JSONFILE = 'database/sample_benchmark2.json'
environments_dict_base = [
'WetSunset_route00024', 'SoftRainSunset_route00000', 'WetNoon_route00024'
]
params = {
'save_dataset': True,
'docker_name': 'carlalatest:latest',
'gpu': 5,
'batch_size': 1,
'remove_wrong_data': False,
'non_rendering_mode': False,
'carla_recording': False # TODO testing
}
agent = NPCAgent()
AGENT_NAME = 'NPCAgent'
# The episodes to be checked must always be sequential
def check_folder(env_name, number_episodes):
""" Check if the folder contain the expected number of episodes
and if they are complete.
"""
path = os.path.join(os.environ["SRL_DATASET_PATH"], 'sample_benchmark2',
env_name)
# List number of folders check if match expected
environments_count = 0
for filename in os.listdir(path):
params = {
'save_dataset': True,
'save_sensors': True,
'save_trajectories': True,
'save_opponents': True,
'save_opp_trajectories': False,
'docker_name': 'carlalatest:latest',
'gpu': 0,
'batch_size': 1,
'remove_wrong_data': False,
'non_rendering_mode': False,
'carla_recording': False
}
# The idea is that the agent class should be completely independent
agent = NPCAgent(sensors_dict=[])
# this could be joined
env_batch = CEXP(
json,
params=params,
execute_all=True,
ignore_previous_execution=True,
port=arguments.port) # THe experience is built, the files necessary
# to load CARLA and the scenarios are made
# Here some docker was set
env_batch.start()
for env in env_batch:
try:
# The policy selected to run this experience vector
# (The class basically) This policy can also learn, just
}
# The idea is that the agent class should be completely independent
agent = NPCAgent(sensors_dict=[{
'type': 'sensor.camera.rgb',
'x': 2.0,
'y': 0.0,
'z': 1.40,
'roll': 0.0,
'pitch': -15.0,
'yaw': 0.0,
'width': 800,
'height': 600,
'fov': 100,
'id': 'rgb'
}, {
'type': 'sensor.camera.rgb',
'x': 2.0,
'y': 0.0,
'z': 15.40,
'roll': 0.0,
'pitch': -30.0,
'yaw': 0.0,
'width': 800,
'height': 600,
'fov': 120,
'id': 'rgb_view'
}],
noise=True)
# this could be joined
env_batch = CEXP(
json,
def collect_data(json_file, params, eliminated_environments, collector_id):
# The idea is that the agent class should be completely independent
# TODO this has to go to a separate file and to be merged with package
agent = NPCAgent(
sensors_dict = [{'type': 'sensor.camera.rgb',
'x': 2.0, 'y': 0.0,
'z': 1.40, 'roll': 0.0,
'pitch': -15.0, 'yaw': 0.0,
'width': 800, 'height': 600,
'fov': 100,
'id': 'rgb_central'},
{'type': 'sensor.camera.semantic_segmentation',
'x': 2.0, 'y': 0.0,
'z': 1.40, 'roll': 0.0,
'pitch': -15.0, 'yaw': 0.0,
'width': 800, 'height': 600,
'fov': 100,
'id': 'labels_central'},
{'type': 'sensor.camera.rgb',
'x': 2.0, 'y': 0.0,
'z': 1.40, 'roll': 0.0,
'pitch': -15.0, 'yaw': -30.0,
'width': 800, 'height': 600,
'fov': 100,
'id': 'rgb_left'},
{'type': 'sensor.camera.semantic_segmentation',
'x': 2.0, 'y': 0.0,
'z': 1.40, 'roll': 0.0,
'pitch': -15.0, 'yaw': -30.0,
'width': 800, 'height': 600,
'fov': 100,
'id': 'labels_left'},
{'type': 'sensor.camera.rgb',
'x': 2.0, 'y': 0.0,
'z': 1.40, 'roll': 0.0,
'pitch': -15.0, 'yaw': 30.0,
'width': 800, 'height': 600,
'fov': 100,
'id': 'rgb_right'},
{'type': 'sensor.camera.semantic_segmentation',
'x': 2.0, 'y': 0.0,
'z': 1.40, 'roll': 0.0,
'pitch': -15.0, 'yaw': 30.0,
'width': 800, 'height': 600,
'fov': 100,
'id': 'labels_right'},
{'type': 'sensor.can_bus',
'reading_frequency': 25,
'id': 'can_bus'
},
{'type': 'sensor.other.gnss',
'x': 0.7, 'y': -0.4, 'z': 1.60,
'id': 'GPS'}
])
# this could be joined
env_batch = CEXP(json_file, params=params, execute_all=True,
eliminated_environments=eliminated_environments)
# THe experience is built, the files necessary
# to load CARLA and the scenarios are made
# Here some docker was set
NPCAgent._name = 'Multi'
env_batch.start(agent_name=NPCAgent._name)
for env in env_batch:
try:
# The policy selected to run this experience vector (The class basically) This policy can also learn, just
# by taking the output from the experience.
# I need a mechanism to test the rewards so I can test the policy gradient strategy
print (" Collector ", collector_id, " Collecting for ", env)
states, rewards = agent.unroll(env)
agent.reinforce(rewards)
except KeyboardInterrupt:
env.stop()
break
except:
traceback.print_exc()
# Just try again
agent.reset()
env.stop()
print(" ENVIRONMENT BROKE trying again.")
env_batch.cleanup()
logging.basicConfig(format='%(levelname)s: %(message)s', level=logging.DEBUG)
# A single loop being made
json_file = 'database/sample_benchmark.json'
# Dictionary with the necessary params related to the execution not the model itself.
params = {
'save_dataset': False,
'docker_name': 'carlalatest:latest',
'gpu': 0,
'batch_size': 1,
'remove_wrong_data': False,
'non_rendering_mode': False,
'carla_recording': True
}
# TODO for now batch size is one
number_of_iterations = 10
# The idea is that the agent class should be completely independent
agent = NPCAgent()
env_batch = CEXP(
json_file,
params,
number_of_iterations,
params['batch_size'],
sequential=True,
debug=True) # THe experience is built, the files necessary
# to load CARLA and the scenarios are made
# Here some docker was set
env_batch.start()
for env in env_batch:
states, rewards = agent.unroll(env)
# if the agent is already un
'batch_size': 1,
'remove_wrong_data': False,
'non_rendering_mode': True,
'carla_recording': False
}
# TODO for now batch size is one
number_of_iterations = 400
# The idea is that the agent class should be completely independent
agent = NPCAgent(sensors_dict=[{
'type': 'sensor.camera.rgb',
'x': 2.0,
'y': 0.0,
'z': 1.40,
'roll': 0.0,
'pitch': -15.0,
'yaw': 0.0,
'width': 1800,
'height': 1200,
'fov': 100,
'id': 'rgb_central'
}])
# this could be joined
env_batch = CEXP(
json, params=params, execute_all=True,
port=arguments.port) # THe experience is built, the files necessary
# to load CARLA and the scenarios are made
# Here some docker was set
env_batch.start()
for env in env_batch: