def _wrapper_fun(iter):
"""
Args:
iter:
Returns:
"""
for i in iter:
executor_num = i
tb_hdfs_path = ''
hdfs_exec_logdir = ''
t = threading.Thread(target=devices._print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t.start()
task_index = None
try:
host = util._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 = _find_index(host_port, cluster)
cluster["task"] = {"type": "worker", "index": task_index}
os.environ["TF_CONFIG"] = json.dumps(cluster)
if task_index == 0:
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs._create_directories(
app_id, run_id, None, 'collective_all_reduce')
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs._init_logger()
tb_hdfs_path, tb_pid = tensorboard._register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir)
gpu_str = '\nChecking for GPUs in the environment' + devices._get_gpu_info(
)
if task_index == 0:
hopshdfs.log(gpu_str)
print(gpu_str)
print('-------------------------------------------------------')
print('Started running task \n')
if task_index == 0:
hopshdfs.log('Started running task')
task_start = datetime.datetime.now()
retval = map_fun()
if task_index == 0:
if retval:
_handle_return(retval, hdfs_exec_logdir)
task_end = datetime.datetime.now()
time_str = 'Finished task - took ' + util._time_diff(
task_start, task_end)
print('\n' + time_str)
print('-------------------------------------------------------')
if task_index == 0:
hopshdfs.log(time_str)
except:
#Always do cleanup
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
raise
finally:
if task_index == 0:
if local_logdir:
local_tb = tensorboard.local_logdir_path
util._store_local_tensorboard(local_tb, hdfs_exec_logdir)
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
def _wrapper_fun(iter):
"""
Args:
iter:
Returns:
"""
for i in iter:
executor_num = i
tb_pid = 0
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:
argcount = six.get_function_code(map_fun).co_argcount
names = six.get_function_code(map_fun).co_varnames
args = []
argIndex = 0
param_string = ''
while argcount > 0:
#Get args for executor and run function
param_name = names[argIndex]
param_val = args_dict[param_name][executor_num]
param_string += str(param_name) + '=' + str(
param_val) + '.'
args.append(param_val)
argcount -= 1
argIndex += 1
param_string = param_string[:-1]
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs._create_directories(
app_id, run_id, param_string, 'launcher')
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs._init_logger()
tb_hdfs_path, tb_pid = tensorboard._register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir)
gpu_str = '\nChecking for GPUs in the environment' + devices._get_gpu_info(
)
hopshdfs.log(gpu_str)
print(gpu_str)
print(
'-------------------------------------------------------')
print('Started running task ' + param_string + '\n')
hopshdfs.log('Started running task ' + param_string)
task_start = datetime.datetime.now()
map_fun(*args)
task_end = datetime.datetime.now()
time_str = 'Finished task ' + param_string + ' - took ' + util._time_diff(
task_start, task_end)
print('\n' + time_str)
print(
'-------------------------------------------------------')
hopshdfs.log(time_str)
else:
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs._create_directories(
app_id, run_id, None, 'launcher')
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs._init_logger()
tb_hdfs_path, tb_pid = tensorboard._register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir)
gpu_str = '\nChecking for GPUs in the environment' + devices._get_gpu_info(
)
hopshdfs.log(gpu_str)
print(gpu_str)
print(
'-------------------------------------------------------')
print('Started running task\n')
hopshdfs.log('Started running task')
task_start = datetime.datetime.now()
retval = map_fun()
task_end = datetime.datetime.now()
if retval:
_handle_return(retval, hdfs_exec_logdir)
time_str = 'Finished task - took ' + util._time_diff(
task_start, task_end)
print('\n' + time_str)
print(
'-------------------------------------------------------')
hopshdfs.log(time_str)
except:
#Always do cleanup
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
raise
finally:
try:
if local_logdir:
local_tb = tensorboard.local_logdir_path
util._store_local_tensorboard(local_tb, hdfs_exec_logdir)
except:
pass
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
def _grid_launch(sc, map_fun, args_dict, direction='max', local_logdir=False, name="no-name"):
"""
Run the wrapper function with each hyperparameter combination as specified by the dictionary
Args:
sc:
map_fun:
args_dict:
direction:
local_logdir:
name:
Returns:
"""
global run_id
app_id = str(sc.applicationId)
num_executions = 1
if direction != 'max' and direction != 'min':
raise ValueError('Invalid direction ' + direction + ', must be max or min')
arg_lists = list(args_dict.values())
currentLen = len(arg_lists[0])
for i in range(len(arg_lists)):
if currentLen != len(arg_lists[i]):
raise ValueError('Length of each function argument list must be equal')
num_executions = len(arg_lists[i])
#Each TF task should be run on 1 executor
nodeRDD = sc.parallelize(range(num_executions), num_executions)
#Make SparkUI intuitive by grouping jobs
sc.setJobGroup("Grid Search", "{} | Hyperparameter Optimization".format(name))
#Force execution on executor, since GPU is located on executor
job_start = datetime.datetime.now()
nodeRDD.foreachPartition(_prepare_func(app_id, run_id, map_fun, args_dict, local_logdir))
job_end = datetime.datetime.now()
job_time_str = util._time_diff(job_start, job_end)
arg_count = six.get_function_code(map_fun).co_argcount
arg_names = six.get_function_code(map_fun).co_varnames
hdfs_appid_dir = hopshdfs._get_experiments_dir() + '/' + app_id
hdfs_runid_dir = _get_logdir(app_id)
max_val, max_hp, min_val, min_hp, avg = _get_best(args_dict, num_executions, arg_names, arg_count, hdfs_appid_dir, run_id)
param_combination = ""
best_val = ""
if direction == 'max':
param_combination = max_hp
best_val = str(max_val)
results = '\n------ Grid Search results ------ direction(' + direction + ') \n' \
'BEST combination ' + max_hp + ' -- metric ' + str(max_val) + '\n' \
'WORST combination ' + min_hp + ' -- metric ' + str(min_val) + '\n' \
'AVERAGE metric -- ' + str(avg) + '\n' \
'Total job time ' + job_time_str + '\n'
_write_result(hdfs_runid_dir, results)
print(results)
elif direction == 'min':
param_combination = min_hp
best_val = str(min_val)
results = '\n------ Grid Search results ------ direction(' + direction + ') \n' \
'BEST combination ' + min_hp + ' -- metric ' + str(min_val) + '\n' \
'WORST combination ' + max_hp + ' -- metric ' + str(max_val) + '\n' \
'AVERAGE metric -- ' + str(avg) + '\n' \
'Total job time ' + job_time_str + '\n'
_write_result(hdfs_runid_dir, results)
print(results)
print('Finished Experiment \n')
return hdfs_runid_dir, param_combination, best_val
def _wrapper_fun(iter):
"""
Args:
iter:
Returns:
"""
for i in iter:
executor_num = i
tb_hdfs_path = ''
hdfs_exec_logdir = ''
t = threading.Thread(target=devices._print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t.start()
role = None
client = parameter_server_reservation.Client(server_addr)
try:
host = util._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 = _find_task_and_index(host_port, cluster)
cluster_spec = {}
cluster_spec["cluster"] = cluster
cluster_spec["task"] = {"type": role, "index": index}
print(cluster_spec)
os.environ["TF_CONFIG"] = json.dumps(cluster_spec)
if role == "chief":
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs._create_directories(
app_id, run_id, None, 'parameter_server')
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs._init_logger()
tb_hdfs_path, tb_pid = tensorboard._register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir)
gpu_str = '\nChecking for GPUs in the environment' + devices._get_gpu_info(
)
if role == "chief":
hopshdfs.log(gpu_str)
print(gpu_str)
print('-------------------------------------------------------')
print('Started running task \n')
if role == "chief":
hopshdfs.log('Started running task')
task_start = datetime.datetime.now()
retval = None
if role == "ps":
ps_thread = threading.Thread(target=lambda: map_fun())
ps_thread.start()
print("waiting for workers")
client.await_all_workers_finished()
print("waiting finished")
else:
retval = map_fun()
if role == "chief":
if retval:
_handle_return(retval, hdfs_exec_logdir)
task_end = datetime.datetime.now()
time_str = 'Finished task - took ' + util._time_diff(
task_start, task_end)
print('\n' + time_str)
print('-------------------------------------------------------')
if role == "chief":
hopshdfs.log(time_str)
except:
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
raise
finally:
if role == "chief":
if local_logdir:
local_tb = tensorboard.local_logdir_path
util._store_local_tensorboard(local_tb, hdfs_exec_logdir)
try:
if role == "worker" or role == "chief":
client.register_worker_finished()
client.close()
except:
pass
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
def _wrapper_fun(iter):
"""
Args:
:iter:
Returns:
"""
for i in iter:
executor_num = i
tb_pid = 0
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:
argcount = six.get_function_code(map_fun).co_argcount
names = six.get_function_code(map_fun).co_varnames
args = []
argIndex = 0
param_string = ''
while argcount > 0:
#Get args for executor and run function
param_name = names[argIndex]
param_val = args_dict[param_name][executor_num]
param_string += str(param_name) + '=' + str(
param_val) + '.'
args.append(param_val)
argcount -= 1
argIndex += 1
param_string = param_string[:-1]
val = _get_metric(param_string, app_id, generation_id, run_id)
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs._create_directories(
app_id,
run_id,
param_string,
'differential_evolution',
sub_type='generation.' + str(generation_id))
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs._init_logger()
tb_hdfs_path, tb_pid = tensorboard._register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir_bool)
gpu_str = '\nChecking for GPUs in the environment' + devices._get_gpu_info(
)
hopshdfs.log(gpu_str)
print(gpu_str)
print(
'-------------------------------------------------------')
print('Started running task ' + param_string + '\n')
if val:
print('Reading returned metric from previous run: ' +
str(val))
hopshdfs.log('Started running task ' + param_string)
task_start = datetime.datetime.now()
if not val:
val = map_fun(*args)
task_end = datetime.datetime.now()
time_str = 'Finished task ' + param_string + ' - took ' + util._time_diff(
task_start, task_end)
print('\n' + time_str)
hopshdfs.log(time_str)
try:
castval = int(val)
except:
raise ValueError(
'Your function needs to return a metric (number) which should be maximized or minimized'
)
metric_file = hdfs_exec_logdir + '/metric'
fs_handle = hopshdfs.get_fs()
try:
fd = fs_handle.open_file(metric_file, mode='w')
except:
fd = fs_handle.open_file(metric_file, flags='w')
fd.write(str(float(val)).encode())
fd.flush()
fd.close()
print('Returning metric ' + str(val))
print(
'-------------------------------------------------------')
except:
#Always do cleanup
if tb_hdfs_path:
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join(20)
raise
finally:
if local_logdir_bool:
local_tb = tensorboard.local_logdir_path
util._store_local_tensorboard(local_tb, hdfs_exec_logdir)
hopshdfs.log('Finished running')
if tb_hdfs_path:
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join(20)
def _launch(sc,
map_fun,
args_dict,
samples,
direction='max',
local_logdir=False,
name="no-name"):
"""
Args:
sc:
map_fun:
args_dict:
local_logdir:
name:
Returns:
"""
global run_id
app_id = str(sc.applicationId)
arg_lists = list(args_dict.values())
for i in range(len(arg_lists)):
if len(arg_lists[i]) != 2:
raise ValueError(
'Boundary list must contain exactly two elements, [lower_bound, upper_bound] for each hyperparameter'
)
hp_names = args_dict.keys()
random_dict = {}
for hp in hp_names:
lower_bound = args_dict[hp][0]
upper_bound = args_dict[hp][1]
assert lower_bound < upper_bound, "lower bound: " + str(
lower_bound) + " must be less than upper bound: " + str(
upper_bound)
random_values = []
if type(lower_bound) == int and type(upper_bound) == int:
for i in range(samples):
random_values.append(random.randint(lower_bound, upper_bound))
elif type(lower_bound) == float and type(upper_bound) == float:
for i in range(samples):
random_values.append(random.uniform(lower_bound, upper_bound))
else:
raise ValueError('Only float and int is currently supported')
random_dict[hp] = random_values
random_dict, new_samples = _remove_duplicates(random_dict, samples)
sc.setJobGroup("Random Search",
"{} | Hyperparameter Optimization".format(name))
#Each TF task should be run on 1 executor
nodeRDD = sc.parallelize(range(new_samples), new_samples)
job_start = datetime.datetime.now()
nodeRDD.foreachPartition(
_prepare_func(app_id, run_id, map_fun, random_dict, local_logdir))
job_end = datetime.datetime.now()
job_time_str = util._time_diff(job_start, job_end)
arg_count = six.get_function_code(map_fun).co_argcount
arg_names = six.get_function_code(map_fun).co_varnames
hdfs_appid_dir = hopshdfs._get_experiments_dir() + '/' + app_id
hdfs_runid_dir = _get_logdir(app_id)
max_val, max_hp, min_val, min_hp, avg = _get_best(random_dict, new_samples,
arg_names, arg_count,
hdfs_appid_dir, run_id)
param_combination = ""
best_val = ""
if direction == 'max':
param_combination = max_hp
best_val = str(max_val)
results = '\n------ Random Search results ------ direction(' + direction + ') \n' \
'BEST combination ' + max_hp + ' -- metric ' + str(max_val) + '\n' \
'WORST combination ' + min_hp + ' -- metric ' + str(min_val) + '\n' \
'AVERAGE metric -- ' + str(avg) + '\n' \
'Total job time ' + job_time_str + '\n'
_write_result(hdfs_runid_dir, results)
print(results)
elif direction == 'min':
param_combination = min_hp
best_val = str(min_val)
results = '\n------ Random Search results ------ direction(' + direction + ') \n' \
'BEST combination ' + min_hp + ' -- metric ' + str(min_val) + '\n' \
'WORST combination ' + max_hp + ' -- metric ' + str(max_val) + '\n' \
'AVERAGE metric -- ' + str(avg) + '\n' \
'Total job time ' + job_time_str + '\n'
_write_result(hdfs_runid_dir, results)
print(results)
print('Finished Experiment \n')
return hdfs_runid_dir, param_combination, best_val
def _wrapper_fun(iter):
"""
Args:
iter:
Returns:
"""
for i in iter:
executor_num = i
tb_pid = 0
tb_hdfs_path = ''
hdfs_exec_logdir = ''
t = threading.Thread(target=devices._print_periodic_gpu_utilization)
if devices.get_num_gpus() > 0:
t.start()
try:
hdfs_exec_logdir, hdfs_appid_logdir = hopshdfs._create_directories(
app_id, run_id, None, 'mirrored')
pydoop.hdfs.dump('',
os.environ['EXEC_LOGFILE'],
user=hopshdfs.project_user())
hopshdfs._init_logger()
tb_hdfs_path, tb_pid = tensorboard._register(
hdfs_exec_logdir,
hdfs_appid_logdir,
executor_num,
local_logdir=local_logdir)
gpu_str = '\nChecking for GPUs in the environment' + devices._get_gpu_info(
)
hopshdfs.log(gpu_str)
print(gpu_str)
print('-------------------------------------------------------')
print('Started running task\n')
hopshdfs.log('Started running task')
task_start = datetime.datetime.now()
retval = map_fun()
task_end = datetime.datetime.now()
if retval:
_handle_return(retval, hdfs_exec_logdir)
time_str = 'Finished task - took ' + util._time_diff(
task_start, task_end)
print('\n' + time_str)
print('-------------------------------------------------------')
hopshdfs.log(time_str)
except:
#Always do cleanup
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
raise
finally:
try:
if local_logdir:
local_tb = tensorboard.local_logdir_path
util._store_local_tensorboard(local_tb, hdfs_exec_logdir)
except:
pass
_cleanup(tb_hdfs_path)
if devices.get_num_gpus() > 0:
t.do_run = False
t.join()
