agent_params.network_wrappers['critic'].input_embedders_parameters.pop('observation') # left camera agent_params.network_wrappers['actor'].input_embedders_parameters['left_camera'] = \ copy.deepcopy(agent_params.network_wrappers['actor'].input_embedders_parameters['forward_camera']) agent_params.network_wrappers['critic'].input_embedders_parameters['left_camera'] = \ copy.deepcopy(agent_params.network_wrappers['critic'].input_embedders_parameters['forward_camera']) # right camera agent_params.network_wrappers['actor'].input_embedders_parameters['right_camera'] = \ copy.deepcopy(agent_params.network_wrappers['actor'].input_embedders_parameters['forward_camera']) agent_params.network_wrappers['critic'].input_embedders_parameters['right_camera'] = \ copy.deepcopy(agent_params.network_wrappers['critic'].input_embedders_parameters['forward_camera']) agent_params.input_filter = CarlaInputFilter() agent_params.input_filter.copy_filters_from_one_observation_to_another( 'forward_camera', 'left_camera') agent_params.input_filter.copy_filters_from_one_observation_to_another( 'forward_camera', 'right_camera') ############### # Environment # ############### env_params = CarlaEnvironmentParameters() env_params.cameras = [CameraTypes.FRONT, CameraTypes.LEFT, CameraTypes.RIGHT] graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=VisualizationParameters())
# 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') vis_params = VisualizationParameters() vis_params.video_dump_methods = [SelectedPhaseOnlyDumpMethod(RunPhase.TEST)] vis_params.dump_mp4 = True graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=vis_params)
schedule_params.evaluation_steps = EnvironmentEpisodes(1) schedule_params.heatup_steps = EnvironmentSteps(1000) ######### # Agent # ######### agent_params = DDQNAgentParameters() agent_params.network_wrappers['main'].learning_rate = 0.00025 agent_params.network_wrappers['main'].heads_parameters = [DuelingQHeadParameters()] agent_params.network_wrappers['main'].middleware_parameters.scheme = MiddlewareScheme.Empty agent_params.network_wrappers['main'].rescale_gradient_from_head_by_factor = [1/math.sqrt(2), 1/math.sqrt(2)] agent_params.network_wrappers['main'].clip_gradients = 10 agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(4) agent_params.network_wrappers['main'].input_embedders_parameters['forward_camera'] = \ agent_params.network_wrappers['main'].input_embedders_parameters.pop('observation') agent_params.output_filter = OutputFilter() agent_params.output_filter.add_action_filter('discretization', BoxDiscretization(5)) ############### # Environment # ############### env_params = CarlaEnvironmentParameters() env_params.level = 'town1' vis_params = VisualizationParameters() vis_params.video_dump_methods = [SelectedPhaseOnlyDumpMethod(RunPhase.TEST), MaxDumpMethod()] vis_params.dump_mp4 = False graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=vis_params)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(20) schedule_params.evaluation_steps = EnvironmentEpisodes(1) schedule_params.heatup_steps = EnvironmentSteps(1000) ######### # Agent # ######### agent_params = DDQNAgentParameters() agent_params.network_wrappers['main'].learning_rate = 0.00025 agent_params.network_wrappers['main'].heads_parameters = \ [DuelingQHeadParameters(rescale_gradient_from_head_by_factor=1/math.sqrt(2))] agent_params.network_wrappers[ 'main'].middleware_parameters.scheme = MiddlewareScheme.Empty agent_params.network_wrappers['main'].clip_gradients = 10 agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(4) agent_params.network_wrappers['main'].input_embedders_parameters['forward_camera'] = \ agent_params.network_wrappers['main'].input_embedders_parameters.pop('observation') agent_params.output_filter = OutputFilter() agent_params.output_filter.add_action_filter('discretization', BoxDiscretization(5)) ############### # Environment # ############### env_params = CarlaEnvironmentParameters() graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=VisualizationParameters())
# critic (q) network parameters agent_params.network_wrappers['q'].heads_parameters[0].network_layers_sizes = ( 32, 32) agent_params.network_wrappers['q'].batch_size = 32 agent_params.network_wrappers['q'].learning_rate = 0.0003 agent_params.network_wrappers['q'].optimizer_epsilon = 1e-5 agent_params.network_wrappers['q'].adam_optimizer_beta2 = 0.999 agent_params.network_wrappers['q'].input_embedders_parameters['forward_camera'] = \ agent_params.network_wrappers['q'].input_embedders_parameters.pop('observation') # actor (policy) network parameters agent_params.network_wrappers['policy'].batch_size = 32 agent_params.network_wrappers['policy'].learning_rate = 0.0003 agent_params.network_wrappers['policy'].middleware_parameters.scheme = [ Dense(32) ] agent_params.network_wrappers['policy'].optimizer_epsilon = 1e-5 agent_params.network_wrappers['policy'].adam_optimizer_beta2 = 0.999 agent_params.network_wrappers['policy'].input_embedders_parameters['forward_camera'] = \ agent_params.network_wrappers['policy'].input_embedders_parameters.pop('observation') ############### # Environment # ############### env_params = CarlaEnvironmentParameters(level='town2') graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=VisualizationParameters())
# TODO: normalize the speed with the maximum speed from the training set speed /= 25 (90 km/h) agent_params.exploration = AdditiveNoiseParameters() agent_params.exploration.noise_percentage_schedule = ConstantSchedule(0) agent_params.exploration.evaluation_noise_percentage = 0 agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(0) agent_params.memory.load_memory_from_file_path = "/home/cvds_lab/Documents/advanced-coach/carla_train_set_replay_buffer.p" agent_params.memory.state_key_with_the_class_index = 'high_level_command' agent_params.memory.num_classes = 4 ############### # Environment # ############### env_params = CarlaEnvironmentParameters() env_params.level = 'town1' env_params.cameras = [CameraTypes.FRONT] env_params.camera_height = 600 env_params.camera_width = 800 env_params.allow_braking = True env_params.quality = CarlaEnvironmentParameters.Quality.EPIC vis_params = VisualizationParameters() vis_params.video_dump_methods = [ SelectedPhaseOnlyDumpMethod(RunPhase.TEST), MaxDumpMethod() ] vis_params.dump_mp4 = True graph_manager = BasicRLGraphManager(agent_params=agent_params,