def evaluate(params):
# file params
experiment_path = os.path.join(params.output_data_dir)
logger.experiment_path = os.path.join(experiment_path, 'evaluation')
params.checkpoint_restore_dir = os.path.join(params.input_data_dir, 'checkpoint')
checkpoint_file = os.path.join(params.checkpoint_restore_dir, 'checkpoint')
inplace_change(checkpoint_file, "/opt/ml/output/data/checkpoint", ".")
# Note that due to a tensorflow issue (https://github.com/tensorflow/tensorflow/issues/9146) we need to replace
# the absolute path for the evaluation-from-a-checkpointed-model to work
vis_params = VisualizationParameters()
vis_params.dump_gifs = True
task_params = TaskParameters(evaluate_only=True, experiment_path=logger.experiment_path)
task_params.__dict__ = add_items_to_dict(task_params.__dict__, params.__dict__)
graph_manager = BasicRLGraphManager(
agent_params=ClippedPPOAgentParameters(),
env_params=GymVectorEnvironment(level='TSP_env:TSPEasyEnv'),
schedule_params=ScheduleParameters(),
vis_params=vis_params
)
graph_manager = graph_manager.create_graph(task_parameters=task_params)
graph_manager.evaluate(EnvironmentSteps(5))
agent_params.network_wrappers['main'].replace_mse_with_huber_loss = False
# ER size
agent_params.memory.max_size = (MemoryGranularity.Transitions, 40000)
# E-Greedy schedule
agent_params.exploration.epsilon_schedule = LinearSchedule(1.0, 0.01, 10000)
agent_params.input_filter = InputFilter()
agent_params.input_filter.add_reward_filter(
'adversarial_inversion', RewardAdversarialInversionFilter(0.3))
###############
# Environment #
###############
env_params = GymVectorEnvironment(level='MountainCar-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 = 250
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.network_wrappers['critic'].middleware_parameters.scheme = [Dense([300])]
agent_params.network_wrappers['critic'].input_embedders_parameters['action'].scheme = EmbedderScheme.Empty
agent_params.input_filter = MujocoInputFilter()
agent_params.input_filter.add_reward_filter("rescale", RewardRescaleFilter(1/10.))
###############
# Environment #
###############
env_params = ControlSuiteEnvironmentParameters()
env_params.level = SingleLevelSelection(control_suite_envs)
vis_params = VisualizationParameters()
vis_params.video_dump_methods = [SelectedPhaseOnlyDumpMethod(RunPhase.TEST), MaxDumpMethod()]
vis_params.dump_mp4 = False
########
# Test #
########
# this preset is too slow to test on a regular basis
# preset_validation_params = PresetValidationParameters()
# preset_validation_params.test = True
# preset_validation_params.min_reward_threshold = 150
# preset_validation_params.max_episodes_to_achieve_reward = 250
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)
from rl_coach.agents.clipped_ppo_agent import ClippedPPOAgentParameters
from rl_coach.core_types import EnvironmentSteps
from rl_coach.environments.gym_environment import GymVectorEnvironment
from rl_coach.graph_managers.basic_rl_graph_manager import BasicRLGraphManager
from rl_coach.graph_managers.graph_manager import SimpleSchedule
graph_manager = BasicRLGraphManager(
agent_params=ClippedPPOAgentParameters(),
env_params=GymVectorEnvironment(level='CartPole-v0'),
schedule_params=SimpleSchedule())
graph_manager.heatup(EnvironmentSteps(100))
graph_manager.train_and_act(EnvironmentSteps(100))
agent_params.network_wrappers['main'].clip_gradients = 40.
agent_params.exploration = CategoricalParameters()
###############
# Environment #
###############
env_params = DoomEnvironmentParameters()
env_params.level = 'basic'
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 = 20
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=vis_params,
preset_validation_params=preset_validation_params)
agent_params.network_wrappers['main'].learning_rate = 0.025
agent_params.exploration.epsilon_schedule = LinearSchedule(1, 0, 500)
###############
# Environment #
###############
level = 'gym_dynamic_multi_armed_bandit.envs:BasicEnv'
env_params = GymVectorEnvironment(level)
########################
# Create Graph Manager #
########################
graph_manager = BasicRLGraphManager(agent_params=agent_params,
env_params=env_params,
schedule_params=schedule)
#######################
# add task parameters #
#######################
log_path = './experiments/log' # training logs are saved
checkpoint_sec = 60 # checkpoints are used to restore the model
if not os.path.exists(log_path):
os.makedirs(log_path)
task_parameters = TaskParameters(evaluate_only=False,
experiment_path=log_path,
checkpoint_save_secs=checkpoint_sec)
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
