def get_experiment_name(initial_experiment_name=''):
match = None
while match is None:
if initial_experiment_name == '':
experiment_name = screen.ask_input(
"Please enter an experiment name: ")
else:
experiment_name = initial_experiment_name
experiment_name = experiment_name.replace(" ", "_")
match = re.match("^$|^[\w -/]{1,100}$", experiment_name)
if match is None:
screen.error(
'Experiment name must be composed only of alphanumeric letters, '
'underscores and dashes and should not be longer than 100 characters.'
)
return match.group(0)
def check_input_and_fill_run_dict(parser):
args = parser.parse_args()
# if no arg is given
if len(sys.argv) == 1:
parser.print_help()
exit(0)
# list available presets
if args.list:
presets_lists = list_all_classes_in_module(presets)
screen.log_title("Available Presets:")
for preset in presets_lists:
print(preset)
sys.exit(0)
# check inputs
try:
# num_workers = int(args.num_workers)
num_workers = int(re.match("^\d+$", args.num_workers).group(0))
except ValueError:
screen.error("Parameter num_workers should be an integer.")
preset_names = list_all_classes_in_module(presets)
if args.preset is not None and args.preset not in preset_names:
screen.error("A non-existing preset was selected. ")
if args.checkpoint_restore_dir is not None and not os.path.exists(args.checkpoint_restore_dir):
screen.error("The requested checkpoint folder to load from does not exist. ")
if args.save_model_sec is not None:
try:
args.save_model_sec = int(args.save_model_sec)
except ValueError:
screen.error("Parameter save_model_sec should be an integer.")
if args.preset is None and (args.agent_type is None or args.environment_type is None
or args.exploration_policy_type is None) and not args.play:
screen.error('When no preset is given for Coach to run, the user is expected to input the desired agent_type,'
' environment_type and exploration_policy_type to assemble a preset. '
'\nAt least one of these parameters was not given.')
elif args.preset is None and args.play and args.environment_type is None:
screen.error('When no preset is given for Coach to run, and the user requests human control over the environment,'
' the user is expected to input the desired environment_type and level.'
'\nAt least one of these parameters was not given.')
elif args.preset is None and args.play and args.environment_type:
args.agent_type = 'Human'
args.exploration_policy_type = 'ExplorationParameters'
# get experiment name and path
experiment_name = logger.get_experiment_name(args.experiment_name)
experiment_path = logger.get_experiment_path(experiment_name)
if args.play and num_workers > 1:
screen.warning("Playing the game as a human is only available with a single worker. "
"The number of workers will be reduced to 1")
num_workers = 1
# fill run_dict
run_dict = dict()
run_dict['agent_type'] = args.agent_type
run_dict['environment_type'] = args.environment_type
run_dict['exploration_policy_type'] = args.exploration_policy_type
run_dict['level'] = args.level
run_dict['preset'] = args.preset
run_dict['custom_parameter'] = args.custom_parameter
run_dict['experiment_path'] = experiment_path
run_dict['framework'] = Frameworks().get(args.framework)
run_dict['play'] = args.play
run_dict['evaluate'] = args.evaluate# or args.play
# multi-threading parameters
run_dict['num_threads'] = num_workers
# checkpoints
run_dict['save_model_sec'] = args.save_model_sec
run_dict['save_model_dir'] = experiment_path if args.save_model_sec is not None else None
run_dict['checkpoint_restore_dir'] = args.checkpoint_restore_dir
# visualization
run_dict['visualization.dump_gifs'] = args.dump_gifs
run_dict['visualization.render'] = args.render
run_dict['visualization.tensorboard'] = args.tensorboard
return args, run_dict
# use the following command line to download and extract the CARLA dataset:
# python rl_coach/utilities/carla_dataset_to_replay_buffer.py
agent_params.memory.load_memory_from_file_path = "./datasets/carla_train_set_replay_buffer.p"
agent_params.memory.state_key_with_the_class_index = 'high_level_command'
agent_params.memory.num_classes = 4
# download dataset if it doesn't exist
if not os.path.exists(agent_params.memory.load_memory_from_file_path):
screen.log_title("The CARLA dataset is not present in the following path: {}"
.format(agent_params.memory.load_memory_from_file_path))
result = screen.ask_yes_no("Do you want to download it now?")
if result:
create_dataset(None, "./datasets/carla_train_set_replay_buffer.p")
else:
screen.error("Please update the path to the CARLA dataset in the CARLA_CIL preset", crash=True)
###############
# Environment #
###############
env_params = CarlaEnvironmentParameters()
env_params.level = 'town1'
env_params.cameras = ['CameraRGB']
env_params.camera_height = 600
env_params.camera_width = 800
env_params.separate_actions_for_throttle_and_brake = True
env_params.allow_braking = True
env_params.quality = CarlaEnvironmentParameters.Quality.EPIC
env_params.experiment_suite = CoRL2017('Town01')
def check_input_and_fill_run_dict(parser):
args = parser.parse_args()
# if no arg is given
if len(sys.argv) == 1:
parser.print_help()
exit(0)
# list available presets
if args.list:
presets_lists = list_all_classes_in_module(presets)
screen.log_title("Available Presets:")
for preset in presets_lists:
print(preset)
sys.exit(0)
# check inputs
try:
# num_workers = int(args.num_workers)
num_workers = int(re.match("^\d+$", args.num_workers).group(0))
except ValueError:
screen.error("Parameter num_workers should be an integer.")
exit(1)
preset_names = list_all_classes_in_module(presets)
if args.preset is not None and args.preset not in preset_names:
screen.error("A non-existing preset was selected. ")
exit(1)
if args.checkpoint_restore_dir is not None and not os.path.exists(
args.checkpoint_restore_dir):
screen.error(
"The requested checkpoint folder to load from does not exist. ")
exit(1)
if args.save_model_sec is not None:
try:
args.save_model_sec = int(args.save_model_sec)
except ValueError:
screen.error("Parameter save_model_sec should be an integer.")
exit(1)
if args.preset is None and (args.agent_type is None
or args.environment_type is None
or args.exploration_policy_type is None):
screen.error(
'When no preset is given for Coach to run, the user is expected to input the desired agent_type,'
' environment_type and exploration_policy_type to assemble a preset. '
'\nAt least one of these parameters was not given.')
exit(1)
experiment_name = args.experiment_name
if args.experiment_name == '':
experiment_name = screen.ask_input("Please enter an experiment name: ")
experiment_name = experiment_name.replace(" ", "_")
match = re.match("^$|^\w{1,100}$", experiment_name)
if match is None:
screen.error(
'Experiment name must be composed only of alphanumeric letters and underscores and should not be '
'longer than 100 characters.')
exit(1)
experiment_path = os.path.join('./experiments/', match.group(0))
experiment_path = get_experiment_path(experiment_path)
# fill run_dict
run_dict = dict()
run_dict['agent_type'] = args.agent_type
run_dict['environment_type'] = args.environment_type
run_dict['exploration_policy_type'] = args.exploration_policy_type
run_dict['preset'] = args.preset
run_dict['custom_parameter'] = args.custom_parameter
run_dict['experiment_path'] = experiment_path
run_dict['framework'] = Frameworks().get(args.framework)
# multi-threading parameters
run_dict['num_threads'] = num_workers
# checkpoints
run_dict['save_model_sec'] = args.save_model_sec
run_dict[
'save_model_dir'] = experiment_path if args.save_model_sec is not None else None
run_dict['checkpoint_restore_dir'] = args.checkpoint_restore_dir
# visualization
run_dict['visualization.dump_gifs'] = args.dump_gifs
run_dict['visualization.render'] = args.render
return args, run_dict
last_num_episodes = csv['Episode #'].values[-1]
# check if reward is enough
if np.any(averaged_rewards >
preset.test_min_return_threshold):
test_passed = True
break
time.sleep(1)
# kill test and print result
os.killpg(os.getpgid(p.pid), signal.SIGTERM)
if test_passed:
screen.success("Passed successfully")
else:
if csv_paths:
screen.error("Failed due to insufficient reward",
crash=False)
screen.error(
"preset.test_max_step_threshold: {}".format(
preset.test_max_step_threshold),
crash=False)
screen.error(
"preset.test_min_return_threshold: {}".format(
preset.test_min_return_threshold),
crash=False)
screen.error(
"averaged_rewards: {}".format(averaged_rewards),
crash=False)
screen.error("episode number: {}".format(
csv['Episode #'].values[-1]),
crash=False)
else:
if csv['Episode #'].shape[0] - last_num_episodes <= 0:
continue
last_num_episodes = csv['Episode #'].values[-1]
# check if reward is enough
if np.any(averaged_rewards >
preset.test_min_return_threshold):
test_passed = True
break
time.sleep(1)
# kill test and print result
os.killpg(os.getpgid(p.pid), signal.SIGTERM)
if test_passed:
screen.success("Passed successfully")
else:
screen.error("Failed due to a mismatch with the golden",
crash=False)
fail_count += 1
shutil.rmtree(test_path)
screen.separator()
if fail_count == 0:
screen.success(" Summary: " + str(test_count) + "/" + str(test_count) +
" tests passed successfully")
else:
screen.error(" Summary: " + str(test_count - fail_count) + "/" +
str(test_count) + " tests passed successfully")