def start_tuner(self, tuner: Tuner, hdl: dict):
self.logger.info(f"Start fine tune task, \nwhich HDL(Hyperparams Descriptions Language) is:\n{hdl}")
self.logger.info(f"which Tuner is:\n{tuner}")
tuner.set_data_manager(self.data_manager)
tuner.set_random_state(self.random_state)
tuner.set_hdl(hdl) # just for get shps of tuner
if estimate_config_space_numbers(tuner.shps) == 1:
self.logger.info("HDL(Hyperparams Descriptions Language) is a constant space, using manual modeling.")
dhp, self.estimator = tuner.shp2model(tuner.shps.sample_configuration())
self.estimator.fit(self.data_manager.X_train, self.data_manager.y_train)
return {"is_manual": True}
n_jobs = tuner.n_jobs
run_limits = [math.ceil(tuner.run_limit / n_jobs)] * n_jobs
is_master_list = [False] * n_jobs
is_master_list[0] = True
initial_configs_list = get_chunks(
tuner.design_initial_configs(n_jobs),
n_jobs)
random_states = np.arange(n_jobs) + self.random_state
if n_jobs > 1 and tuner.search_method != "grid":
sync_dict = Manager().dict()
sync_dict["exit_processes"] = tuner.exit_processes
else:
sync_dict = None
self.resource_manager.close_trials_db()
self.resource_manager.clear_pid_list()
self.resource_manager.close_redis()
resource_managers = [deepcopy(self.resource_manager) for i in range(n_jobs)]
tuners = [deepcopy(tuner) for i in range(n_jobs)]
with joblib.parallel_backend(n_jobs=n_jobs, backend="multiprocessing"):
joblib.Parallel()(
joblib.delayed(self.run)
(tuner, resource_manager, run_limit, initial_configs, is_master, random_state, sync_dict)
for tuner, resource_manager, run_limit, initial_configs, is_master, random_state in
zip(tuners, resource_managers, run_limits, initial_configs_list, is_master_list, random_states)
)
return {"is_manual": False}
def __init__(
self,
tuner: Union[Tuner, List[Tuner], None, dict] = None,
hdl_constructor: Union[HDL_Constructor, List[HDL_Constructor], None, dict] = None,
resource_manager: Union[ResourceManager, str] = None,
ensemble_builder: Union[StackEnsembleBuilder, None, bool, int] = None,
random_state=42
):
# ---logger------------------------------------
self.logger = get_logger(__name__)
# ---random_state-----------------------------------
self.random_state = random_state
# ---ensemble_builder-----------------------------------
if ensemble_builder is None:
self.logger.info("Using default StackEnsembleBuilder.")
ensemble_builder = StackEnsembleBuilder()
elif ensemble_builder == False:
self.logger.info("Not using EnsembleBuilder, will select the best estimator.")
else:
ensemble_builder = StackEnsembleBuilder(set_model=ensemble_builder)
self.ensemble_builder = ensemble_builder
# ---tuners-----------------------------------
if not tuner:
tuner = Tuner()
if not isinstance(tuner, (list, tuple)):
tuner = [tuner]
self.tuners: List[Tuner] = tuner
# ---hdl_constructors-----------------------------------
if not hdl_constructor:
hdl_constructor = HDL_Constructor()
if not isinstance(hdl_constructor, (list, tuple)):
hdl_constructor = [hdl_constructor]
self.hdl_constructors = hdl_constructor
# ---resource_manager-----------------------------------
if resource_manager is None:
resource_manager = ResourceManager()
self.resource_manager = resource_manager
# ---member_variable------------------------------------
self.estimator = None
import pandas as pd
from sklearn.model_selection import ShuffleSplit
from hyperflow.estimator.base import HyperFlowEstimator
from hyperflow.hdl.hdl_constructor import HDL_Constructor
from hyperflow.tuner.tuner import Tuner
df = pd.read_csv("../examples/classification/train_classification.csv")
ss = ShuffleSplit(n_splits=1, random_state=0, test_size=0.25)
train_ix, test_ix = next(ss.split(df))
df_train = df.iloc[train_ix, :]
df_test = df.iloc[test_ix, :]
tuner = Tuner(
initial_runs=5,
run_limit=12,
)
hdl_constructor = HDL_Constructor(
DAG_descriptions={
"nan->imp": "impute.fill_abnormal",
"imp->{cat_name=cat,num_name=num}": "operate.split.cat_num",
"cat->num": "encode.cat_boost|scale.standardize",
"num->target": "reduce.pca|lightgbm"
})
hyperflow_pipeline = HyperFlowEstimator(tuner, hdl_constructor)
column_descriptions = {
"id": "PassengerId",
"target": "Survived",
"ignore": "Name"
}
import pandas as pd
from sklearn.model_selection import ShuffleSplit
from hyperflow.estimator.base import HyperFlowEstimator
from hyperflow.tuner.tuner import Tuner
df = pd.read_csv("../examples/classification/train_classification.csv")
ss = ShuffleSplit(n_splits=1, random_state=0, test_size=0.25)
train_ix, test_ix = next(ss.split(df))
df_train = df.iloc[train_ix, :]
df_test = df.iloc[test_ix, :]
tuner = Tuner(
initial_runs=1,
run_limit=100,
n_jobs=1,
search_method_params={"anneal_func":"lambda x:1*(1/(-(3*(x-1))))"}
)
hyperflow_pipeline = HyperFlowEstimator(tuner)
column_descriptions = {
"id": "PassengerId",
"target": "Survived",
"ignore": "Name"
}
hyperflow_pipeline.fit(
X=df_train, X_test=df_test, column_descriptions=column_descriptions
)
df = pd.read_csv("../examples/classification/train_classification.csv")
ss = ShuffleSplit(n_splits=1, random_state=0, test_size=0.25)
train_ix, test_ix = next(ss.split(df))
df_train = df.iloc[train_ix, :]
df_test = df.iloc[test_ix, :]
hdl_constructor = HDL_Constructor(
DAG_descriptions={
"nan->{highR=highR_nan,lowR=lowR_nan}": "operate.split.nan",
"lowR_nan->nan": "impute.fill_abnormal",
"highR_nan->nan": "operate.drop",
"all->{cat_name=cat,num_name=num}": "operate.split.cat_num",
"cat->num": "encode.cat_boost",
"num->target": {
"_name": "lightgbm",
"_vanilla": True
}
})
tuner = Tuner(run_limit=-1, search_method="grid")
hyperflow_pipeline = HyperFlowEstimator(tuner, hdl_constructor)
column_descriptions = {
"id": "PassengerId",
"target": "Survived",
"ignore": "Name"
}
hyperflow_pipeline.fit(X=df_train,
X_test=df_test,
column_descriptions=column_descriptions)
"lowR_nan->nan": "impute.fill_abnormal",
"highR_nan->nan": "operate.drop",
"all->{cat_name=cat,num_name=num}": "operate.split.cat_num",
"cat->num": "encode.label",
"num->num": {"_name": "<placeholder>",
"_select_percent": {"_type": "quniform", "_value": [1, 100, 0.5],
"_default": 80}},
"num->target": {"_name": "lightgbm", "_vanilla": True}
}
),
]
tuners = [
Tuner(
run_limit=-1,
search_method="grid",
n_jobs=3
),
Tuner(
run_limit=50,
initial_runs=10,
search_method="smac",
n_jobs=3
),
]
hyperflow_pipeline = HyperFlowEstimator(tuners, hdl_constructors)
column_descriptions = {
"id": "PassengerId",
"target": "Survived",
"ignore": "Name"
}
import pandas as pd
from sklearn.model_selection import ShuffleSplit
from hyperflow.estimator.base import HyperFlowEstimator
from hyperflow.hdl.hdl_constructor import HDL_Constructor
from hyperflow.tuner.tuner import Tuner
df = pd.read_csv("../examples/classification/train_classification.csv")
ss = ShuffleSplit(n_splits=1, random_state=0, test_size=0.25)
train_ix, test_ix = next(ss.split(df))
df_train = df.iloc[train_ix, :]
df_test = df.iloc[test_ix, :]
tuner = Tuner(
initial_runs=30,
run_limit=0,
)
hdl_constructor = HDL_Constructor(
DAG_descriptions={
"nan->imp": "impute.fill_abnormal",
"imp->{cat_name=cat,num_name=num}": "operate.split.cat_num",
"cat->num": "encode.cat_boost",
"over_sample": [
"balance.under_sample.all_knn",
"balance.under_sample.cluster_centroids",
"balance.under_sample.condensed_nearest_neighbour",
"balance.under_sample.edited_nearest_neighbours",
"balance.under_sample.instance_hardness_threshold",
"balance.under_sample.near_miss",
"balance.under_sample.neighbourhood_cleaning_rule",
"balance.under_sample.one_sided_selection",
import pandas as pd
from hyperflow.estimator.base import HyperFlowEstimator
from hyperflow.hdl.hdl_constructor import HDL_Constructor
from hyperflow.tuner.tuner import Tuner
df = pd.read_csv("../data/QSAR.csv")
hdl_constructor = HDL_Constructor(
DAG_descriptions={
"num->var": "compress.variance",
"var->pea": {
"_name": "compress.pearson",
"n_jobs": 6
},
"pea->target": "logistic_regression"
})
tuner = Tuner(run_limit=12, initial_runs=12, search_method="smac")
hyperflow_pipeline = HyperFlowEstimator(tuner, hdl_constructor)
column_descriptions = {"id": "Name", "target": "labels"}
hyperflow_pipeline.fit(X=df, column_descriptions=column_descriptions, n_jobs=3)
if ind.any():
temp = newfeature[:, i]
a = temp[~np.isnan(temp)].mean()
newfeature[:, i][np.isnan(temp)] = a
# 标准化
stdScale = StandardScaler().fit(newfeature)
newfeaturenorm = stdScale.transform(newfeature)
# 区间化
bins = [-9, -5, -3, -1, 1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 24]
new_range = pd.cut(df.Label, bins)
newlabel = np.array(df.Label)
return newfeaturenorm, newlabel, new_range
x_train, y_train, y_range = data_preprocessing()
tuner = Tuner(
initial_runs=12,
run_limit=120,
)
hdl_constructor = HDL_Constructor(DAG_descriptions={"num->target": "lightgbm"})
resource_manager = ResourceManager(os.getcwd() + "/for_hxw_result")
hyperflow_pipeline = HyperFlowEstimator(tuner,
hdl_constructor,
ensemble_builder=False)
hyperflow_pipeline.fit(X=x_train, y=y_train, n_jobs=3)
joblib.dump(hyperflow_pipeline, "hyperflow_pipeline_for_hxw.bz")