0} members. ({1}% full)'.format(len(self.__experiences['actions']), percent_full), flush=True)
if (percent_full >= 100.0):
return
except ConnectorException:
print('Lost connection to car while fillling replay memory. Attempting to reconnect...', flush=True)
self.__car_connector.connect()
# Training step
def __train_model(self):
try:
#Generate a series of training examples by driving the vehicle in AirSim
print('Running Airsim Epoch.', flush=True)
experiences, frame_count = self.__run_training_epoch(False)
# If we didn't immediately crash, train on the gathered experiences
if (frame_count > 0):
print('Generating {0} minibatches...'.format(frame_count), flush=True)
print('Sampling Experiences.', flush=True)
# Sample experiences from the replay memory
sampled_experiences = self.__sample_experiences(experiences, frame_count, True)
# If we successfully sampled, train on the collected minibatches and send the gradients to the trainer node
if (len(sampled_experiences) > 0):
print('Publishing training epoch results...', flush=True)
self.__publish_batch_and_update_model(sampled_experiences, frame_count)
# Occasionally, the AirSim exe will stop working.
# For example, if a user connects to the node to visualize progress.
# In that case, attempt to reconnect.
except ConnectorException:
print('Lost connection to car while training. Attempting to reconnect...', flush=True)
self.__car_connector.connect()
# Runs an interation of data generation from AirSim.
# Data will be saved in the replay memory.
def __run_training_epoch(self, always_random):
print('Training epoch.', flush=True)
self.__car_connector.reset_car()
# Initialize the state buffer.
# For now, save 4 images at 0.01 second intervals.
state_buffer_len = 4
state_buffer = []
wait_delta_sec = 0.01
# Start initializing the state buffer
stop_run_time = datetime.datetime.now() + datetime.timedelta(seconds=2)
while(datetime.datetime.now() < stop_run_time):
time.sleep(wait_delta_sec)
state_buffer = append_to_ring_buffer(self.__car_connector.get_image(), state_buffer, state_buffer_len)
done = False
actions = []
pre_states = []
post_states = []
rewards = []
predicted_rewards = []
# car_state = self.__car_connector.car_state()
start_time = datetime.datetime.utcnow()
end_time = start_time + datetime.timedelta(seconds=self.__max_epoch_runtime_sec)
num_random = 0
far_off = False
# Main data collection loop
# Check for terminal conditions:
# 1) Car has collided
# 2) Car is stopped (or car_state.speed < 2) DELETED
# 3) The run has been running for longer than max_epoch_runtime_sec.
# This constraint is so the model doesn't end up having to churn through huge chunks of data, slowing down training
while not (self.__car_connector.has_collided() or datetime.datetime.utcnow() > end_time or far_off):
# The Agent should occasionally pick random action instead of best action
do_greedy = np.random.random_sample()
pre_state = copy.deepcopy(state_buffer)
if (do_greedy < self.__epsilon or always_random):
num_random += 1
next_state = self.__model.get_random_state()
predicted_reward = 0
else:
next_state, predicted_reward = self.__model.predict_state(pre_state)
print('Model predicts {0}'.format(next_state), flush=True)
self.__car_connector.execute_action(next_state)
# Wait for a short period of time to see outcome
time.sleep(wait_delta_sec)
# Observe outcome and compute reward from action
state_buffer = append_to_ring_buffer(self.__car_connector.get_image(), state_buffer, state_buffer_len)
car_state = self.__car_connector.car_state()
collision_info = self.__car_connector.has_collided()
reward, far_off = self.__rewarder.compute_reward(collision_info, car_state)
print('Actual reward: {0} - Predicted : {1}'.format(reward, predicted_reward))
# Add the experience to the set of examples from this iteration
pre_states.append(pre_state)
post_states.append(state_buffer)
rewards.append(reward)
predicted_rewards.append(predicted_reward)
actions.append(next_state)
if (datetime.datetime.utcnow() < end_time):
rewards = self.__rewarder.discount_rewards(rewards)
print('Start time: {0}, end time: {1}'.format(start_time, datetime.datetime.utcnow()), file=sys.stderr)
if (datetime.datetime.utcnow() > end_time):
print('Timed out.')
print('Full autonomous run finished at {0}'.format(datetime.datetime.utcnow()), file=sys.stderr)
sys.stderr.flush()
# Only the last state is a terminal state.
is_not_terminal = [1 for i in range(0, len(actions) - 1, 1)]
is_not_terminal.append(0)
# Add all of the states from this iteration to the replay memory
self.__add_to_replay_memory('pre_states', pre_states)
self.__add_to_replay_memory('post_states', post_states)
self.__add_to_replay_memory('actions', actions)
self.__add_to_replay_memory('rewards', rewards)
self.__add_to_replay_memory('predicted_rewards', predicted_rewards)
self.__add_to_replay_memory('is_not_terminal', is_not_terminal)
print('Percent random actions: {0}'.format(num_random / max(1, len(actions))), flush=True)
print('Num total actions: {0}'.format(len(actions)), flush=True)
return self.__experiences, len(actions)
# Adds a set of examples to the replay memory
def __add_to_replay_memory(self, field_name, data):
if field_name not in self.__experiences:
self.__experiences[field_name] = data
else:
self.__experiences[field_name] += data
start_index = max(0, len(self.__experiences[field_name]) - self.__replay_memory_size)
self.__experiences[field_name] = self.__experiences[field_name][start_index:]
# Sample experiences from the replay memory
def __sample_experiences(self, experiences, frame_count, sample_randomly):
sampled_experiences = {}
sampled_experiences['pre_states'] = []
sampled_experiences['post_states'] = []
sampled_experiences['actions'] = []
sampled_experiences['rewards'] = []
sampled_experiences['predicted_rewards'] = []
sampled_experiences['is_not_terminal'] = []
# Compute the surprise factor, which is the difference between the predicted an the actual Q value for each state.
# We can use that to weight examples so that we are more likely to train on examples that the model got wrong.
suprise_factor = np.abs(np.array(experiences['rewards'], dtype=np.dtype(float)) - np.array(experiences['predicted_rewards'], dtype=np.dtype(float)))
suprise_factor_normalizer = np.sum(suprise_factor)
suprise_factor /= float(suprise_factor_normalizer)
# Generate one minibatch for each frame of the run
for _ in range(0, frame_count, 1):
if sample_randomly:
idx_set = set(np.random.choice(list(range(0, suprise_factor.shape[0], 1)), size=(self.__batch_size), replace=False))
else:
idx_set = set(np.random.choice(list(range(0, suprise_factor.shape[0], 1)), size=(self.__batch_size), replace=False, p=suprise_factor))
sampled_experiences['pre_states'] += [experiences['pre_states'][i] for i in idx_set]
sampled_experiences['post_states'] += [experiences['post_states'][i] for i in idx_set]
sampled_experiences['actions'] += [experiences['actions'][i] for i in idx_set]
sampled_experiences['rewards'] += [experiences['rewards'][i] for i in idx_set]
sampled_experiences['predicted_rewards'] += [experiences['predicted_rewards'][i] for i in idx_set]
sampled_experiences['is_not_terminal'] += [experiences['is_not_terminal'][i] for i in idx_set]
return sampled_experiences
# Train the model on minibatches and post to the trainer node.
# The trainer node will respond with the latest version of the model that will be used in further data generation iterations.
def __publish_batch_and_update_model(self, batches, batches_count):
# Train and get the gradients
print('Publishing epoch data and getting latest model from parameter server...', flush=True)
gradients = self.__model.get_gradient_update_from_batches(batches)
if gradients is None:
return
# Post the data to the trainer node
post_data = {}
post_data['gradients'] = gradients
post_data['batch_count'] = batches_count
response = requests.post('http://{0}/gradient_update'.format(self.__trainer_ip_address), json=post_data)
print('Response:', flush=True)
print(response, flush=True)
new_model_parameters = response.json()
# Update the existing model with the new parameters
self.__model.from_packet(new_model_parameters)
#If the trainer sends us a epsilon, allow it to override our local value
if ('epsilon' in new_model_parameters):
new_epsilon = float(new_model_parameters['epsilon'])
print('Overriding local epsilon with {0}, which was sent from trainer'.format(new_epsilon), flush=True)
self.__epsilon = new_epsilon
# Gets the latest model from the trainer node
def __get_latest_model(self):
print('Getting latest model from parameter server...', flush=True)
response = requests.get('http://{0}/latest'.format(self.__trainer_ip_address)).json()
self.__model.from_packet(response)
class DistributedAgent():
def __init__(self, parameters):
required_parameters = [
'data_dir', 'max_epoch_runtime_sec', 'replay_memory_size',
'batch_size', 'min_epsilon', 'per_iter_epsilon_reduction',
'experiment_name', 'train_conv_layers', 'start_x', 'start_y',
'start_z', 'log_path'
]
for required_parameter in required_parameters:
if required_parameter not in parameters:
raise ValueError('Missing required parameter {0}'.format(
required_parameter))
parameters['role_type'] = 'agent'
print('Starting time: {0}'.format(datetime.datetime.utcnow()),
file=sys.stderr)
self.__model_buffer = None
self.__model = None
self.__airsim_started = False
self.__data_dir = parameters['data_dir']
self.__per_iter_epsilon_reduction = float(
parameters['per_iter_epsilon_reduction'])
self.__min_epsilon = float(parameters['min_epsilon'])
self.__max_epoch_runtime_sec = float(
parameters['max_epoch_runtime_sec'])
self.__replay_memory_size = int(parameters['replay_memory_size'])
self.__batch_size = int(parameters['batch_size'])
self.__experiment_name = parameters['experiment_name']
self.__train_conv_layers = bool(
(parameters['train_conv_layers'].lower().strip() == 'true'))
self.__epsilon = 1
self.__num_batches_run = 0
self.__last_checkpoint_batch_count = 0
if 'batch_update_frequency' in parameters:
self.__batch_update_frequency = int(
parameters['batch_update_frequency'])
if 'weights_path' in parameters:
self.__weights_path = parameters['weights_path']
else:
self.__weights_path = None
if 'airsim_path' in parameters:
self.__airsim_path = parameters['airsim_path']
else:
self.__airsim_path = None
self.__local_run = 'local_run' in parameters
self.__car_client = None
self.__car_controls = None
self.__minibatch_dir = os.path.join(self.__data_dir, 'minibatches')
self.__output_model_dir = os.path.join(self.__data_dir, 'models')
self.__make_dir_if_not_exist(self.__minibatch_dir)
self.__make_dir_if_not_exist(self.__output_model_dir)
self.__last_model_file = ''
self.__possible_ip_addresses = []
self.__trainer_ip_address = None
self.__experiences = {}
self.__start_point = [
float(parameters['start_x']),
float(parameters['start_y']),
float(parameters['start_z'])
]
self.__log_file = parameters['log_path']
# initiate coverage map
self.__coverage_map = CoverageMap(start_point=self.__start_point,
map_size=12000,
scale_ratio=20,
state_size=6000,
input_size=20,
height_threshold=0.9,
reward_norm=30,
paint_radius=15)
# create txt file
if not os.path.isdir(
os.path.join(self.__data_dir, '\\checkpoint',
self.__experiment_name)):
os.makedirs(
os.path.join(self.__data_dir, '\\checkpoint',
self.__experiment_name))
self.__rewards_log = open(
os.path.join(self.__data_dir, '\\checkpoint',
self.__experiment_name, "rewards.txt"), "w")
self.__rewards_log.write("Timestamp\tSum\tMean\n")
self.__rewards_log.close()
# create starting points list
#self.__starting_points = self.__get_starting_points()
self.__starting_points = [self.__start_point]
# Starts the agent
def start(self):
self.__run_function()
# The function that will be run during training.
# It will initialize the connection to the trainer, start AirSim, and continuously run training iterations.
def __run_function(self):
print('Starting run function')
# Once the trainer is online, it will write its IP to a file in (data_dir)\trainer_ip\trainer_ip.txt
# Wait for that file to exist
if not self.__local_run:
print('Waiting for trainer to come online')
while True:
trainer_ip_dir = os.path.join(
os.path.join(self.__data_dir, 'trainer_ip'),
self.__experiment_name)
print('Checking {0}...'.format(trainer_ip_dir))
if os.path.isdir(trainer_ip_dir):
with open(os.path.join(trainer_ip_dir, 'trainer_ip.txt'),
'r') as f:
self.__possible_ip_addresses.append(f.read().replace(
'\n', ''))
break
print('Not online yet. Sleeping...')
time.sleep(5)
# We now have the IP address for the trainer. Attempt to ping the trainer.
ping_idx = -1
while True:
ping_idx += 1
print('Attempting to ping trainer...')
try:
print('\tPinging {0}...'.format(
self.__possible_ip_addresses[ping_idx % len(
self.__possible_ip_addresses)]))
response = requests.get('http://{0}:80/ping'.format(
self.__possible_ip_addresses[ping_idx % len(
self.__possible_ip_addresses)])).json()
if response['message'] != 'pong':
raise ValueError(
'Received unexpected message: {0}'.format(
response))
print('Success!')
self.__trainer_ip_address = self.__possible_ip_addresses[
ping_idx % len(self.__possible_ip_addresses)]
break
except Exception as e:
print('Could not get response. Message is {0}'.format(e))
if (ping_idx % len(self.__possible_ip_addresses) == 0):
print('Waiting 5 seconds and trying again...')
time.sleep(5)
# Get the latest model from the trainer
print('Getting model from the trainer')
sys.stdout.flush()
buffer_len = 4
self.__model = RlModel(weights_path=self.__weights_path,
train_conv_layers=self.__train_conv_layers,
buffer_len=buffer_len)
self.__get_latest_model()
else:
print('Run is local. Skipping connection to trainer.')
buffer_len = 4
self.__model = RlModel(weights_path=self.__weights_path,
train_conv_layers=self.__train_conv_layers,
buffer_len=buffer_len)
# Connect to the AirSim exe
self.__connect_to_airsim()
# Fill the replay memory by driving randomly.
print('Filling replay memory...')
while True:
print('Running Airsim Epoch.')
try:
_, num_of_actions = self.__run_airsim_epoch(True)
if num_of_actions > 0:
percent_full = 100.0 * len(self.__experiences['actions']
) / self.__replay_memory_size
print(
'Replay memory now contains {0} members. ({1}% full)'.
format(len(self.__experiences['actions']),
percent_full))
if (percent_full >= 100.0):
break
except msgpackrpc.error.TimeoutError:
print(
'Lost connection to AirSim while filling replay memory. Attempting to reconnect.'
)
self.__connect_to_airsim()
# Get the latest model. Other agents may have finished before us.
print('Replay memory filled. Starting main loop...')
if not self.__local_run:
self.__get_latest_model()
while True:
try:
if (self.__model is not None):
#Generate a series of training examples by driving the vehicle in AirSim
print('Running Airsim Epoch.')
experiences, frame_count = self.__run_airsim_epoch(False)
# If we didn't immediately crash, train on the gathered experiences
if (frame_count > 0):
print('Generating {0} minibatches...'.format(
frame_count))
print('Sampling Experiences.')
# Sample experiences from the replay memory
sampled_experiences = self.__sample_experiences(
experiences, frame_count, True)
self.__num_batches_run += frame_count
# If we successfully sampled, train on the collected minibatches and send the gradients to the trainer node
if (len(sampled_experiences) > 0):
print('Publishing AirSim Epoch.')
# write all rewards to log file
self.__rewards_log = open(
os.path.join(self.__data_dir, 'checkpoint',
self.__experiment_name,
"rewards.txt"), "a+")
rewards_sum = 0
for reward in sampled_experiences['rewards']:
rewards_sum += reward
self.__rewards_log.write("{}\t{}\t{}\n".format(
time.time(), rewards_sum, rewards_sum /
len(sampled_experiences['rewards'])))
self.__rewards_log.close()
self.__publish_batch_and_update_model(
sampled_experiences, frame_count)
# Occasionally, the AirSim exe will stop working.
# For example, if a user connects to the node to visualize progress.
# In that case, attempt to reconnect.
except msgpackrpc.error.TimeoutError:
print('Lost connection to AirSim. Attempting to reconnect.')
self.__connect_to_airsim()
# Connects to the AirSim Exe.
# Assume that it is already running. After 10 successive attempts, attempt to restart the executable.
def __connect_to_airsim(self):
attempt_count = 0
while True:
try:
print('Attempting to connect to AirSim (attempt {0})'.format(
attempt_count))
self.__car_client = airsim.CarClient()
self.__car_client.confirmConnection()
self.__car_client.enableApiControl(True)
self.__car_controls = airsim.CarControls()
self.__coverage_map.set_client(
client=self.__car_client) # update client on coverage map
print('Connected!')
return
except:
print('Failed to connect.')
attempt_count += 1
if (attempt_count % 10 == 0):
print(
'10 consecutive failures to connect. Attempting to start AirSim on my own.'
)
if self.__local_run:
os.system('START "" powershell.exe {0}'.format(
os.path.join(
self.__airsim_path,
'AD_Cookbook_Start_AirSim.ps1 neighborhood -windowed'
)))
else:
os.system(
'START "" powershell.exe D:\\AD_Cookbook_AirSim\\Scripts\\DistributedRL\\restart_airsim_if_agent.ps1'
)
print('Waiting a few seconds.')
time.sleep(10)
# Appends a sample to a ring buffer.
# If the appended example takes the size of the buffer over buffer_size, the example at the front will be removed.
def __append_to_ring_buffer(self, item, buffer, buffer_size):
if (len(buffer) >= buffer_size):
buffer = buffer[1:]
buffer.append(item)
return buffer
# Runs an interation of data generation from AirSim.
# Data will be saved in the replay memory.
def __run_airsim_epoch(self, always_random):
print('Running AirSim epoch.')
# reset coverage map
self.__coverage_map.reset()
# Pick a random starting point on the roads
starting_points, starting_direction = self.__get_next_starting_point()
# Initialize the state buffer.
# For now, save 4 images at 0.01 second intervals.
state_buffer_len = 4
state_buffer = []
wait_delta_sec = 0.01
print('Getting Pose')
self.__car_client.simSetVehiclePose(
airsim.Pose(
airsim.Vector3r(starting_points[0], starting_points[1],
starting_points[2]),
toQuaternion(starting_direction[0], starting_direction[1],
starting_direction[2])), True)
# Currently, simSetVehiclePose does not allow us to set the velocity.
# So, if we crash and call simSetVehiclePose, the car will be still moving at its previous velocity.
# We need the car to stop moving, so push the brake and wait for a few seconds.
print('Waiting for momentum to die')
self.__car_controls.steering = 0
self.__car_controls.throttle = 0
self.__car_controls.brake = 1
self.__car_client.setCarControls(self.__car_controls)
time.sleep(4)
print('Resetting')
self.__car_client.simSetVehiclePose(
airsim.Pose(
airsim.Vector3r(starting_points[0], starting_points[1],
starting_points[2]),
toQuaternion(starting_direction[0], starting_direction[1],
starting_direction[2])), True)
#Start the car rolling so it doesn't get stuck
print('Running car for a few seconds...')
self.__car_controls.steering = 0
self.__car_controls.throttle = 0.4
self.__car_controls.brake = 0
self.__car_client.setCarControls(self.__car_controls)
# While the car is rolling, start initializing the state buffer
stop_run_time = datetime.datetime.now() + datetime.timedelta(seconds=1)
while (datetime.datetime.now() < stop_run_time):
time.sleep(wait_delta_sec)
image, _ = self.__get_image()
state_buffer = self.__append_to_ring_buffer(
image, state_buffer, state_buffer_len)
done = False
actions = [] #records the state we go to
pre_states = []
post_states = []
rewards = []
predicted_rewards = []
car_state = self.__car_client.getCarState()
# slow down a bit
self.__car_controls.throttle = 0.3
self.__car_client.setCarControls(self.__car_controls)
start_time = datetime.datetime.utcnow()
end_time = start_time + datetime.timedelta(
seconds=self.__max_epoch_runtime_sec)
num_random = 0
# Main data collection loop
while not done:
collision_info = self.__car_client.simGetCollisionInfo()
utc_now = datetime.datetime.utcnow()
# Check for terminal conditions:
# 1) Car has collided
# 2) Car is stopped
# 3) The run has been running for longer than max_epoch_runtime_sec.
# This constraint is so the model doesn't end up having to churn through huge chunks of data, slowing down training
if (collision_info.has_collided or abs(car_state.speed) < 0.02
or utc_now > end_time):
print('Start time: {0}, end time: {1}'.format(
start_time, utc_now),
file=sys.stderr)
if (utc_now > end_time):
print('timed out.')
print(
'Full autonomous run finished at {0}'.format(utc_now),
file=sys.stderr)
done = True
sys.stderr.flush()
else:
# The Agent should occasionally pick random action instead of best action
do_greedy = np.random.random_sample()
pre_state = copy.deepcopy(state_buffer)
if (do_greedy < self.__epsilon or always_random):
num_random += 1
next_state = self.__model.get_random_state()
predicted_reward = 0
else:
next_state, predicted_reward, _ = self.__model.predict_state(
pre_state)
print('Model predicts {0}'.format(next_state))
# Convert the selected state to a control signal
next_steering, next_brake = self.__model.state_to_control_signals(
next_state, self.__car_client.getCarState())
# Take the action
self.__car_controls.steering = next_steering
self.__car_controls.brake = next_brake
self.__car_client.setCarControls(self.__car_controls)
# Wait for a short period of time to see outcome
time.sleep(wait_delta_sec)
# Observe outcome and compute reward from action
post_image, cov_reward = self.__get_image()
state_buffer = self.__append_to_ring_buffer(
post_image, state_buffer, state_buffer_len)
car_state = self.__car_client.getCarState()
collision_info = self.__car_client.simGetCollisionInfo()
reward = self.__compute_reward(collision_info, car_state,
cov_reward, next_state)
# Add the experience to the set of examples from this iteration
pre_states.append(pre_state)
post_states.append(state_buffer)
rewards.append(reward)
predicted_rewards.append(predicted_reward)
actions.append(next_state)
# Only the last state is a terminal state.
is_not_terminal = [1 for i in range(0, len(actions) - 1, 1)]
is_not_terminal.append(0)
# only add to the replay memory if have enough data
if len(actions) > 30:
# Add all of the states from this iteration to the replay memory
self.__add_to_replay_memory('pre_states', pre_states)
self.__add_to_replay_memory('post_states', post_states)
self.__add_to_replay_memory('actions', actions)
self.__add_to_replay_memory('rewards', rewards)
self.__add_to_replay_memory('predicted_rewards', predicted_rewards)
self.__add_to_replay_memory('is_not_terminal', is_not_terminal)
print('Percent random actions: {0}'.format(num_random /
max(1, len(actions))))
print('Num total actions: {0}'.format(len(actions)))
# If we are in the main loop, reduce the epsilon parameter so that the model will be called more often
# Note: this will be overwritten by the trainer's epsilon if running in distributed mode
if not always_random:
self.__epsilon -= self.__per_iter_epsilon_reduction
self.__epsilon = max(self.__epsilon, self.__min_epsilon)
return self.__experiences, len(actions)
else:
return self.__experiences, 0
# Adds a set of examples to the replay memory
def __add_to_replay_memory(self, field_name, data):
if field_name not in self.__experiences:
self.__experiences[field_name] = data
else:
self.__experiences[field_name] += data
start_index = max(
0,
len(self.__experiences[field_name]) -
self.__replay_memory_size)
self.__experiences[field_name] = self.__experiences[field_name][
start_index:]
# Sample experiences from the replay memory
def __sample_experiences(self, experiences, frame_count, sample_randomly):
sampled_experiences = {}
sampled_experiences['pre_states'] = []
sampled_experiences['post_states'] = []
sampled_experiences['actions'] = []
sampled_experiences['rewards'] = []
sampled_experiences['predicted_rewards'] = []
sampled_experiences['is_not_terminal'] = []
# Compute the surprise factor, which is the difference between the predicted an the actual Q value for each state.
# We can use that to weight examples so that we are more likely to train on examples that the model got wrong.
suprise_factor = np.abs(
np.array(experiences['rewards'], dtype=np.dtype(float)) -
np.array(experiences['predicted_rewards'], dtype=np.dtype(float)))
suprise_factor_normalizer = np.sum(suprise_factor)
suprise_factor /= float(suprise_factor_normalizer)
# Generate one minibatch for each frame of the run
for _ in range(0, frame_count, 1):
if sample_randomly:
idx_set = set(
np.random.choice(list(range(0, suprise_factor.shape[0],
1)),
size=(self.__batch_size),
replace=False))
else:
idx_set = set(
np.random.choice(list(range(0, suprise_factor.shape[0],
1)),
size=(self.__batch_size),
replace=False,
p=suprise_factor))
sampled_experiences['pre_states'] += [
experiences['pre_states'][i] for i in idx_set
]
sampled_experiences['post_states'] += [
experiences['post_states'][i] for i in idx_set
]
sampled_experiences['actions'] += [
experiences['actions'][i] for i in idx_set
]
sampled_experiences['rewards'] += [
experiences['rewards'][i] for i in idx_set
]
sampled_experiences['predicted_rewards'] += [
experiences['predicted_rewards'][i] for i in idx_set
]
sampled_experiences['is_not_terminal'] += [
experiences['is_not_terminal'][i] for i in idx_set
]
return sampled_experiences
# Train the model on minibatches and post to the trainer node.
# The trainer node will respond with the latest version of the model that will be used in further data generation iterations.
def __publish_batch_and_update_model(self, batches, batches_count):
# Train and get the gradients
print(
'Publishing epoch data and getting latest model from parameter server...'
)
gradients = self.__model.get_gradient_update_from_batches(batches)
# Post the data to the trainer node
if not self.__local_run:
post_data = {}
post_data['gradients'] = gradients
post_data['batch_count'] = batches_count
response = requests.post('http://{0}:80/gradient_update'.format(
self.__trainer_ip_address),
json=post_data)
print('Response:')
print(response)
new_model_parameters = response.json()
# Update the existing model with the new parameters
self.__model.from_packet(new_model_parameters)
#If the trainer sends us a epsilon, allow it to override our local value
if ('epsilon' in new_model_parameters):
new_epsilon = float(new_model_parameters['epsilon'])
print(
'Overriding local epsilon with {0}, which was sent from trainer'
.format(new_epsilon))
self.__epsilon = new_epsilon
else:
if (self.__num_batches_run > self.__batch_update_frequency +
self.__last_checkpoint_batch_count):
self.__model.update_critic()
checkpoint = {}
checkpoint['model'] = self.__model.to_packet(get_target=True)
checkpoint['batch_count'] = batches_count
checkpoint_str = json.dumps(checkpoint)
checkpoint_dir = os.path.join(
os.path.join(self.__data_dir, 'checkpoint'),
self.__experiment_name)
if not os.path.isdir(checkpoint_dir):
try:
os.makedirs(checkpoint_dir)
except OSError as e:
if e.errno != errno.EEXIST:
raise
file_name = os.path.join(
checkpoint_dir, '{0}.json'.format(self.__num_batches_run))
with open(file_name, 'w') as f:
print('Checkpointing to {0}'.format(file_name))
f.write(checkpoint_str)
self.__last_checkpoint_batch_count = self.__num_batches_run
# Gets the latest model from the trainer node
def __get_latest_model(self):
print('Getting latest model from parameter server...')
response = requests.get('http://{0}:80/latest'.format(
self.__trainer_ip_address)).json()
self.__model.from_packet(response)
# Gets a coverage image from AirSim
def __get_cov_image(self):
state, cov_reward = self.__coverage_map.get_state_from_pose()
# debug only
#im = Image.fromarray(np.uint8(state))
#im.save("DistributedRL\\debug\\{}.png".format(time.time()))
# normalize state
state = state / 255.0
return state, cov_reward
# Gets an image from AirSim
def __get_image(self):
responses = self.__car_client.simGetImages([
airsim.ImageRequest("RCCamera", airsim.ImageType.DepthPerspective,
True, False)
])
img1d = np.array(responses[0].image_data_float, dtype=np.float)
if img1d.size > 1:
img1d = 255 / np.maximum(np.ones(img1d.size), img1d)
img2d = np.reshape(img1d,
(responses[0].height, responses[0].width))
image = Image.fromarray(img2d)
# debug only
#image_png = image.convert('RGB')
#image_png.save("DistributedRL\\debug\\{}.png".format(time.time()))
depth_im = np.array(image.resize((84, 84)).convert('L'))
depth_im = depth_im / 255.0
else:
depth_im = np.zeros((84, 84)).astype(float)
cov_im, cov_reward = self.__get_cov_image()
depth_im[:cov_im.shape[0], :cov_im.shape[1]] = cov_im
#image = Image.fromarray(depth_im)
#image.save("DistributedRL\\debug\\{}.png".format(time.time()))
return depth_im, cov_reward
"""
image_response = self.__car_client.simGetImages([airsim.ImageRequest("RCCamera", airsim.ImageType.Scene, False, False)])[0]
image1d = np.fromstring(image_response.image_data_uint8, dtype=np.uint8)
if image1d.size > 1:
image_rgba = image1d.reshape(image_response.height, image_response.width, 4)
#im = Image.fromarray(np.uint8(image_rgba))
#im.save("DistributedRL\\debug\\{}.png".format(time.time()))
image_rgba = image_rgba / 255.0
return image_rgba[60:144,86:170,0:3].astype(float)
return np.zeros((84,84,3)).astype(float)
"""
# Computes the reward functinon based on collision.
def __compute_reward(self, collision_info, car_state, cov_reward, action):
alpha = 1.0
# If the car has collided, the reward is always zero
if (collision_info.has_collided):
return 0.0
# If the car has stopped for some reason, the reward is always zero
if abs(car_state.speed) < 0.02:
return 0.0
# If there is no new coverage, there is no reward
if cov_reward < 0.1:
return 0.0
# straight will be rewarded as 1.0, semi straight as 0.5
direction_reward = float(2 - abs(action - 2)) / 2.0
# final reward
reward = alpha * cov_reward + (1 - alpha) * direction_reward
#print("cov reward: {}, reward: {}".format(cov_reward, reward))
return reward
# prepare starting points list
def __get_starting_points(self):
starting_points_file = open(
os.path.join(self.__data_dir, 'data\\starting_points.txt'))
starting_points_list = []
for line in starting_points_file:
starting_points_list.append(
[float(x) for x in line.split(' ')[:3]])
return starting_points_list
# get most newly generated random point
def __get_next_generated_random_point(self):
"""
# grab the newest line with generated random point
newest_rp = "None"
# keep searching until the simulation is giving something
while newest_rp == "None":
# notify user
print("Searching for a random point...")
# open log file
log_file = open(self.__log_file, "r")
# search for the newest generated random point line
for line in log_file:
if "RandomPoint" in line:
newest_rp = line
# notify user
print("Found random point.")
# filter random point from line
random_point = [float(newest_rp.split(" ")[-3].split("=")[1]), float(newest_rp.split(" ")[-2].split("=")[1]), float(newest_rp.split(" ")[-1].split("=")[1])]
return random_point
"""
idx = randint(0, len(self.__starting_points) - 1)
return self.__starting_points[idx]
# Randomly selects a starting point on the road
# Used for initializing an iteration of data generation from AirSim
def __get_next_starting_point(self):
# get random start point from log file, and make it relative to agent's starting point
random_start_point = self.__get_next_generated_random_point()
random_start_point = [
random_start_point[0] - self.__start_point[0],
random_start_point[1] - self.__start_point[1],
random_start_point[2] - self.__start_point[2]
]
random_start_point = [x / 100.0 for x in random_start_point]
# draw random orientation
#random_direction = (0, 0, np.random.uniform(-math.pi,math.pi))
random_direction = (0, 0, 0)
# Get the current state of the vehicle
car_state = self.__car_client.getCarState()
# The z coordinate is always zero
random_start_point[2] = -0
return (random_start_point, random_direction)
# A helper function to make a directory if it does not exist
def __make_dir_if_not_exist(self, directory):
if not (os.path.exists(directory)):
try:
os.makedirs(directory)
except OSError as e:
if e.errno != errno.EEXIST:
raise
class DistributedAgent():
def __init__(self, parameters):
required_parameters = ['data_dir', 'max_epoch_runtime_sec', 'replay_memory_size', 'batch_size', 'min_epsilon', 'per_iter_epsilon_reduction', 'experiment_name', 'train_conv_layers']
for required_parameter in required_parameters:
if required_parameter not in parameters:
raise ValueError('Missing required parameter {0}'.format(required_parameter))
parameters['role_type'] = 'agent'
self.__model_buffer = None
self.__model = None
self.__airsim_started = False
self.__data_dir = parameters['data_dir']
self.__per_iter_epsilon_reduction = float(parameters['per_iter_epsilon_reduction'])
self.__min_epsilon = float(parameters['min_epsilon'])
self.__max_epoch_runtime_sec = float(parameters['max_epoch_runtime_sec'])
self.__replay_memory_size = int(parameters['replay_memory_size'])
self.__batch_size = int(parameters['batch_size'])
self.__experiment_name = parameters['experiment_name']
self.__train_conv_layers = bool(parameters['train_conv_layers'])
self.__epsilon = 1
self.__num_batches_run = 0
self.__last_checkpoint_batch_count = 0
if 'batch_update_frequency' in parameters:
self.__batch_update_frequency = int(parameters['batch_update_frequency'])
if 'weights_path' in parameters:
self.__weights_path = parameters['weights_path']
else:
self.__weights_path = None
if 'airsim_path' in parameters:
self.__airsim_path = parameters['airsim_path']
else:
self.__airsim_path = None
self.__local_run = 'local_run' in parameters
self.__car_client = None
self.__car_controls = None
self.__minibatch_dir = os.path.join(self.__data_dir, 'minibatches')
self.__output_model_dir = os.path.join(self.__data_dir, 'models')
self.__make_dir_if_not_exist(self.__minibatch_dir)
self.__make_dir_if_not_exist(self.__output_model_dir)
self.__last_model_file = ''
self.__possible_ip_addresses = []
self.__trainer_ip_address = None
self.__experiences = {}
self.__init_road_points()
self.__init_reward_points()
# Starts the agent
def start(self):
self.__run_function()
# The function that will be run during training.
# It will initialize the connection to the trainer, start AirSim, and continuously run training iterations.
def __run_function(self):
print('Starting run function')
# Once the trainer is online, it will write its IP to a file in (data_dir)\trainer_ip\trainer_ip.txt
# Wait for that file to exist
if not self.__local_run:
print('Waiting for trainer to come online')
while True:
trainer_ip_dir = os.path.join(os.path.join(self.__data_dir, 'trainer_ip'), self.__experiment_name)
print('Checking {0}...'.format(trainer_ip_dir))
if os.path.isdir(trainer_ip_dir):
with open(os.path.join(trainer_ip_dir, 'trainer_ip.txt'), 'r') as f:
self.__possible_ip_addresses.append(f.read().replace('\n', ''))
break
print('Not online yet. Sleeping...')
time.sleep(5)
# We now have the IP address for the trainer. Attempt to ping the trainer.
ping_idx = -1
while True:
ping_idx += 1
print('Attempting to ping trainer...')
try:
print('\tPinging {0}...'.format(self.__possible_ip_addresses[ping_idx % len(self.__possible_ip_addresses)]))
response = requests.get('http://{0}:80/ping'.format(self.__possible_ip_addresses[ping_idx % len(self.__possible_ip_addresses)])).json()
if response['message'] != 'pong':
raise ValueError('Received unexpected message: {0}'.format(response))
print('Success!')
self.__trainer_ip_address = self.__possible_ip_addresses[ping_idx % len(self.__possible_ip_addresses)]
break
except Exception as e:
print('Could not get response. Message is {0}'.format(e))
if (ping_idx % len(self.__possible_ip_addresses) == 0):
print('Waiting 5 seconds and trying again...')
time.sleep(5)
# Get the latest model from the trainer
print('Getting model from the trainer')
sys.stdout.flush()
self.__model = RlModel(self.__weights_path, self.__train_conv_layers)
self.__get_latest_model()
else:
print('Run is local. Skipping connection to trainer.')
self.__model = RlModel(self.__weights_path, self.__train_conv_layers)
# Connect to the AirSim exe
self.__connect_to_airsim()
# Fill the replay memory by driving randomly.
print('Filling replay memory...')
while True:
print('Running Airsim Epoch.')
try:
self.__run_airsim_epoch(True)
percent_full = 100.0 * len(self.__experiences['actions'])/self.__replay_memory_size
print('Replay memory now contains {0} members. ({1}% full)'.format(len(self.__experiences['actions']), percent_full))
if (percent_full >= 100.0):
break
except msgpackrpc.error.TimeoutError:
print('Lost connection to AirSim while fillling replay memory. Attempting to reconnect.')
self.__connect_to_airsim()
# Get the latest model. Other agents may have finished before us.
print('Replay memory filled. Starting main loop...')
if not self.__local_run:
self.__get_latest_model()
while True:
try:
if (self.__model is not None):
#Generate a series of training examples by driving the vehicle in AirSim
print('Running Airsim Epoch.')
experiences, frame_count = self.__run_airsim_epoch(False)
# If we didn't immediately crash, train on the gathered experiences
if (frame_count > 0):
print('Generating {0} minibatches...'.format(frame_count))
print('Sampling Experiences.')
# Sample experiences from the replay memory
sampled_experiences = self.__sample_experiences(experiences, frame_count, True)
self.__num_batches_run += frame_count
# If we successfully sampled, train on the collected minibatches and send the gradients to the trainer node
if (len(sampled_experiences) > 0):
print('Publishing AirSim Epoch.')
self.__publish_batch_and_update_model(sampled_experiences, frame_count)
# Occasionally, the AirSim exe will stop working.
# For example, if a user connects to the node to visualize progress.
# In that case, attempt to reconnect.
except msgpackrpc.error.TimeoutError:
print('Lost connection to AirSim. Attempting to reconnect.')
self.__connect_to_airsim()
# Connects to the AirSim Exe.
# Assume that it is already running. After 10 successive attempts, attempt to restart the executable.
def __connect_to_airsim(self):
attempt_count = 0
while True:
try:
print('Attempting to connect to AirSim (attempt {0})'.format(attempt_count))
self.__car_client = CarClient()
self.__car_client.confirmConnection()
self.__car_client.enableApiControl(True)
self.__car_controls = CarControls()
print('Connected!')
return
except:
print('Failed to connect.')
attempt_count += 1
if (attempt_count % 10 == 0):
print('10 consecutive failures to connect. Attempting to start AirSim on my own.')
if self.__local_run:
os.system('START "" powershell.exe {0}'.format(os.path.join(self.__airsim_path, 'AD_Cookbook_Start_AirSim.ps1 neighborhood -windowed')))
else:
os.system('START "" powershell.exe D:\\AD_Cookbook_AirSim\\Scripts\\DistributedRL\\restart_airsim_if_agent.ps1')
print('Waiting a few seconds.')
time.sleep(10)
# Appends a sample to a ring buffer.
# If the appended example takes the size of the buffer over buffer_size, the example at the front will be removed.
def __append_to_ring_buffer(self, item, buffer, buffer_size):
if (len(buffer) >= buffer_size):
buffer = buffer[1:]
buffer.append(item)
return buffer
# Runs an interation of data generation from AirSim.
# Data will be saved in the replay memory.
def __run_airsim_epoch(self, always_random):
print('Running AirSim epoch.')
# Pick a random starting point on the roads
starting_points, starting_direction = self.__get_next_starting_point()
# Initialize the state buffer.
# For now, save 4 images at 0.01 second intervals.
state_buffer_len = 4
state_buffer = []
wait_delta_sec = 0.01
print('Getting Pose')
self.__car_client.simSetPose(Pose(Vector3r(starting_points[0], starting_points[1], starting_points[2]), AirSimClientBase.toQuaternion(starting_direction[0], starting_direction[1], starting_direction[2])), True)
# Currently, simSetPose does not allow us to set the velocity.
# So, if we crash and call simSetPose, the car will be still moving at its previous velocity.
# We need the car to stop moving, so push the brake and wait for a few seconds.
print('Waiting for momentum to die')
self.__car_controls.steering = 0
self.__car_controls.throttle = 0
self.__car_controls.brake = 1
self.__car_client.setCarControls(self.__car_controls)
time.sleep(4)
print('Resetting')
self.__car_client.simSetPose(Pose(Vector3r(starting_points[0], starting_points[1], starting_points[2]), AirSimClientBase.toQuaternion(starting_direction[0], starting_direction[1], starting_direction[2])), True)
#Start the car rolling so it doesn't get stuck
print('Running car for a few seconds...')
self.__car_controls.steering = 0
self.__car_controls.throttle = 1
self.__car_controls.brake = 0
self.__car_client.setCarControls(self.__car_controls)
# While the car is rolling, start initializing the state buffer
stop_run_time =datetime.datetime.now() + datetime.timedelta(seconds=2)
while(datetime.datetime.now() < stop_run_time):
time.sleep(wait_delta_sec)
state_buffer = self.__append_to_ring_buffer(self.__get_image(), state_buffer, state_buffer_len)
done = False
actions = [] #records the state we go to
pre_states = []
post_states = []
rewards = []
predicted_rewards = []
car_state = self.__car_client.getCarState()
start_time = datetime.datetime.utcnow()
end_time = start_time + datetime.timedelta(seconds=self.__max_epoch_runtime_sec)
num_random = 0
far_off = False
# Main data collection loop
while not done:
collision_info = self.__car_client.getCollisionInfo()
utc_now = datetime.datetime.utcnow()
# Check for terminal conditions:
# 1) Car has collided
# 2) Car is stopped
# 3) The run has been running for longer than max_epoch_runtime_sec.
# This constraint is so the model doesn't end up having to churn through huge chunks of data, slowing down training
# 4) The car has run off the road
if (collision_info.has_collided or car_state.speed < 2 or utc_now > end_time or far_off):
if (utc_now > end_time):
print('timed out.')
done = True
else:
# The Agent should occasionally pick random action instead of best action
do_greedy = np.random.random_sample()
pre_state = copy.deepcopy(state_buffer)
if (do_greedy < self.__epsilon or always_random):
num_random += 1
next_state = self.__model.get_random_state()
predicted_reward = 0
else:
next_state, predicted_reward = self.__model.predict_state(pre_state)
print('Model predicts {0}'.format(next_state))
# Convert the selected state to a control signal
next_control_signals = self.__model.state_to_control_signals(next_state, self.__car_client.getCarState())
# Take the action
self.__car_controls.steering = next_control_signals[0]
self.__car_controls.throttle = next_control_signals[1]
self.__car_controls.brake = next_control_signals[2]
self.__car_client.setCarControls(self.__car_controls)
# Wait for a short period of time to see outcome
time.sleep(wait_delta_sec)
# Observe outcome and compute reward from action
state_buffer = self.__append_to_ring_buffer(self.__get_image(), state_buffer, state_buffer_len)
car_state = self.__car_client.getCarState()
collision_info = self.__car_client.getCollisionInfo()
reward, far_off = self.__compute_reward(collision_info, car_state)
# Add the experience to the set of examples from this iteration
pre_states.append(pre_state)
post_states.append(state_buffer)
rewards.append(reward)
predicted_rewards.append(predicted_reward)
actions.append(next_state)
# Only the last state is a terminal state.
is_not_terminal = [1 for i in range(0, len(actions)-1, 1)]
is_not_terminal.append(0)
# Add all of the states from this iteration to the replay memory
self.__add_to_replay_memory('pre_states', pre_states)
self.__add_to_replay_memory('post_states', post_states)
self.__add_to_replay_memory('actions', actions)
self.__add_to_replay_memory('rewards', rewards)
self.__add_to_replay_memory('predicted_rewards', predicted_rewards)
self.__add_to_replay_memory('is_not_terminal', is_not_terminal)
print('Percent random actions: {0}'.format(num_random / max(1, len(actions))))
print('Num total actions: {0}'.format(len(actions)))
# If we are in the main loop, reduce the epsilon parameter so that the model will be called more often
if not always_random:
self.__epsilon -= self.__per_iter_epsilon_reduction
self.__epsilon = max(self.__epsilon, self.__min_epsilon)
return self.__experiences, len(actions)
# Adds a set of examples to the replay memory
def __add_to_replay_memory(self, field_name, data):
if field_name not in self.__experiences:
self.__experiences[field_name] = data
else:
self.__experiences[field_name] += data
start_index = max(0, len(self.__experiences[field_name]) - self.__replay_memory_size)
self.__experiences[field_name] = self.__experiences[field_name][start_index:]
# Sample experiences from the replay memory
def __sample_experiences(self, experiences, frame_count, sample_randomly):
sampled_experiences = {}
sampled_experiences['pre_states'] = []
sampled_experiences['post_states'] = []
sampled_experiences['actions'] = []
sampled_experiences['rewards'] = []
sampled_experiences['predicted_rewards'] = []
sampled_experiences['is_not_terminal'] = []
# Compute the surprise factor, which is the difference between the predicted an the actual Q value for each state.
# We can use that to weight examples so that we are more likely to train on examples that the model got wrong.
suprise_factor = np.abs(np.array(experiences['rewards'], dtype=np.dtype(float)) - np.array(experiences['predicted_rewards'], dtype=np.dtype(float)))
suprise_factor_normalizer = np.sum(suprise_factor)
suprise_factor /= float(suprise_factor_normalizer)
# Generate one minibatch for each frame of the run
for _ in range(0, frame_count, 1):
if sample_randomly:
idx_set = set(np.random.choice(list(range(0, suprise_factor.shape[0], 1)), size=(self.__batch_size), replace=False))
else:
idx_set = set(np.random.choice(list(range(0, suprise_factor.shape[0], 1)), size=(self.__batch_size), replace=False, p=suprise_factor))
sampled_experiences['pre_states'] += [experiences['pre_states'][i] for i in idx_set]
sampled_experiences['post_states'] += [experiences['post_states'][i] for i in idx_set]
sampled_experiences['actions'] += [experiences['actions'][i] for i in idx_set]
sampled_experiences['rewards'] += [experiences['rewards'][i] for i in idx_set]
sampled_experiences['predicted_rewards'] += [experiences['predicted_rewards'][i] for i in idx_set]
sampled_experiences['is_not_terminal'] += [experiences['is_not_terminal'][i] for i in idx_set]
return sampled_experiences
# Train the model on minibatches and post to the trainer node.
# The trainer node will respond with the latest version of the model that will be used in further data generation iterations.
def __publish_batch_and_update_model(self, batches, batches_count):
# Train and get the gradients
print('Publishing epoch data and getting latest model from parameter server...')
gradients = self.__model.get_gradient_update_from_batches(batches)
# Post the data to the trainer node
if not self.__local_run:
post_data = {}
post_data['gradients'] = gradients
post_data['batch_count'] = batches_count
new_model_parameters = requests.post('http://{0}:80/gradient_update'.format(self.__trainer_ip_address), json=post_data)
print('New params:')
print(new_model_parameters)
# Update the existing model with the new parameters
self.__model.from_packet(new_model_parameters.json())
else:
if (self.__num_batches_run > self.__batch_update_frequency + self.__last_checkpoint_batch_count):
self.__model.update_critic()
checkpoint = {}
checkpoint['model'] = self.__model.to_packet(get_target=True)
checkpoint['batch_count'] = batches_count
checkpoint_str = json.dumps(checkpoint)
checkpoint_dir = os.path.join(os.path.join(self.__data_dir, 'checkpoint'), self.__experiment_name)
if not os.path.isdir(checkpoint_dir):
try:
os.makedirs(checkpoint_dir)
except OSError as e:
if e.errno != errno.EEXIST:
raise
file_name = os.path.join(checkpoint_dir,'{0}.json'.format(self.__num_batches_run))
with open(file_name, 'w') as f:
print('Checkpointing to {0}'.format(file_name))
f.write(checkpoint_str)
self.__last_checkpoint_batch_count = self.__num_batches_run
# Gets the latest model from the trainer node
def __get_latest_model(self):
print('Getting latest model from parameter server...')
response = requests.get('http://{0}:80/latest'.format(self.__trainer_ip_address)).json()
self.__model.from_packet(response)
# Gets an image from AirSim
def __get_image(self):
image_response = self.__car_client.simGetImages([ImageRequest(0, AirSimImageType.Scene, False, False)])[0]
image1d = np.fromstring(image_response.image_data_uint8, dtype=np.uint8)
image_rgba = image1d.reshape(image_response.height, image_response.width, 4)
return image_rgba[76:135,0:255,0:3].astype(float)
# Computes the reward functinon based on the car position.
def __compute_reward(self, collision_info, car_state):
#Define some constant parameters for the reward function
THRESH_DIST = 3.5 # The maximum distance from the center of the road to compute the reward function
DISTANCE_DECAY_RATE = 1.2 # The rate at which the reward decays for the distance function
CENTER_SPEED_MULTIPLIER = 2.0 # The ratio at which we prefer the distance reward to the speed reward
# If the car has collided, the reward is always zero
if (collision_info.has_collided):
return 0.0, True
# If the car is stopped, the reward is always zero
speed = car_state.speed
if (speed < 2):
return 0.0, True
#Get the car position
position_key = bytes('position', encoding='utf8')
x_val_key = bytes('x_val', encoding='utf8')
y_val_key = bytes('y_val', encoding='utf8')
car_point = np.array([car_state.kinematics_true[position_key][x_val_key], car_state.kinematics_true[position_key][y_val_key], 0])
# Distance component is exponential distance to nearest line
distance = 999
#Compute the distance to the nearest center line
for line in self.__reward_points:
local_distance = 0
length_squared = ((line[0][0]-line[1][0])**2) + ((line[0][1]-line[1][1])**2)
if (length_squared != 0):
t = max(0, min(1, np.dot(car_point-line[0], line[1]-line[0]) / length_squared))
proj = line[0] + (t * (line[1]-line[0]))
local_distance = np.linalg.norm(proj - car_point)
distance = min(local_distance, distance)
distance_reward = math.exp(-(distance * DISTANCE_DECAY_RATE))
return distance_reward, distance > THRESH_DIST
# Initializes the points used for determining the starting point of the vehicle
def __init_road_points(self):
self.__road_points = []
car_start_coords = [12961.722656, 6660.329102, 0]
with open(os.path.join(os.path.join(self.__data_dir, 'data'), 'road_lines.txt'), 'r') as f:
for line in f:
points = line.split('\t')
first_point = np.array([float(p) for p in points[0].split(',')] + [0])
second_point = np.array([float(p) for p in points[1].split(',')] + [0])
self.__road_points.append(tuple((first_point, second_point)))
# Points in road_points.txt are in unreal coordinates
# But car start coordinates are not the same as unreal coordinates
for point_pair in self.__road_points:
for point in point_pair:
point[0] -= car_start_coords[0]
point[1] -= car_start_coords[1]
point[0] /= 100
point[1] /= 100
# Initializes the points used for determining the optimal position of the vehicle during the reward function
def __init_reward_points(self):
self.__reward_points = []
with open(os.path.join(os.path.join(self.__data_dir, 'data'), 'reward_points.txt'), 'r') as f:
for line in f:
point_values = line.split('\t')
first_point = np.array([float(point_values[0]), float(point_values[1]), 0])
second_point = np.array([float(point_values[2]), float(point_values[3]), 0])
self.__reward_points.append(tuple((first_point, second_point)))
# Randomly selects a starting point on the road
# Used for initializing an iteration of data generation from AirSim
def __get_next_starting_point(self):
# Get the current state of the vehicle
car_state = self.__car_client.getCarState()
# Pick a random road.
random_line_index = np.random.randint(0, high=len(self.__road_points))
# Pick a random position on the road.
# Do not start too close to either end, as the car may crash during the initial run.
random_interp = (np.random.random_sample() * 0.4) + 0.3
# Pick a random direction to face
random_direction_interp = np.random.random_sample()
# Compute the starting point of the car
random_line = self.__road_points[random_line_index]
random_start_point = list(random_line[0])
random_start_point[0] += (random_line[1][0] - random_line[0][0])*random_interp
random_start_point[1] += (random_line[1][1] - random_line[0][1])*random_interp
# Compute the direction that the vehicle will face
# Vertical line
if (np.isclose(random_line[0][1], random_line[1][1])):
if (random_direction_interp > 0.5):
random_direction = (0,0,0)
else:
random_direction = (0, 0, math.pi)
# Horizontal line
elif (np.isclose(random_line[0][0], random_line[1][0])):
if (random_direction_interp > 0.5):
random_direction = (0,0,math.pi/2)
else:
random_direction = (0,0,-1.0 * math.pi/2)
# The z coordinate is always zero
random_start_point[2] = -0
return (random_start_point, random_direction)
# A helper function to make a directory if it does not exist
def __make_dir_if_not_exist(self, directory):
if not (os.path.exists(directory)):
try:
os.makedirs(directory)
except OSError as e:
if e.errno != errno.EEXIST:
raise
class DistributedAgent():
def __init__(self):
self.__model_buffer = None
self.__model = None
self.__airsim_started = False
self.__data_dir = 'data/'
self.__handle_dir = 'data/handle_image/'
self.__per_iter_epsilon_reduction = 0.003
self.__min_epsilon = 0.1
self.__max_epoch_runtime_sec = float(30)
self.__replay_memory_size = 50
self.__batch_size = 32
self.__experiment_name = 'local_run'
self.__train_conv_layers = False
self.__epsilon = 1
self.__percent_full = 0
self.__num_batches_run = 0
self.__last_checkpoint_batch_count = 0
self.__handles = {}
self.__batch_update_frequency = 10
self.__weights_path = None
self.__airsim_path = '../AD_Cookbook_AirSim/'
self.__local_run = True
self.__car_client = None
self.__car_controls = None
# self.__minibatch_dir = os.path.join(self.__data_dir, 'minibatches')
# self.__output_model_dir = os.path.join(self.__data_dir, 'models')
# self.__make_dir_if_not_exist(self.__minibatch_dir)
# self.__make_dir_if_not_exist(self.__output_model_dir)
self.__last_model_file = ''
self.__possible_ip_addresses = []
self.__trainer_ip_address = None
self.__experiences = {}
self.prev_steering = 0
self.__init_road_points()
self.__init_reward_points()
self.__init_handle_images()
self.__best_drive = datetime.timedelta(seconds=-1) #added 2021-03-09 by kang
self.__best_model = None #added 2021-03-09 by kang
self.__num_of_trial = 0
def start(self):
self.__run_function()
def __run_function(self):
self.__model = RlModel(self.__weights_path, self.__train_conv_layers)
self.__connect_to_airsim()
print('Filling replay memory...')
while True:
print('Running Airsim Epoch.')
try:
if self.__percent_full < 100:
self.__run_airsim_epoch(True)
self.__percent_full = 100.0 * len(self.__experiences['actions'])/self.__replay_memory_size
print('Replay memory now contains {0} members. ({1}% full)'.format(len(self.__experiences['actions']), self.__percent_full))
else:
if (self.__model is not None):
print('Running Airsim Epoch.')
experiences, frame_count, drive_time = self.__run_airsim_epoch(False)
# If we didn't immediately crash, train on the gathered experiences
if (frame_count > 0):
print('Generating {0} minibatches...'.format(frame_count))
print('Sampling Experiences.')
# Sample experiences from the replay memory
sampled_experiences = self.__sample_experiences(experiences, frame_count, True)
self.__num_batches_run += frame_count
# If we successfully sampled, train on the collected minibatches and send the gradients to the trainer node
if (len(sampled_experiences) > 0):
print('Publishing AirSim Epoch.')
self.__publish_batch_and_update_model(sampled_experiences, frame_count, drive_time)
except msgpackrpc.error.TimeoutError:
print('Lost connection to AirSim while fillling replay memory. Attempting to reconnect.')
self.__connect_to_airsim()
def __connect_to_airsim(self):
attempt_count = 0
while True:
try:
print('Attempting to connect to AirSim (attempt {0})'.format(attempt_count))
self.__car_client = CarClient()
self.__car_client.confirmConnection()
self.__car_client.enableApiControl(True)
self.__car_controls = CarControls()
print('Connected!')
return
except:
print('Failed to connect.')
attempt_count += 1
if (attempt_count % 10 == 0):
print('10 consecutive failures to connect. Attempting to start AirSim on my own.')
os.system('START "" powershell.exe {0}'.format(os.path.join(self.__airsim_path, 'AD_Cookbook_Start_AirSim.ps1 neighborhood -windowed')))
print('Waiting a few seconds.')
time.sleep(10)
def __run_airsim_epoch(self, always_random):
starting_points, starting_direction = self.__get_next_starting_point()
# state_buffer_len = 4 changed by kang 2021-03-09 cuz of no use
state_buffer = []
wait_delta_sec = 0.01
self.__car_client.simSetPose(Pose(Vector3r(starting_points[0], starting_points[1], starting_points[2]), AirSimClientBase.toQuaternion(starting_direction[0], starting_direction[1], starting_direction[2])), True)
self.__car_controls.steering = 0
self.__car_controls.throttle = 1
self.__car_controls.brake = 0
self.prev_steering = 0
self.__car_client.setCarControls(self.__car_controls)
time.sleep(1.5)
state_buffer = self.__get_image()
done = False
actions = []
pre_states = []
post_states = []
rewards = []
predicted_rewards = []
car_state = self.__car_client.getCarState()
start_time = datetime.datetime.utcnow()
end_time = start_time + datetime.timedelta(seconds=self.__max_epoch_runtime_sec)
num_random = 0
far_off = False
while not done:
collision_info = self.__car_client.getCollisionInfo()
utc_now = datetime.datetime.utcnow()
if (collision_info.has_collided or car_state.speed < 1 or utc_now > end_time or far_off):
print('Start time: {0}, end time: {1}'.format(start_time, utc_now), file=sys.stderr)
self.__car_controls.steering = 0
self.__car_controls.throttle = 0
self.__car_controls.brake = 1
self.__car_client.setCarControls(self.__car_controls)
time.sleep(4)
if (utc_now > end_time):
print('timed out.')
print('Full autonomous run finished at {0}'.format(utc_now), file=sys.stderr)
done = True
sys.stderr.flush()
else:
# The Agent should occasionally pick random action instead of best action
do_greedy = np.random.random_sample()
pre_state = copy.deepcopy(state_buffer)
angle = -int(self.prev_steering/0.05*4)
pre_handle = self.__handles[angle].reshape(59,255,1)
pre_state = np.concatenate([pre_state, pre_handle], axis=2)
if (do_greedy < self.__epsilon or always_random):
num_random += 1
next_state = self.__model.get_random_state()
predicted_reward = 0
else:
next_state, predicted_reward = self.__model.predict_state(pre_state)
print('Model predicts {0}'.format(next_state))
# Convert the selected state to a control signal
next_control_signals = self.__model.state_to_control_signals(next_state, self.__car_client.getCarState())
# Take the action
self.__car_controls.steering = self.prev_steering + next_control_signals[0]
if self.__car_controls.steering > 1.0:
self.__car_controls.steering = 1.0
elif self.__car_controls.steering < -1.0:
self.__car_controls.steering = -1.0
self.prev_steering = self.__car_controls.steering
print('change steering : ', self.prev_steering)
self.__car_controls.throttle = next_control_signals[1]
self.__car_controls.brake = next_control_signals[2]
self.__car_client.setCarControls(self.__car_controls)
# Wait for a short period of time to see outcome
time.sleep(wait_delta_sec)
# Observe outcome and compute reward from action
state_buffer = self.__get_image()
angle = -int(self.prev_steering/0.05*4)
post_handle = self.__handles[angle].reshape(59,255,1)
post_state = np.concatenate([state_buffer, post_handle],axis=2)
car_state = self.__car_client.getCarState()
collision_info = self.__car_client.getCollisionInfo()
reward, far_off = self.__compute_reward(collision_info, car_state)
# Add the experience to the set of examples from this iteration
pre_states.append(pre_state)
post_states.append(post_state)
rewards.append(reward)
predicted_rewards.append(predicted_reward)
actions.append(next_state)
# action수가 너무 적을경우, 그 회차의 학습을 진행하지 않음. #added 2021-03-09 by kang
if len(actions) < 10:
return self.__experiences, 0, 0
is_not_terminal = [1 for i in range(0, len(actions)-1, 1)]
is_not_terminal.append(0)
self.__add_to_replay_memory('pre_states', pre_states)
self.__add_to_replay_memory('post_states', post_states)
self.__add_to_replay_memory('actions', actions)
self.__add_to_replay_memory('rewards', rewards)
self.__add_to_replay_memory('predicted_rewards', predicted_rewards)
self.__add_to_replay_memory('is_not_terminal', is_not_terminal)
print('Percent random actions: {0}'.format(num_random / max(1, len(actions))))
print('Num total actions: {0}'.format(len(actions)))
if not always_random:
self.__epsilon -= self.__per_iter_epsilon_reduction
self.__epsilon = max(self.__epsilon, self.__min_epsilon)
return self.__experiences, len(actions), utc_now - start_time
def __add_to_replay_memory(self, field_name, data):
if field_name not in self.__experiences:
self.__experiences[field_name] = data
else:
self.__experiences[field_name] += data
start_index = max(0, len(self.__experiences[field_name]) - self.__replay_memory_size)
self.__experiences[field_name] = self.__experiences[field_name][start_index:]
def __sample_experiences(self, experiences, frame_count, sample_randomly):
sampled_experiences = {}
sampled_experiences['pre_states'] = []
sampled_experiences['post_states'] = []
sampled_experiences['actions'] = []
sampled_experiences['rewards'] = []
sampled_experiences['predicted_rewards'] = []
sampled_experiences['is_not_terminal'] = []
# Compute the surprise factor, which is the difference between the predicted an the actual Q value for each state.
# We can use that to weight examples so that we are more likely to train on examples that the model got wrong.
suprise_factor = np.abs(np.array(experiences['rewards'], dtype=np.dtype(float)) - np.array(experiences['predicted_rewards'], dtype=np.dtype(float)))
suprise_factor_normalizer = np.sum(suprise_factor)
suprise_factor /= float(suprise_factor_normalizer)
# Generate one minibatch for each frame of the run
for _ in range(0, frame_count, 1):
if sample_randomly:
idx_set = set(np.random.choice(list(range(0, suprise_factor.shape[0], 1)), size=(self.__batch_size), replace=False))
else:
idx_set = set(np.random.choice(list(range(0, suprise_factor.shape[0], 1)), size=(self.__batch_size), replace=False, p=suprise_factor))
sampled_experiences['pre_states'] += [experiences['pre_states'][i] for i in idx_set]
sampled_experiences['post_states'] += [experiences['post_states'][i] for i in idx_set]
sampled_experiences['actions'] += [experiences['actions'][i] for i in idx_set]
sampled_experiences['rewards'] += [experiences['rewards'][i] for i in idx_set]
sampled_experiences['predicted_rewards'] += [experiences['predicted_rewards'][i] for i in idx_set]
sampled_experiences['is_not_terminal'] += [experiences['is_not_terminal'][i] for i in idx_set]
return sampled_experiences
def __publish_batch_and_update_model(self, batches, batches_count, drive_time): # added 2021-03-09 by kang
# Train and get the gradients
print('Publishing epoch data and getting latest model from parameter server...')
gradients = self.__model.get_gradient_update_from_batches(batches)
if (self.__num_batches_run > self.__batch_update_frequency + self.__last_checkpoint_batch_count):
self.__model.update_critic()
checkpoint = {}
checkpoint['model'] = self.__model.to_packet(get_target=True)
checkpoint['batch_count'] = batches_count
checkpoint_str = json.dumps(checkpoint)
checkpoint_dir = os.path.join(os.path.join(self.__data_dir, 'checkpoint'), self.__experiment_name)
if not os.path.isdir(checkpoint_dir):
try:
os.makedirs(checkpoint_dir)
except OSError as e:
if e.errno != errno.EEXIST:
raise
file_name = os.path.join(checkpoint_dir,'{0}.json'.format(self.__num_batches_run))
with open(file_name, 'w') as f:
print('Checkpointing to {0}'.format(file_name))
f.write(checkpoint_str)
self.__last_checkpoint_batch_count = self.__num_batches_run
# 운행시간을 이용해서 가장 오래 걸린 시간을 best policy로 보고, best policy를 따로 저장. #added 2021-03-09 by kang
if drive_time > self.__best_drive:
print("="*30)
print("New Best Policy!!!!!!")
print("="*30)
self.__best_drive = drive_time
bestpoint_dir = os.path.join(os.path.join(self.__data_dir, 'bestpoint'), self.__experiment_name)
if not os.path.isdir(bestpoint_dir):
try:
os.makedirs(bestpoint_dir)
except OSError as e:
if e.errno != errno.EEXIST:
raise
file_name = os.path.join(bestpoint_dir,'{0}.json'.format(self.__num_batches_run))
with open(file_name, 'w') as f:
print('Add Best Policy to {0}'.format(file_name))
f.write(checkpoint_str)
self.__best_model = self.__model
self.__best_experiences = self.__experiences
elif self.__num_of_trial > 10:
print("="*30)
print("Reload best Model")
print("="*30)
self.__model = self.__best_model
self.__experiences = self.__best_experiences
self.__num_of_trial = 0
self.__num_of_trial += 1
def __compute_reward(self, collision_info, car_state):
#Define some constant parameters for the reward function
THRESH_DIST = 3.5 # The maximum distance from the center of the road to compute the reward function
DISTANCE_DECAY_RATE = 1.2 # The rate at which the reward decays for the distance function
CENTER_SPEED_MULTIPLIER = 2.0 # The ratio at which we prefer the distance reward to the speed reward
# If the car has collided, the reward is always zero
# 충돌 시 reward를 음수로 줘보았음.
if (collision_info.has_collided):