def get_experiment_from_string(experiment, running=False):
"""
Returns an Experiment object given a string of either a json file or a name of an already existing eperiment.
Args:
experiment (str): The string to look for (either local json file or local experiment's name)
running (bool): Whether this experiment is already running.
Returns:
The found Experiment object.
"""
file = "experiments/{}/{}.json". \
format(experiment, "experiment" if not running else "experiment_running")
if not os.path.exists(file):
if running:
raise util.TFCliError(
"ERROR: Experiment {} does not seem to be running right now! You have to create, then"
"start it with 'experiment new/start'.".format(experiment))
else:
raise util.TFCliError(
"ERROR: Experiment {} not found! You have to create it first with 'experiment new'."
.format(experiment))
# get the experiment object from its json file
with open(file) as f:
spec = json.load(f)
exp_obj = Experiment(**spec)
return exp_obj
def cmd_cluster_delete(args):
if args.all:
clusters = util.get_cluster_specs()
for c in clusters.values():
cluster = Cluster(**c)
cluster.delete()
print("+ All clusters deleted.")
else:
if not args.cluster:
raise util.TFCliError("ERROR: Cluster name (-c option) not given!")
print("+ Looking for clusters ...")
cluster = get_cluster_from_string(args.cluster)
if not isinstance(cluster, Cluster):
raise util.TFCliError("ERROR: No cluster with name {} found!".format(args.cluster))
cluster.delete()
def get_cluster_from_string(cluster, running_clusters=None):
"""
Returns a Cluster object given a string of either a json file or an already running remote cluster's name.
Args:
cluster (str): The string to look for (either local json file or remote cluster's name)
running_clusters (dict): Specs for already running cloud clusters by cluster name.
Returns:
The found Cluster object.
"""
# no running clusters given -> get them now
if not running_clusters:
running_clusters = util.get_cluster_specs()
# json file (get spec)
if re.search(r'\.json$', cluster):
cluster = Cluster(running_clusters=running_clusters,
file=cluster,
**util.read_json_spec(cluster, "clusters"))
# cluster name (cluster must already exists in cloud)
else:
cluster_name = re.sub(r'_', '-', cluster)
if cluster_name in running_clusters:
cluster = Cluster(running_clusters=running_clusters,
**running_clusters[cluster_name])
else:
raise util.TFCliError(
"ERROR: Given cluster {} not found in cloud!".format(
cluster_name))
return cluster
def create(self):
"""
Create the Kubernetes cluster with the options given in self.
This also sets up the local kubectl app to point to the new cluster automatically.
"""
print("+ Creating cluster: {}. This may take a few minutes ...".format(
self.name_hyphenated))
if self.num_gpus == 0:
out = util.syscall(
"gcloud container clusters create {} -m {} --disk-size {} --num-nodes {} {}"
.format(self.name_hyphenated, self.machine_type,
self.disk_size, self.num_nodes,
"--zone " + self.location if self.location else ""),
return_outputs="as_str")
else:
out = util.syscall(
"gcloud container clusters create {} --enable-cloud-logging --enable-cloud-monitoring "
"--accelerator type={},count={} {} -m {} --disk-size {} --enable-kubernetes-alpha "
"--image-type UBUNTU --num-nodes {} --cluster-version 1.9.2-gke.1 --quiet"
.format(self.name_hyphenated, self.gpu_type,
self.gpus_per_node,
"--zone " + self.location if self.location else "",
self.machine_type, self.disk_size, self.num_nodes),
return_outputs="as_str")
# check output of cluster generating code
if re.search(r'error', out, re.IGNORECASE):
raise util.TFCliError(out)
else:
print("+ Successfully created cluster.")
self.instances, self.primary_name = util.get_compute_instance_specs(
self.name_hyphenated)
self.started = True
# install NVIDIA drivers on machines per local kubectl
if self.num_gpus > 0:
print("+ Installing NVIDIA GPU drivers and k8s device plugins ...")
util.syscall(
"kubectl create -f https://raw.githubusercontent.com/GoogleCloudPlatform/"
"container-engine-accelerators/k8s-1.9/daemonset.yaml")
util.syscall(
"kubectl delete -f https://raw.githubusercontent.com/kubernetes/kubernetes/"
"release-1.9/cluster/addons/device-plugins/nvidia-gpu/daemonset.yaml"
)
util.syscall(
"kubectl create -f https://raw.githubusercontent.com/kubernetes/kubernetes/"
"release-1.9/cluster/addons/device-plugins/nvidia-gpu/daemonset.yaml"
)
print("+ Done. Cluster: {} created.".format(self.name_hyphenated))
def cmd_experiment_new(args, project_id=None):
# check for experiment already existing
experiments = get_local_experiments()
# setup the Experiment object
experiment = Experiment(**args.__dict__)
if experiment.name in experiments:
print("ERROR: An experiment with the name {} already exists in this project! "
"Use `experiment start` to start it.".format(experiment.name))
return
# write experiment files to local disk
experiment.generate_locally()
# and start the experiment?
if args.start:
if not project_id:
raise util.TFCliError("ERROR: Cannot start experiment without remote project ID!")
print("+ New experiment created. Starting ...")
experiment.start(project_id)
else:
print("+ New experiment created. Use 'experiment start' to run.")
def _ssh_parallel_target(self, node, silent, items):
for item in items:
# an ssh command to execute on the node
if isinstance(item, str):
_ = util.syscall(
"gcloud compute ssh {} {} --command \"{}\"".format(
node,
"--zone=" + self.location if self.location else "",
item),
return_outputs=silent)
# an scp command (copy from ... to ...)
elif isinstance(item, (list, tuple)) and len(item) == 2:
item = list(map(lambda i: re.sub(r'_NODE_', node, i), item))
_ = util.syscall("gcloud compute scp {} {} {}".format(
"--zone=" + self.location if self.location else "",
item[0], item[1]),
return_outputs=silent)
else:
raise util.TFCliError(
"ERROR: unknown ssh command structure. Needs to be str (ssh-command) "
"or list/tuple of exactly 2 str (scp).")
def cmd_init(args):
# check if there is already a .tensorforce file in this folder
if os.path.isfile(".tensorforce.json"):
# TODO: read .tensorforce.json file to display project's name and other data
if not args.force:
print("WARNING: This directory already contains a tensorforce project. Would you like to overwrite it?")
response = input(">")
if response.upper() != "Y":
quit()
# erase the existing project and create a new one
shutil.rmtree(".tensorforce.json")
print("+ Creating project paths and copying sample spec files.")
# add sub-dirs to it and write the main project file
if not os.path.isdir("clusters"):
os.makedirs("clusters")
if not os.path.isdir("experiments"):
os.makedirs("experiments")
# copy all json example spec files from cloned github repo
import tensorforce_client
p = tensorforce_client.__path__[0] + "/configs"
shutil.rmtree("configs/", ignore_errors=True)
shutil.copytree("{}".format(p), "configs/")
# add the experiment jinja file (k8s yaml template) into project's config dir
shutil.copy("{}/experiment.yaml.jinja".format(p), "configs/")
print("+ Checking requirements (gcloud and kubectl installations).")
# check for installations of gcloud, then kubectl
try:
out = syscall("gcloud --version", return_outputs="as_str", merge_err=True)
# Linux: fake Win command not found error
except OSError:
out = "not recognized as an internal"
if re.match(r'not recognized as an internal', out):
print("INIT ERROR: Installation of gcloud command line tool required.\nPlease install first:"
" https://cloud.google.com/sdk/docs/quickstarts")
quit()
# we can install kubectl via gcloud: `gcloud components install kubectl`
try:
out = syscall("kubectl version", return_outputs="as_str", merge_err=True)
# Linux: fake Win command not found error
except OSError:
out = "not recognized as an internal"
if re.match(r'not recognized as an internal', out):
print("++ Installing missing kubectl command line tool (this is necessary to manage your clusters via the"
" Kubernetes tool):")
syscall("gcloud components install kubectl")
# login to google cloud
print("+ Logging you into google cloud account.")
while True:
print("Please enter your remote project's google cloud service account (full email address) here:")
service_account = input(">")
if not re.match(r'^[\w\-\.]+\@[\w\-\.]+\.[a-z]+', service_account):
print("ERROR: The service account needs to be an email address.")
else:
break
while True:
print("Please enter the location of your private key file associated with this service account:")
key_file = input(">")
if not os.path.isfile(key_file):
print("ERROR: The key_file you entered does not exist or is not a file.")
else:
break
# kubernetes-account@introkubernetes-191608.iam.gserviceaccount.com
# l:/programming/privatekeys/MaRLEnE-bbad55cddab1.json
syscall("gcloud auth activate-service-account {} --key-file={}".format(service_account, key_file))
remote_projects_by_name, remote_projects_by_id = util.get_remote_projects()
# if remote given -> only check for that one and exit if doesn't exist
if args.remote_project_id:
print("+ Checking for existing remote-project ID ({}).".format(args.remote_project_id))
if args.remote_project_id not in remote_projects_by_id:
print("ERROR: No remote project ID {} found in cloud!".format(args.remote_project_id))
quit()
print("+ Found remote project ID {}.".format(args.remote_project_id))
remote_project_id = args.remote_project_id
remote_project_name = remote_projects_by_id[args.remote_project_id]["project-name"]
# if no name -> take remote's name
if not args.name:
args.name = remote_project_name
# look for existing project in cloud with same name and ask whether to use that one. If not, user can specify
# a remote project name that may differ from the local project folder's name
else:
# if no name -> take folder name
if not args.name:
cwd = os.getcwd()
args.name = re.sub(r'.+[/\\](\w+)$', '\\1', cwd)
print("+ Name not given. Assuming folder name ({}) is project's name.".format(args.name))
print("+ Checking name ({}) against existing projects in the cloud.".format(args.name))
if args.name in remote_projects_by_name:
remote_project_name = args.name
remote_project_id = remote_projects_by_name[remote_project_name]["project-id"]
print("++ Given project name ({}) already exists as a project in google cloud. "
"Will use remote project {} with ID {}.".format(args.name, args.name, remote_project_id))
# TODO: service accounts cannot create projects without a parent
else:
remote_project_id = re.sub(r'_', '-', args.name) # replace all underscores with hyphens
remote_project_name = remote_project_id
# what if id already exists -> use a modified one
if remote_project_id in remote_projects_by_id:
remote_project_id = remote_project_id + str(time.time())
print("+ Project '{}' does not exist in cloud yet. Will create new project (with ID {}).".
format(args.name, remote_project_id))
out, err = syscall("gcloud projects create {} --name {} --set-as-default".
format(remote_project_id, remote_project_name), return_outputs=True, merge_err=False)
err_msg = err.raw.readall().decode("latin-1")
if err_msg != b"":
raise util.TFCliError(err_msg)
# write project settings into .tensorforce dir
util.write_project_file(args.name, remote_project_name, remote_project_id)
def __init__(self, **kwargs):
"""
Keyword Args:
file (str): The Experiment's json spec file (can contain all other args).
name (str): The name of the Experiment. This is also the name of the folder where it is stored.
environment (str): The filename of the json env-spec file to use (see TensorForce documentation).
agent (str): The filename of the json agent-spec file to use (see TensorForce documentation).
network (str): The filename of the json network-spec file to use (see TensorForce documentation).
cluster (str): The filename of the json cluster-spec file to use (see class `Cluster`).
episodes (int): The total number of episodes to run (all parallel agents).
total_timesteps (int): The max. total number of timesteps to run (all parallel agents).
max_timesteps_per_episode (int): The max. number of timesteps to run in each episode.
deterministic (bool): Whether to not(!) use stochastic exploration on top of plain action outputs.
repeat_actions (int): The number of actions to repeat for each action selection (by calling agent.act()).
debug_logging (bool): Whether to switch on debug logging (default: False).
run_mode (str): Which runner mode to use. Valid values are only 'single', 'multi-threaded' and
'distributed'.
num_workers (int): The number of worker processes to use (see `distributed` and `multi-threaded`
run_modes).
num_parameter_servers (int): The number of parameter servers to use (see distributed tensorflow).
saver_frequency (str): The frequency with which to save the model. This is a combination of an int
and a unit (e.g. "600s"), where unit can be "s" (seconds), "e" (episodes), or "t" (timesteps).
summary_frequency (str): The frequency with which to save a tensorboard summary.
This is a combination of an int and a unit (e.g. "600s"), where unit can be "s" (seconds)
or "t" (timesteps). The episode unit (e) is not allowed here.
"""
# see whether we have a json (yaml?) file for the experiment
# TODO: yaml support
self.file = kwargs.get("file")
if self.file:
from_json = util.read_json_spec(self.file, "experiments")
# get all attributes from kwargs
else:
from_json = {}
# From here on, give kwargs priority over spec (from file), so that single settings in the json file can be
# overwritten by command line.
# sanity check name
self.name = kwargs.get("name") or from_json.get("name", "")
if not re.match(r'^\w+$', self.name):
raise util.TFCliError(
"ERROR: Name of experiment needs to be all alphanumeric characters"
)
self.name_hyphenated = re.sub(r'_', '-', self.name)
self.path = "experiments/{}/".format(self.name)
self.k8s_config = "{}experiment.yaml".format(self.path)
# read in sub-spec files (to JSON)
self.environment = kwargs.get("environment") or from_json.get(
"environment")
if isinstance(self.environment, str):
self.environment = util.read_json_spec(self.environment,
"environments")
if self.environment.get(
"remote") and not self.environment.get("image"):
raise util.TFCliError(
"WARNING: Defining a remote environment without a docker image in experiment spec! "
"Use field `image` to define a docker image for the remote env."
)
self.network = kwargs.get("network") or from_json.get("network")
if isinstance(self.network, str):
self.network = util.read_json_spec(self.network, "networks")
self.agent = kwargs.get("agent") or from_json.get("agent")
if isinstance(self.agent, str):
self.agent = util.read_json_spec(self.agent, "agents")
self.cluster = kwargs.get("cluster") or from_json.get("cluster")
if isinstance(self.cluster, str):
cluster = get_cluster_from_string(self.cluster)
self.cluster = cluster.get_spec()
elif not isinstance(self.cluster, dict):
raise util.TFCliError(
"ERROR: Cluster (-c option) has to be given as json filename.")
self.episodes = kwargs.get("episodes") or from_json.get(
"episodes", 10000)
self.total_timesteps = kwargs.get("total_timesteps") or from_json.get(
"total_timesteps", 1000000)
self.max_timesteps_per_episode = kwargs.get("max_timesteps_per_episode") or \
from_json.get("max_timesteps_per_episode", 1000)
self.deterministic = kwargs.get("deterministic")
if self.deterministic is None:
self.deterministic = from_json.get("deterministic", False)
self.repeat_actions = kwargs.get("repeat_actions") or from_json.get(
"repeat_actions", 1)
self.num_workers = kwargs.get("num_workers") or from_json.get(
"num_workers", 3)
self.num_parameter_servers = kwargs.get(
"num_parameter_servers") or from_json.get("num_parameter_servers",
1)
# update our json file pointer and write us into the experiment's dir
self.file = "{}experiment.json".format(self.path)
self.debug_logging = kwargs.get("debug_logging") or from_json.get(
"debug_logging", False)
# the experiment's run type
self.run_mode = kwargs.get("run_mode") or from_json.get(
"run_mode", "distributed")
assert self.run_mode in ["distributed", "multi-threaded", "single"],\
"ERROR: run-type needs to be one of distributed|multi-threaded|single!"
if self.run_mode == "distributed" and self.num_parameter_servers <= 0:
raise util.TFCliError(
"ERROR: Cannot create experiment of run-mode=distributed and zero parameter servers!"
)
self.saver_frequency = kwargs.get("saver_frequency")\
or from_json.get("saver_frequency", "600s" if self.run_mode == "distributed" else "100e")
self.summary_frequency = kwargs.get("summary_frequency")\
or from_json.get("summary_frequency", "120s" if self.run_mode == "distributed" else "10e")
# whether this experiment runs on a dedicated cluster
self.has_dedicated_cluster = kwargs.get(
"has_dedicated_cluster") or from_json.get("has_dedicated_cluster",
True)
# status (running, paused, stopped, etc..)
self.status = kwargs.get("status") or from_json.get("status", None)
# json file specific to a certain experiment 'run' (e.g. cluster may differ from experiment's base config)
self.running_json_file = "experiment_running.json"
def setup_cluster(self, cluster, project_id, start=False):
"""
Given a cluster name (or None) and a remote project-ID,
sets up the cluster settings for this Experiment locally.
Also starts the cluster if start is set to True.
Args:
cluster (str): The name of the cluster. If None, will get cluster-spec from the Experiment, or create a
default Cluster object.
project_id (str): The remote gcloud project ID.
start (bool): Whether to already create (start) the cluster in the cloud.
Returns: The Cluster object.
"""
clusters = util.get_cluster_specs()
# cluster is given (separate from experiment's own cluster)
if cluster:
cluster = get_cluster_from_string(cluster,
running_clusters=clusters)
self.has_dedicated_cluster = False
# use experiment's own cluster
elif self.cluster:
cluster = Cluster(running_clusters=clusters, **self.cluster)
self.has_dedicated_cluster = True
# use a default cluster
else:
cluster = Cluster(name=self.name_hyphenated)
self.has_dedicated_cluster = True
# start cluster if not up yet
if start and not cluster.started:
cluster.create()
# cluster up but not in good state
elif clusters[cluster.name_hyphenated]["status"] != "RUNNING":
raise util.TFCliError(
"ERROR: Given cluster {} is not in status RUNNING (but in status {})!"
.format(cluster.name_hyphenated,
clusters[cluster.name_hyphenated]["status"]))
# check cluster vs experiment setup and warn or abort if something doesn't match
if self.run_mode != "distributed" and cluster.num_nodes > 1:
warn(
"WARNING: Running non-distributed experiment on cluster with more than 1 node. Make sure you are "
"not wasting costly resources!")
num_gpus = cluster.num_nodes * cluster.gpus_per_node
if self.run_mode == "distributed" and self.num_workers + self.num_parameter_servers > num_gpus:
warn(
"WARNING: Running distributed experiment with {} processes total on cluster with only {} GPUs! "
"This could lead to K8s scheduling problems.".format(
self.num_workers + self.num_parameter_servers, num_gpus))
print("+ Setting up credentials to connect to cluster {}.".format(
cluster.name_hyphenated))
util.syscall(
"gcloud container clusters get-credentials {} --zone {} --project {}"
.format(cluster.name_hyphenated, cluster.location, project_id))
print("+ Setting kubectl to point to cluster {}.".format(
cluster.name_hyphenated))
util.syscall("kubectl config set-cluster {} --server={}".format(
cluster.name_hyphenated,
clusters[cluster.name_hyphenated]["master_ip"]))
self.cluster = cluster.get_spec()
return cluster
def __init__(self, **kwargs):
r"""
A cloud cluster object specifying things like: number of nodes, GPUs per node and GPU type, memory per node,
disk size, zone, etc..
Args:
kwargs (any): See below.
Keyword Args:
file (str): The filename of a cluster spec json file to use. Single settings in this file
can be overwritten by specifying these in further kwargs to this c'tor.
name (str): The name of the cluster.
machine_type (str): The machine type to use for all nodes in the cluster. Machine types can
either be gcloud-accepted strings such as everything listed in `gcloud compute machine-types list`
or custom strings that conform to these rules:
https://cloud.google.com/compute/docs/instances/creating-instance-with-custom-machine-type.
When the kwargs `cpus_per_node` and `memory_per_node` are given,
tensorforce-client will automatically create the correct machine-type.
cpus_per_node (int): The number of vCPUs per node.
gpus_per_node (int): The number of (physical) GPUs per node.
gpu_type (str): The GPU type to use. Supported are only 'nvidia-tesla-k80' and 'nvidia-tesla-p100'.
memory_per_node (int): The memory (in Gb) per node.
num_nodes (int): The number of nodes for the cluster.
disk_size (int): The amount of disk space per node in Gb.
location (str): The location of the cluster. Default us the gcloud/project set default zone.
"""
self.file = kwargs.get("file")
if self.file:
from_json = util.read_json_spec(self.file, "clusters")
# get all attributes from kwargs
else:
from_json = {}
self.name = kwargs.get("name") or from_json.get("name")
if not self.name:
raise util.TFCliError("ERROR: Cluster requires a name!")
self.name_hyphenated = re.sub(r'_', '-', self.name)
self.machine_type = kwargs.get("machine_type") or from_json.get(
"machine_type")
# alternative to machine_type -> provide `cpus_per_node` and `memory_per_node`
if not self.machine_type:
cpus = kwargs.get("cpus_per_node") or from_json.get(
"cpus_per_node")
mem = kwargs.get("memory_per_node") or from_json.get(
"memory_per_node")
if not cpus or not mem:
raise util.TFCliError(
"ERROR: no vCPUs_per_node OR no memory_per_node given for cluster {}"
.format(self.name))
self.machine_type = "custom-{}-{}".format(cpus, mem * 1024)
self.num_nodes = kwargs.get("num_nodes") or from_json.get(
"num_nodes", 3)
self.num_gpus = 0
self.gpu_type = None
self.gpus_per_node = kwargs.get("gpus_per_node") or from_json.get(
"gpus_per_node", 0)
if self.gpus_per_node > 0:
self.gpu_type = kwargs.get("gpu_type") or from_json.get(
"gpu_type", "nvidia-tesla-k80")
self.num_gpus = self.gpus_per_node * self.num_nodes
# size of single disks (one per node)
self.disk_size = kwargs.get("disk_size") or from_json.get(
"disk_size", 100)
self.location = kwargs.get("location") or from_json.get("location")
# add information from running clusters
if "running_clusters" in kwargs:
self.instances, self.primary_name = util.get_compute_instance_specs(
self.name_hyphenated)
self.started = True if self.instances else False
# cluster is not running yet
else:
self.instances = None
self.primary_name = None
self.started = False
self.deleted = False # is terminated or being shut down right now?
