from rl_coach.schedules import LinearSchedule #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps(10000000) schedule_params.steps_between_evaluation_periods = EnvironmentSteps(2048) schedule_params.evaluation_steps = EnvironmentEpisodes(5) schedule_params.heatup_steps = EnvironmentSteps(0) ######### # Agent # ######### agent_params = ClippedPPOAgentParameters() agent_params.network_wrappers['main'].learning_rate = 0.0003 agent_params.network_wrappers['main'].input_embedders_parameters[ 'observation'].activation_function = 'tanh' agent_params.network_wrappers['main'].input_embedders_parameters[ 'observation'].scheme = [Dense(64)] agent_params.network_wrappers['main'].middleware_parameters.scheme = [ Dense(64) ] agent_params.network_wrappers[ 'main'].middleware_parameters.activation_function = 'tanh' agent_params.network_wrappers['main'].batch_size = 64 agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5 agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999
from rl_coach.schedules import LinearSchedule #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps(10000000) schedule_params.steps_between_evaluation_periods = EnvironmentSteps(2048) schedule_params.evaluation_steps = EnvironmentEpisodes(5) schedule_params.heatup_steps = EnvironmentSteps(0) ######### # Agent # ######### agent_params = ClippedPPOAgentParameters() agent_params.network_wrappers['main'].learning_rate = 0.0001 # 0.0003 agent_params.network_wrappers['main'].input_embedders_parameters[ 'observation'].activation_function = 'tanh' agent_params.network_wrappers['main'].input_embedders_parameters[ 'observation'].scheme = [Dense(64)] agent_params.network_wrappers['main'].middleware_parameters.scheme = [ Dense(64) ] agent_params.network_wrappers[ 'main'].middleware_parameters.activation_function = 'tanh' agent_params.network_wrappers['main'].batch_size = 64 agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5 agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999
#from rl_coach.environments.environment import SelectedPhaseOnlyDumpMethod, MaxDumpMethod from rl_coach import logger from rl_coach.base_parameters import TaskParameters from rl_coach.base_parameters import VisualizationParameters ################ # Environment # ################ env_params = GymVectorEnvironment(level='VRP_abstract_env:VRPEasyEnv') ######### # Agent # ######### agent_params = ClippedPPOAgentParameters() ################# # Visualization # ################# env_params.frame_skip = 5 #to make sure the gifs work without skipping steps vis_params = VisualizationParameters() vis_params.dump_gifs = True #vis_params.video_dump_methods = [SelectedPhaseOnlyDumpMethod(RunPhase.TEST), MaxDumpMethod()] #experiment_name = "TSPEasy" #experiment_name = logger.get_experiment_name(experiment_name) #experiment_path = logger.get_experiment_path(experiment_name)
from rl_coach.schedules import LinearSchedule #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps(int(5e5)) schedule_params.steps_between_evaluation_periods = EnvironmentSteps(50000) schedule_params.evaluation_steps = EnvironmentEpisodes(5) schedule_params.heatup_steps = EnvironmentSteps(150000) ######### # Agent # ######### agent_params = ClippedPPOAgentParameters() agent_params.network_wrappers['main'].learning_rate = 0.0003 agent_params.network_wrappers['main'].input_embedders_parameters['observation'].activation_function = 'tanh' agent_params.network_wrappers['main'].input_embedders_parameters['observation'].scheme = [Dense(64)] agent_params.network_wrappers['main'].middleware_parameters.scheme = [Dense(64)] agent_params.network_wrappers['main'].middleware_parameters.activation_function = 'tanh' agent_params.network_wrappers['main'].batch_size = 64 agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5 agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999 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 = 0 agent_params.algorithm.gae_lambda = 0.95 agent_params.algorithm.discount = 0.99
from rl_coach.memories.memory import MemoryGranularity #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps(10000000) schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(60) schedule_params.evaluation_steps = EnvironmentEpisodes(5) schedule_params.heatup_steps = EnvironmentSteps(0) ######### # Agent # ######### agent_params = ClippedPPOAgentParameters() agent_params.memory.max_size = (MemoryGranularity.Transitions, 10**5) agent_params.network_wrappers['main'].learning_rate = 0.00027 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 = 96 agent_params.network_wrappers['main'].optimizer_epsilon = 1e-8 agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999 agent_params.algorithm.clip_likelihood_ratio_using_epsilon = 0.2 agent_params.algorithm.clipping_decay_schedule = LinearSchedule( 1.0, 0, 1000000)
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 # ############### 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 = 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
from rl_coach.schedules import LinearSchedule #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps(2000) schedule_params.steps_between_evaluation_periods = EnvironmentSteps(1000) schedule_params.evaluation_steps = EnvironmentEpisodes(2) schedule_params.heatup_steps = EnvironmentSteps(0) ######### # Agent # ######### agent_params = ClippedPPOAgentParameters() agent_params.network_wrappers["main"].learning_rate = 0.0003 agent_params.network_wrappers["main"].input_embedders_parameters[ "observation"].activation_function = "tanh" agent_params.network_wrappers["main"].input_embedders_parameters[ "observation"].scheme = [Dense(64)] agent_params.network_wrappers["main"].middleware_parameters.scheme = [ Dense(64) ] agent_params.network_wrappers[ "main"].middleware_parameters.activation_function = "tanh" agent_params.network_wrappers["main"].batch_size = 64 agent_params.network_wrappers["main"].optimizer_epsilon = 1e-5 agent_params.network_wrappers["main"].adam_optimizer_beta2 = 0.999
from rl_coach.filters.observation.observation_to_uint8_filter import ObservationToUInt8Filter #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps(10000000) 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.memory.max_size = (MemoryGranularity.Transitions, 10**5) agent_params.network_wrappers['main'].learning_rate = 0.0003 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 = 64 agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5 agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999 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 = 0.01 # also try 0.001 agent_params.algorithm.gae_lambda = 0.95 agent_params.algorithm.discount = 0.999
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
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["lr_decay_rate"] = float(hp_dict.get("lr_decay_rate", 0)) params["lr_decay_steps"] = float(hp_dict.get("lr_decay_steps", 0)) 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["tensorboard"] = hp_dict.get("tensorboard", False) params["dump_mp4"] = hp_dict.get("dump_mp4", False) params["dump_gifs"] = hp_dict.get("dump_gifs", False) 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'].learning_rate_decay_rate = params["lr_decay_rate"] agent_params.network_wrappers['main'].learning_rate_decay_steps = params["lr_decay_steps"] agent_params.network_wrappers['main'].input_embedders_parameters['observation'].activation_function = 'relu' # Replace the default CNN with single layer Conv2d(32, 3, 1) # agent_params.network_wrappers['main'].input_embedders_parameters['observation'].scheme = EmbedderScheme.Shallow # agent_params.network_wrappers['main'].input_embedders_parameters['observation'].dropout_rate = 0.3 agent_params.network_wrappers['main'].middleware_parameters.activation_function = 'relu' # agent_params.network_wrappers['main'].middleware_parameters.scheme = MiddlewareScheme.Shallow # agent_params.network_wrappers['main'].middleware_parameters.dropout_rate = 0.3 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 # agent_params.network_wrappers['main'].l2_regularization = 2e-5 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) # Add an observation image pertubation for many aspects # DeepRacerInputFilter.add_observation_filter('observation', 'perturb_color', ObservationColorPerturbation(0.2)) # Rescale to much smaller input when using shallow networks to avoid OOM # DeepRacerInputFilter.add_observation_filter('observation', 'rescaling', # ObservationRescaleToSizeFilter(ImageObservationSpace(np.array([84, 84, 3]), # high=255))) 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.tensorboard = params["tensorboard"] vis_params.dump_mp4 = params["dump_mp4"] vis_params.dump_gifs = params["dump_gifs"] # AlwaysDumpFilter, MaxDumpFilter, EveryNEpisodesDumpFilter, SelectedPhaseOnlyDumpFilter vis_params.video_dump_filters = [AlwaysDumpFilter()] ######## # 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