def construct_memory_params(json: dict):
if json['store_type'] == 'redispubsub':
memory_params = RedisPubSubMemoryBackendParameters(
json['redis_address'], json['redis_port'], channel=json.get('channel', ''), run_type=json['run_type']
)
return memory_params
def handle_distributed_coach_orchestrator(graph_manager, args):
ckpt_inside_container = "/checkpoint"
rollout_command = ['python3', 'rl_coach/coach.py', '--distributed_coach_run_type', str(RunType.ROLLOUT_WORKER)] + sys.argv[1:]
trainer_command = ['python3', 'rl_coach/coach.py', '--distributed_coach_run_type', str(RunType.TRAINER)] + sys.argv[1:]
if '--experiment_name' not in rollout_command:
rollout_command = rollout_command + ['--experiment_name', args.experiment_name]
if '--experiment_name' not in trainer_command:
trainer_command = trainer_command + ['--experiment_name', args.experiment_name]
memory_backend_params = None
if args.memory_backend == "redispubsub":
memory_backend_params = RedisPubSubMemoryBackendParameters()
ds_params_instance = None
if args.data_store == "s3":
ds_params = DataStoreParameters("s3", "", "")
ds_params_instance = S3DataStoreParameters(ds_params=ds_params, end_point=args.s3_end_point, bucket_name=args.s3_bucket_name,
creds_file=args.s3_creds_file, checkpoint_dir=ckpt_inside_container)
elif args.data_store == "nfs":
ds_params = DataStoreParameters("nfs", "kubernetes", "")
ds_params_instance = NFSDataStoreParameters(ds_params)
worker_run_type_params = RunTypeParameters(args.image, rollout_command, run_type=str(RunType.ROLLOUT_WORKER), num_replicas=args.num_workers)
trainer_run_type_params = RunTypeParameters(args.image, trainer_command, run_type=str(RunType.TRAINER))
orchestration_params = KubernetesParameters([worker_run_type_params, trainer_run_type_params],
kubeconfig='~/.kube/config',
memory_backend_parameters=memory_backend_params,
data_store_params=ds_params_instance)
orchestrator = Kubernetes(orchestration_params)
if not orchestrator.setup():
print("Could not setup.")
return
if orchestrator.deploy_trainer():
print("Successfully deployed trainer.")
else:
print("Could not deploy trainer.")
return
if orchestrator.deploy_worker():
print("Successfully deployed rollout worker(s).")
else:
print("Could not deploy rollout worker(s).")
return
try:
orchestrator.trainer_logs()
except KeyboardInterrupt:
pass
orchestrator.undeploy()
def main():
screen.set_use_colors(False)
parser = argparse.ArgumentParser()
parser.add_argument('-pk',
'--preset_s3_key',
help="(string) Name of a preset to download from S3",
type=str,
required=False)
parser.add_argument(
'-ek',
'--environment_s3_key',
help="(string) Name of an environment file to download from S3",
type=str,
required=False)
parser.add_argument('--model_metadata_s3_key',
help="(string) Model Metadata File S3 Key",
type=str,
required=False)
parser.add_argument(
'-c',
'--checkpoint-dir',
help=
'(string) Path to a folder containing a checkpoint to write the model to.',
type=str,
default='./checkpoint')
parser.add_argument(
'--pretrained-checkpoint-dir',
help='(string) Path to a folder for downloading a pre-trained model',
type=str,
default=PRETRAINED_MODEL_DIR)
parser.add_argument('--s3_bucket',
help='(string) S3 bucket',
type=str,
default=os.environ.get(
"SAGEMAKER_SHARED_S3_BUCKET_PATH", "gsaur-test"))
parser.add_argument('--s3_prefix',
help='(string) S3 prefix',
type=str,
default='sagemaker')
parser.add_argument('--framework',
help='(string) tensorflow or mxnet',
type=str,
default='tensorflow')
parser.add_argument('--pretrained_s3_bucket',
help='(string) S3 bucket for pre-trained model',
type=str)
parser.add_argument('--pretrained_s3_prefix',
help='(string) S3 prefix for pre-trained model',
type=str,
default='sagemaker')
parser.add_argument('--aws_region',
help='(string) AWS region',
type=str,
default=os.environ.get("AWS_REGION", "us-east-1"))
args, unknown = parser.parse_known_args()
s3_client = SageS3Client(bucket=args.s3_bucket,
s3_prefix=args.s3_prefix,
aws_region=args.aws_region)
# Load the model metadata
model_metadata_local_path = os.path.join(CUSTOM_FILES_PATH,
'model_metadata.json')
load_model_metadata(s3_client, args.model_metadata_s3_key,
model_metadata_local_path)
s3_client.upload_file(
os.path.normpath("%s/model/model_metadata.json" % args.s3_prefix),
model_metadata_local_path)
shutil.copy2(model_metadata_local_path, SM_MODEL_OUTPUT_DIR)
success_custom_environment = False
if args.environment_s3_key:
environment_local_path = "./markov/environments/deepracer_racetrack_env.py"
success_custom_environment = s3_client.download_file(
s3_key=args.environment_s3_key, local_path=environment_local_path)
if not success_custom_environment:
print(
"Could not download the environment file. Using the default DeepRacer environment."
)
else:
success_custom_environment = s3_client.upload_file(
s3_key=os.path.normpath(
"%s/environments/deepracer_racetrack_env.py" %
args.s3_prefix),
local_path=environment_local_path)
if success_custom_environment:
print("Using environment: %s" % args.environment_s3_key)
# Import to register the environment with Gym
import markov.environments
success_custom_preset = False
if args.preset_s3_key:
preset_local_path = "./markov/presets/preset.py"
success_custom_preset = s3_client.download_file(
s3_key=args.preset_s3_key, local_path=preset_local_path)
if not success_custom_preset:
print(
"Could not download the preset file. Using the default DeepRacer preset."
)
else:
preset_location = "markov.presets.preset:graph_manager"
graph_manager = short_dynamic_import(preset_location,
ignore_module_case=True)
success_custom_preset = s3_client.upload_file(
s3_key=os.path.normpath("%s/presets/preset.py" %
args.s3_prefix),
local_path=preset_local_path)
if success_custom_preset:
print("Using preset: %s" % args.preset_s3_key)
if not success_custom_preset:
from markov.sagemaker_graph_manager import get_graph_manager
params_blob = os.environ.get('SM_TRAINING_ENV', '')
if params_blob:
params = json.loads(params_blob)
sm_hyperparams_dict = params["hyperparameters"]
else:
sm_hyperparams_dict = {}
graph_manager, robomaker_hyperparams_json = get_graph_manager(
**sm_hyperparams_dict)
s3_client.upload_hyperparameters(robomaker_hyperparams_json)
print("Uploaded hyperparameters.json to S3")
host_ip_address = get_ip_from_host()
s3_client.write_ip_config(host_ip_address)
print("Uploaded IP address information to S3: %s" % host_ip_address)
use_pretrained_model = False
if args.pretrained_s3_bucket and args.pretrained_s3_prefix:
s3_client_pretrained = SageS3Client(
bucket=args.pretrained_s3_bucket,
s3_prefix=args.pretrained_s3_prefix,
aws_region=args.aws_region)
s3_client_pretrained.download_model(args.pretrained_checkpoint_dir)
use_pretrained_model = True
memory_backend_params = RedisPubSubMemoryBackendParameters(
redis_address="localhost",
redis_port=6379,
run_type='trainer',
channel=args.s3_prefix)
graph_manager.agent_params.memory.register_var('memory_backend_params',
memory_backend_params)
ds_params_instance = S3BotoDataStoreParameters(
bucket_name=args.s3_bucket,
checkpoint_dir=args.checkpoint_dir,
aws_region=args.aws_region,
s3_folder=args.s3_prefix)
graph_manager.data_store_params = ds_params_instance
data_store = S3BotoDataStore(ds_params_instance)
data_store.graph_manager = graph_manager
graph_manager.data_store = data_store
training_worker(graph_manager=graph_manager,
checkpoint_dir=args.checkpoint_dir,
use_pretrained_model=use_pretrained_model,
framework=args.framework)
def handle_distributed_coach_orchestrator(args):
from rl_coach.orchestrators.kubernetes_orchestrator import KubernetesParameters, Kubernetes, \
RunTypeParameters
ckpt_inside_container = "/checkpoint"
arg_list = sys.argv[1:]
try:
i = arg_list.index('--distributed_coach_run_type')
arg_list.pop(i)
arg_list.pop(i)
except ValueError:
pass
trainer_command = [
'python3', 'rl_coach/coach.py', '--distributed_coach_run_type',
str(RunType.TRAINER)
] + arg_list
rollout_command = [
'python3', 'rl_coach/coach.py', '--distributed_coach_run_type',
str(RunType.ROLLOUT_WORKER)
] + arg_list
if '--experiment_name' not in rollout_command:
rollout_command = rollout_command + [
'--experiment_name', args.experiment_name
]
if '--experiment_name' not in trainer_command:
trainer_command = trainer_command + [
'--experiment_name', args.experiment_name
]
memory_backend_params = None
if args.memory_backend == "redispubsub":
memory_backend_params = RedisPubSubMemoryBackendParameters()
ds_params_instance = None
if args.data_store == "s3":
ds_params = DataStoreParameters("s3", "", "")
ds_params_instance = S3DataStoreParameters(
ds_params=ds_params,
end_point=args.s3_end_point,
bucket_name=args.s3_bucket_name,
creds_file=args.s3_creds_file,
checkpoint_dir=ckpt_inside_container,
expt_dir=args.experiment_path)
elif args.data_store == "nfs":
ds_params = DataStoreParameters("nfs", "kubernetes", "")
ds_params_instance = NFSDataStoreParameters(ds_params)
worker_run_type_params = RunTypeParameters(args.image,
rollout_command,
run_type=str(
RunType.ROLLOUT_WORKER),
num_replicas=args.num_workers)
trainer_run_type_params = RunTypeParameters(args.image,
trainer_command,
run_type=str(RunType.TRAINER))
orchestration_params = KubernetesParameters(
[worker_run_type_params, trainer_run_type_params],
kubeconfig='~/.kube/config',
memory_backend_parameters=memory_backend_params,
data_store_params=ds_params_instance)
orchestrator = Kubernetes(orchestration_params)
if not orchestrator.setup():
print("Could not setup.")
return 1
if orchestrator.deploy_trainer():
print("Successfully deployed trainer.")
else:
print("Could not deploy trainer.")
return 1
if orchestrator.deploy_worker():
print("Successfully deployed rollout worker(s).")
else:
print("Could not deploy rollout worker(s).")
return 1
if args.dump_worker_logs:
screen.log_title("Dumping rollout worker logs in: {}".format(
args.experiment_path))
orchestrator.worker_logs(path=args.experiment_path)
exit_code = 1
try:
exit_code = orchestrator.trainer_logs()
except KeyboardInterrupt:
pass
orchestrator.undeploy()
return exit_code
def main():
screen.set_use_colors(False)
parser = argparse.ArgumentParser()
parser.add_argument('-pk', '--preset_s3_key',
help="(string) Name of a preset to download from S3",
type=str,
required=False)
parser.add_argument('-ek', '--environment_s3_key',
help="(string) Name of an environment file to download from S3",
type=str,
required=False)
parser.add_argument('--model_metadata_s3_key',
help="(string) Model Metadata File S3 Key",
type=str,
required=False)
parser.add_argument('-c', '--checkpoint-dir',
help='(string) Path to a folder containing a checkpoint to write the model to.',
type=str,
default='./checkpoint')
parser.add_argument('--pretrained-checkpoint-dir',
help='(string) Path to a folder for downloading a pre-trained model',
type=str,
default=PRETRAINED_MODEL_DIR)
parser.add_argument('--s3_bucket',
help='(string) S3 bucket',
type=str,
default=os.environ.get("SAGEMAKER_SHARED_S3_BUCKET_PATH", "gsaur-test"))
parser.add_argument('--s3_prefix',
help='(string) S3 prefix',
type=str,
default='sagemaker')
parser.add_argument('--framework',
help='(string) tensorflow or mxnet',
type=str,
default='tensorflow')
parser.add_argument('--pretrained_s3_bucket',
help='(string) S3 bucket for pre-trained model',
type=str)
parser.add_argument('--pretrained_s3_prefix',
help='(string) S3 prefix for pre-trained model',
type=str,
default='sagemaker')
parser.add_argument('--aws_region',
help='(string) AWS region',
type=str,
default=os.environ.get("AWS_REGION", "us-east-1"))
start_redis_server()
args, _ = parser.parse_known_args()
s3_client = SageS3Client(bucket=args.s3_bucket, s3_prefix=args.s3_prefix, aws_region=args.aws_region)
# Load the model metadata
model_metadata_local_path = os.path.join(CUSTOM_FILES_PATH, 'model_metadata.json')
utils.load_model_metadata(s3_client, args.model_metadata_s3_key, model_metadata_local_path)
s3_client.upload_file(os.path.normpath("%s/model/model_metadata.json" % args.s3_prefix), model_metadata_local_path)
shutil.copy2(model_metadata_local_path, SM_MODEL_OUTPUT_DIR)
success_custom_preset = False
if args.preset_s3_key:
preset_local_path = "./markov/presets/preset.py"
success_custom_preset = s3_client.download_file(s3_key=args.preset_s3_key, local_path=preset_local_path)
if not success_custom_preset:
logger.info("Could not download the preset file. Using the default DeepRacer preset.")
else:
preset_location = "markov.presets.preset:graph_manager"
graph_manager = short_dynamic_import(preset_location, ignore_module_case=True)
success_custom_preset = s3_client.upload_file(
s3_key=os.path.normpath("%s/presets/preset.py" % args.s3_prefix), local_path=preset_local_path)
if success_custom_preset:
logger.info("Using preset: %s" % args.preset_s3_key)
if not success_custom_preset:
params_blob = os.environ.get('SM_TRAINING_ENV', '')
if params_blob:
params = json.loads(params_blob)
sm_hyperparams_dict = params["hyperparameters"]
else:
sm_hyperparams_dict = {}
#! TODO each agent should have own config
agent_config = {'model_metadata': model_metadata_local_path,
'car_ctrl_cnfig': {ConfigParams.LINK_NAME_LIST.value: [],
ConfigParams.VELOCITY_LIST.value : {},
ConfigParams.STEERING_LIST.value : {},
ConfigParams.CHANGE_START.value : None,
ConfigParams.ALT_DIR.value : None,
ConfigParams.ACTION_SPACE_PATH.value : 'custom_files/model_metadata.json',
ConfigParams.REWARD.value : None,
ConfigParams.AGENT_NAME.value : 'racecar'}}
agent_list = list()
agent_list.append(create_training_agent(agent_config))
#agent_list.append(create_training_agent(agent_config))
graph_manager, robomaker_hyperparams_json = get_graph_manager(sm_hyperparams_dict, agent_list)
s3_client.upload_hyperparameters(robomaker_hyperparams_json)
logger.info("Uploaded hyperparameters.json to S3")
host_ip_address = utils.get_ip_from_host()
s3_client.write_ip_config(host_ip_address)
logger.info("Uploaded IP address information to S3: %s" % host_ip_address)
use_pretrained_model = args.pretrained_s3_bucket and args.pretrained_s3_prefix
if use_pretrained_model:
# Handle backward compatibility
_, _, version = parse_model_metadata(model_metadata_local_path)
if float(version) < float(utils.SIMAPP_VERSION) and \
not utils.has_current_ckpnt_name(args.pretrained_s3_bucket, args.pretrained_s3_prefix, args.aws_region):
utils.make_compatible(args.pretrained_s3_bucket, args.pretrained_s3_prefix,
args.aws_region, SyncFiles.TRAINER_READY.value)
ds_params_instance_pretrained = S3BotoDataStoreParameters(aws_region=args.aws_region,
bucket_name=args.pretrained_s3_bucket,
checkpoint_dir=args.pretrained_checkpoint_dir,
s3_folder=args.pretrained_s3_prefix)
data_store_pretrained = S3BotoDataStore(ds_params_instance_pretrained)
data_store_pretrained.load_from_store()
memory_backend_params = RedisPubSubMemoryBackendParameters(redis_address="localhost",
redis_port=6379,
run_type=str(RunType.TRAINER),
channel=args.s3_prefix)
graph_manager.memory_backend_params = memory_backend_params
ds_params_instance = S3BotoDataStoreParameters(aws_region=args.aws_region,
bucket_name=args.s3_bucket,
checkpoint_dir=args.checkpoint_dir,
s3_folder=args.s3_prefix)
graph_manager.data_store_params = ds_params_instance
data_store = S3BotoDataStore(ds_params_instance)
data_store.graph_manager = graph_manager
graph_manager.data_store = data_store
task_parameters = TaskParameters()
task_parameters.experiment_path = SM_MODEL_OUTPUT_DIR
task_parameters.checkpoint_save_secs = 20
if use_pretrained_model:
task_parameters.checkpoint_restore_path = args.pretrained_checkpoint_dir
task_parameters.checkpoint_save_dir = args.checkpoint_dir
training_worker(
graph_manager=graph_manager,
task_parameters=task_parameters,
user_batch_size=json.loads(robomaker_hyperparams_json)["batch_size"],
user_episode_per_rollout=json.loads(robomaker_hyperparams_json)["num_episodes_between_training"]
)
def main():
screen.set_use_colors(False)
try:
parser = argparse.ArgumentParser()
parser.add_argument(
'-pk',
'--preset_s3_key',
help="(string) Name of a preset to download from S3",
type=str,
required=False)
parser.add_argument(
'-ek',
'--environment_s3_key',
help="(string) Name of an environment file to download from S3",
type=str,
required=False)
parser.add_argument('--model_metadata_s3_key',
help="(string) Model Metadata File S3 Key",
type=str,
required=False)
parser.add_argument(
'-c',
'--checkpoint-dir',
help=
'(string) Path to a folder containing a checkpoint to write the model to.',
type=str,
default='./checkpoint')
parser.add_argument(
'--pretrained-checkpoint-dir',
help=
'(string) Path to a folder for downloading a pre-trained model',
type=str,
default=PRETRAINED_MODEL_DIR)
parser.add_argument('--s3_bucket',
help='(string) S3 bucket',
type=str,
default=os.environ.get(
"SAGEMAKER_SHARED_S3_BUCKET_PATH",
"gsaur-test"))
parser.add_argument('--s3_prefix',
help='(string) S3 prefix',
type=str,
default='sagemaker')
parser.add_argument('--framework',
help='(string) tensorflow or mxnet',
type=str,
default='tensorflow')
parser.add_argument('--pretrained_s3_bucket',
help='(string) S3 bucket for pre-trained model',
type=str)
parser.add_argument('--pretrained_s3_prefix',
help='(string) S3 prefix for pre-trained model',
type=str,
default='sagemaker')
parser.add_argument('--aws_region',
help='(string) AWS region',
type=str,
default=os.environ.get("AWS_REGION", "us-east-1"))
args, unknown = parser.parse_known_args()
start_redis_server()
s3_client = SageS3Client(bucket=args.s3_bucket,
s3_prefix=args.s3_prefix,
aws_region=args.aws_region)
# Load the model metadata
model_metadata_local_path = os.path.join(CUSTOM_FILES_PATH,
'model_metadata.json')
load_model_metadata(s3_client, args.model_metadata_s3_key,
model_metadata_local_path)
s3_client.upload_file(
os.path.normpath("%s/model/model_metadata.json" % args.s3_prefix),
model_metadata_local_path)
shutil.copy2(model_metadata_local_path, SM_MODEL_OUTPUT_DIR)
# Register the gym enviroment, this will give clients the ability to creat the enviroment object
register(id=defaults.ENV_ID,
entry_point=defaults.ENTRY_POINT,
max_episode_steps=defaults.MAX_STEPS,
reward_threshold=defaults.THRESHOLD)
user_batch_size, user_episode_per_rollout = None, None
success_custom_preset = False
if args.preset_s3_key:
preset_local_path = "./markov/presets/preset.py"
success_custom_preset = s3_client.download_file(
s3_key=args.preset_s3_key, local_path=preset_local_path)
if not success_custom_preset:
logger.info(
"Could not download the preset file. Using the default DeepRacer preset."
)
else:
preset_location = "markov.presets.preset:graph_manager"
graph_manager = short_dynamic_import(preset_location,
ignore_module_case=True)
success_custom_preset = s3_client.upload_file(
s3_key=os.path.normpath("%s/presets/preset.py" %
args.s3_prefix),
local_path=preset_local_path)
if success_custom_preset:
agent_param_loc = "markov.presets.preset:agent_params"
agent_params = short_dynamic_import(
agent_param_loc, ignore_module_case=True)
user_batch_size = agent_params.network_wrappers[
'main'].batch_size
user_episode_per_rollout = agent_params.algorithm.num_consecutive_playing_steps.num_steps
logger.info("Using preset: %s" % args.preset_s3_key)
if not success_custom_preset:
from markov.sagemaker_graph_manager import get_graph_manager
user_batch_size = json.loads(
robomaker_hyperparams_json)["batch_size"],
user_episode_per_rollout = json.loads(
robomaker_hyperparams_json)["num_episodes_between_training"]
params_blob = os.environ.get('SM_TRAINING_ENV', '')
if params_blob:
params = json.loads(params_blob)
sm_hyperparams_dict = params["hyperparameters"]
else:
sm_hyperparams_dict = {}
graph_manager, robomaker_hyperparams_json = get_graph_manager(
**sm_hyperparams_dict)
s3_client.upload_hyperparameters(robomaker_hyperparams_json)
logger.info("Uploaded hyperparameters.json to S3")
host_ip_address = get_ip_from_host()
s3_client.write_ip_config(host_ip_address)
logger.info("Uploaded IP address information to S3: %s" %
host_ip_address)
use_pretrained_model = args.pretrained_s3_bucket and args.pretrained_s3_prefix
if use_pretrained_model:
s3_client_pretrained = SageS3Client(
bucket=args.pretrained_s3_bucket,
s3_prefix=args.pretrained_s3_prefix,
aws_region=args.aws_region)
s3_client_pretrained.download_model(args.pretrained_checkpoint_dir)
memory_backend_params = RedisPubSubMemoryBackendParameters(
redis_address="localhost",
redis_port=6379,
run_type='trainer',
channel=args.s3_prefix)
ds_params_instance = S3BotoDataStoreParameters(
bucket_name=args.s3_bucket,
checkpoint_dir=args.checkpoint_dir,
aws_region=args.aws_region,
s3_folder=args.s3_prefix)
graph_manager.data_store_params = ds_params_instance
data_store = S3BotoDataStore(ds_params_instance)
data_store.graph_manager = graph_manager
graph_manager.data_store = data_store
training_worker(graph_manager=graph_manager,
checkpoint_dir=args.checkpoint_dir,
use_pretrained_model=use_pretrained_model,
framework=args.framework,
memory_backend_params=memory_backend_params,
user_batch_size=user_batch_size,
user_episode_per_rollout=user_episode_per_rollout)
except Exception as ex:
utils.json_format_logger(
"Training worker exited with exception: {}".format(ex),
**utils.build_system_error_dict(
utils.SIMAPP_TRAINING_WORKER_EXCEPTION,
utils.SIMAPP_EVENT_ERROR_CODE_500))
utils.simapp_exit_gracefully()
def main():
screen.set_use_colors(False)
parser = argparse.ArgumentParser()
parser.add_argument(
'-c',
'--checkpoint_dir',
help=
'(string) Path to a folder containing a checkpoint to restore the model from.',
type=str,
default='./checkpoint')
parser.add_argument('--s3_bucket',
help='(string) S3 bucket',
type=str,
default=os.environ.get("SAGEMAKER_SHARED_S3_BUCKET",
"gsaur-test"))
parser.add_argument('--s3_prefix',
help='(string) S3 prefix',
type=str,
default=os.environ.get("SAGEMAKER_SHARED_S3_PREFIX",
"sagemaker"))
parser.add_argument(
'--num-workers',
help="(int) The number of workers started in this pool",
type=int,
default=1)
parser.add_argument('-r',
'--redis_ip',
help="(string) IP or host for the redis server",
default='localhost',
type=str)
parser.add_argument('-rp',
'--redis_port',
help="(int) Port of the redis server",
default=6379,
type=int)
parser.add_argument('--aws_region',
help='(string) AWS region',
type=str,
default=os.environ.get("APP_REGION", "us-east-1"))
parser.add_argument('--reward_file_s3_key',
help='(string) Reward File S3 Key',
type=str,
default=os.environ.get("REWARD_FILE_S3_KEY", None))
parser.add_argument('--model_metadata_s3_key',
help='(string) Model Metadata File S3 Key',
type=str,
default=os.environ.get("MODEL_METADATA_FILE_S3_KEY",
None))
args = parser.parse_args()
s3_client = SageS3Client(bucket=args.s3_bucket,
s3_prefix=args.s3_prefix,
aws_region=args.aws_region)
print("S3 bucket: %s" % args.s3_bucket)
print("S3 prefix: %s" % args.s3_prefix)
# Load the model metadata
model_metadata_local_path = os.path.join(CUSTOM_FILES_PATH,
'model_metadata.json')
load_model_metadata(s3_client, args.model_metadata_s3_key,
model_metadata_local_path)
redis_ip = s3_client.get_ip()
print("Received IP from SageMaker successfully: %s" % redis_ip)
# Download hyperparameters from SageMaker
hyperparameters_file_success = False
hyperparams_s3_key = os.path.normpath(args.s3_prefix +
"/ip/hyperparameters.json")
hyperparameters_file_success = s3_client.download_file(
s3_key=hyperparams_s3_key, local_path="hyperparameters.json")
sm_hyperparams_dict = {}
if hyperparameters_file_success:
print("Received Sagemaker hyperparameters successfully!")
with open("hyperparameters.json") as fp:
sm_hyperparams_dict = json.load(fp)
else:
print("SageMaker hyperparameters not found.")
preset_file_success = False
environment_file_success = False
preset_file_success, environment_file_success = download_custom_files_if_present(
s3_client, args.s3_prefix)
if not environment_file_success:
# Download reward function if environment file is not downloaded
if not args.reward_file_s3_key:
raise ValueError("Customer reward S3 key not supplied!")
download_customer_reward_function(s3_client, args.reward_file_s3_key)
import markov.environments
print("Using default environment!")
else:
register_custom_environments()
print("Using custom environment!")
if preset_file_success:
preset_location = os.path.join(CUSTOM_FILES_PATH, "preset.py")
preset_location += ":graph_manager"
graph_manager = short_dynamic_import(preset_location,
ignore_module_case=True)
print("Using custom preset file!")
else:
from markov.sagemaker_graph_manager import get_graph_manager
graph_manager, _ = get_graph_manager(**sm_hyperparams_dict)
memory_backend_params = RedisPubSubMemoryBackendParameters(
redis_address=redis_ip,
redis_port=6379,
run_type='worker',
channel=args.s3_prefix)
graph_manager.agent_params.memory.register_var('memory_backend_params',
memory_backend_params)
ds_params_instance = S3BotoDataStoreParameters(
bucket_name=args.s3_bucket,
checkpoint_dir=args.checkpoint_dir,
aws_region=args.aws_region,
s3_folder=args.s3_prefix)
data_store = S3BotoDataStore(ds_params_instance)
data_store.graph_manager = graph_manager
graph_manager.data_store = data_store
rollout_worker(
graph_manager=graph_manager,
checkpoint_dir=args.checkpoint_dir,
data_store=data_store,
num_workers=args.num_workers,
)
def main():
screen.set_use_colors(False)
parser = argparse.ArgumentParser()
parser.add_argument(
'-c',
'--checkpoint_dir',
help=
'(string) Path to a folder containing a checkpoint to restore the model from.',
type=str,
default='./checkpoint')
parser.add_argument('--s3_bucket',
help='(string) S3 bucket',
type=str,
default=os.environ.get("SAGEMAKER_SHARED_S3_BUCKET",
"gsaur-test"))
parser.add_argument('--s3_prefix',
help='(string) S3 prefix',
type=str,
default=os.environ.get("SAGEMAKER_SHARED_S3_PREFIX",
"sagemaker"))
parser.add_argument(
'--num-workers',
help="(int) The number of workers started in this pool",
type=int,
default=1)
parser.add_argument('-r',
'--redis_ip',
help="(string) IP or host for the redis server",
default='localhost',
type=str)
parser.add_argument('-rp',
'--redis_port',
help="(int) Port of the redis server",
default=6379,
type=int)
parser.add_argument('--aws_region',
help='(string) AWS region',
type=str,
default=os.environ.get("APP_REGION", "us-east-1"))
parser.add_argument('--reward_file_s3_key',
help='(string) Reward File S3 Key',
type=str,
default=os.environ.get("REWARD_FILE_S3_KEY", None))
parser.add_argument('--model_metadata_s3_key',
help='(string) Model Metadata File S3 Key',
type=str,
default=os.environ.get("MODEL_METADATA_FILE_S3_KEY",
None))
parser.add_argument('--aws_endpoint_url',
help='(string) AWS region',
type=str,
default=os.environ.get("AWS_ENDPOINT_URL", None))
args = parser.parse_args()
s3_client = SageS3Client(bucket=args.s3_bucket,
s3_prefix=args.s3_prefix,
aws_region=args.aws_region,
endpoint_url=args.aws_endpoint_url)
logger.info("S3 bucket: %s" % args.s3_bucket)
logger.info("S3 prefix: %s" % args.s3_prefix)
# Load the model metadata
model_metadata_local_path = os.path.join(CUSTOM_FILES_PATH,
'model_metadata.json')
load_model_metadata(s3_client, args.model_metadata_s3_key,
model_metadata_local_path)
# Download reward function
if not args.reward_file_s3_key:
utils.json_format_logger(
"Reward function code S3 key not available for S3 bucket {} and prefix {}"
.format(args.s3_bucket, args.s3_prefix),
**utils.build_system_error_dict(
utils.SIMAPP_SIMULATION_WORKER_EXCEPTION,
utils.SIMAPP_EVENT_ERROR_CODE_500))
traceback.print_exc()
utils.simapp_exit_gracefully()
download_customer_reward_function(s3_client, args.reward_file_s3_key)
# Register the gym enviroment, this will give clients the ability to creat the enviroment object
register(id=defaults.ENV_ID,
entry_point=defaults.ENTRY_POINT,
max_episode_steps=defaults.MAX_STEPS,
reward_threshold=defaults.THRESHOLD)
redis_ip = s3_client.get_ip()
logger.info("Received IP from SageMaker successfully: %s" % redis_ip)
# Download hyperparameters from SageMaker
hyperparameters_file_success = False
hyperparams_s3_key = os.path.normpath(args.s3_prefix +
"/ip/hyperparameters.json")
hyperparameters_file_success = s3_client.download_file(
s3_key=hyperparams_s3_key, local_path="hyperparameters.json")
sm_hyperparams_dict = {}
if hyperparameters_file_success:
logger.info("Received Sagemaker hyperparameters successfully!")
with open("hyperparameters.json") as fp:
sm_hyperparams_dict = json.load(fp)
else:
logger.info("SageMaker hyperparameters not found.")
preset_file_success, _ = download_custom_files_if_present(
s3_client, args.s3_prefix)
if preset_file_success:
preset_location = os.path.join(CUSTOM_FILES_PATH, "preset.py")
preset_location += ":graph_manager"
graph_manager = short_dynamic_import(preset_location,
ignore_module_case=True)
logger.info("Using custom preset file!")
else:
from markov.sagemaker_graph_manager import get_graph_manager
graph_manager, _ = get_graph_manager(**sm_hyperparams_dict)
logger.info("Connecting to redis at %s:%d" % (redis_ip, args.redis_port))
memory_backend_params = RedisPubSubMemoryBackendParameters(
redis_address=redis_ip,
redis_port=6379,
run_type='worker',
channel=args.s3_prefix)
logger.info("Connecting to s3 boto data store at %s" %
args.aws_endpoint_url)
ds_params_instance = S3BotoDataStoreParameters(
bucket_name=args.s3_bucket,
checkpoint_dir=args.checkpoint_dir,
aws_region=args.aws_region,
s3_folder=args.s3_prefix,
aws_endpoint_url=args.aws_endpoint_url)
data_store = S3BotoDataStore(ds_params_instance)
data_store.graph_manager = graph_manager
graph_manager.data_store = data_store
rollout_worker(graph_manager=graph_manager,
checkpoint_dir=args.checkpoint_dir,
data_store=data_store,
num_workers=args.num_workers,
memory_backend_params=memory_backend_params)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--markov-preset-file',
help="(string) Name of a preset file to run in Markov's preset directory.",
type=str,
default=os.environ.get("MARKOV_PRESET_FILE", "object_tracker.py"))
parser.add_argument('-c', '--local-model-directory',
help='(string) Path to a folder containing a checkpoint to restore the model from.',
type=str,
default=os.environ.get("LOCAL_MODEL_DIRECTORY", "./checkpoint"))
parser.add_argument('-n', '--num-rollout-workers',
help="(int) Number of workers for multi-process based agents, e.g. A3C",
default=os.environ.get("NUMBER_OF_ROLLOUT_WORKERS", 1),
type=int)
parser.add_argument('--model-s3-bucket',
help='(string) S3 bucket where trained models are stored. It contains model checkpoints.',
type=str,
default=os.environ.get("MODEL_S3_BUCKET"))
parser.add_argument('--model-s3-prefix',
help='(string) S3 prefix where trained models are stored. It contains model checkpoints.',
type=str,
default=os.environ.get("MODEL_S3_PREFIX"))
parser.add_argument('--aws-region',
help='(string) AWS region',
type=str,
default=os.environ.get("ROS_AWS_REGION", "us-west-2"))
args = parser.parse_args()
data_store_params_instance = S3BotoDataStoreParameters(bucket_name=args.model_s3_bucket,
s3_folder=args.model_s3_prefix,
checkpoint_dir=args.local_model_directory,
aws_region=args.aws_region)
data_store = S3BotoDataStore(data_store_params_instance)
# Get the IP of the trainer machine
trainer_ip = data_store.get_ip()
print("Received IP from SageMaker successfully: %s" % trainer_ip)
preset_file_success = data_store.download_presets_if_present(PRESET_LOCAL_PATH)
if preset_file_success:
environment_file_success = data_store.download_environments_if_present(ENVIRONMENT_LOCAL_PATH)
path_and_module = PRESET_LOCAL_PATH + args.markov_preset_file + ":graph_manager"
graph_manager = short_dynamic_import(path_and_module, ignore_module_case=True)
if environment_file_success:
import robomaker.environments
print("Using custom preset file!")
elif args.markov_preset_file:
markov_path = imp.find_module("markov")[1]
preset_location = os.path.join(markov_path, "presets", args.markov_preset_file)
path_and_module = preset_location + ":graph_manager"
graph_manager = short_dynamic_import(path_and_module, ignore_module_case=True)
print("Using custom preset file from Markov presets directory!")
else:
raise ValueError("Unable to determine preset file")
memory_backend_params = RedisPubSubMemoryBackendParameters(redis_address=trainer_ip,
redis_port=TRAINER_REDIS_PORT,
run_type='worker',
channel=args.model_s3_prefix)
graph_manager.agent_params.memory.register_var('memory_backend_params', memory_backend_params)
graph_manager.data_store_params = data_store_params_instance
graph_manager.data_store = data_store
utils.wait_for_checkpoint(checkpoint_dir=args.local_model_directory, data_store=data_store)
rollout_worker(
graph_manager=graph_manager,
checkpoint_dir=args.local_model_directory,
data_store=data_store,
num_workers=args.num_rollout_workers
)
def main():
screen.set_use_colors(False)
parser = argparse.ArgumentParser()
parser.add_argument(
'-c',
'--checkpoint-dir',
help=
'(string) Path to a local folder containing a checkpoint to write the model to.',
type=str,
default='./checkpoint')
parser.add_argument(
'--pretrained-checkpoint-dir',
help=
'(string) Path to a local folder for downloading a pre-trained model',
type=str,
default=PRETRAINED_MODEL_DIR)
parser.add_argument('--s3_bucket',
help='(string) S3 bucket',
type=str,
default=os.environ.get(
"SAGEMAKER_SHARED_S3_BUCKET_PATH", "gsaur-test"))
parser.add_argument('--s3_prefix',
help='(string) S3 prefix',
type=str,
default='sagemaker')
parser.add_argument('--framework',
help='(string) tensorflow or mxnet',
type=str,
default='tensorflow')
parser.add_argument('--pretrained_s3_bucket',
help='(string) S3 bucket for pre-trained model',
type=str)
parser.add_argument('--pretrained_s3_prefix',
help='(string) S3 prefix for pre-trained model',
type=str,
default='sagemaker')
parser.add_argument('--RLCOACH_PRESET',
help='(string) Default preset to use',
type=str,
default='object_tracker')
parser.add_argument('--aws_region',
help='(string) AWS region',
type=str,
required=True)
args, unknown = parser.parse_known_args()
s3_client = SageS3Client(bucket=args.s3_bucket,
s3_prefix=args.s3_prefix,
aws_region=args.aws_region)
# Import to register the environment with Gym
import robomaker.environments
preset_location = "robomaker.presets.%s:graph_manager" % args.RLCOACH_PRESET
graph_manager = short_dynamic_import(preset_location,
ignore_module_case=True)
host_ip_address = get_ip_from_host()
s3_client.write_ip_config(host_ip_address)
print("Uploaded IP address information to S3: %s" % host_ip_address)
use_pretrained_model = False
if args.pretrained_s3_bucket and args.pretrained_s3_prefix:
s3_client_pretrained = SageS3Client(
bucket=args.pretrained_s3_bucket,
s3_prefix=args.pretrained_s3_prefix,
aws_region=args.aws_region)
s3_client_pretrained.download_model(PRETRAINED_MODEL_DIR)
use_pretrained_model = True
memory_backend_params = RedisPubSubMemoryBackendParameters(
redis_address="localhost",
redis_port=6379,
run_type='trainer',
channel=args.s3_prefix)
graph_manager.agent_params.memory.register_var('memory_backend_params',
memory_backend_params)
ds_params_instance = S3BotoDataStoreParameters(
bucket_name=args.s3_bucket,
checkpoint_dir=args.checkpoint_dir,
s3_folder=args.s3_prefix,
aws_region=args.aws_region)
graph_manager.data_store_params = ds_params_instance
data_store = S3BotoDataStore(ds_params_instance)
data_store.graph_manager = graph_manager
graph_manager.data_store = data_store
training_worker(graph_manager=graph_manager,
checkpoint_dir=args.checkpoint_dir,
use_pretrained_model=use_pretrained_model,
framework=args.framework)
def main():
screen.set_use_colors(False)
parser = argparse.ArgumentParser()
parser.add_argument(
'-c',
'--checkpoint_dir',
help=
'(string) Path to a folder containing a checkpoint to restore the model from.',
type=str,
default='./checkpoint')
parser.add_argument('--s3_bucket',
help='(string) S3 bucket',
type=str,
default=rospy.get_param("SAGEMAKER_SHARED_S3_BUCKET",
"gsaur-test"))
parser.add_argument('--s3_prefix',
help='(string) S3 prefix',
type=str,
default=rospy.get_param("SAGEMAKER_SHARED_S3_PREFIX",
"sagemaker"))
parser.add_argument('--s3_endpoint_url',
help='(string) S3 endpoint URL',
type=str,
default=rospy.get_param("S3_ENDPOINT_URL", None))
parser.add_argument(
'--num-workers',
help="(int) The number of workers started in this pool",
type=int,
default=1)
parser.add_argument('-r',
'--redis_ip',
help="(string) IP or host for the redis server",
default='localhost',
type=str)
parser.add_argument('-rp',
'--redis_port',
help="(int) Port of the redis server",
default=6379,
type=int)
parser.add_argument('--aws_region',
help='(string) AWS region',
type=str,
default=rospy.get_param("AWS_REGION", "us-east-1"))
parser.add_argument('--reward_file_s3_key',
help='(string) Reward File S3 Key',
type=str,
default=rospy.get_param("REWARD_FILE_S3_KEY", None))
parser.add_argument('--model_metadata_s3_key',
help='(string) Model Metadata File S3 Key',
type=str,
default=rospy.get_param("MODEL_METADATA_FILE_S3_KEY",
None))
# For training job, reset is not allowed. penalty_seconds, off_track_penalty, and
# collision_penalty will all be 0 be default
parser.add_argument('--number_of_resets',
help='(integer) Number of resets',
type=int,
default=int(rospy.get_param("NUMBER_OF_RESETS", 0)))
parser.add_argument('--penalty_seconds',
help='(float) penalty second',
type=float,
default=float(rospy.get_param("PENALTY_SECONDS", 0.0)))
parser.add_argument('--job_type',
help='(string) job type',
type=str,
default=rospy.get_param("JOB_TYPE", "TRAINING"))
parser.add_argument('--is_continuous',
help='(boolean) is continous after lap completion',
type=bool,
default=utils.str2bool(
rospy.get_param("IS_CONTINUOUS", False)))
parser.add_argument('--race_type',
help='(string) Race type',
type=str,
default=rospy.get_param("RACE_TYPE", "TIME_TRIAL"))
parser.add_argument('--off_track_penalty',
help='(float) off track penalty second',
type=float,
default=float(rospy.get_param("OFF_TRACK_PENALTY",
0.0)))
parser.add_argument('--collision_penalty',
help='(float) collision penalty second',
type=float,
default=float(rospy.get_param("COLLISION_PENALTY",
0.0)))
args = parser.parse_args()
s3_client = SageS3Client(bucket=args.s3_bucket,
s3_prefix=args.s3_prefix,
aws_region=args.aws_region,
s3_endpoint_url=args.s3_endpoint_url)
logger.info("S3 bucket: %s", args.s3_bucket)
logger.info("S3 prefix: %s", args.s3_prefix)
logger.info("S3 endpoint URL: %s" % args.s3_endpoint_url)
# Load the model metadata
model_metadata_local_path = os.path.join(CUSTOM_FILES_PATH,
'model_metadata.json')
utils.load_model_metadata(s3_client, args.model_metadata_s3_key,
model_metadata_local_path)
# Download and import reward function
if not args.reward_file_s3_key:
log_and_exit(
"Reward function code S3 key not available for S3 bucket {} and prefix {}"
.format(args.s3_bucket, args.s3_prefix),
SIMAPP_SIMULATION_WORKER_EXCEPTION, SIMAPP_EVENT_ERROR_CODE_500)
download_customer_reward_function(s3_client, args.reward_file_s3_key)
try:
from custom_files.customer_reward_function import reward_function
except Exception as e:
log_and_exit("Failed to import user's reward_function: {}".format(e),
SIMAPP_SIMULATION_WORKER_EXCEPTION,
SIMAPP_EVENT_ERROR_CODE_400)
# Instantiate Cameras
configure_camera(namespaces=['racecar'])
preset_file_success, _ = download_custom_files_if_present(
s3_client, args.s3_prefix)
#! TODO each agent should have own config
_, _, version = utils_parse_model_metadata.parse_model_metadata(
model_metadata_local_path)
agent_config = {
'model_metadata': model_metadata_local_path,
ConfigParams.CAR_CTRL_CONFIG.value: {
ConfigParams.LINK_NAME_LIST.value:
LINK_NAMES,
ConfigParams.VELOCITY_LIST.value:
VELOCITY_TOPICS,
ConfigParams.STEERING_LIST.value:
STEERING_TOPICS,
ConfigParams.CHANGE_START.value:
utils.str2bool(rospy.get_param('CHANGE_START_POSITION', True)),
ConfigParams.ALT_DIR.value:
utils.str2bool(
rospy.get_param('ALTERNATE_DRIVING_DIRECTION', False)),
ConfigParams.ACTION_SPACE_PATH.value:
'custom_files/model_metadata.json',
ConfigParams.REWARD.value:
reward_function,
ConfigParams.AGENT_NAME.value:
'racecar',
ConfigParams.VERSION.value:
version,
ConfigParams.NUMBER_OF_RESETS.value:
args.number_of_resets,
ConfigParams.PENALTY_SECONDS.value:
args.penalty_seconds,
ConfigParams.NUMBER_OF_TRIALS.value:
None,
ConfigParams.IS_CONTINUOUS.value:
args.is_continuous,
ConfigParams.RACE_TYPE.value:
args.race_type,
ConfigParams.COLLISION_PENALTY.value:
args.collision_penalty,
ConfigParams.OFF_TRACK_PENALTY.value:
args.off_track_penalty
}
}
#! TODO each agent should have own s3 bucket
metrics_s3_config = {
MetricsS3Keys.METRICS_BUCKET.value:
rospy.get_param('METRICS_S3_BUCKET'),
MetricsS3Keys.METRICS_KEY.value:
rospy.get_param('METRICS_S3_OBJECT_KEY'),
MetricsS3Keys.ENDPOINT_URL.value:
rospy.get_param('S3_ENDPOINT_URL', None),
MetricsS3Keys.REGION.value:
rospy.get_param('AWS_REGION'),
MetricsS3Keys.STEP_BUCKET.value:
rospy.get_param('SAGEMAKER_SHARED_S3_BUCKET'),
MetricsS3Keys.STEP_KEY.value:
os.path.join(rospy.get_param('SAGEMAKER_SHARED_S3_PREFIX'),
TRAINING_SIMTRACE_DATA_S3_OBJECT_KEY)
}
metrics_s3_model_cfg = {
MetricsS3Keys.METRICS_BUCKET.value:
args.s3_bucket,
MetricsS3Keys.METRICS_KEY.value:
os.path.join(args.s3_prefix, DEEPRACER_CHKPNT_KEY_SUFFIX),
MetricsS3Keys.REGION.value:
args.aws_region
}
run_phase_subject = RunPhaseSubject()
agent_list = list()
agent_list.append(
create_rollout_agent(
agent_config,
TrainingMetrics('agent', metrics_s3_config, metrics_s3_model_cfg,
args.checkpoint_dir, run_phase_subject),
run_phase_subject))
agent_list.append(create_obstacles_agent())
agent_list.append(create_bot_cars_agent())
# ROS service to indicate all the robomaker markov packages are ready for consumption
signal_robomaker_markov_package_ready()
PhaseObserver('/agent/training_phase', run_phase_subject)
aws_region = rospy.get_param('AWS_REGION', args.aws_region)
simtrace_s3_bucket = rospy.get_param('SIMTRACE_S3_BUCKET', None)
mp4_s3_bucket = rospy.get_param('MP4_S3_BUCKET', None)
if simtrace_s3_bucket:
simtrace_s3_object_prefix = rospy.get_param('SIMTRACE_S3_PREFIX')
if mp4_s3_bucket:
mp4_s3_object_prefix = rospy.get_param('MP4_S3_OBJECT_PREFIX')
s3_writer_job_info = []
if simtrace_s3_bucket:
s3_writer_job_info.append(
IterationData(
'simtrace', simtrace_s3_bucket, simtrace_s3_object_prefix,
aws_region,
os.path.join(
ITERATION_DATA_LOCAL_FILE_PATH, 'agent',
IterationDataLocalFileNames.SIM_TRACE_TRAINING_LOCAL_FILE.
value)))
if mp4_s3_bucket:
s3_writer_job_info.extend([
IterationData(
'pip', mp4_s3_bucket, mp4_s3_object_prefix, aws_region,
os.path.join(
ITERATION_DATA_LOCAL_FILE_PATH, 'agent',
IterationDataLocalFileNames.
CAMERA_PIP_MP4_VALIDATION_LOCAL_PATH.value)),
IterationData(
'45degree', mp4_s3_bucket, mp4_s3_object_prefix, aws_region,
os.path.join(
ITERATION_DATA_LOCAL_FILE_PATH, 'agent',
IterationDataLocalFileNames.
CAMERA_45DEGREE_MP4_VALIDATION_LOCAL_PATH.value)),
IterationData(
'topview', mp4_s3_bucket, mp4_s3_object_prefix, aws_region,
os.path.join(
ITERATION_DATA_LOCAL_FILE_PATH, 'agent',
IterationDataLocalFileNames.
CAMERA_TOPVIEW_MP4_VALIDATION_LOCAL_PATH.value))
])
s3_writer = S3Writer(job_info=s3_writer_job_info,
s3_endpoint_url=args.s3_endpoint_url)
redis_ip = s3_client.get_ip()
logger.info("Received IP from SageMaker successfully: %s", redis_ip)
# Download hyperparameters from SageMaker
hyperparameters_file_success = False
hyperparams_s3_key = os.path.normpath(args.s3_prefix +
"/ip/hyperparameters.json")
hyperparameters_file_success = s3_client.download_file(
s3_key=hyperparams_s3_key, local_path="hyperparameters.json")
sm_hyperparams_dict = {}
if hyperparameters_file_success:
logger.info("Received Sagemaker hyperparameters successfully!")
with open("hyperparameters.json") as filepointer:
sm_hyperparams_dict = json.load(filepointer)
else:
logger.info("SageMaker hyperparameters not found.")
if preset_file_success:
preset_location = os.path.join(CUSTOM_FILES_PATH, "preset.py")
preset_location += ":graph_manager"
graph_manager = short_dynamic_import(preset_location,
ignore_module_case=True)
logger.info("Using custom preset file!")
else:
graph_manager, _ = get_graph_manager(
hp_dict=sm_hyperparams_dict,
agent_list=agent_list,
run_phase_subject=run_phase_subject)
memory_backend_params = RedisPubSubMemoryBackendParameters(
redis_address=redis_ip,
redis_port=6379,
run_type=str(RunType.ROLLOUT_WORKER),
channel=args.s3_prefix)
graph_manager.memory_backend_params = memory_backend_params
ds_params_instance = S3BotoDataStoreParameters(
aws_region=args.aws_region,
bucket_names={'agent': args.s3_bucket},
base_checkpoint_dir=args.checkpoint_dir,
s3_folders={'agent': args.s3_prefix},
s3_endpoint_url=args.s3_endpoint_url)
graph_manager.data_store = S3BotoDataStore(ds_params_instance,
graph_manager)
task_parameters = TaskParameters()
task_parameters.checkpoint_restore_path = args.checkpoint_dir
rollout_worker(graph_manager=graph_manager,
num_workers=args.num_workers,
task_parameters=task_parameters,
s3_writer=s3_writer)
def main():
screen.set_use_colors(False)
parser = argparse.ArgumentParser()
parser.add_argument(
'-c',
'--checkpoint_dir',
help=
'(string) Path to a folder containing a checkpoint to restore the model from.',
type=str,
default='./checkpoint')
parser.add_argument('--s3_bucket',
help='(string) S3 bucket',
type=str,
default=rospy.get_param("SAGEMAKER_SHARED_S3_BUCKET",
"gsaur-test"))
parser.add_argument('--s3_prefix',
help='(string) S3 prefix',
type=str,
default=rospy.get_param("SAGEMAKER_SHARED_S3_PREFIX",
"sagemaker"))
parser.add_argument(
'--num-workers',
help="(int) The number of workers started in this pool",
type=int,
default=1)
parser.add_argument('-r',
'--redis_ip',
help="(string) IP or host for the redis server",
default='localhost',
type=str)
parser.add_argument('-rp',
'--redis_port',
help="(int) Port of the redis server",
default=6379,
type=int)
parser.add_argument('--aws_region',
help='(string) AWS region',
type=str,
default=rospy.get_param("AWS_REGION", "us-east-1"))
parser.add_argument('--reward_file_s3_key',
help='(string) Reward File S3 Key',
type=str,
default=rospy.get_param("REWARD_FILE_S3_KEY", None))
parser.add_argument('--model_metadata_s3_key',
help='(string) Model Metadata File S3 Key',
type=str,
default=rospy.get_param("MODEL_METADATA_FILE_S3_KEY",
None))
args = parser.parse_args()
s3_client = SageS3Client(bucket=args.s3_bucket,
s3_prefix=args.s3_prefix,
aws_region=args.aws_region)
logger.info("S3 bucket: %s" % args.s3_bucket)
logger.info("S3 prefix: %s" % args.s3_prefix)
# Load the model metadata
model_metadata_local_path = os.path.join(CUSTOM_FILES_PATH,
'model_metadata.json')
utils.load_model_metadata(s3_client, args.model_metadata_s3_key,
model_metadata_local_path)
# Download and import reward function
if not args.reward_file_s3_key:
utils.log_and_exit(
"Reward function code S3 key not available for S3 bucket {} and prefix {}"
.format(args.s3_bucket,
args.s3_prefix), utils.SIMAPP_SIMULATION_WORKER_EXCEPTION,
utils.SIMAPP_EVENT_ERROR_CODE_500)
download_customer_reward_function(s3_client, args.reward_file_s3_key)
try:
from custom_files.customer_reward_function import reward_function
except Exception as e:
utils.log_and_exit(
"Failed to import user's reward_function: {}".format(e),
utils.SIMAPP_SIMULATION_WORKER_EXCEPTION,
utils.SIMAPP_EVENT_ERROR_CODE_400)
# Instantiate Cameras
configure_camera()
redis_ip = s3_client.get_ip()
logger.info("Received IP from SageMaker successfully: %s" % redis_ip)
# Download hyperparameters from SageMaker
hyperparameters_file_success = False
hyperparams_s3_key = os.path.normpath(args.s3_prefix +
"/ip/hyperparameters.json")
hyperparameters_file_success = s3_client.download_file(
s3_key=hyperparams_s3_key, local_path="hyperparameters.json")
sm_hyperparams_dict = {}
if hyperparameters_file_success:
logger.info("Received Sagemaker hyperparameters successfully!")
with open("hyperparameters.json") as fp:
sm_hyperparams_dict = json.load(fp)
else:
logger.info("SageMaker hyperparameters not found.")
preset_file_success, _ = download_custom_files_if_present(
s3_client, args.s3_prefix)
#! TODO each agent should have own config
_, _, version = utils_parse_model_metadata.parse_model_metadata(
model_metadata_local_path)
agent_config = {
'model_metadata': model_metadata_local_path,
'car_ctrl_cnfig': {
ConfigParams.LINK_NAME_LIST.value:
LINK_NAMES,
ConfigParams.VELOCITY_LIST.value:
VELOCITY_TOPICS,
ConfigParams.STEERING_LIST.value:
STEERING_TOPICS,
ConfigParams.CHANGE_START.value:
utils.str2bool(rospy.get_param('CHANGE_START_POSITION', True)),
ConfigParams.ALT_DIR.value:
utils.str2bool(
rospy.get_param('ALTERNATE_DRIVING_DIRECTION', False)),
ConfigParams.ACTION_SPACE_PATH.value:
'custom_files/model_metadata.json',
ConfigParams.REWARD.value:
reward_function,
ConfigParams.AGENT_NAME.value:
'racecar',
ConfigParams.VERSION.value:
version
}
}
#! TODO each agent should have own s3 bucket
metrics_s3_config = {
MetricsS3Keys.METRICS_BUCKET.value:
rospy.get_param('METRICS_S3_BUCKET'),
MetricsS3Keys.METRICS_KEY.value:
rospy.get_param('METRICS_S3_OBJECT_KEY'),
MetricsS3Keys.REGION.value:
rospy.get_param('AWS_REGION'),
MetricsS3Keys.STEP_BUCKET.value:
rospy.get_param('SAGEMAKER_SHARED_S3_BUCKET'),
MetricsS3Keys.STEP_KEY.value:
os.path.join(rospy.get_param('SAGEMAKER_SHARED_S3_PREFIX'),
TRAINING_SIMTRACE_DATA_S3_OBJECT_KEY)
}
agent_list = list()
agent_list.append(
create_rollout_agent(agent_config, TrainingMetrics(metrics_s3_config)))
agent_list.append(create_obstacles_agent())
agent_list.append(create_bot_cars_agent())
if preset_file_success:
preset_location = os.path.join(CUSTOM_FILES_PATH, "preset.py")
preset_location += ":graph_manager"
graph_manager = short_dynamic_import(preset_location,
ignore_module_case=True)
logger.info("Using custom preset file!")
else:
graph_manager, _ = get_graph_manager(sm_hyperparams_dict, agent_list)
memory_backend_params = RedisPubSubMemoryBackendParameters(
redis_address=redis_ip,
redis_port=6379,
run_type=str(RunType.ROLLOUT_WORKER),
channel=args.s3_prefix)
graph_manager.memory_backend_params = memory_backend_params
ds_params_instance = S3BotoDataStoreParameters(
aws_region=args.aws_region,
bucket_name=args.s3_bucket,
checkpoint_dir=args.checkpoint_dir,
s3_folder=args.s3_prefix)
data_store = S3BotoDataStore(ds_params_instance)
data_store.graph_manager = graph_manager
graph_manager.data_store = data_store
task_parameters = TaskParameters()
task_parameters.checkpoint_restore_path = args.checkpoint_dir
rollout_worker(graph_manager=graph_manager,
data_store=data_store,
num_workers=args.num_workers,
task_parameters=task_parameters)
