def _wrapper_fun(iter):
"""
Args:
iter:
Returns:
"""
for i in iter:
executor_num = i
experiment_utils._set_ml_id(app_id, run_id)
tb_hdfs_path = ''
hdfs_exec_logdir = ''
t = threading.Thread(target=devices._print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t.start()
try:
#Arguments
if args_dict:
param_string, params, args = experiment_utils.build_parameters(
map_fun, executor_num, args_dict)
hdfs_exec_logdir, hdfs_appid_logdir = experiment_utils._create_experiment_subdirectories(
app_id,
run_id,
param_string,
'random_search',
params=params)
logfile = experiment_utils._init_logger(hdfs_exec_logdir)
tb_hdfs_path, tb_pid = tensorboard._register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir)
print(devices._get_gpu_info())
print(
'-------------------------------------------------------')
print('Started running task ' + param_string)
task_start = time.time()
retval = map_fun(*args)
task_end = time.time()
experiment_utils._handle_return(retval, hdfs_exec_logdir,
optimization_key, logfile)
time_str = 'Finished task ' + param_string + ' - took ' + experiment_utils._time_diff(
task_start, task_end)
print(time_str)
print('Returning metric ' + str(retval))
print(
'-------------------------------------------------------')
except:
raise
finally:
experiment_utils._cleanup(tensorboard, t)
def _wrapper_fun(iter):
"""
Args:
:iter:
Returns:
"""
for i in iter:
executor_num = i
experiment_utils._set_ml_id(app_id, run_id)
tb_hdfs_path = ''
hdfs_exec_logdir = ''
t = threading.Thread(target=devices._print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t.start()
global local_logdir_bool
try:
#Arguments
if args_dict:
param_string, params, args = experiment_utils.build_parameters(map_fun, executor_num, args_dict)
val = _get_return_file(param_string, app_id, generation_id, run_id)
hdfs_exec_logdir, hdfs_appid_logdir = experiment_utils._create_experiment_subdirectories(app_id, run_id, param_string, 'differential_evolution', sub_type='generation.' + str(generation_id), params=params)
logfile = experiment_utils._init_logger(hdfs_exec_logdir)
tb_hdfs_path, tb_pid = tensorboard._register(hdfs_exec_logdir, hdfs_appid_logdir, executor_num, local_logdir=local_logdir_bool)
print(devices._get_gpu_info())
print('-------------------------------------------------------')
print('Started running task ' + param_string)
if val is not None:
val = json.loads(val)
task_start = time.time()
if val is None:
val = map_fun(*args)
task_end = time.time()
time_str = 'Finished task ' + param_string + ' - took ' + experiment_utils._time_diff(task_start, task_end)
print(time_str)
experiment_utils._handle_return(val, hdfs_exec_logdir, opt_key, logfile)
print('Returning metric ' + str(val))
print('-------------------------------------------------------')
except:
raise
finally:
experiment_utils._cleanup(tensorboard, t)
def _wrapper_fun(iter):
"""
Args:
iter:
Returns:
"""
for i in iter:
executor_num = i
experiment_utils._set_ml_id(app_id, run_id)
t = threading.Thread(target=devices._print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t.start()
is_chief = False
logdir = None
tb_hdfs_path = None
try:
host = experiment_utils._get_ip_address()
tmp_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
tmp_socket.bind(('', 0))
port = tmp_socket.getsockname()[1]
client = allreduce_reservation.Client(server_addr)
host_port = host + ":" + str(port)
client.register({"worker": host_port, "index": executor_num})
cluster = client.await_reservations()
tmp_socket.close()
client.close()
task_index = experiment_utils._find_index(host_port, cluster)
if task_index == -1:
cluster["task"] = {"type": "chief", "index": 0}
else:
cluster["task"] = {"type": "worker", "index": task_index}
evaluator_node = None
if evaluator:
last_worker_index = len(cluster["cluster"]["worker"]) - 1
evaluator_node = cluster["cluster"]["worker"][
last_worker_index]
cluster["cluster"]["evaluator"] = [evaluator_node]
del cluster["cluster"]["worker"][last_worker_index]
if evaluator_node == host_port:
cluster["task"] = {"type": "evaluator", "index": 0}
print('TF_CONFIG: {} '.format(cluster))
if num_executors > 1:
os.environ["TF_CONFIG"] = json.dumps(cluster)
is_chief = (cluster["task"]["type"] == "chief")
is_evaluator = (cluster["task"]["type"] == "evaluator")
if is_chief:
logdir = experiment_utils._get_logdir(app_id, run_id)
tb_hdfs_path, tb_pid = tensorboard._register(
logdir, logdir, executor_num, local_logdir=local_logdir)
elif is_evaluator:
logdir = experiment_utils._get_logdir(app_id, run_id)
tensorboard.events_logdir = logdir
logfile = experiment_utils._init_logger(
experiment_utils._get_logdir(app_id, run_id),
role=cluster["task"]["type"],
index=cluster["task"]["index"])
print(devices._get_gpu_info())
print('-------------------------------------------------------')
print('Started running task')
task_start = time.time()
retval = map_fun()
if is_chief:
experiment_utils._handle_return_simple(
retval, experiment_utils._get_logdir(app_id, run_id),
logfile)
task_end = time.time()
time_str = 'Finished task - took ' + experiment_utils._time_diff(
task_start, task_end)
print(time_str)
print('-------------------------------------------------------')
except:
raise
finally:
experiment_utils._cleanup(tensorboard, t)
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
def _wrapper_fun(iter):
"""
Args:
iter:
Returns:
"""
for i in iter:
executor_num = i
experiment_utils._set_ml_id(app_id, run_id)
t = threading.Thread(target=devices._print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t.start()
role = None
logdir = None
tb_hdfs_path = None
client = parameter_server_reservation.Client(server_addr)
try:
host = experiment_utils._get_ip_address()
tmp_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
tmp_socket.bind(('', 0))
port = tmp_socket.getsockname()[1]
host_port = host + ":" + str(port)
exec_spec = {}
if executor_num < num_ps:
exec_spec["task_type"] = "ps"
else:
exec_spec["task_type"] = "worker"
exec_spec["host_port"] = host_port
exec_spec["gpus_present"] = devices.get_num_gpus() > 0
client.register(exec_spec)
cluster = client.await_reservations()
tmp_socket.close()
role, index = experiment_utils._find_task_and_index(host_port, cluster)
cluster_spec = {}
cluster_spec["cluster"] = cluster
cluster_spec["task"] = {"type": role, "index": index}
evaluator_node = None
if evaluator:
last_worker_index = len(cluster_spec["cluster"]["worker"])-1
evaluator_node = cluster_spec["cluster"]["worker"][last_worker_index]
cluster_spec["cluster"]["evaluator"] = [evaluator_node]
del cluster_spec["cluster"]["worker"][last_worker_index]
if evaluator_node == host_port:
role = "evaluator"
cluster_spec["task"] = {"type": "evaluator", "index": 0}
print('TF_CONFIG: {} '.format(cluster_spec))
os.environ["TF_CONFIG"] = json.dumps(cluster_spec)
logfile = experiment_utils._init_logger(experiment_utils._get_logdir(app_id, run_id), role=role, index=cluster_spec["task"]["index"])
dist_logdir = experiment_utils._get_logdir(app_id, run_id) + '/logdir'
is_chief = (cluster["task"]["type"] == "chief")
if is_chief:
hdfs.mkdir(dist_logdir)
tensorboard._register(dist_logdir, experiment_utils._get_logdir(app_id, run_id), executor_num, local_logdir=local_logdir)
else:
tensorboard.events_logdir = dist_logdir
print(devices._get_gpu_info())
print('-------------------------------------------------------')
print('Started running task')
task_start = time.time()
retval=None
if role == "ps":
ps_thread = threading.Thread(target=lambda: map_fun())
ps_thread.start()
client.await_all_workers_finished()
else:
retval = map_fun()
if role == "chief":
experiment_utils._handle_return_simple(retval, experiment_utils._get_logdir(app_id, run_id), logfile)
task_end = time.time()
time_str = 'Finished task - took ' + experiment_utils._time_diff(task_start, task_end)
print(time_str)
print('-------------------------------------------------------')
except:
raise
finally:
if role != "ps":
client.register_worker_finished()
client.close()
experiment_utils._cleanup(tensorboard, t)