def get_graph_manager(**hp_dict): #################### # All Default Parameters # #################### params = {} params["batch_size"] = int(hp_dict.get("batch_size", 64)) params["num_epochs"] = int(hp_dict.get("num_epochs", 10)) params["stack_size"] = int(hp_dict.get("stack_size", 1)) params["lr"] = float(hp_dict.get("lr", 0.0003)) params["exploration_type"] = (hp_dict.get("exploration_type", "huber")).lower() params["e_greedy_value"] = float(hp_dict.get("e_greedy_value", .05)) params["epsilon_steps"] = int(hp_dict.get("epsilon_steps", 10000)) params["beta_entropy"] = float(hp_dict.get("beta_entropy", .01)) params["discount_factor"] = float(hp_dict.get("discount_factor", .999)) params["loss_type"] = hp_dict.get("loss_type", "Mean squared error").lower() params["num_episodes_between_training"] = int(hp_dict.get("num_episodes_between_training", 20)) params["term_cond_max_episodes"] = int(hp_dict.get("term_cond_max_episodes", 100000)) params["term_cond_avg_score"] = float(hp_dict.get("term_cond_avg_score", 100000)) params_json = json.dumps(params, indent=2, sort_keys=True) print("Using the following hyper-parameters", params_json, sep='\n') #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps(params["term_cond_max_episodes"]) schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(40) schedule_params.evaluation_steps = EnvironmentEpisodes(5) schedule_params.heatup_steps = EnvironmentSteps(0) ######### # Agent # ######### agent_params = ClippedPPOAgentParameters() agent_params.network_wrappers['main'].learning_rate = params["lr"] agent_params.network_wrappers['main'].input_embedders_parameters['observation'].activation_function = 'relu' agent_params.network_wrappers['main'].middleware_parameters.activation_function = 'relu' agent_params.network_wrappers['main'].batch_size = params["batch_size"] agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5 agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999 if params["loss_type"] == "huber": agent_params.network_wrappers['main'].replace_mse_with_huber_loss = True agent_params.algorithm.clip_likelihood_ratio_using_epsilon = 0.2 agent_params.algorithm.clipping_decay_schedule = LinearSchedule(1.0, 0, 1000000) agent_params.algorithm.beta_entropy = params["beta_entropy"] agent_params.algorithm.gae_lambda = 0.95 agent_params.algorithm.discount = params["discount_factor"] agent_params.algorithm.optimization_epochs = params["num_epochs"] agent_params.algorithm.estimate_state_value_using_gae = True agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentEpisodes( params["num_episodes_between_training"]) agent_params.algorithm.num_consecutive_playing_steps = EnvironmentEpisodes(params["num_episodes_between_training"]) agent_params.algorithm.distributed_coach_synchronization_type = DistributedCoachSynchronizationType.SYNC if params["exploration_type"] == "categorical": agent_params.exploration = CategoricalParameters() else: agent_params.exploration = EGreedyParameters() agent_params.exploration.epsilon_schedule = LinearSchedule(1.0, params["e_greedy_value"], params["epsilon_steps"]) ############### # Environment # ############### SilverstoneInputFilter = InputFilter(is_a_reference_filter=True) SilverstoneInputFilter.add_observation_filter('observation', 'to_grayscale', ObservationRGBToYFilter()) SilverstoneInputFilter.add_observation_filter('observation', 'to_uint8', ObservationToUInt8Filter(0, 255)) SilverstoneInputFilter.add_observation_filter('observation', 'stacking', ObservationStackingFilter(params["stack_size"])) env_params = GymVectorEnvironment() env_params.default_input_filter = SilverstoneInputFilter env_params.level = 'SilverstoneRacetrack-Discrete-v0' vis_params = VisualizationParameters() vis_params.dump_mp4 = False ######## # Test # ######## preset_validation_params = PresetValidationParameters() preset_validation_params.test = True preset_validation_params.min_reward_threshold = 400 preset_validation_params.max_episodes_to_achieve_reward = 1000 graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=vis_params, preset_validation_params=preset_validation_params) return graph_manager, params_json
agent_params.algorithm.clipping_decay_schedule = LinearSchedule(1.0, 0, 1000000) agent_params.algorithm.beta_entropy = 0 agent_params.algorithm.gae_lambda = 0.95 agent_params.algorithm.discount = 0.99 agent_params.algorithm.optimization_epochs = 10 agent_params.algorithm.estimate_state_value_using_gae = True agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentSteps(2048) # Distributed Coach synchronization type. agent_params.algorithm.distributed_coach_synchronization_type = DistributedCoachSynchronizationType.SYNC agent_params.exploration = EGreedyParameters() agent_params.exploration.epsilon_schedule = LinearSchedule(1.0, 0.01, 10000) ############### # Environment # ############### env_params = GymVectorEnvironment(level='autoscalesim:SimpleScalableWebserviceSim') ######## # Test # ######## preset_validation_params = PresetValidationParameters() preset_validation_params.test = True preset_validation_params.min_reward_threshold = 150 preset_validation_params.max_episodes_to_achieve_reward = 400 graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=VisualizationParameters(), preset_validation_params=preset_validation_params)
# E-Greedy schedule experience_generating_agent_params.exploration.epsilon_schedule = LinearSchedule( 1.0, 0.01, DATASET_SIZE) # experience_generating_agent_params.exploration.epsilon_schedule = LinearSchedule(1.0, 0.1, DATASET_SIZE//10) experience_generating_agent_params.exploration.evaluation_epsilon = 0 ################ # Environment # ################ env_params = GymVectorEnvironment(level='CartPole-v0') ######## # Test # ######## preset_validation_params = PresetValidationParameters() preset_validation_params.test = True preset_validation_params.min_reward_threshold = 150 preset_validation_params.max_episodes_to_achieve_reward = 50 preset_validation_params.read_csv_tries = 500 graph_manager = BatchRLGraphManager( agent_params=agent_params, experience_generating_agent_params=experience_generating_agent_params, experience_generating_schedule_params=experience_generating_schedule_params, env_params=env_params, schedule_params=schedule_params, vis_params=VisualizationParameters(dump_signals_to_csv_every_x_episodes=1), preset_validation_params=preset_validation_params, reward_model_num_epochs=30, train_to_eval_ratio=0.4)
agent_params.network_wrappers['main'].input_embedders_parameters['observation'].scheme = [Conv2d(32, [1, 3], 1)] agent_params.network_wrappers['main'].middleware_parameters.scheme = MiddlewareScheme.Empty agent_params.network_wrappers['main'].batch_size = 64 agent_params.algorithm.clipping_decay_schedule = LinearSchedule(1.0, 0, 150000) agent_params.algorithm.discount = 0.99 agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentSteps(2048) # Distributed Coach synchronization type. agent_params.algorithm.distributed_coach_synchronization_type = DistributedCoachSynchronizationType.SYNC agent_params.exploration = EGreedyParameters() agent_params.exploration.epsilon_schedule = LinearSchedule(1.0, 0.01, 10000) ############### # Environment # ############### env_params = GymVectorEnvironment(level='portfolio_env:PortfolioEnv') env_params.__dict__['observation_space_type'] = ObservationSpaceType.Tensor ######## # Test # ######## preset_validation_params = PresetValidationParameters() preset_validation_params.test = True graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=VisualizationParameters(), preset_validation_params=preset_validation_params)
'action'].scheme = EmbedderScheme.Empty agent_params.output_filter = \ OutputFilter( action_filters=OrderedDict([ ('discretization', BoxDiscretization(num_bins_per_dimension=int(1e6))) ]), is_a_reference_filter=False ) ############### # Environment # ############### env_params = GymVectorEnvironment(level=SingleLevelSelection(mujoco_v2)) ######## # Test # ######## preset_validation_params = PresetValidationParameters() preset_validation_params.test = True preset_validation_params.min_reward_threshold = 500 preset_validation_params.max_episodes_to_achieve_reward = 1000 preset_validation_params.reward_test_level = 'inverted_pendulum' preset_validation_params.trace_test_levels = ['inverted_pendulum'] graph_manager = BasicRLGraphManager( agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=VisualizationParameters(), preset_validation_params=preset_validation_params)
################ # Environment # ################ env_params = GymVectorEnvironment(level="patient_envs:PatientMountainCar") ################# # Visualization # ################# vis_params = VisualizationParameters() vis_params.dump_gifs = True vis_params.video_dump_filters = [ SelectedPhaseOnlyDumpFilter(RunPhase.TEST), MaxDumpFilter() ] ######## # Test # ######## preset_validation_params = PresetValidationParameters() preset_validation_params.test = True preset_validation_params.min_reward_threshold = -200 preset_validation_params.max_episodes_to_achieve_reward = 125 graph_manager = BasicRLGraphManager( agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=vis_params, preset_validation_params=preset_validation_params, )
Conv2d([32, 4, 2])] agent_params.network_wrappers['main'].middleware_parameters.scheme = [ Dense([256]) ] ############### # Environment # ############### env_params = Atari() env_params.level = SingleLevelSelection(atari_deterministic_v4) vis_params = VisualizationParameters() vis_params.video_dump_methods = [ SelectedPhaseOnlyDumpMethod(RunPhase.TEST), MaxDumpMethod() ] vis_params.dump_mp4 = False ######## # Test # ######## preset_validation_params = PresetValidationParameters() preset_validation_params.trace_test_levels = ['breakout', 'pong', 'alien'] graph_manager = BasicRLGraphManager( agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=vis_params, preset_validation_params=preset_validation_params)
schedule_params.evaluation_steps = EnvironmentEpisodes(3) schedule_params.heatup_steps = EnvironmentSteps(2000) ######### # Agent # ######### agent_params = NECAgentParameters() agent_params.network_wrappers['main'].learning_rate = 0.00001 agent_params.input_filter = AtariInputFilter() agent_params.input_filter.remove_reward_filter('clipping') ############### # Environment # ############### env_params = Atari(level=SingleLevelSelection(atari_deterministic_v4)) env_params.random_initialization_steps = 1 ######## # Test # ######## preset_validation_params = PresetValidationParameters() preset_validation_params.test_using_a_trace_test = False graph_manager = BasicRLGraphManager( agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=VisualizationParameters(), preset_validation_params=preset_validation_params)
def get_graph_manager(hp_dict, agent_list, run_phase_subject, enable_domain_randomization=False, done_condition=any, run_type=str(RunType.ROLLOUT_WORKER), pause_physics=None, unpause_physics=None): #################### # Hyperparameters # #################### training_algorithm = agent_list[ 0].ctrl.model_metadata.training_algorithm if agent_list else None params = get_updated_hyper_parameters(hp_dict, training_algorithm) params_json = json.dumps(params, indent=2, sort_keys=True) print("Using the following hyper-parameters", params_json, sep='\n') #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps( params[HyperParameterKeys.TERMINATION_CONDITION_MAX_EPISODES.value]) schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(40) schedule_params.evaluation_steps = EnvironmentEpisodes(5) schedule_params.heatup_steps = EnvironmentSteps(0) ######### # Agent # ######### trainable_agents_list = list() non_trainable_agents_list = list() for agent in agent_list: if agent.network_settings: if TrainingAlgorithm.SAC.value == training_algorithm: agent_params = get_sac_params(DeepRacerSACAgentParams(), agent, params, run_type) else: agent_params = get_clipped_ppo_params( DeepRacerClippedPPOAgentParams(), agent, params) agent_params.env_agent = agent input_filter = InputFilter(is_a_reference_filter=True) for observation in agent.network_settings['input_embedders'].keys( ): if observation == Input.LEFT_CAMERA.value or observation == Input.CAMERA.value or\ observation == Input.OBSERVATION.value: input_filter.add_observation_filter( observation, 'to_grayscale', ObservationRGBToYFilter()) input_filter.add_observation_filter( observation, 'to_uint8', ObservationToUInt8Filter(0, 255)) input_filter.add_observation_filter( observation, 'stacking', ObservationStackingFilter(1)) if observation == Input.STEREO.value: input_filter.add_observation_filter( observation, 'to_uint8', ObservationToUInt8Filter(0, 255)) if observation == Input.LIDAR.value: input_filter.add_observation_filter( observation, 'clipping', ObservationClippingFilter(0.15, 1.0)) if observation == Input.SECTOR_LIDAR.value: input_filter.add_observation_filter( observation, 'binary', ObservationBinarySectorFilter()) agent_params.input_filter = input_filter() trainable_agents_list.append(agent_params) else: non_trainable_agents_list.append(agent) ############### # Environment # ############### env_params = DeepRacerRacetrackEnvParameters() env_params.agents_params = trainable_agents_list env_params.non_trainable_agents = non_trainable_agents_list env_params.level = 'DeepRacerRacetrackEnv-v0' env_params.run_phase_subject = run_phase_subject env_params.enable_domain_randomization = enable_domain_randomization env_params.done_condition = done_condition env_params.pause_physics = pause_physics env_params.unpause_physics = unpause_physics vis_params = VisualizationParameters() vis_params.dump_mp4 = False ######## # Test # ######## preset_validation_params = PresetValidationParameters() preset_validation_params.test = True preset_validation_params.min_reward_threshold = 400 preset_validation_params.max_episodes_to_achieve_reward = 10000 graph_manager = MultiAgentGraphManager( agents_params=trainable_agents_list, env_params=env_params, schedule_params=schedule_params, vis_params=vis_params, preset_validation_params=preset_validation_params, done_condition=done_condition) return graph_manager, params_json
agent_params.network_wrappers['main'].batch_size) agent_params.network_wrappers['main'].learning_rate = 0.0001 agent_params.exploration.epsilon_schedule = LinearSchedule(0.5, 0, 10000) agent_params.exploration.evaluation_epsilon = 0 # this works better than the default which is 64 agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(1) agent_params.algorithm.use_accumulated_reward_as_measurement = True agent_params.algorithm.goal_vector = [0, 1] # ammo, accumulated_reward agent_params.algorithm.handling_targets_after_episode_end = HandlingTargetsAfterEpisodeEnd.LastStep ############### # Environment # ############### env_params = DoomEnvironmentParameters(level='basic') ######## # Test # ######## preset_validation_params = PresetValidationParameters() preset_validation_params.trace_max_env_steps = 2000 graph_manager = BasicRLGraphManager( agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=VisualizationParameters(), preset_validation_params=preset_validation_params)
######### agent_params = NAFAgentParameters() agent_params.network_wrappers['main'].input_embedders_parameters['observation'].scheme = [Dense(200)] agent_params.network_wrappers['main'].middleware_parameters.scheme = [Dense(200)] agent_params.network_wrappers['main'].clip_gradients = 1000 agent_params.network_wrappers['main'].gradients_clipping_method = GradientClippingMethod.ClipByValue ############### # Environment # ############### env_params = GymVectorEnvironment(level='LunarLanderContinuous-v2') # this preset is currently broken - no test ######## # Test # ######## preset_validation_params = PresetValidationParameters() # preset_validation_params.test = True # preset_validation_params.min_reward_threshold = 200 # preset_validation_params.max_episodes_to_achieve_reward = 600 # preset_validation_params.reward_test_level = 'inverted_pendulum' preset_validation_params.trace_test_levels = ['inverted_pendulum', 'hopper'] graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=VisualizationParameters(), preset_validation_params=preset_validation_params)
def get_graph_manager(**hp_dict): #################### # All Default Parameters # #################### params = {} params["batch_size"] = int(hp_dict.get("batch_size", 64)) params["num_epochs"] = int(hp_dict.get("num_epochs", 10)) params["stack_size"] = int(hp_dict.get("stack_size", 1)) params["lr"] = float(hp_dict.get("lr", 0.0003)) params["exploration_type"] = (hp_dict.get("exploration_type", "categorical")).lower() params["e_greedy_value"] = float(hp_dict.get("e_greedy_value", .05)) params["epsilon_steps"] = int(hp_dict.get("epsilon_steps", 10000)) params["beta_entropy"] = float(hp_dict.get("beta_entropy", .01)) params["discount_factor"] = float(hp_dict.get("discount_factor", .999)) params["loss_type"] = hp_dict.get("loss_type", "Mean squared error").lower() params["num_episodes_between_training"] = int( hp_dict.get("num_episodes_between_training", 20)) params["term_cond_max_episodes"] = int( hp_dict.get("term_cond_max_episodes", 100000)) params["term_cond_avg_score"] = float( hp_dict.get("term_cond_avg_score", 100000)) params_json = json.dumps(params, indent=2, sort_keys=True) print("Using the following hyper-parameters", params_json, sep='\n') #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps( params["term_cond_max_episodes"]) schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(40) schedule_params.evaluation_steps = EnvironmentEpisodes(5) schedule_params.heatup_steps = EnvironmentSteps(0) ######### # Agent # ######### agent_params = ClippedPPOAgentParameters() agent_params.network_wrappers['main'].learning_rate = params["lr"] agent_params.network_wrappers['main'].input_embedders_parameters[ 'observation'].activation_function = 'relu' agent_params.network_wrappers[ 'main'].middleware_parameters.activation_function = 'relu' agent_params.network_wrappers['main'].batch_size = params["batch_size"] agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5 agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999 if params["loss_type"] == "huber": agent_params.network_wrappers[ 'main'].replace_mse_with_huber_loss = True agent_params.algorithm.clip_likelihood_ratio_using_epsilon = 0.2 agent_params.algorithm.clipping_decay_schedule = LinearSchedule( 1.0, 0, 1000000) agent_params.algorithm.beta_entropy = params["beta_entropy"] agent_params.algorithm.gae_lambda = 0.95 agent_params.algorithm.discount = params["discount_factor"] agent_params.algorithm.optimization_epochs = params["num_epochs"] agent_params.algorithm.estimate_state_value_using_gae = True agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentEpisodes( params["num_episodes_between_training"]) agent_params.algorithm.num_consecutive_playing_steps = EnvironmentEpisodes( params["num_episodes_between_training"]) agent_params.algorithm.distributed_coach_synchronization_type = DistributedCoachSynchronizationType.SYNC if params["exploration_type"] == "categorical": agent_params.exploration = CategoricalParameters() else: agent_params.exploration = EGreedyParameters() agent_params.exploration.epsilon_schedule = LinearSchedule( 1.0, params["e_greedy_value"], params["epsilon_steps"]) ############### # Environment # ############### DeepRacerInputFilter = InputFilter(is_a_reference_filter=True) DeepRacerInputFilter.add_observation_filter('observation', 'to_grayscale', ObservationRGBToYFilter()) DeepRacerInputFilter.add_observation_filter( 'observation', 'to_uint8', ObservationToUInt8Filter(0, 255)) DeepRacerInputFilter.add_observation_filter( 'observation', 'stacking', ObservationStackingFilter(params["stack_size"])) env_params = GymVectorEnvironment() env_params.default_input_filter = DeepRacerInputFilter env_params.level = 'DeepRacerRacetrackCustomActionSpaceEnv-v0' vis_params = VisualizationParameters() vis_params.dump_mp4 = False ######## # Test # ######## preset_validation_params = PresetValidationParameters() preset_validation_params.test = True preset_validation_params.min_reward_threshold = 400 preset_validation_params.max_episodes_to_achieve_reward = 10000 graph_manager = BasicRLGraphManager( agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=vis_params, preset_validation_params=preset_validation_params) return graph_manager, params_json
agent_params.network_wrappers["main"].middleware_parameters.scheme = [ Dense(512) ] # ER size agent_params.memory.max_size = (MemoryGranularity.Transitions, 10000) # E-Greedy schedule agent_params.exploration.epsilon_schedule = LinearSchedule(1.0, 0.01, 40000) ############# # Environment ############# env_params = GymVectorEnvironment(level="trading_env:TradingEnv") ################## # Manage resources ################## preset_validation_params = PresetValidationParameters() preset_validation_params.test = True graph_manager = BasicRLGraphManager( agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=VisualizationParameters(), preset_validation_params=preset_validation_params, )
def get_graph_manager(hp_dict, agent_list, run_phase_subject, enable_domain_randomization=False, done_condition=any, run_type=str(RunType.ROLLOUT_WORKER), pause_physics=None, unpause_physics=None): #################### # Hyperparameters # #################### # Note: The following three line hard-coded to pick the first agent's trainig algorithm # and dump the hyper parameters for the particular training algorithm into json # for training jobs (so that the console display the training hyperparameters correctly) # since right now, we only support training one model at a time. # TODO: clean these lines up when we support multi-agent training. training_algorithm = agent_list[ 0].ctrl.model_metadata.training_algorithm if agent_list else None params = get_updated_hyper_parameters(hp_dict, training_algorithm) params_json = json.dumps(params, indent=2, sort_keys=True) print("Using the following hyper-parameters", params_json, sep='\n') #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps( params[HyperParameterKeys.TERMINATION_CONDITION_MAX_EPISODES.value]) schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(40) schedule_params.evaluation_steps = EnvironmentEpisodes(5) schedule_params.heatup_steps = EnvironmentSteps(0) ######### # Agent # ######### trainable_agents_list = list() non_trainable_agents_list = list() for agent in agent_list: if agent.network_settings: training_algorithm = agent.ctrl.model_metadata.training_algorithm params = get_updated_hyper_parameters(hp_dict, training_algorithm) if TrainingAlgorithm.SAC.value == training_algorithm: agent_params = get_sac_params(DeepRacerSACAgentParams(), agent, params, run_type) else: agent_params = get_clipped_ppo_params( DeepRacerClippedPPOAgentParams(), agent, params) agent_params.env_agent = agent input_filter = InputFilter(is_a_reference_filter=True) for observation in agent.network_settings['input_embedders'].keys( ): if observation == Input.LEFT_CAMERA.value or observation == Input.CAMERA.value or \ observation == Input.OBSERVATION.value: input_filter.add_observation_filter( observation, 'to_grayscale', ObservationRGBToYFilter()) input_filter.add_observation_filter( observation, 'to_uint8', ObservationToUInt8Filter(0, 255)) input_filter.add_observation_filter( observation, 'stacking', ObservationStackingFilter(1)) if observation == Input.STEREO.value: input_filter.add_observation_filter( observation, 'to_uint8', ObservationToUInt8Filter(0, 255)) if observation == Input.LIDAR.value: input_filter.add_observation_filter( observation, 'clipping', ObservationClippingFilter(0.15, 1.0)) if observation == Input.SECTOR_LIDAR.value: sector_binary_filter = ObservationSectorDiscretizeFilter( num_sectors=NUMBER_OF_LIDAR_SECTORS, num_values_per_sector=1, clipping_dist=SECTOR_LIDAR_CLIPPING_DIST) input_filter.add_observation_filter( observation, 'binary', sector_binary_filter) if observation == Input.DISCRETIZED_SECTOR_LIDAR.value: num_sectors = agent.ctrl.model_metadata.lidar_num_sectors num_values_per_sector = agent.ctrl.model_metadata.lidar_num_values_per_sector clipping_dist = agent.ctrl.model_metadata.lidar_clipping_dist sector_discretize_filter = ObservationSectorDiscretizeFilter( num_sectors=num_sectors, num_values_per_sector=num_values_per_sector, clipping_dist=clipping_dist) input_filter.add_observation_filter( observation, 'discrete', sector_discretize_filter) agent_params.input_filter = input_filter() trainable_agents_list.append(agent_params) else: non_trainable_agents_list.append(agent) ############### # Environment # ############### env_params = DeepRacerRacetrackEnvParameters() env_params.agents_params = trainable_agents_list env_params.non_trainable_agents = non_trainable_agents_list env_params.level = 'DeepRacerRacetrackEnv-v0' env_params.run_phase_subject = run_phase_subject env_params.enable_domain_randomization = enable_domain_randomization env_params.done_condition = done_condition env_params.pause_physics = pause_physics env_params.unpause_physics = unpause_physics vis_params = VisualizationParameters() vis_params.dump_mp4 = False ######## # Test # ######## preset_validation_params = PresetValidationParameters() preset_validation_params.test = True preset_validation_params.min_reward_threshold = 400 preset_validation_params.max_episodes_to_achieve_reward = 10000 graph_manager = MultiAgentGraphManager( agents_params=trainable_agents_list, env_params=env_params, schedule_params=schedule_params, vis_params=vis_params, preset_validation_params=preset_validation_params, done_condition=done_condition) return graph_manager, params_json
agent_params.pre_network_filter.add_observation_filter('desired_goal', 'normalize_desired_goal', ObservationNormalizationFilter(name='normalize_desired_goal')) ############### # Environment # ############### env_params = Mujoco() env_params.level = SingleLevelSelection(fetch_v1) env_params.custom_reward_threshold = -49 vis_params = VisualizationParameters() vis_params.video_dump_methods = [SelectedPhaseOnlyDumpMethod(RunPhase.TEST), MaxDumpMethod()] vis_params.dump_mp4 = False ######## # Test # ######## preset_validation_params = PresetValidationParameters() # preset_validation_params.test = True # preset_validation_params.min_reward_threshold = 200 # preset_validation_params.max_episodes_to_achieve_reward = 600 # preset_validation_params.reward_test_level = 'inverted_pendulum' preset_validation_params.trace_test_levels = ['slide', 'pick_and_place', 'push', 'reach'] graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params, schedule_params=schedule_params, vis_params=vis_params, preset_validation_params=preset_validation_params)