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
def get_graph_manager(hp_dict, agent_list, run_phase_subject):
####################
# 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 #
#########
trainable_agents_list = list()
non_trainable_agents_list = list()
for agent in agent_list:
agent_params = DeepRacerAgentParams()
if agent.network_settings:
agent_params.env_agent = agent
agent_params.network_wrappers['main'].learning_rate = params["lr"]
agent_params.network_wrappers['main'].input_embedders_parameters = \
create_input_embedder(agent.network_settings['input_embedders'],
agent.network_settings['embedder_type'],
agent.network_settings['activation_function'])
agent_params.network_wrappers['main'].middleware_parameters = \
create_middle_embedder(agent.network_settings['middleware_embedders'],
agent.network_settings['embedder_type'],
agent.network_settings['activation_function'])
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()
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"])
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
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)
return graph_manager, params_json
def __init__(self, improve_steps=TrainingSteps(10000000000)):
super().__init__()
self.heatup_steps = EnvironmentSteps(0)
self.evaluation_steps = EnvironmentEpisodes(0)
self.steps_between_evaluation_periods = improve_steps
self.improve_steps = improve_steps
def rollout_worker(graph_manager, num_workers, rollout_idx, task_parameters, s3_writer):
"""
wait for first checkpoint then perform rollouts using the model
"""
if not graph_manager.data_store:
raise AttributeError("None type for data_store object")
data_store = graph_manager.data_store
checkpoint_dir = task_parameters.checkpoint_restore_path
wait_for_checkpoint(checkpoint_dir, data_store)
wait_for_trainer_ready(checkpoint_dir, data_store)
# Make the clients that will allow us to pause and unpause the physics
rospy.wait_for_service('/gazebo/pause_physics')
rospy.wait_for_service('/gazebo/unpause_physics')
rospy.wait_for_service('/racecar/save_mp4/subscribe_to_save_mp4')
rospy.wait_for_service('/racecar/save_mp4/unsubscribe_from_save_mp4')
pause_physics = ServiceProxyWrapper('/gazebo/pause_physics', Empty)
unpause_physics = ServiceProxyWrapper('/gazebo/unpause_physics', Empty)
subscribe_to_save_mp4 = ServiceProxyWrapper('/racecar/save_mp4/subscribe_to_save_mp4', Empty)
unsubscribe_from_save_mp4 = ServiceProxyWrapper('/racecar/save_mp4/unsubscribe_from_save_mp4', Empty)
graph_manager.create_graph(task_parameters=task_parameters, stop_physics=pause_physics,
start_physics=unpause_physics, empty_service_call=EmptyRequest)
chkpt_state_reader = CheckpointStateReader(checkpoint_dir, checkpoint_state_optional=False)
last_checkpoint = chkpt_state_reader.get_latest().num
# this worker should play a fraction of the total playing steps per rollout
episode_steps_per_rollout = graph_manager.agent_params.algorithm.num_consecutive_playing_steps.num_steps
act_steps = int(episode_steps_per_rollout / num_workers)
if rollout_idx < episode_steps_per_rollout % num_workers:
act_steps += 1
act_steps = EnvironmentEpisodes(act_steps)
configure_environment_randomizer()
for _ in range((graph_manager.improve_steps / act_steps.num_steps).num_steps):
graph_manager.phase = RunPhase.TRAIN
exit_if_trainer_done(checkpoint_dir, s3_writer, rollout_idx)
unpause_physics(EmptyRequest())
graph_manager.reset_internal_state(True)
graph_manager.act(act_steps, wait_for_full_episodes=graph_manager.agent_params.algorithm.act_for_full_episodes)
graph_manager.reset_internal_state(True)
time.sleep(1)
pause_physics(EmptyRequest())
graph_manager.phase = RunPhase.UNDEFINED
new_checkpoint = -1
if graph_manager.agent_params.algorithm.distributed_coach_synchronization_type\
== DistributedCoachSynchronizationType.SYNC:
unpause_physics(EmptyRequest())
is_save_mp4_enabled = rospy.get_param('MP4_S3_BUCKET', None) and rollout_idx == 0
if is_save_mp4_enabled:
subscribe_to_save_mp4(EmptyRequest())
if rollout_idx == 0:
for _ in range(MIN_EVAL_TRIALS):
graph_manager.evaluate(EnvironmentSteps(1))
while new_checkpoint < last_checkpoint + 1:
exit_if_trainer_done(checkpoint_dir, s3_writer, rollout_idx)
if rollout_idx == 0:
graph_manager.evaluate(EnvironmentSteps(1))
new_checkpoint = data_store.get_chkpoint_num('agent')
if is_save_mp4_enabled:
unsubscribe_from_save_mp4(EmptyRequest())
s3_writer.upload_to_s3()
pause_physics(EmptyRequest())
data_store.load_from_store(expected_checkpoint_number=last_checkpoint+1)
graph_manager.restore_checkpoint()
if graph_manager.agent_params.algorithm.distributed_coach_synchronization_type\
== DistributedCoachSynchronizationType.ASYNC:
if new_checkpoint > last_checkpoint:
graph_manager.restore_checkpoint()
last_checkpoint = new_checkpoint
def rollout_worker(graph_manager, num_workers, rollout_idx, task_parameters,
s3_writer):
"""
wait for first checkpoint then perform rollouts using the model
"""
if not graph_manager.data_store:
raise AttributeError("None type for data_store object")
data_store = graph_manager.data_store
checkpoint_dir = task_parameters.checkpoint_restore_path
wait_for_checkpoint(checkpoint_dir, data_store)
wait_for_trainer_ready(checkpoint_dir, data_store)
# Make the clients that will allow us to pause and unpause the physics
rospy.wait_for_service('/gazebo/pause_physics_dr')
rospy.wait_for_service('/gazebo/unpause_physics_dr')
rospy.wait_for_service('/racecar/save_mp4/subscribe_to_save_mp4')
rospy.wait_for_service('/racecar/save_mp4/unsubscribe_from_save_mp4')
pause_physics = ServiceProxyWrapper('/gazebo/pause_physics_dr', Empty)
unpause_physics = ServiceProxyWrapper('/gazebo/unpause_physics_dr', Empty)
subscribe_to_save_mp4 = ServiceProxyWrapper(
'/racecar/save_mp4/subscribe_to_save_mp4', Empty)
unsubscribe_from_save_mp4 = ServiceProxyWrapper(
'/racecar/save_mp4/unsubscribe_from_save_mp4', Empty)
graph_manager.create_graph(task_parameters=task_parameters,
stop_physics=pause_physics,
start_physics=unpause_physics,
empty_service_call=EmptyRequest)
chkpt_state_reader = CheckpointStateReader(checkpoint_dir,
checkpoint_state_optional=False)
last_checkpoint = chkpt_state_reader.get_latest().num
# this worker should play a fraction of the total playing steps per rollout
episode_steps_per_rollout = graph_manager.agent_params.algorithm.num_consecutive_playing_steps.num_steps
act_steps = int(episode_steps_per_rollout / num_workers)
if rollout_idx < episode_steps_per_rollout % num_workers:
act_steps += 1
act_steps = EnvironmentEpisodes(act_steps)
configure_environment_randomizer()
for _ in range(
(graph_manager.improve_steps / act_steps.num_steps).num_steps):
# Collect profiler information only IS_PROFILER_ON is true
with utils.Profiler(s3_bucket=PROFILER_S3_BUCKET,
s3_prefix=PROFILER_S3_PREFIX,
output_local_path=ROLLOUT_WORKER_PROFILER_PATH,
enable_profiling=IS_PROFILER_ON):
graph_manager.phase = RunPhase.TRAIN
exit_if_trainer_done(checkpoint_dir, s3_writer, rollout_idx)
unpause_physics(EmptyRequest())
graph_manager.reset_internal_state(True)
graph_manager.act(act_steps,
wait_for_full_episodes=graph_manager.
agent_params.algorithm.act_for_full_episodes)
graph_manager.reset_internal_state(True)
time.sleep(1)
pause_physics(EmptyRequest())
graph_manager.phase = RunPhase.UNDEFINED
new_checkpoint = -1
if graph_manager.agent_params.algorithm.distributed_coach_synchronization_type\
== DistributedCoachSynchronizationType.SYNC:
is_save_mp4_enabled = rospy.get_param(
'MP4_S3_BUCKET', None) and rollout_idx == 0
if is_save_mp4_enabled:
subscribe_to_save_mp4(EmptyRequest())
if rollout_idx == 0:
unpause_physics(EmptyRequest())
for _ in range(int(rospy.get_param('MIN_EVAL_TRIALS',
'5'))):
graph_manager.evaluate(EnvironmentSteps(1))
while new_checkpoint < last_checkpoint + 1:
exit_if_trainer_done(checkpoint_dir, s3_writer,
rollout_idx)
if rollout_idx == 0:
print(
"Additional evaluation. New Checkpoint: {}, Last Checkpoint: {}"
.format(new_checkpoint, last_checkpoint))
graph_manager.evaluate(EnvironmentSteps(1))
else:
time.sleep(5)
new_checkpoint = data_store.get_chkpoint_num('agent')
if is_save_mp4_enabled:
unsubscribe_from_save_mp4(EmptyRequest())
logger.info(
"Completed iteration tasks. Writing results to S3.")
s3_writer.upload_to_s3()
pause_physics(EmptyRequest())
logger.info(
"Preparing to load checkpoint {}".format(last_checkpoint +
1))
data_store.load_from_store(
expected_checkpoint_number=last_checkpoint + 1)
graph_manager.restore_checkpoint()
if graph_manager.agent_params.algorithm.distributed_coach_synchronization_type\
== DistributedCoachSynchronizationType.ASYNC:
if new_checkpoint > last_checkpoint:
graph_manager.restore_checkpoint()
last_checkpoint = new_checkpoint
logger.info("Exited main loop. Done.")
context_vector = attention_weights * conv
context_vector = tf.reduce_sum(context_vector, axis=1)
return context_vector
def __str__(self):
return "Convolution (num filters = {}, kernel size = {}, stride = {})"\
.format(self.num_filters, self.kernel_size, self.strides)
####################
# 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.network_wrappers['main'].input_embedders_parameters = {
# 'left_camera': InputEmbedderParameters(activation_function='relu', dropout_rate=0.3),
# 'stereo': InputEmbedderParameters(activation_function='relu', dropout_rate=0.3)
# }
agent_params.network_wrappers['main'].input_embedders_parameters = {
'left_camera':
from rl_coach.base_parameters import VisualizationParameters, PresetValidationParameters from rl_coach.core_types import TrainingSteps, EnvironmentEpisodes, EnvironmentSteps from rl_coach.environments.doom_environment import DoomEnvironmentParameters from rl_coach.graph_managers.basic_rl_graph_manager import BasicRLGraphManager from rl_coach.graph_managers.graph_manager import ScheduleParameters from rl_coach.memories.memory import MemoryGranularity from rl_coach.filters.filter import InputFilter from rl_coach.filters.reward.reward_rescale_filter import RewardRescaleFilter #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps(10000000000) schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(10) schedule_params.evaluation_steps = EnvironmentEpisodes(1) schedule_params.heatup_steps = EnvironmentSteps(0) ######### # Agent # ######### agent_params = ACERAgentParameters() agent_params.algorithm.num_steps_between_gradient_updates = 30 agent_params.algorithm.apply_gradients_every_x_episodes = 1 agent_params.network_wrappers['main'].learning_rate = 0.0001 agent_params.algorithm.ratio_of_replay = 4 agent_params.algorithm.num_transitions_to_start_replay = 2000 agent_params.memory.max_size = (MemoryGranularity.Transitions, 100000) agent_params.input_filter = InputFilter()
def train_using_experience_agent(env_params, n_epochs, dataset_size):
tf.reset_default_graph(
) # just to clean things up; only needed for the tutorial
# Experience Generating Agent parameters
experience_generating_agent_params = DDQNAgentParameters()
# schedule parameters
experience_generating_schedule_params = ScheduleParameters()
experience_generating_schedule_params.heatup_steps = EnvironmentSteps(1000)
experience_generating_schedule_params.improve_steps = TrainingSteps(
dataset_size -
experience_generating_schedule_params.heatup_steps.num_steps)
experience_generating_schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(
10)
experience_generating_schedule_params.evaluation_steps = EnvironmentEpisodes(
1)
# DQN params
experience_generating_agent_params.algorithm.num_steps_between_copying_online_weights_to_target = EnvironmentSteps(
100)
experience_generating_agent_params.algorithm.discount = 0.99
experience_generating_agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(
1)
# NN configuration
experience_generating_agent_params.network_wrappers[
'main'].learning_rate = 0.0001
experience_generating_agent_params.network_wrappers[
'main'].batch_size = 128
experience_generating_agent_params.network_wrappers[
'main'].replace_mse_with_huber_loss = False
experience_generating_agent_params.network_wrappers['main'].heads_parameters = \
[QHeadParameters(output_bias_initializer=tf.constant_initializer(-100))]
# experience_generating_agent_params.network_wrappers['main'].heads_parameters = \
# [QHeadParameters(output_bias_initializer=tf.constant_initializer(0))]
# ER size
experience_generating_agent_params.memory = EpisodicExperienceReplayParameters(
)
experience_generating_agent_params.memory.max_size = \
(MemoryGranularity.Transitions,
experience_generating_schedule_params.heatup_steps.num_steps +
experience_generating_schedule_params.improve_steps.num_steps)
# E-Greedy schedule
experience_generating_agent_params.exploration.epsilon_schedule = LinearSchedule(
1.0, 0.01, DATASET_SIZE)
experience_generating_agent_params.exploration.evaluation_epsilon = 0
schedule_params = set_schedule_params(n_epochs, dataset_size)
# set the agent params as before
# agent_params = set_agent_params(DDQNAgentParameters)
agent_params = set_agent_params(DDQNBCQAgentParameters)
agent_params.algorithm.action_drop_method_parameters = NNImitationModelParameters(
)
# 50 epochs of training (the entire dataset is used each epoch)
# schedule_params.improve_steps = TrainingSteps(50)
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),
reward_model_num_epochs=30,
train_to_eval_ratio=0.5)
graph_manager.create_graph(task_parameters)
graph_manager.improve()
return
from rl_coach.core_types import EnvironmentEpisodes, EnvironmentSteps
from rl_coach.environments.gym_environment import MujocoInputFilter, GymEnvironmentParameters, MujocoOutputFilter
from rl_coach.exploration_policies.additive_noise import AdditiveNoiseParameters
from rl_coach.exploration_policies.truncated_normal import TruncatedNormalParameters
from rl_coach.schedules import ConstantSchedule, PieceWiseSchedule, ExponentialSchedule
from rl_coach.memories.memory import MemoryGranularity
from rl_coach.base_parameters import EmbedderScheme
from rl_coach.architectures.tensorflow_components.architecture import Dense
steps_per_episode = 13
####################
# Block Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = EnvironmentEpisodes(400)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(
3) #3) # Neta: (1000)
schedule_params.evaluation_steps = EnvironmentEpisodes(1) #1) # Neta: 0
schedule_params.heatup_steps = EnvironmentSteps(
2) #120*steps_per_episode) # Neta (2)
#####################
# DDPG Agent Params #
#####################
agent_params = DDPGAgentParameters()
agent_params.network_wrappers['actor'].input_embedders_parameters[
'observation'].scheme = [Dense([300])]
agent_params.network_wrappers['actor'].middleware_parameters.scheme = [
Dense([300])
]
from rl_coach.graph_managers.graph_manager import ScheduleParameters
from rl_coach.graph_managers.hac_graph_manager import HACGraphManager
from rl_coach.memories.episodic.episodic_hindsight_experience_replay import HindsightGoalSelectionMethod, \
EpisodicHindsightExperienceReplayParameters
from rl_coach.memories.episodic.episodic_hrl_hindsight_experience_replay import \
EpisodicHRLHindsightExperienceReplayParameters
from rl_coach.memories.memory import MemoryGranularity
from rl_coach.schedules import ConstantSchedule
from rl_coach.spaces import GoalsSpace, ReachingGoal
####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = EnvironmentEpisodes(40 * 4 * 64) # 40 epochs
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(
4 * 64) # 4 small batches of 64 episodes
schedule_params.evaluation_steps = EnvironmentEpisodes(64)
schedule_params.heatup_steps = EnvironmentSteps(0)
polar_coordinates = False
#########
# Agent #
#########
if polar_coordinates:
distance_from_goal_threshold = np.array([0.075, 0.75])
else:
distance_from_goal_threshold = np.array([0.075, 0.075, 0.75])
def rollout_worker(graph_manager, num_workers, rollout_idx, task_parameters,
simtrace_video_s3_writers):
"""
wait for first checkpoint then perform rollouts using the model
"""
if not graph_manager.data_store:
raise AttributeError("None type for data_store object")
data_store = graph_manager.data_store
#TODO change agent to specific agent name for multip agent case
checkpoint_dir = os.path.join(task_parameters.checkpoint_restore_path,
"agent")
graph_manager.data_store.wait_for_checkpoints()
graph_manager.data_store.wait_for_trainer_ready()
# Make the clients that will allow us to pause and unpause the physics
rospy.wait_for_service('/gazebo/pause_physics_dr')
rospy.wait_for_service('/gazebo/unpause_physics_dr')
rospy.wait_for_service('/racecar/save_mp4/subscribe_to_save_mp4')
rospy.wait_for_service('/racecar/save_mp4/unsubscribe_from_save_mp4')
pause_physics = ServiceProxyWrapper('/gazebo/pause_physics_dr', Empty)
unpause_physics = ServiceProxyWrapper('/gazebo/unpause_physics_dr', Empty)
subscribe_to_save_mp4 = ServiceProxyWrapper(
'/racecar/save_mp4/subscribe_to_save_mp4', Empty)
unsubscribe_from_save_mp4 = ServiceProxyWrapper(
'/racecar/save_mp4/unsubscribe_from_save_mp4', Empty)
graph_manager.create_graph(task_parameters=task_parameters,
stop_physics=pause_physics,
start_physics=unpause_physics,
empty_service_call=EmptyRequest)
chkpt_state_reader = CheckpointStateReader(checkpoint_dir,
checkpoint_state_optional=False)
last_checkpoint = chkpt_state_reader.get_latest().num
# this worker should play a fraction of the total playing steps per rollout
episode_steps_per_rollout = graph_manager.agent_params.algorithm.num_consecutive_playing_steps.num_steps
act_steps = int(episode_steps_per_rollout / num_workers)
if rollout_idx < episode_steps_per_rollout % num_workers:
act_steps += 1
act_steps = EnvironmentEpisodes(act_steps)
configure_environment_randomizer()
for _ in range(
(graph_manager.improve_steps / act_steps.num_steps).num_steps):
# Collect profiler information only IS_PROFILER_ON is true
with utils.Profiler(s3_bucket=PROFILER_S3_BUCKET,
s3_prefix=PROFILER_S3_PREFIX,
output_local_path=ROLLOUT_WORKER_PROFILER_PATH,
enable_profiling=IS_PROFILER_ON):
graph_manager.phase = RunPhase.TRAIN
exit_if_trainer_done(checkpoint_dir, simtrace_video_s3_writers,
rollout_idx)
unpause_physics(EmptyRequest())
graph_manager.reset_internal_state(True)
graph_manager.act(act_steps,
wait_for_full_episodes=graph_manager.
agent_params.algorithm.act_for_full_episodes)
graph_manager.reset_internal_state(True)
time.sleep(1)
pause_physics(EmptyRequest())
graph_manager.phase = RunPhase.UNDEFINED
new_checkpoint = -1
if graph_manager.agent_params.algorithm.distributed_coach_synchronization_type\
== DistributedCoachSynchronizationType.SYNC:
unpause_physics(EmptyRequest())
is_save_mp4_enabled = rospy.get_param(
'MP4_S3_BUCKET', None) and rollout_idx == 0
if is_save_mp4_enabled:
subscribe_to_save_mp4(EmptyRequest())
if rollout_idx == 0:
for _ in range(MIN_EVAL_TRIALS):
graph_manager.evaluate(EnvironmentSteps(1))
while new_checkpoint < last_checkpoint + 1:
exit_if_trainer_done(checkpoint_dir,
simtrace_video_s3_writers,
rollout_idx)
if rollout_idx == 0:
graph_manager.evaluate(EnvironmentSteps(1))
new_checkpoint = data_store.get_coach_checkpoint_number(
'agent')
if is_save_mp4_enabled:
unsubscribe_from_save_mp4(EmptyRequest())
# upload simtrace and mp4 into s3 bucket
for s3_writer in simtrace_video_s3_writers:
s3_writer.persist(utils.get_s3_kms_extra_args())
pause_physics(EmptyRequest())
data_store.load_from_store(
expected_checkpoint_number=last_checkpoint + 1)
graph_manager.restore_checkpoint()
if graph_manager.agent_params.algorithm.distributed_coach_synchronization_type\
== DistributedCoachSynchronizationType.ASYNC:
if new_checkpoint > last_checkpoint:
graph_manager.restore_checkpoint()
last_checkpoint = new_checkpoint
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
def get_clipped_ppo_params(agent_params, agent, params):
"""This function is algorithm specific settings required for Clipped PPO algorithm
Args:
agent_params (DeepRacerClippedPPOAgentParams): the agent parameters that will be used to create the RL agent
agent (Agent): The agent object that was created either as part of create_rollout_agent or create_training_agent
params (dict): dictionary of hyperparameters
Returns:
DeepRacerClippedPPOAgentParams: updated agent params object with hyperparameters and other required details
"""
agent_params.network_wrappers['main'].learning_rate = params[
HyperParameterKeys.LEARNING_RATE.value]
agent_params.network_wrappers['main'].input_embedders_parameters = \
create_input_embedder(agent.network_settings['input_embedders'],
agent.network_settings['embedder_type'],
agent.network_settings['activation_function'])
agent_params.network_wrappers['main'].middleware_parameters = \
create_middle_embedder(agent.network_settings['middleware_embedders'],
agent.network_settings['embedder_type'],
agent.network_settings['activation_function'])
agent_params.network_wrappers['main'].batch_size = params[
HyperParameterKeys.BATCH_SIZE.value]
agent_params.network_wrappers['main'].optimizer_epsilon = 1e-5
agent_params.network_wrappers['main'].adam_optimizer_beta2 = 0.999
if params[HyperParameterKeys.LOSS_TYPE.value] == LossTypes.HUBER.value:
agent_params.network_wrappers[
'main'].replace_mse_with_huber_loss = True
agent_params.algorithm.clip_likelihood_ratio_using_epsilon = 0.2
agent_params.algorithm.beta_entropy = params[
HyperParameterKeys.BETA_ENTROPY.value]
agent_params.algorithm.gae_lambda = 0.95
agent_params.algorithm.discount = params[
HyperParameterKeys.DISCOUNT_FACTOR.value]
agent_params.algorithm.optimization_epochs = params[
HyperParameterKeys.NUM_EPOCHS.value]
agent_params.algorithm.estimate_state_value_using_gae = True
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = \
EnvironmentEpisodes(params[HyperParameterKeys.NUM_EPISODES_BETWEEN_TRAINING.value])
agent_params.algorithm.num_consecutive_playing_steps = \
EnvironmentEpisodes(params[HyperParameterKeys.NUM_EPISODES_BETWEEN_TRAINING.value])
agent_params.algorithm.distributed_coach_synchronization_type = \
DistributedCoachSynchronizationType.SYNC
if params[HyperParameterKeys.EXPLORATION_TYPE.value].lower().strip(
) == ExplorationTypes.CATEGORICAL.value:
agent_params.exploration = {
DiscreteActionSpace:
DeepRacerCategoricalParameters(
use_stochastic_evaluation_policy=False),
ScalableBoxActionSpace:
AdditiveNoiseParameters()
}
elif params[HyperParameterKeys.EXPLORATION_TYPE.value].lower().strip(
) == ExplorationTypes.E_GREEDY.value:
agent_params.exploration = {
DiscreteActionSpace: EGreedyParameters(),
ScalableBoxActionSpace: AdditiveNoiseParameters()
}
agent_params.exploration[DiscreteActionSpace].epsilon_schedule = \
LinearSchedule(1.0,
params[HyperParameterKeys.E_GREEDY_VALUE.value],
params[HyperParameterKeys.EPSILON_STEPS.value])
else:
log_and_exit(
"Unknown exploration_type found in hyper parameters. \
exploration_type: {}".format(params[
HyperParameterKeys.EXPLORATION_TYPE.value].lower().strip()),
SIMAPP_SIMULATION_WORKER_EXCEPTION, SIMAPP_EVENT_ERROR_CODE_500)
agent_params.memory = DeepRacerMemoryParameters()
return agent_params
from rl_coach.graph_managers.basic_rl_graph_manager import BasicRLGraphManager from rl_coach.graph_managers.graph_manager import ScheduleParameters from rl_coach.schedules import LinearSchedule from rl_coach.exploration_policies.e_greedy import EGreedyParameters from rl_coach.filters.filter import NoInputFilter, NoOutputFilter, InputFilter from rl_coach.filters.observation.observation_stacking_filter import ObservationStackingFilter from rl_coach.filters.observation.observation_to_uint8_filter import ObservationToUInt8Filter from rl_coach.memories.memory import MemoryGranularity from markov import environments #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps(100000) #Changing to 100K schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(40) schedule_params.evaluation_steps = EnvironmentEpisodes(5) schedule_params.heatup_steps = EnvironmentSteps(0) ######### # Agent # ######### agent_params = DQNAgentParameters() 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
def rollout_worker(graph_manager, checkpoint_dir, data_store, num_workers,
memory_backend_params):
"""
wait for first checkpoint then perform rollouts using the model
"""
wait_for_checkpoint(checkpoint_dir, data_store)
task_parameters = TaskParameters()
task_parameters.__dict__['checkpoint_restore_dir'] = checkpoint_dir
graph_manager.create_graph(task_parameters)
graph_manager.reset_internal_state()
for level in graph_manager.level_managers:
for agent in level.agents.values():
agent.memory.memory_backend = deepracer_memory.DeepRacerRolloutBackEnd(
memory_backend_params, graph_manager.agent_params.algorithm.
num_consecutive_playing_steps)
with graph_manager.phase_context(RunPhase.TRAIN):
last_checkpoint = 0
act_steps = math.ceil(
(graph_manager.agent_params.algorithm.
num_consecutive_playing_steps.num_steps) / num_workers)
for i in range(int(graph_manager.improve_steps.num_steps / act_steps)):
if should_stop(checkpoint_dir):
break
try:
# This will only work for DeepRacerRacetrackEnv enviroments
graph_manager.top_level_manager.environment.env.env.set_allow_servo_step_signals(
True)
except Exception as ex:
utils.json_format_logger(
"Method not defined in enviroment class: {}".format(ex),
**utils.build_system_error_dict(
utils.SIMAPP_SIMULATION_WORKER_EXCEPTION,
utils.SIMAPP_EVENT_ERROR_CODE_500))
if type(graph_manager.agent_params.algorithm.
num_consecutive_playing_steps) == EnvironmentSteps:
graph_manager.act(EnvironmentSteps(num_steps=act_steps),
wait_for_full_episodes=graph_manager.
agent_params.algorithm.act_for_full_episodes)
elif type(graph_manager.agent_params.algorithm.
num_consecutive_playing_steps) == EnvironmentEpisodes:
graph_manager.act(EnvironmentEpisodes(num_steps=act_steps))
try:
# This will only work for DeepRacerRacetrackEnv enviroments
graph_manager.top_level_manager.environment.env.env.set_allow_servo_step_signals(
False)
graph_manager.top_level_manager.environment.env.env.stop_car()
except Exception as ex:
utils.json_format_logger(
"Method not defined in enviroment class: {}".format(ex),
**utils.build_system_error_dict(
utils.SIMAPP_SIMULATION_WORKER_EXCEPTION,
utils.SIMAPP_EVENT_ERROR_CODE_500))
new_checkpoint = get_latest_checkpoint(checkpoint_dir)
if graph_manager.agent_params.algorithm.distributed_coach_synchronization_type == DistributedCoachSynchronizationType.SYNC:
while new_checkpoint is None or new_checkpoint <= last_checkpoint:
if should_stop(checkpoint_dir):
break
if data_store:
data_store.load_from_store(
expected_checkpoint_number=new_checkpoint)
new_checkpoint = get_latest_checkpoint(checkpoint_dir)
graph_manager.restore_checkpoint()
if graph_manager.agent_params.algorithm.distributed_coach_synchronization_type == DistributedCoachSynchronizationType.ASYNC:
if new_checkpoint is not None and new_checkpoint > last_checkpoint:
graph_manager.restore_checkpoint()
if new_checkpoint is not None:
last_checkpoint = new_checkpoint
from rl_coach.architectures.head_parameters import QHeadParameters
from rl_coach.agents.ddqn_bcq_agent import DDQNBCQAgentParameters
from rl_coach.agents.ddqn_bcq_agent import KNNParameters, NNImitationModelParameters
DATASET_SIZE = 200000
####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
# schedule_params.improve_steps = TrainingSteps(10000000000)
schedule_params.improve_steps = TrainingSteps(400) # 400 epochs
schedule_params.steps_between_evaluation_periods = TrainingSteps(1)
schedule_params.evaluation_steps = EnvironmentEpisodes(10)
schedule_params.heatup_steps = EnvironmentSteps(DATASET_SIZE)
#########
# Agent #
#########
agent_params = DDQNBCQAgentParameters()
agent_params.network_wrappers['main'].batch_size = 128
# TODO cross-DL framework abstraction for a constant initializer?
agent_params.network_wrappers['main'].heads_parameters = [
QHeadParameters(output_bias_initializer=tf.constant_initializer(-100))
]
agent_params.algorithm.num_steps_between_copying_online_weights_to_target = TrainingSteps(
100)
from rl_coach.base_parameters import VisualizationParameters, PresetValidationParameters from rl_coach.core_types import TrainingSteps, EnvironmentEpisodes, EnvironmentSteps, RunPhase from rl_coach.environments.environment import MaxDumpMethod, SelectedPhaseOnlyDumpMethod, SingleLevelSelection from rl_coach.environments.gym_environment import Mujoco, mujoco_v2, MujocoInputFilter from rl_coach.exploration_policies.continuous_entropy import ContinuousEntropyParameters from rl_coach.filters.observation.observation_normalization_filter import ObservationNormalizationFilter from rl_coach.graph_managers.basic_rl_graph_manager import BasicRLGraphManager from rl_coach.graph_managers.graph_manager import ScheduleParameters #################### # Graph Scheduling # #################### schedule_params = ScheduleParameters() schedule_params.improve_steps = TrainingSteps(10000000000) schedule_params.steps_between_evaluation_periods = EnvironmentSteps(2000) schedule_params.evaluation_steps = EnvironmentEpisodes(1) schedule_params.heatup_steps = EnvironmentSteps(0) ######### # Agent # ######### agent_params = PPOAgentParameters() agent_params.network_wrappers['actor'].learning_rate = 0.001 agent_params.network_wrappers['critic'].learning_rate = 0.001 agent_params.network_wrappers['actor'].input_embedders_parameters['observation'].scheme = [Dense([64])] agent_params.network_wrappers['actor'].middleware_parameters.scheme = [Dense([64])] agent_params.network_wrappers['critic'].input_embedders_parameters['observation'].scheme = [Dense([64])] agent_params.network_wrappers['critic'].middleware_parameters.scheme = [Dense([64])] agent_params.input_filter = MujocoInputFilter()
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
vis_params.dump_csv = True
vis_params.print_networks_summary = True
vis_params.dump_parameters_documentation = False
vis_params.dump_signals_to_csv_every_x_episodes = 1
vis_params.tensorboard = True
########
# 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.agents.td3_agent import TD3AgentParameters
from rl_coach.architectures.layers import Dense
from rl_coach.base_parameters import VisualizationParameters, PresetValidationParameters, EmbedderScheme
from rl_coach.core_types import EnvironmentEpisodes, EnvironmentSteps
from rl_coach.environments.environment import SingleLevelSelection
from rl_coach.environments.gym_environment import GymVectorEnvironment, mujoco_v2
from rl_coach.graph_managers.basic_rl_graph_manager import BasicRLGraphManager
from rl_coach.graph_managers.graph_manager import ScheduleParameters
from rl_coach.exploration_policies.truncated_normal import TruncatedNormalParameters
from rl_coach.exploration_policies.additive_noise import AdditiveNoiseParameters
####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = EnvironmentEpisodes(800)
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(5000)
schedule_params.evaluation_steps = EnvironmentEpisodes(0) # Neta: 0
schedule_params.heatup_steps = EnvironmentEpisodes(100)
#########
# Agent #
#########
agent_params = TD3AgentParameters()
agent_params.network_wrappers['actor'].input_embedders_parameters[
'observation'].scheme = [Dense(400)]
agent_params.network_wrappers['actor'].middleware_parameters.scheme = [
Dense(300)
]
agent_params.network_wrappers['critic'].input_embedders_parameters[
def rollout_worker(graph_manager, num_workers, rollout_idx, task_parameters,
simtrace_video_s3_writers, pause_physics, unpause_physics):
"""
wait for first checkpoint then perform rollouts using the model
"""
if not graph_manager.data_store:
raise AttributeError("None type for data_store object")
is_sageonly = check_is_sageonly()
data_store = graph_manager.data_store
#TODO change agent to specific agent name for multip agent case
checkpoint_dir = os.path.join(task_parameters.checkpoint_restore_path,
"agent")
graph_manager.data_store.wait_for_checkpoints()
graph_manager.data_store.wait_for_trainer_ready()
# wait for the required cancel services to become available
# Do this only for Robomaker job.
# if not is_sageonly:
# rospy.wait_for_service('/robomaker/job/cancel') # Make the clients that will allow us to pause and unpause the physics
rospy.wait_for_service('/racecar/save_mp4/subscribe_to_save_mp4')
rospy.wait_for_service('/racecar/save_mp4/unsubscribe_from_save_mp4')
subscribe_to_save_mp4 = ServiceProxyWrapper(
'/racecar/save_mp4/subscribe_to_save_mp4', Empty)
unsubscribe_from_save_mp4 = ServiceProxyWrapper(
'/racecar/save_mp4/unsubscribe_from_save_mp4', Empty)
graph_manager.create_graph(task_parameters=task_parameters,
stop_physics=pause_physics,
start_physics=unpause_physics,
empty_service_call=EmptyRequest)
chkpt_state_reader = CheckpointStateReader(checkpoint_dir,
checkpoint_state_optional=False)
last_checkpoint = chkpt_state_reader.get_latest().num
# this worker should play a fraction of the total playing steps per rollout
episode_steps_per_rollout = graph_manager.agent_params.algorithm.num_consecutive_playing_steps.num_steps
act_steps = int(episode_steps_per_rollout / num_workers)
if rollout_idx < episode_steps_per_rollout % num_workers:
act_steps += 1
act_steps = EnvironmentEpisodes(act_steps)
configure_environment_randomizer()
for _ in range(
(graph_manager.improve_steps / act_steps.num_steps).num_steps):
# Collect profiler information only IS_PROFILER_ON is true
with utils.Profiler(s3_bucket=PROFILER_S3_BUCKET,
s3_prefix=PROFILER_S3_PREFIX,
output_local_path=ROLLOUT_WORKER_PROFILER_PATH,
enable_profiling=IS_PROFILER_ON):
graph_manager.phase = RunPhase.TRAIN
exit_if_trainer_done(checkpoint_dir, simtrace_video_s3_writers,
rollout_idx)
unpause_physics(EmptyRequest())
graph_manager.reset_internal_state(True)
graph_manager.act(act_steps,
wait_for_full_episodes=graph_manager.
agent_params.algorithm.act_for_full_episodes)
graph_manager.reset_internal_state(True)
time.sleep(1)
pause_physics(EmptyRequest())
graph_manager.phase = RunPhase.UNDEFINED
new_checkpoint = -1
if graph_manager.agent_params.algorithm.distributed_coach_synchronization_type\
== DistributedCoachSynchronizationType.SYNC:
unpause_physics(EmptyRequest())
is_save_mp4_enabled = rospy.get_param(
'MP4_S3_BUCKET', None) and rollout_idx == 0
if is_save_mp4_enabled:
subscribe_to_save_mp4(EmptyRequest())
if rollout_idx == 0:
for _ in range(int(rospy.get_param('MIN_EVAL_TRIALS',
'5'))):
graph_manager.evaluate(EnvironmentSteps(1))
# For sageonly job for better performance only run limited number of evaluations.
# Pausing the physics makes its performance same as RoboMaker + SageMaker
if is_sageonly:
if is_save_mp4_enabled:
unsubscribe_from_save_mp4(EmptyRequest())
# upload simtrace and mp4 into s3 bucket
for s3_writer in simtrace_video_s3_writers:
s3_writer.persist(utils.get_s3_kms_extra_args())
graph_manager.phase = RunPhase.WAITING
pause_physics(EmptyRequest())
while new_checkpoint < last_checkpoint + 1:
exit_if_trainer_done(checkpoint_dir,
simtrace_video_s3_writers,
rollout_idx)
# Continously run the evaluation only for SageMaker + RoboMaker job
if not is_sageonly and rollout_idx == 0:
print(
"Additional evaluation. New Checkpoint: {}, Last Checkpoint: {}"
.format(new_checkpoint, last_checkpoint))
graph_manager.evaluate(EnvironmentSteps(1))
else:
time.sleep(5)
new_checkpoint = data_store.get_coach_checkpoint_number(
'agent')
# Save the mp4 for Robo+Sage jobs
if not is_sageonly:
if is_save_mp4_enabled:
unsubscribe_from_save_mp4(EmptyRequest())
# upload simtrace and mp4 into s3 bucket
for s3_writer in simtrace_video_s3_writers:
s3_writer.persist(utils.get_s3_kms_extra_args())
pause_physics(EmptyRequest())
data_store.load_from_store(
expected_checkpoint_number=last_checkpoint + 1)
graph_manager.restore_checkpoint()
if graph_manager.agent_params.algorithm.distributed_coach_synchronization_type\
== DistributedCoachSynchronizationType.ASYNC:
if new_checkpoint > last_checkpoint:
graph_manager.restore_checkpoint()
last_checkpoint = new_checkpoint
logger.info("Exited main loop. Done.")
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 ScheduleParameters
from rl_coach.memories.episodic.episodic_hindsight_experience_replay import \
EpisodicHindsightExperienceReplayParameters, HindsightGoalSelectionMethod
from rl_coach.memories.memory import MemoryGranularity
from rl_coach.schedules import ConstantSchedule
from rl_coach.spaces import GoalsSpace, ReachingGoal
bit_length = 20
####################
# Graph Scheduling #
####################
schedule_params = ScheduleParameters()
schedule_params.improve_steps = EnvironmentEpisodes(16 * 50 * 200) # 200 epochs
schedule_params.steps_between_evaluation_periods = EnvironmentEpisodes(16 * 50) # 50 cycles
schedule_params.evaluation_steps = EnvironmentEpisodes(10)
schedule_params.heatup_steps = EnvironmentSteps(0)
#########
# Agent #
#########
agent_params = DQNAgentParameters()
agent_params.network_wrappers['main'].learning_rate = 0.001
agent_params.network_wrappers['main'].batch_size = 128
agent_params.network_wrappers['main'].middleware_parameters.scheme = [Dense(256)]
agent_params.network_wrappers['main'].input_embedders_parameters = {
'state': InputEmbedderParameters(scheme=EmbedderScheme.Empty),
'desired_goal': InputEmbedderParameters(scheme=EmbedderScheme.Empty)}
agent_params.algorithm.discount = 0.98
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
