def _finalize_experiment(
experiment_json,
metric,
app_id,
run_id,
state,
duration,
logdir,
best_logdir,
optimization_key,
):
"""Attaches the experiment outcome as xattr metadata to the app directory.
"""
outputs = _build_summary_json(logdir)
if outputs:
hopshdfs.dump(outputs, logdir + "/.summary.json")
if best_logdir:
experiment_json["bestDir"] = best_logdir[len(hopshdfs.project_path()):]
experiment_json["optimizationKey"] = optimization_key
experiment_json["metric"] = metric
experiment_json["state"] = state
experiment_json["duration"] = duration
experiment_utils._attach_experiment_xattr(app_id, run_id, experiment_json,
"REPLACE")
def _handle_return_simple(retval, hdfs_exec_logdir, logfile):
"""
Args:
val:
hdfs_exec_logdir:
Returns:
"""
return_file = hdfs_exec_logdir + '/.outputs.json'
if not retval:
if logfile is not None:
retval = {'log': logfile}
hdfs.dump(dumps(retval), return_file)
return
_upload_file_output(retval, hdfs_exec_logdir)
# Validation
if type(retval) is not dict:
try:
retval = {'metric': retval}
except:
pass
retval['log'] = hdfs_exec_logdir.replace(hdfs.project_path(),
'') + '/output.log'
hdfs.dump(dumps(retval), return_file)
def write_featureframe(self):
"""
Writes a dataframe of data as a training dataset on HDFS in the hdf5 format
Returns:
None
Raises:
:ValueError: if the user supplied a write mode that is not supported
:HDF5DatasetFormatNotSupportedForExternalTrainingDatasets: if the user tries to write an
external training dataset in the .hdf5 format.
"""
if self.training_dataset.training_dataset_type == constants.REST_CONFIG.JSON_TRAINING_DATASET_EXTERNAL_TYPE:
raise HDF5DatasetFormatNotSupportedForExternalTrainingDatasets("The .hdf5 dataset format is not "
"supported for external training datasets.")
if (self.write_mode == constants.SPARK_CONFIG.SPARK_APPEND_MODE):
raise ValueError(
"Append is not supported for training datasets stored in .hdf5 format, only overwrite, "
"set the optional argument write_mode='overwrite'")
if not isinstance(self.df, np.ndarray):
if isinstance(self.df, DataFrame) or isinstance(self.df, RDD):
df = np.array(self.df.collect())
if isinstance(self.df, pd.DataFrame):
df = df.values
if isinstance(self.df, list):
df = np.array(self.df)
tf = TemporaryFile()
tf.seek(0)
hdf5_file = h5py.File(tf)
tf.seek(0)
hdf5_file.create_dataset(self.training_dataset.name, data=df)
tf.seek(0)
hdf5_file.close()
tf.seek(0)
hdfs.dump(tf.read(), self.path + constants.FEATURE_STORE.TRAINING_DATASET_HDF5_SUFFIX)
def init_logger(self, trial_log_file):
"""Initializes the trial log file
"""
self.trial_log_file = trial_log_file
# Open trial log file descriptor
if not hopshdfs.exists(self.trial_log_file):
hopshdfs.dump("", self.trial_log_file)
self.trial_fd = hopshdfs.open_file(self.trial_log_file, flags="w")
def _create_experiment_subdirectories(app_id,
run_id,
param_string,
type,
sub_type=None,
params=None):
"""
Creates directories for an experiment, if Experiments folder exists it will create directories
below it, otherwise it will create them in the Logs directory.
Args:
:app_id: YARN application ID of the experiment
:run_id: Experiment ID
:param_string: name of the new directory created under parent directories
:type: type of the new directory parent, e.g differential_evolution
:sub_type: type of sub directory to parent, e.g generation
:params: dict of hyperparameters
Returns:
The new directories for the yarn-application and for the execution (hdfs_exec_logdir, hdfs_appid_logdir)
"""
pyhdfs_handle = hdfs.get()
hdfs_events_parent_dir = hdfs.project_path() + "Experiments"
hdfs_experiment_dir = hdfs_events_parent_dir + "/" + app_id + "_" + str(
run_id)
# determine directory structure based on arguments
if sub_type:
hdfs_exec_logdir = hdfs_experiment_dir + "/" + str(
sub_type) + '/' + str(param_string)
if pyhdfs_handle.exists(hdfs_exec_logdir):
hdfs.delete(hdfs_exec_logdir, recursive=True)
elif not param_string and not sub_type:
if pyhdfs_handle.exists(hdfs_experiment_dir):
hdfs.delete(hdfs_experiment_dir, recursive=True)
hdfs_exec_logdir = hdfs_experiment_dir + '/'
else:
hdfs_exec_logdir = hdfs_experiment_dir + '/' + str(param_string)
if pyhdfs_handle.exists(hdfs_exec_logdir):
hdfs.delete(hdfs_exec_logdir, recursive=True)
# Need to remove directory if it exists (might be a task retry)
# create the new directory
pyhdfs_handle.create_directory(hdfs_exec_logdir)
return_file = hdfs_exec_logdir + '/.hparams.json'
hdfs.dump(dumps(params), return_file)
return hdfs_exec_logdir, hdfs_experiment_dir
def initialize_logger(self, exp_dir):
"""Initialize logger of optimizer
:param exp_dir: path of experiment directory
:rtype exp_dir: str
"""
# configure logger
self.log_file = exp_dir + "/pruner.log"
if not hdfs.exists(self.log_file):
hdfs.dump("", self.log_file)
self.fd = hdfs.open_file(self.log_file, flags="w")
self._log("Initialized Pruner Logger")
def _handle_return(retval, hdfs_exec_logdir, optimization_key, logfile):
"""
Args:
val:
hdfs_exec_logdir:
Returns:
"""
_upload_file_output(retval, hdfs_exec_logdir)
# Validation
if not optimization_key and type(retval) is dict and len(
retval.keys()) > 1:
raise Exception(
'Missing optimization_key argument, when returning multiple values in a dict the optimization_key argument must be set.'
)
elif type(retval) is dict and optimization_key not in retval and len(
retval.keys()) >= 1:
raise Exception(
'optimization_key not in returned dict, when returning multiple values in a dict the optimization_key argument must be set to indicate which key the optimization algorithm should maximize or minimize on.'
)
elif type(retval) is dict and len(retval.keys()) == 0:
raise Exception(
'Returned dict is empty, must contain atleast 1 metric to maximize or minimize.'
)
# Validate that optimization_key is a number
if type(retval) is dict and len(retval.keys()) > 1:
opt_val = retval[optimization_key]
opt_val = _validate_optimization_value(opt_val)
retval[optimization_key] = opt_val
elif type(retval) is dict and len(retval.keys()) == 1:
opt_val = retval[list(retval.keys())[0]]
opt_val = _validate_optimization_value(opt_val)
retval[list(retval.keys())[0]] = opt_val
else:
opt_val = _validate_optimization_value(retval)
retval = {'metric': opt_val}
retval['log'] = logfile
return_file = hdfs_exec_logdir + '/.outputs.json'
hdfs.dump(dumps(retval), return_file)
metric_file = hdfs_exec_logdir + '/.metric'
hdfs.dump(str(opt_val), metric_file)
def _store_tf_record_schema_hdfs(tfrecord_schema, hdfs_path):
"""
Stores a tfrecord json schema to HDFS
Args:
:tfrecord_schema: the tfrecord schema to store
:hdfs_path: the hdfs path to store it
Returns:
None
"""
json_str = json.dumps(tfrecord_schema)
hdfs.dump(
json_str, hdfs_path + constants.DELIMITERS.SLASH_DELIMITER +
constants.FEATURE_STORE.TRAINING_DATASET_TF_RECORD_SCHEMA_FILE_NAME)
def _finalize_experiment(experiment_json, metric, app_id, run_id, state,
duration, logdir, bestLogdir, optimization_key):
summary_file = _build_summary_json(logdir)
if summary_file:
hdfs.dump(summary_file, logdir + '/.summary.json')
if bestLogdir:
experiment_json['bestDir'] = bestLogdir[len(hdfs.project_path()):]
experiment_json['optimizationKey'] = optimization_key
experiment_json['metric'] = metric
experiment_json['state'] = state
experiment_json['duration'] = duration
_attach_experiment_xattr(app_id, run_id, experiment_json, 'REPLACE')
def __init__(self, log_file, partition_id, task_attempt, print_executor):
self.metric = None
self.lock = threading.RLock()
self.stop = False
self.trial_id = None
self.trial_log_file = None
self.logs = ""
self.log_file = log_file
self.partition_id = partition_id
self.task_attempt = task_attempt
self.print_executor = print_executor
# Open executor log file descriptor
# This log is for all maggy system related log messages
if not hopshdfs.exists(log_file):
hopshdfs.dump("", log_file)
self.fd = hopshdfs.open_file(log_file, flags="w")
self.trial_fd = None
def _handle_return_val(return_val, log_dir, optimization_key, log_file):
"""Handles the return value of the user defined training function.
"""
experiment_utils._upload_file_output(return_val, log_dir)
# Return type validation
if not optimization_key:
raise ValueError("Optimization key cannot be None.")
if not return_val:
raise exceptions.ReturnTypeError(optimization_key, return_val)
if not isinstance(return_val, constants.USER_FCT.RETURN_TYPES):
raise exceptions.ReturnTypeError(optimization_key, return_val)
if isinstance(return_val, dict) and optimization_key not in return_val:
raise KeyError(
"Returned dictionary does not contain optimization key with the "
"provided name: {}".format(optimization_key))
# validate that optimization metric is numeric
if isinstance(return_val, dict):
opt_val = return_val[optimization_key]
else:
opt_val = return_val
return_val = {optimization_key: opt_val}
if not isinstance(opt_val, constants.USER_FCT.NUMERIC_TYPES):
raise exceptions.MetricTypeError(optimization_key, opt_val)
# for key, value in return_val.items():
# return_val[key] = value if isinstance(value, str) else str(value)
return_val["log"] = log_file.replace(hopshdfs.project_path(), "")
return_file = log_dir + "/.outputs.json"
hopshdfs.dump(json.dumps(return_val, default=json_default_numpy),
return_file)
metric_file = log_dir + "/.metric"
hopshdfs.dump(json.dumps(opt_val, default=json_default_numpy), metric_file)
return opt_val
def _handle_return_simple(retval, hdfs_exec_logdir, logfile):
"""
Args:
val:
hdfs_exec_logdir:
Returns:
"""
return_file = hdfs_exec_logdir + '/.outputs.json'
_upload_file_output(retval, hdfs_exec_logdir)
# Validation
if type(retval) is not dict:
try:
retval = {'metric': retval}
except:
pass
retval['log'] = logfile
hdfs.dump(dumps(retval), return_file)
def dump(self, data, hdfs_path):
return hopshdfs.dump(data, hdfs_path)
def _register(hdfs_exec_dir, endpoint_dir, exec_num, local_logdir=False):
"""
Args:
hdfs_exec_dir:
endpoint_dir:
exec_num:
local_logdir:
Returns:
"""
global tb_pid
if tb_pid != 0:
subprocess.Popen(["kill", str(tb_pid)])
_reset_global()
global events_logdir
events_logdir = hdfs_exec_dir
global local_logdir_bool
local_logdir_bool = local_logdir
if tb_pid == 0:
global pypath
pypath = os.getenv("PYSPARK_PYTHON")
#find free port
tb_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
tb_socket.bind(('', 0))
global tb_port
tb_addr, tb_port = tb_socket.getsockname()
global tb_path
tb_path = experiment_utils._find_tensorboard()
tb_socket.close()
tb_env = _init_tb_env()
global local_logdir_path
if local_logdir:
local_logdir_path = os.getcwd() + '/local_logdir'
if os.path.exists(local_logdir_path):
shutil.rmtree(local_logdir_path)
os.makedirs(local_logdir_path)
else:
os.makedirs(local_logdir_path)
local_logdir_path = local_logdir_path + '/'
tb_proc = subprocess.Popen(
[
pypath, tb_path,
"--logdir=%s" % local_logdir_path,
"--port=%d" % tb_port,
"--host=%s" % "0.0.0.0"
],
env=tb_env,
preexec_fn=util._on_executor_exit('SIGTERM'))
else:
tb_proc = subprocess.Popen(
[
pypath, tb_path,
"--logdir=%s" % events_logdir,
"--port=%d" % tb_port,
"--host=%s" % "0.0.0.0"
],
env=tb_env,
preexec_fn=util._on_executor_exit('SIGTERM'))
tb_pid = tb_proc.pid
host = socket.getfqdn()
global tb_url
tb_url = "http://{0}:{1}".format(host, tb_port)
global endpoint
endpoint = endpoint_dir + "/TensorBoard.task" + str(exec_num)
#dump tb host:port to hdfs
hopshdfs.dump(tb_url, endpoint)
return endpoint, tb_pid
def _wrapper_fun(iter):
"""
Wraps the user supplied training function in order to be passed to the
Spark Executors.
Args:
iter:
Returns:
"""
experiment_utils._set_ml_id(app_id, run_id)
# get task context information to determine executor identifier
partition_id, task_attempt = util.get_partition_attempt_id()
client = rpc.Client(server_addr, partition_id, task_attempt,
hb_interval, secret)
log_file = (log_dir + "/executor_" + str(partition_id) + "_" +
str(task_attempt) + ".log")
# save the builtin print
original_print = __builtin__.print
reporter = Reporter(log_file, partition_id, task_attempt,
original_print)
def maggy_print(*args, **kwargs):
"""Maggy custom print() function."""
original_print(*args, **kwargs)
reporter.log(" ".join(str(x) for x in args), True)
# override the builtin print
__builtin__.print = maggy_print
try:
client_addr = client.client_addr
host_port = client_addr[0] + ":" + str(client_addr[1])
exec_spec = {}
exec_spec["partition_id"] = partition_id
exec_spec["task_attempt"] = task_attempt
exec_spec["host_port"] = host_port
exec_spec["trial_id"] = None
reporter.log("Registering with experiment driver", False)
client.register(exec_spec)
client.start_heartbeat(reporter)
# blocking
trial_id, parameters = client.get_suggestion(reporter)
while not client.done:
if experiment_type == "ablation":
ablation_params = {
"ablated_feature":
parameters.get("ablated_feature", "None"),
"ablated_layer":
parameters.get("ablated_layer", "None"),
}
parameters.pop("ablated_feature")
parameters.pop("ablated_layer")
tb_logdir = log_dir + "/" + trial_id
trial_log_file = tb_logdir + "/output.log"
reporter.set_trial_id(trial_id)
# If trial is repeated, delete trial directory, except log file
if hopshdfs.exists(tb_logdir):
util._clean_dir(tb_logdir, [trial_log_file])
else:
hopshdfs.mkdir(tb_logdir)
reporter.init_logger(trial_log_file)
tensorboard._register(tb_logdir)
if experiment_type == "ablation":
hopshdfs.dump(
json.dumps(ablation_params,
default=util.json_default_numpy),
tb_logdir + "/.hparams.json",
)
else:
hopshdfs.dump(
json.dumps(parameters,
default=util.json_default_numpy),
tb_logdir + "/.hparams.json",
)
try:
reporter.log("Starting Trial: {}".format(trial_id), False)
reporter.log("Trial Configuration: {}".format(parameters),
False)
if experiment_type == "optimization":
tensorboard._write_hparams(parameters, trial_id)
sig = inspect.signature(map_fun)
if sig.parameters.get("reporter", None):
retval = map_fun(**parameters, reporter=reporter)
else:
retval = map_fun(**parameters)
if experiment_type == "optimization":
tensorboard._write_session_end()
retval = util._handle_return_val(retval, tb_logdir,
optimization_key,
trial_log_file)
except exceptions.EarlyStopException as e:
retval = e.metric
reporter.log("Early Stopped Trial.", False)
reporter.log("Finished Trial: {}".format(trial_id), False)
reporter.log("Final Metric: {}".format(retval), False)
client.finalize_metric(retval, reporter)
# blocking
trial_id, parameters = client.get_suggestion(reporter)
except: # noqa: E722
reporter.log(traceback.format_exc(), False)
raise
finally:
reporter.close_logger()
client.stop()
client.close()
def __init__(self, experiment_type, **kwargs):
global driver_secret
# COMMON EXPERIMENT SETUP
self._final_store = []
self._trial_store = {}
self.num_executors = kwargs.get("num_executors")
self._message_q = queue.Queue()
self.name = kwargs.get("name")
self.experiment_done = False
self.worker_done = False
self.hb_interval = kwargs.get("hb_interval")
self.description = kwargs.get("description")
self.experiment_type = experiment_type
self.es_interval = kwargs.get("es_interval")
self.es_min = kwargs.get("es_min")
# TYPE-SPECIFIC EXPERIMENT SETUP
if self.experiment_type == "optimization":
# set up an optimization experiment
self.num_trials = kwargs.get("num_trials", 1)
searchspace = kwargs.get("searchspace")
if isinstance(searchspace, Searchspace):
self.searchspace = searchspace
elif searchspace is None:
self.searchspace = Searchspace()
else:
raise Exception(
"The experiment's search space should be an instance of maggy.Searchspace, "
"but it is {0} (of type '{1}').".format(
str(searchspace),
type(searchspace).__name__))
optimizer = kwargs.get("optimizer")
if optimizer is None:
if len(self.searchspace.names()) == 0:
self.optimizer = SingleRun()
else:
raise Exception(
"Searchspace has to be empty or None to use without optimizer"
)
elif isinstance(optimizer, str):
if optimizer.lower() == "randomsearch":
self.optimizer = RandomSearch()
elif optimizer.lower() == "asha":
self.optimizer = Asha()
elif optimizer.lower() == "none":
if len(self.searchspace.names()) == 0:
self.optimizer = SingleRun()
else:
raise Exception(
"Searchspace has to be empty or None to use without Optimizer."
)
else:
raise Exception(
"Unknown Optimizer. Can't initialize experiment driver."
)
elif isinstance(optimizer, AbstractOptimizer):
self.optimizer = optimizer
print("Custom Optimizer initialized.")
else:
raise Exception(
"The experiment's optimizer should either be an string indicating the name "
"of an implemented optimizer (such as 'randomsearch') or an instance of "
"maggy.optimizer.AbstractOptimizer, "
"but it is {0} (of type '{1}').".format(
str(optimizer),
type(optimizer).__name__))
direction = kwargs.get("direction", "max")
if isinstance(direction,
str) and direction.lower() in ["min", "max"]:
self.direction = direction.lower()
else:
raise Exception(
"The experiment's direction should be an string (either 'min' or 'max') "
"but it is {0} (of type '{1}').".format(
str(direction),
type(direction).__name__))
es_policy = kwargs.get("es_policy")
if isinstance(es_policy, str):
if es_policy.lower() == "median":
self.earlystop_check = MedianStoppingRule.earlystop_check
elif es_policy.lower() == "none":
self.earlystop_check = NoStoppingRule.earlystop_check
else:
raise Exception(
"The experiment's early stopping policy should either be a string ('median' or 'none') "
"or a custom policy that is an instance of maggy.earlystop.AbstractEarlyStop, "
"but it is {0} (of type '{1}').".format(
str(es_policy),
type(es_policy).__name__))
elif isinstance(es_policy, AbstractEarlyStop):
self.earlystop_check = es_policy.earlystop_check
print("Custom Early Stopping policy initialized.")
else:
raise Exception(
"The experiment's early stopping policy should either be a string ('median' or 'none') "
"or a custom policy that is an instance of maggy.earlystop.AbstractEarlyStop, "
"but it is {0} (of type '{1}').".format(
str(es_policy),
type(es_policy).__name__))
self.es_interval = kwargs.get("es_interval")
self.es_min = kwargs.get("es_min")
self.result = {
"best_val": "n.a.",
"num_trials": 0,
"early_stopped": 0
}
elif self.experiment_type == "ablation":
# set up an ablation study experiment
self.earlystop_check = NoStoppingRule.earlystop_check
ablation_study = kwargs.get("ablation_study")
if isinstance(ablation_study, AblationStudy):
self.ablation_study = ablation_study
else:
raise Exception(
"The experiment's ablation study configuration should be an instance of "
"maggy.ablation.AblationStudy, "
"but it is {0} (of type '{1}').".format(
str(ablation_study),
type(ablation_study).__name__))
searchspace = kwargs.get("searchspace")
if not searchspace:
self.searchspace = Searchspace()
else:
raise Exception(
"The experiment's search space should be None for ablation experiments, "
"but it is {0} (of type '{1}').".format(
str(searchspace),
type(searchspace).__name__))
ablator = kwargs.get("ablator")
if isinstance(ablator, str):
if ablator.lower() == "loco":
self.ablator = LOCO(ablation_study, self._final_store)
self.num_trials = self.ablator.get_number_of_trials()
if self.num_executors > self.num_trials:
self.num_executors = self.num_trials
else:
raise Exception(
"The experiment's ablation study policy should either be a string ('loco') "
"or a custom policy that is an instance of maggy.ablation.ablation.AbstractAblator, "
"but it is {0} (of type '{1}').".format(
str(ablator),
type(ablator).__name__))
elif isinstance(ablator, AbstractAblator):
self.ablator = ablator
print("Custom Ablator initialized. \n")
else:
raise Exception(
"The experiment's ablation study policy should either be a string ('loco') "
"or a custom policy that is an instance of maggy.ablation.ablation.AbstractAblator, "
"but it is {0} (of type '{1}').".format(
str(ablator),
type(ablator).__name__))
self.result = {
"best_val": "n.a.",
"num_trials": 0,
"early_stopped": "n.a"
}
else:
raise Exception(
"Unknown experiment type. experiment_type should be either 'optimization' or 'ablation', "
"but it is {0}.".format(str(self.experiment_type)))
# FINALIZE EXPERIMENT SETUP
self.server = rpc.Server(self.num_executors)
if not driver_secret:
driver_secret = self._generate_secret(
ExperimentDriver.SECRET_BYTES)
self._secret = driver_secret
self.job_start = datetime.now()
self.executor_logs = ""
self.maggy_log = ""
self.log_lock = threading.RLock()
self.log_file = kwargs.get("log_dir") + "/maggy.log"
self.log_dir = kwargs.get("log_dir")
self.exception = None
# Open File desc for HDFS to log
if not hopshdfs.exists(self.log_file):
hopshdfs.dump("", self.log_file)
self.fd = hopshdfs.open_file(self.log_file, flags="w")
def _target_function(self):
try:
time_earlystop_check = (
time.time()
) # only used by earlystop-supporting experiments
while not self.worker_done:
trial = None
# get a message
try:
msg = self._message_q.get_nowait()
except queue.Empty:
msg = {"type": None}
if self.earlystop_check != NoStoppingRule.earlystop_check:
if (time.time() -
time_earlystop_check) >= self.es_interval:
time_earlystop_check = time.time()
# pass currently running trials to early stop component
if len(self._final_store) > self.es_min:
self._log("Check for early stopping.")
try:
to_stop = self.earlystop_check(
self._trial_store,
self._final_store,
self.direction,
)
except Exception as e:
self._log(e)
to_stop = []
if len(to_stop) > 0:
self._log(
"Trials to stop: {}".format(to_stop))
for trial_id in to_stop:
self.get_trial(trial_id).set_early_stop()
# depending on message do the work
# 1. METRIC
if msg["type"] == "METRIC":
# append executor logs if in the message
logs = msg.get("logs", None)
if logs is not None:
with self.log_lock:
self.executor_logs = self.executor_logs + logs
if msg["trial_id"] is not None and msg[
"data"] is not None:
self.get_trial(msg["trial_id"]).append_metric(
msg["data"])
# 2. BLACKLIST the trial
elif msg["type"] == "BLACK":
trial = self.get_trial(msg["trial_id"])
with trial.lock:
trial.status = Trial.SCHEDULED
self.server.reservations.assign_trial(
msg["partition_id"], msg["trial_id"])
# 3. FINAL
elif msg["type"] == "FINAL":
# set status
# get trial only once
trial = self.get_trial(msg["trial_id"])
logs = msg.get("logs", None)
if logs is not None:
with self.log_lock:
self.executor_logs = self.executor_logs + logs
# finalize the trial object
with trial.lock:
trial.status = Trial.FINALIZED
trial.final_metric = msg["data"]
trial.duration = experiment_utils._seconds_to_milliseconds(
time.time() - trial.start)
# move trial to the finalized ones
self._final_store.append(trial)
self._trial_store.pop(trial.trial_id)
# update result dictionary
self._update_result(trial)
# keep for later in case tqdm doesn't work
self.maggy_log = self._update_maggy_log()
self._log(self.maggy_log)
hopshdfs.dump(
trial.to_json(),
self.log_dir + "/" + trial.trial_id +
"/trial.json",
)
# assign new trial
if self.experiment_type == "optimization":
trial = self.optimizer.get_suggestion(trial)
elif self.experiment_type == "ablation":
trial = self.ablator.get_trial(trial)
if trial is None:
self.server.reservations.assign_trial(
msg["partition_id"], None)
self.experiment_done = True
elif trial == "IDLE":
self.add_message({
"type":
"IDLE",
"partition_id":
msg["partition_id"],
"idle_start":
time.time(),
})
self.server.reservations.assign_trial(
msg["partition_id"], None)
else:
with trial.lock:
trial.start = time.time()
trial.status = Trial.SCHEDULED
self.server.reservations.assign_trial(
msg["partition_id"], trial.trial_id)
self.add_trial(trial)
# 4. Let executor be idle
elif msg["type"] == "IDLE":
# execute only every 0.1 seconds but do not block thread
if (self.experiment_type == "optimization"
and time.time() - msg["idle_start"] > 0.1):
trial = self.optimizer.get_suggestion()
if trial is None:
self.server.reservations.assign_trial(
msg["partition_id"], None)
self.experiment_done = True
elif trial == "IDLE":
# reset timeout
msg["idle_start"] = time.time()
self.add_message(msg)
else:
with trial.lock:
trial.start = time.time()
trial.status = Trial.SCHEDULED
self.server.reservations.assign_trial(
msg["partition_id"], trial.trial_id)
self.add_trial(trial)
elif self.experiment_type == "optimization":
self.add_message(msg)
# 4. REG
elif msg["type"] == "REG":
if self.experiment_type == "optimization":
trial = self.optimizer.get_suggestion()
elif self.experiment_type == "ablation":
trial = self.ablator.get_trial()
if trial is None:
self.server.reservations.assign_trial(
msg["partition_id"], None)
self.experiment_done = True
elif trial == "IDLE":
# reset timeout
msg["idle_start"] = time.time()
self.add_message(msg)
else:
with trial.lock:
trial.start = time.time()
trial.status = Trial.SCHEDULED
self.server.reservations.assign_trial(
msg["partition_id"], trial.trial_id)
self.add_trial(trial)
except Exception as exc:
# Exception can't be propagated to parent thread
# therefore log the exception and fail experiment
self._log(exc)
self.exception = exc
self.server.stop()
def finalize(self, job_end):
results = ""
if self.experiment_type == "optimization":
_ = self.optimizer.finalize_experiment(self._final_store)
self.job_end = job_end
self.duration = experiment_utils._seconds_to_milliseconds(
self.job_end - self.job_start)
self.duration_str = experiment_utils._time_diff(
self.job_start, self.job_end)
results = ("\n------ " + self.optimizer.name() +
" Results ------ direction(" + self.direction + ") \n"
"BEST combination " +
json.dumps(self.result["best_hp"]) + " -- metric " +
str(self.result["best_val"]) + "\n"
"WORST combination " +
json.dumps(self.result["worst_hp"]) + " -- metric " +
str(self.result["worst_val"]) + "\n"
"AVERAGE metric -- " + str(self.result["avg"]) + "\n"
"EARLY STOPPED Trials -- " +
str(self.result["early_stopped"]) + "\n"
"Total job time " + self.duration_str + "\n")
elif self.experiment_type == "ablation":
_ = self.ablator.finalize_experiment(self._final_store)
self.job_end = job_end
self.duration = experiment_utils._seconds_to_milliseconds(
self.job_end - self.job_start)
self.duration_str = experiment_utils._time_diff(
self.job_start, self.job_end)
results = ("\n------ " + self.ablator.name() +
" Results ------ \n" + "BEST Config Excludes " +
json.dumps(self.result["best_config"]) + " -- metric " +
str(self.result["best_val"]) + "\n" +
"WORST Config Excludes " +
json.dumps(self.result["worst_config"]) +
" -- metric " + str(self.result["worst_val"]) + "\n" +
"AVERAGE metric -- " + str(self.result["avg"]) + "\n" +
"Total Job Time " + self.duration_str + "\n")
print(results)
self._log(results)
hopshdfs.dump(
json.dumps(self.result, default=util.json_default_numpy),
self.log_dir + "/result.json",
)
sc = hopsutil._find_spark().sparkContext
hopshdfs.dump(self.json(sc), self.log_dir + "/maggy.json")
return self.result