def test_JobConfig():
dfX, y, auto_ml_config = get_automl_config()
job_config = JobConfig()
job_config.guess_cv(auto_ml_config)
assert job_config.cv is not None
job_config.guess_scoring(auto_ml_config)
assert isinstance(job_config.scoring, list)
def test_JobConfig(num_only):
dfX, y, auto_ml_config = get_automl_config(num_only)
job_config = JobConfig()
job_config.guess_cv(auto_ml_config)
assert job_config.cv is not None
assert hasattr(job_config.cv, "split")
job_config.guess_scoring(auto_ml_config)
assert isinstance(job_config.scoring, list)
assert hasattr(job_config, "allow_approx_cv")
assert hasattr(job_config, "start_with_default")
assert hasattr(job_config, "do_blocks_search")
assert isinstance(job_config.allow_approx_cv, bool)
assert isinstance(job_config.start_with_default, bool)
assert isinstance(job_config.do_blocks_search, bool)
with pytest.raises(ValueError):
job_config.cv = "this is not a cv"
def test_create_everything_sequentially(num_only, tmpdir):
# DataPersister
data_persister = FolderDataPersister(base_folder=tmpdir)
# Data
dfX, y = loader(num_only)
# Auto Ml Config
auto_ml_config = AutoMlConfig(dfX, y)
auto_ml_config.guess_everything()
assert auto_ml_config
# Job Config
job_config = JobConfig()
job_config.guess_scoring(auto_ml_config)
job_config.guess_cv(auto_ml_config)
assert job_config
# Result Reader
result_reader = AutoMlResultReader(data_persister)
assert result_reader # just verify the object was created
# Auto ml guider
auto_ml_guider = AutoMlModelGuider(result_reader=result_reader,
job_config=job_config,
metric_transformation="default",
avg_metric=True)
assert auto_ml_guider
# Job Controller
job_controller = MlJobManager(
auto_ml_config=auto_ml_config,
job_config=job_config,
auto_ml_guider=auto_ml_guider,
data_persister=data_persister,
seed=None,
)
assert job_controller
# Job Runner
job_runner = MlJobRunner(
dfX=dfX,
y=y,
groups=None,
auto_ml_config=auto_ml_config,
job_config=job_config,
data_persister=data_persister,
seed=None,
)
assert job_runner
### Do one iteration of the job_controller
for i, (temp_job_id,
temp_job_param) in enumerate(job_controller.iterate()):
if i > 0:
break # I need to complete a loop, so I need the break to be AFTER second yield
job_id = temp_job_id
job_param = temp_job_param
assert isinstance(job_id, str)
assert isinstance(job_param, dict)
### retriveve job by worker
for worker_job_id, worker_job_param in job_runner.iterate():
break
assert isinstance(worker_job_id, str)
assert isinstance(worker_job_param, dict)
assert worker_job_id == job_id
assert worker_job_param == job_param
class MlMachineLauncher(object):
""" object to handle the ml machine processes,
it will hold the configurations,
and allow the creation of
* worker(s)
* controller
* creation of results
* ...
"""
class Commands(Enum):
""" enumeration of accepted commands """
run = "run"
init = "init"
worker = "worker"
controller = "controller"
stop = "stop"
result = "result"
def __init__(self, base_folder, name=None, loader=None, set_configs=None):
self.base_folder = base_folder
self.name = name
self.job_runner = None
self.job_controller = None
self.result_reader = None
self.auto_ml_guider = None
self.data_persister = None
self.job_config = None
self.auto_ml_config = None
self.loader = loader
self.set_configs = set_configs
self.dfX = None
self.y = None
self.groups = None
self._seed = None # to initialize the attribute
self._nbworks = None
self.command = None
@staticmethod
def _get_args_parse():
parser = argparse.ArgumentParser(description="Ml Machine launcher")
parser.add_argument(
"command",
help=
"choice among 'run', 'init', 'controller', 'worker', 'stop', 'result'",
type=str)
parser.add_argument("--nbworkers",
"-n",
help="number of workers to start",
type=int,
default=1)
parser.add_argument("--seed",
help="force seed of worker(s) or controllers",
type=int,
default=None)
return parser
def _process_command_arguments(self, parser=None):
""" process the command arguments to save in the manager the command to launch and its parameters """
if parser is None:
parser = self._get_args_parse()
if not hasattr(parser, "parse_args"):
raise TypeError("'parser' should have a 'parse_args' method")
args = parser.parse_args()
if args.command is None:
raise ValueError(
"Unknown command please choose among 'start','stop','controller', 'worker' or 'result'"
)
self.command = self.Commands(args.command.lower())
self._nbworkers = args.nbworkers
self._seed = args.seed
# self._noinit = args.noinit
return self
def execute_processed_command_argument(self, parser=None):
self._process_command_arguments(parser)
if self.command == self.Commands.run:
self.run_command()
elif self.command == self.Commands.init:
self.init_command()
elif self.command == self.Commands.stop:
self.stop_command()
elif self.command == self.Commands.worker:
self.worker_command()
elif self.command == self.Commands.controller:
self.controller_command()
elif self.command == self.Commands.result:
self.result_command()
else:
raise ValueError("Unknown command %s" % self.command.value)
def init_command(self):
""" this command is to initialized a new task WITHOUT any controller and worker.
This command can be executed using the 'init' keyword in the command argument.
It will :
* load data base using self.loader, if dfX and y are not present
* initialize the configurations (if 'initialize_done' is not True)
* initialize the persister
* call 'launcher_modif' to apply use change
* persist everyting so that it can be reloaded later
"""
print("...starting INIT command...")
print("...initialized")
self.initialize()
print("...persist")
self.persist()
def run_command(self):
""" this command is to start a new task WITH controller and worker(s)
It can be executed using the 'start' keyoword in the command argument.
It will :
* load data base using self.loader, if dfX and y are not present
* initialize the configurations
* initialize the persister
* persister everything so that it can be reloaded later
* start a controller in a SUB process
* start 'nbworkers - 1' worker(s) in SUB process(es)
* start the last worker in main process
"""
print("...start command...")
self.init_command()
print("...starting controller in subprocess...")
self.start_controller_subprocess()
if self._nbworkers > 1:
print("...starting %d workers in subprocess..." %
(self._nbworkers - 1))
self.start_worker_subprocess(self._nbworkers - 1)
if self._nbworkers > 0:
print("...start worker in main process...")
self.start_worker()
def worker_command(self):
""" this command is to add new worker(s) to an existing task.
It can be executed using the 'worker' keyword in the command argument.
It will :
* reload configurations and data
* start 'nbworkers - 1' worker(s) in SUB process(es)
* start the last worker in main process
"""
print("...worker command...")
print("...reload...")
self.reload()
if self._nbworkers > 1:
print("...starting workers in subprocess...")
self.start_worker_subprocess(self._nbworkers - 1)
print("...start worker in main process...")
self.start_worker()
def controller_command(self):
""" this command is to start a controller on an existing task.
It can be executed using the 'controller' keyword in the command argument.
It will:
* reload configurations and data
* start a controller in main process
"""
print("...controller command...")
self.reload()
self.start_controller()
def stop_command(self):
""" this command is to stop the automl processes.
It can be executed using the 'stop' keyword in the command argument.
It will:
* create stop file in controller/worker folder to stop them
"""
self.data_persister = FolderDataPersister(base_folder=self.base_folder)
self.data_persister.write("",
key="stop",
path="mljobmanager_workers",
write_type=SavingType.txt)
self.data_persister.write("",
key="stop",
path="mljobrunner_workers",
write_type=SavingType.txt)
def result_command(self):
""" this command is to launch aggregat of result.
It can be executed using the 'result' keyword in the command argument.
It will:
* load result, params and error
* merge them
* save everything into two excel files
"""
self.data_persister = FolderDataPersister(base_folder=self.base_folder)
self.result_reader = AutoMlResultReader(self.data_persister)
df_results = self.result_reader.load_all_results()
df_params = self.result_reader.load_all_params()
df_errors = self.result_reader.load_all_errors()
df_params_other = self.result_reader.load_all_other_params()
df_merged_result = pd.merge(df_params,
df_results,
how="inner",
on="job_id")
df_merged_result = pd.merge(df_merged_result,
df_params_other,
how="inner",
on="job_id")
df_merged_error = pd.merge(df_params,
df_errors,
how="inner",
on="job_id")
# df_merged_result2 = pd.merge( df_params_other, df_results, how = "inner",on = "job_id")
# df_merged_result2 = df_merged_result2.sort_values(by="job_creation_time")
try:
df_merged_result.to_excel(self.base_folder + "/result.xlsx",
index=False)
except OSError:
print("I couldn't save excel file")
try:
df_merged_error.to_excel(self.base_folder + "/result_error.xlsx",
index=False)
except OSError:
print("I couldn't save excel file")
return df_merged_result, df_merged_error
# In[]
def reload(self):
""" method to reload dfX, y, auto_ml_config and job_config """
self.data_persister = FolderDataPersister(base_folder=self.base_folder)
self.job_config = self.data_persister.read(
key="job_config", write_type=SavingType.pickle)
self.auto_ml_config = self.data_persister.read(
key="auto_ml_config", write_type=SavingType.pickle)
self.dfX = self.data_persister.read(key="dfX",
write_type=SavingType.pickle)
self.y = self.data_persister.read(key="y",
write_type=SavingType.pickle)
self.groups = self.data_persister.read(key="groups",
write_type=SavingType.pickle)
def initialize(self):
""" method to initialize auto_ml_config and job_config """
##################################
### ** load data and target ** ###
##################################
if self.dfX is None or self.y is None:
temp = self.loader()
if len(temp) == 2:
self.dfX, self.y = temp
self.groups = None
else:
self.dfX, self.y, self.groups = temp
###########################################
### ** create database configuration ** ###
###########################################
if self.auto_ml_config is None:
self.auto_ml_config = AutoMlConfig(dfX=self.dfX,
y=self.y,
groups=self.groups,
name=self.name)
self.auto_ml_config.guess_everything()
#####################################
### ** create job configuation ** ###
#####################################
if self.job_config is None:
self.job_config = JobConfig()
self.job_config.guess_cv(auto_ml_config=self.auto_ml_config,
n_splits=10)
self.job_config.guess_scoring(auto_ml_config=self.auto_ml_config)
###################################
### ** create data persister ** ###
###################################
if self.data_persister is None:
self.data_persister = FolderDataPersister(
base_folder=self.base_folder)
#################################
### ** apply custom config ** ###
#################################
if self.set_configs is not None:
self.set_configs(self)
def persist(self):
""" method to persist 'auto_ml_config', 'job_config', 'dfX' and 'y' """
##########################################
### ** Persist everything if needed ** ###
##########################################
self.auto_ml_config.dfX = None
self.auto_ml_config.y = None
self.auto_ml_config.groups = None
self.data_persister.write(data=self.job_config,
key="job_config",
write_type=SavingType.pickle)
self.data_persister.write(data=self.auto_ml_config,
key="auto_ml_config",
write_type=SavingType.pickle)
self.data_persister.write(data=self.dfX,
key="dfX",
write_type=SavingType.pickle)
self.data_persister.write(data=self.y,
key="y",
write_type=SavingType.pickle)
self.data_persister.write(data=self.groups,
key="groups",
write_type=SavingType.pickle)
###############
## Controler ##
###############
def _create_controller(self):
""" create a controller object, if it doesn't exist, but doesn't start it """
###############################
### ** Create controller ** ###
###############################
if self.job_controller is None:
self.result_reader = AutoMlResultReader(self.data_persister)
self.auto_ml_guider = AutoMlModelGuider(
result_reader=self.result_reader,
job_config=self.job_config,
metric_transformation="default",
avg_metric=True,
)
self.job_controller = MlJobManager(
auto_ml_config=self.auto_ml_config,
job_config=self.job_config,
auto_ml_guider=self.auto_ml_guider,
data_persister=self.data_persister,
seed=self._seed,
)
return self
def start_controller(self):
""" create a controller and start it in MAIN process """
self._create_controller()
self.job_controller.run()
def start_controller_subprocess(self):
""" create a controller and start it in a SUB process """
self._create_controller()
self.controller_process = Process(target=start_runner,
args=(self.job_controller, ))
self.controller_process.start()
time.sleep(
1
) # to prevent lock issue if everything is created at the same time...
############
## Worker ##
############
def _create_worker(self):
""" create a worker, if it doesn't exist, but doesn't start it """
if self.job_runner is None:
##############################
### ** Create worker(s) ** ###
##############################
self.job_runner = MlJobRunner(
dfX=self.dfX,
y=self.y,
groups=self.groups,
auto_ml_config=self.auto_ml_config,
job_config=self.job_config,
data_persister=self.data_persister,
seed=self._seed,
)
return self
def start_worker(self):
""" create a worker and start it in MAIN process """
self._create_worker()
self.job_runner.run()
def start_worker_subprocess(self, n=1):
""" create a worker and start 'n' copy of it in SUB processes """
self._create_worker()
self.all_job_runner_process = []
for other_worker in range(n):
self.all_job_runner_process.append(
Process(target=start_runner, args=(self.job_runner, )))
for job_runner_process in self.all_job_runner_process:
job_runner_process.start()
time.sleep(1) # small pause to prevent lock mistakes