def test_everything_with_rules(self):
def a_func():
return 0
def b_func():
return 0
def c_func():
return 0
x1 = WorkflowListTask(is_ordered=False,
name="x1",
tasks=["x1__p1", "x1__p2", "x1__p3"])
x2 = WorkflowListTask(is_ordered=False,
name="x2",
tasks=["x2__p1", "x2__p2", "x2__p3"])
start = WorkflowChoiceScenario(name="root", scenarios=[x1, x2])
sampler = {
"x1__p1":
Parameter("x1__p1", [0, 1], "uniform", "float"),
"x1__p2":
Parameter("x1__p2", [1, 2], "choice", "int"),
"x1__p3":
Parameter("x1__p3", ["v", "w", "x", "y", "z"], "choice", "string"),
"x2__p1":
Parameter("x2__p1", ["a", "b", "c"], "choice", "string"),
"x2__p2":
Parameter("x2__p2", [a_func, b_func, c_func], "choice", "int"),
"x2__p3":
Parameter("x2__p3", "lol", "constant", "string"),
}
rules = [
ChildRule(applied_to=["x2__p2"], parent="x2__p1", value=["a",
"c"]),
ValueRule(constraints=[("x1__p2", 1), ("x1__p3", "v")]),
ValueRule(constraints=[("x1__p2", 2), ("x1__p3", "w")])
]
def evaluate(config, bestconfig):
return random.uniform(0, 1)
env = Env(evaluate, scenario=start, sampler=sampler, rules=rules)
mcts = MCTS(env=env)
mcts.run(n=100)
def __init__(self,
environment,
time_budget=3600,
seed=1,
policy_arg={},
exec_dir=None,
verbose=False):
"""Initialization algorithm.
:param environment: environment class extending AbstractEnvironment
:param time_budget: overall time budget
:param seed: random seed
:param policy_arg: specific option for MCTS policy
:param exec_dir: directory to store tmp files
"""
# config logger
self.logger = logging.getLogger('mcts')
self.logger.setLevel(logging.DEBUG)
# execution directory
if exec_dir is None:
exec_dir = tempfile.mkdtemp()
else:
os.makedirs(exec_dir)
hdlr = logging.FileHandler(os.path.join(exec_dir, "mcts.log"),
mode='w')
formatter = logging.Formatter(
'%(asctime)s :: %(levelname)s :: %(funcName)s :: %(message)s')
hdlr.setFormatter(formatter)
self.logger.addHandler(hdlr)
if verbose:
handler = logging.StreamHandler(sys.stdout)
handler.setLevel(logging.DEBUG)
handler.setFormatter(formatter)
self.logger.addHandler(handler)
env = environment
self.mcts = MCTS(env=env,
time_budget=time_budget,
policy_arg=policy_arg,
exec_dir=exec_dir)
np.random.seed(seed)
def __init__(self,
environment,
time_budget=3600,
verbose=False,
exec_dir=None,
bandit_policy=None,
seed=1,
coef_progressive_widening=0.6):
"""Init method.
"""
# config logger
self.logger = logging.getLogger('mcts')
self.logger.setLevel(logging.DEBUG)
# Default bandit policy
if bandit_policy is None:
bandit_policy = {"policy_name": "uct", "c_uct": np.sqrt(2)}
# execution directory
if exec_dir is None:
exec_dir = tempfile.mkdtemp()
else:
os.makedirs(exec_dir)
hdlr = logging.FileHandler(os.path.join(exec_dir, "mcts.log"),
mode='w')
formatter = logging.Formatter(
'%(asctime)s :: %(levelname)s :: %(funcName)s :: %(message)s')
hdlr.setFormatter(formatter)
self.logger.addHandler(hdlr)
if verbose:
handler = logging.StreamHandler(sys.stdout)
handler.setLevel(logging.DEBUG)
handler.setFormatter(formatter)
self.logger.addHandler(handler)
self.mcts = MCTS(env=environment,
time_budget=time_budget,
exec_dir=exec_dir,
bandit_policy=bandit_policy,
coef_progressive_widening=coef_progressive_widening)
np.random.seed(seed)
def init_mcts(self):
def a_func():
return 0
def b_func():
return 0
def c_func():
return 0
x1 = WorkflowListTask(is_ordered=False,
name="x1",
tasks=["x1__p1", "x1__p2"])
x2 = WorkflowListTask(is_ordered=True,
name="x2",
tasks=["x2__p1", "x2__p2"])
start = WorkflowChoiceScenario(name="root", scenarios=[x1, x2])
sampler = {
"x1__p1":
Parameter("x1__p1", [0, 1], "uniform", "float"),
"x1__p2":
Parameter("x1__p2", [1, 2], "choice", "int"),
"x2__p1":
Parameter("x2__p1", ["a", "b", "c"], "choice", "string"),
"x2__p2":
Parameter("x2__p2", [a_func, b_func, c_func], "choice", "int")
}
env = Env(a_func, scenario=start, sampler=sampler)
def evaluate(config, bestconfig):
return random.uniform(0, 1)
Env.evaluate = evaluate
return MCTS(env=env)
class Search:
"""
Search optimal pipeline using Monte-Carlo Tree Search
Parameters:
----------
environment: object
environment class extending AbstractEnvironment
time_budget: int
overall time budget
seed: int
random seed
bandit_policy: dict
bandit policy used in MCTS. Available choice are uct, besa, puct.
Example {"policy_name": "uct", "c_ub": 1.41}, {"policy_name": "besa"}
exec_dir: str
directory to store tmp files
Attributes
----------
logger: class <logging>
Logger used
mcts : class <mosaic.MCTS>
object that run MCTS algorithm
"""
def __init__(self,
environment,
time_budget=3600,
verbose=False,
exec_dir=None,
bandit_policy=None,
seed=1,
coef_progressive_widening=0.6):
"""Init method.
"""
# config logger
self.logger = logging.getLogger('mcts')
self.logger.setLevel(logging.DEBUG)
# Default bandit policy
if bandit_policy is None:
bandit_policy = {"policy_name": "uct", "c_uct": np.sqrt(2)}
# execution directory
if exec_dir is None:
exec_dir = tempfile.mkdtemp()
else:
os.makedirs(exec_dir)
hdlr = logging.FileHandler(os.path.join(exec_dir, "mcts.log"),
mode='w')
formatter = logging.Formatter(
'%(asctime)s :: %(levelname)s :: %(funcName)s :: %(message)s')
hdlr.setFormatter(formatter)
self.logger.addHandler(hdlr)
if verbose:
handler = logging.StreamHandler(sys.stdout)
handler.setLevel(logging.DEBUG)
handler.setFormatter(formatter)
self.logger.addHandler(handler)
self.mcts = MCTS(env=environment,
time_budget=time_budget,
exec_dir=exec_dir,
bandit_policy=bandit_policy,
coef_progressive_widening=coef_progressive_widening)
np.random.seed(seed)
def run(self,
nb_simulation=10,
initial_configurations=[],
step_to_generate_img=-1):
"""Run MCTS algorithm
Parameters:
----------
nb_simulation: int
number of MCTS simulation to run (default is 10)
initial_configurations: list of object
set of configuration to start with (default is [])
step_to_generate_img: int or None
set of initial configuration (default -1, generate image for each MCTS iteration)
Do not generate images if None.
Returns:
----------
configuration: object
best configuration
"""
self.logger.info("# Run {0} iterations of MCTS".format(nb_simulation))
self.mcts.run(nb_simulation, initial_configurations,
step_to_generate_img)
return self.mcts.best_config, self.mcts.best_score
class Search:
"""
Main class to tune pipeline using Monte-Carlo Tree Search
...
Attributes
----------
mcts : class <mosaic.MCTS>
object that run MCTS algorithm
Methods
-------
run(nb_simulation=10, initial_configurations=[], nb_iter_to_generate_img=-1)
Run nb_simulation of MCTS and initialize with initial_configurations
"""
def __init__(self,
environment,
time_budget=3600,
seed=1,
policy_arg={},
exec_dir=None,
verbose=False):
"""Initialization algorithm.
:param environment: environment class extending AbstractEnvironment
:param time_budget: overall time budget
:param seed: random seed
:param policy_arg: specific option for MCTS policy
:param exec_dir: directory to store tmp files
"""
# config logger
self.logger = logging.getLogger('mcts')
self.logger.setLevel(logging.DEBUG)
# execution directory
if exec_dir is None:
exec_dir = tempfile.mkdtemp()
else:
os.makedirs(exec_dir)
hdlr = logging.FileHandler(os.path.join(exec_dir, "mcts.log"), mode='w')
formatter = logging.Formatter('%(asctime)s :: %(levelname)s :: %(funcName)s :: %(message)s')
hdlr.setFormatter(formatter)
self.logger.addHandler(hdlr)
if verbose:
handler = logging.StreamHandler(sys.stdout)
handler.setLevel(logging.DEBUG)
handler.setFormatter(formatter)
self.logger.addHandler(handler)
env = environment
self.mcts = MCTS(env=env,
time_budget=time_budget,
policy_arg=policy_arg,
exec_dir=exec_dir)
np.random.seed(seed)
def run(self, nb_simulation=10, initial_configurations=[], nb_iter_to_generate_img=1, pb=None, status_txt=None, chart=None, status_txt2=None):
"""Run MCTS algorithm
:param nb_simulation: number of MCTS simulation to run
:param initial_configurations: path for generated image , optional
:param nb_iter_to_generate_img: set of initial configuration, optional
:return:
configuration: best configuration found
"""
self.logger.info("# Run {0} iterations of MCTS".format(nb_simulation))
self.mcts.run(nb_simulation, initial_configurations, nb_iter_to_generate_img, pb, status_txt, chart, status_txt2)
return self.mcts.bestconfig, self.mcts.bestscore