def _handleOverwrite(self, path):
from pyspark.ml.wrapper import JavaWrapper
_java_obj = JavaWrapper._new_java_obj(
"org.apache.spark.ml.util.FileSystemOverwrite")
wrapper = JavaWrapper(_java_obj)
wrapper._call_java("handleOverwrite", path, True, self.sc._jsc.sc())
def _handleOverwrite(self, path: str) -> None:
from pyspark.ml.wrapper import JavaWrapper
_java_obj = JavaWrapper._new_java_obj( # type: ignore[attr-defined]
"org.apache.spark.ml.util.FileSystemOverwrite")
wrapper = JavaWrapper(_java_obj)
wrapper._call_java( # type: ignore[attr-defined]
"handleOverwrite", path, True, self.sparkSession._jsparkSession)
def __setCreds(conf, auth):
jconf = conf._jconf
field = jconf.getClass().getDeclaredField("nonJVMClientCreds")
field.setAccessible(True)
from pyspark.ml.wrapper import JavaWrapper
creds = JavaWrapper._new_java_obj(
"org.apache.spark.h2o.utils.FlowCredentials", auth[0], auth[1])
someCreds = JavaWrapper._new_java_obj("scala.Some", creds)
field.set(jconf, someCreds)
def _from_java_impl(cls, java_stage):
"""
Return Python estimator, estimatorParamMaps, and evaluator from a Java ValidatorParams.
"""
# Load information from java_stage to the instance.
estimator = JavaWrapper._from_java(java_stage.getEstimator())
evaluator = JavaWrapper._from_java(java_stage.getEvaluator())
epms = [estimator._transfer_param_map_from_java(epm)
for epm in java_stage.getEstimatorParamMaps()]
return estimator, epms, evaluator
def _from_java_impl(cls, java_stage):
"""
Return Python estimator, estimatorParamMaps, and evaluator from a Java ValidatorParams.
"""
# Load information from java_stage to the instance.
estimator = JavaWrapper._from_java(java_stage.getEstimator())
evaluator = JavaWrapper._from_java(java_stage.getEvaluator())
epms = [
estimator._transfer_param_map_from_java(epm)
for epm in java_stage.getEstimatorParamMaps()
]
return estimator, epms, evaluator
def meta_estimator_transfer_param_maps_to_java(pyEstimator, pyParamMaps):
pyStages = MetaAlgorithmReadWrite.getAllNestedStages(pyEstimator)
stagePairs = list(
map(lambda stage: (stage, stage._to_java()), pyStages))
sc = SparkContext._active_spark_context
paramMapCls = SparkContext._jvm.org.apache.spark.ml.param.ParamMap
javaParamMaps = SparkContext._gateway.new_array(
paramMapCls, len(pyParamMaps))
for idx, pyParamMap in enumerate(pyParamMaps):
javaParamMap = JavaWrapper._new_java_obj(
"org.apache.spark.ml.param.ParamMap")
for pyParam, pyValue in pyParamMap.items():
javaParam = None
for pyStage, javaStage in stagePairs:
if pyStage._testOwnParam(pyParam.parent, pyParam.name):
javaParam = javaStage.getParam(pyParam.name)
break
if javaParam is None:
raise ValueError(
'Resolve param in estimatorParamMaps failed: ' +
str(pyParam))
if isinstance(pyValue, Params) and hasattr(
pyValue, "_to_java"):
javaValue = pyValue._to_java()
else:
javaValue = _py2java(sc, pyValue)
pair = javaParam.w(javaValue)
javaParamMap.put([pair])
javaParamMaps[idx] = javaParamMap
return javaParamMaps
def metrics(*metrics):
"""
Given a list of metrics, provides a builder that it turns computes metrics from a column.
See the documentation of [[Summarizer]] for an example.
The following metrics are accepted (case sensitive):
- mean: a vector that contains the coefficient-wise mean.
- variance: a vector tha contains the coefficient-wise variance.
- count: the count of all vectors seen.
- numNonzeros: a vector with the number of non-zeros for each coefficients
- max: the maximum for each coefficient.
- min: the minimum for each coefficient.
- normL2: the Euclidean norm for each coefficient.
- normL1: the L1 norm of each coefficient (sum of the absolute values).
:param metrics:
metrics that can be provided.
:return:
an object of :py:class:`pyspark.ml.stat.SummaryBuilder`
Note: Currently, the performance of this interface is about 2x~3x slower then using the RDD
interface.
"""
sc = SparkContext._active_spark_context
js = JavaWrapper._new_java_obj("org.apache.spark.ml.stat.Summarizer.metrics",
_to_seq(sc, metrics))
return SummaryBuilder(js)
def metrics(*metrics):
"""
Given a list of metrics, provides a builder that it turns computes metrics from a column.
See the documentation of [[Summarizer]] for an example.
The following metrics are accepted (case sensitive):
- mean: a vector that contains the coefficient-wise mean.
- sum: a vector that contains the coefficient-wise sum.
- variance: a vector tha contains the coefficient-wise variance.
- std: a vector tha contains the coefficient-wise standard deviation.
- count: the count of all vectors seen.
- numNonzeros: a vector with the number of non-zeros for each coefficients
- max: the maximum for each coefficient.
- min: the minimum for each coefficient.
- normL2: the Euclidean norm for each coefficient.
- normL1: the L1 norm of each coefficient (sum of the absolute values).
:param metrics:
metrics that can be provided.
:return:
an object of :py:class:`pyspark.ml.stat.SummaryBuilder`
Note: Currently, the performance of this interface is about 2x~3x slower then using the RDD
interface.
"""
sc = SparkContext._active_spark_context
js = JavaWrapper._new_java_obj(
"org.apache.spark.ml.stat.Summarizer.metrics",
_to_seq(sc, metrics))
return SummaryBuilder(js)
def to_java_params(sc, model, pyParamMap):
paramMap = JavaWrapper._new_java_obj("org.apache.spark.ml.param.ParamMap")
for param, value in pyParamMap.items():
java_param = model._java_obj.getParam(param.name)
java_value = _py2java(sc, value)
paramMap.put([java_param.w(java_value)])
return paramMap
def _get_single_metric(col: Column, weightCol: Optional[Column],
metric: str) -> Column:
col, weightCol = Summarizer._check_param(col, weightCol)
return Column(
JavaWrapper._new_java_obj(
"org.apache.spark.ml.stat.Summarizer." + metric, col._jc,
weightCol._jc))
def _stages_java2py(java_stages):
"""
Transforms the parameter Python stages from a list of Java stages.
:param java_stages: An array of Java stages.
:return: An array of Python stages.
"""
return [JavaWrapper._transfer_stage_from_java(stage) for stage in java_stages]
def _stages_java2py(java_stages):
"""
Transforms the parameter Python stages from a list of Java stages.
:param java_stages: An array of Java stages.
:return: An array of Python stages.
"""
return [
JavaWrapper._transfer_stage_from_java(stage) for stage in java_stages
]
def _transfer_param_map_to_java(self, pyParamMap):
"""
Transforms a Python ParamMap into a Java ParamMap.
"""
paramMap = JavaWrapper._new_java_obj("org.apache.spark.ml.param.ParamMap")
for param in self.params:
if param in pyParamMap:
pair = self._make_java_param_pair(param, pyParamMap[param])
paramMap.put([pair])
return paramMap
def _from_java(cls, java_stage):
"""
Given a Java PipelineModel, create and return a Python wrapper of it.
Used for ML persistence.
"""
# Load information from java_stage to the instance.
py_stages = [JavaWrapper._from_java(s) for s in java_stage.stages()]
# Create a new instance of this stage.
py_stage = cls(py_stages)
py_stage._resetUid(java_stage.uid())
return py_stage
def _py2j(self, arg):
if isinstance(arg, dict):
return ScalaMap(arg)._to_java()
elif isinstance(arg, StructType):
return JavaWrapper._new_java_obj(
"org.apache.spark.sql.types.DataType.fromJson",
json.dumps(arg.jsonValue()))
elif isinstance(arg, SageMakerJavaWrapper):
return arg._to_java()
else:
return arg
def _from_java(cls, java_stage):
"""
Given a Java PipelineModel, create and return a Python wrapper of it.
Used for ML persistence.
"""
# Load information from java_stage to the instance.
py_stages = [JavaWrapper._from_java(s) for s in java_stage.stages()]
# Create a new instance of this stage.
py_stage = cls(py_stages)
py_stage._resetUid(java_stage.uid())
return py_stage
def test_new_java_array(self):
# test array of strings
str_list = ["a", "b", "c"]
java_class = self.sc._gateway.jvm.java.lang.String
java_array = JavaWrapper._new_java_array(str_list, java_class)
self.assertEqual(_java2py(self.sc, java_array), str_list)
# test array of integers
int_list = [1, 2, 3]
java_class = self.sc._gateway.jvm.java.lang.Integer
java_array = JavaWrapper._new_java_array(int_list, java_class)
self.assertEqual(_java2py(self.sc, java_array), int_list)
# test array of floats
float_list = [0.1, 0.2, 0.3]
java_class = self.sc._gateway.jvm.java.lang.Double
java_array = JavaWrapper._new_java_array(float_list, java_class)
self.assertEqual(_java2py(self.sc, java_array), float_list)
# test array of bools
bool_list = [False, True, True]
java_class = self.sc._gateway.jvm.java.lang.Boolean
java_array = JavaWrapper._new_java_array(bool_list, java_class)
self.assertEqual(_java2py(self.sc, java_array), bool_list)
# test array of Java DenseVectors
v1 = DenseVector([0.0, 1.0])
v2 = DenseVector([1.0, 0.0])
vec_java_list = [_py2java(self.sc, v1), _py2java(self.sc, v2)]
java_class = self.sc._gateway.jvm.org.apache.spark.ml.linalg.DenseVector
java_array = JavaWrapper._new_java_array(vec_java_list, java_class)
self.assertEqual(_java2py(self.sc, java_array), [v1, v2])
# test empty array
java_class = self.sc._gateway.jvm.java.lang.Integer
java_array = JavaWrapper._new_java_array([], java_class)
self.assertEqual(_java2py(self.sc, java_array), [])
def test_new_java_array(self):
# test array of strings
str_list = ["a", "b", "c"]
java_class = self.sc._gateway.jvm.java.lang.String
java_array = JavaWrapper._new_java_array(str_list, java_class)
self.assertEqual(_java2py(self.sc, java_array), str_list)
# test array of integers
int_list = [1, 2, 3]
java_class = self.sc._gateway.jvm.java.lang.Integer
java_array = JavaWrapper._new_java_array(int_list, java_class)
self.assertEqual(_java2py(self.sc, java_array), int_list)
# test array of floats
float_list = [0.1, 0.2, 0.3]
java_class = self.sc._gateway.jvm.java.lang.Double
java_array = JavaWrapper._new_java_array(float_list, java_class)
self.assertEqual(_java2py(self.sc, java_array), float_list)
# test array of bools
bool_list = [False, True, True]
java_class = self.sc._gateway.jvm.java.lang.Boolean
java_array = JavaWrapper._new_java_array(bool_list, java_class)
self.assertEqual(_java2py(self.sc, java_array), bool_list)
# test array of Java DenseVectors
v1 = DenseVector([0.0, 1.0])
v2 = DenseVector([1.0, 0.0])
vec_java_list = [_py2java(self.sc, v1), _py2java(self.sc, v2)]
java_class = self.sc._gateway.jvm.org.apache.spark.ml.linalg.DenseVector
java_array = JavaWrapper._new_java_array(vec_java_list, java_class)
self.assertEqual(_java2py(self.sc, java_array), [v1, v2])
# test empty array
java_class = self.sc._gateway.jvm.java.lang.Integer
java_array = JavaWrapper._new_java_array([], java_class)
self.assertEqual(_java2py(self.sc, java_array), [])
def _from_java(cls, java_stage):
"""
Given a Java CrossValidatorModel, create and return a Python wrapper of it.
Used for ML persistence.
"""
# Load information from java_stage to the instance.
bestModel = JavaWrapper._from_java(java_stage.bestModel())
estimator, epms, evaluator = super(CrossValidatorModel, cls)._from_java_impl(java_stage)
# Create a new instance of this stage.
py_stage = cls(bestModel=bestModel)\
.setEstimator(estimator).setEstimatorParamMaps(epms).setEvaluator(evaluator)
py_stage._resetUid(java_stage.uid())
return py_stage
def _new_java_obj(self, java_class, *args):
"""
Creates a java object. We convert SageMakerJavaClass arguments
to their java versions and then hand over to JavaWrapper
:param java_class: Java ClassName
:param args: constructor arguments
:return: Java Instance
"""
java_args = []
for arg in args:
java_args.append(self._py2j(arg))
return JavaWrapper._new_java_obj(java_class, *java_args)
def _from_java(cls, java_stage):
"""
Given a Java TrainValidationSplitModel, create and return a Python wrapper of it.
Used for ML persistence.
"""
# Load information from java_stage to the instance.
bestModel = JavaWrapper._from_java(java_stage.bestModel())
estimator, epms, evaluator = \
super(TrainValidationSplitModel, cls)._from_java_impl(java_stage)
# Create a new instance of this stage.
py_stage = cls(bestModel=bestModel)\
.setEstimator(estimator).setEstimatorParamMaps(epms).setEvaluator(evaluator)
py_stage._resetUid(java_stage.uid())
return py_stage
def _to_java(self):
"""
Transfer this instance to a Java PipelineModel. Used for ML persistence.
:return: Java object equivalent to this instance.
"""
gateway = SparkContext._gateway
cls = SparkContext._jvm.org.apache.spark.ml.Transformer
java_stages = gateway.new_array(cls, len(self.stages))
for idx, stage in enumerate(self.stages):
java_stages[idx] = stage._to_java()
_java_obj =\
JavaWrapper._new_java_obj("org.apache.spark.ml.PipelineModel", self.uid, java_stages)
return _java_obj
def _to_java(self):
"""
Transfer this instance to a Java PipelineModel. Used for ML persistence.
:return: Java object equivalent to this instance.
"""
gateway = SparkContext._gateway
cls = SparkContext._jvm.org.apache.spark.ml.Transformer
java_stages = gateway.new_array(cls, len(self.stages))
for idx, stage in enumerate(self.stages):
java_stages[idx] = stage._to_java()
_java_obj =\
JavaWrapper._new_java_obj("org.apache.spark.ml.PipelineModel", self.uid, java_stages)
return _java_obj
def _to_java(self):
"""
Transfer this instance to a Java CrossValidator. Used for ML persistence.
:return: Java object equivalent to this instance.
"""
estimator, epms, evaluator = super(CrossValidator, self)._to_java_impl()
_java_obj = JavaWrapper._new_java_obj("org.apache.spark.ml.tuning.CrossValidator", self.uid)
_java_obj.setEstimatorParamMaps(epms)
_java_obj.setEvaluator(evaluator)
_java_obj.setEstimator(estimator)
_java_obj.setSeed(self.getSeed())
_java_obj.setNumFolds(self.getNumFolds())
return _java_obj
def _to_java(self):
"""
Transfer this instance to a Java TrainValidationSplitModel. Used for ML persistence.
:return: Java object equivalent to this instance.
"""
sc = SparkContext._active_spark_context
_java_obj = JavaWrapper._new_java_obj(
"org.apache.spark.ml.tuning.TrainValidationSplitModel", self.uid,
self.bestModel._to_java(), _py2java(sc, []))
estimator, epms, evaluator = super(TrainValidationSplitModel,
self)._to_java_impl()
_java_obj.set("evaluator", evaluator)
_java_obj.set("estimator", estimator)
_java_obj.set("estimatorParamMaps", epms)
return _java_obj
def _to_java(self):
"""
Transfer this instance to a Java CrossValidatorModel. Used for ML persistence.
:return: Java object equivalent to this instance.
"""
sc = SparkContext._active_spark_context
_java_obj = JavaWrapper._new_java_obj("org.apache.spark.ml.tuning.CrossValidatorModel",
self.uid,
self.bestModel._to_java(),
_py2java(sc, []))
estimator, epms, evaluator = super(CrossValidatorModel, self)._to_java_impl()
_java_obj.set("evaluator", evaluator)
_java_obj.set("estimator", estimator)
_java_obj.set("estimatorParamMaps", epms)
return _java_obj
def _to_java(self):
"""
Transfer this instance to a Java TrainValidationSplit. Used for ML persistence.
:return: Java object equivalent to this instance.
"""
estimator, epms, evaluator = super(TrainValidationSplit,
self)._to_java_impl()
_java_obj = JavaWrapper._new_java_obj(
"org.apache.spark.ml.tuning.TrainValidationSplit", self.uid)
_java_obj.setEstimatorParamMaps(epms)
_java_obj.setEvaluator(evaluator)
_java_obj.setEstimator(estimator)
_java_obj.setTrainRatio(self.getTrainRatio())
_java_obj.setSeed(self.getSeed())
return _java_obj
def main():
try:
# init spark
spark = get_spark(app_name="pyspark-xgb")
# get logger
logger = get_logger(spark, "app")
# load data
train = spark.read.schema(get_btrain_schema()).option('header', True).csv(
DATASET_PATH + '/emp_train.csv')
test = spark.read.schema(get_btrain_schema()).option('header', True).csv(
DATASET_PATH + '/emp_test.csv')
# preprocess
LABEL = 'Attrition'
FEATURES = 'features'
features = [c for c in train.columns if c != LABEL]
assembler = VectorAssembler(inputCols=features, outputCol=FEATURES)
train = assembler.transform(train).select(FEATURES, LABEL)
test = assembler.transform(test).select(FEATURES, LABEL)
# training
logger.info('training')
xgb_params = {
"eta": 0.1, "gamma": 0, "max_depth": 4,
"num_round": 100, "num_early_stopping_rounds": 10,
"num_workers": 1, "use_external_memory": False, "missing": np.nan,
"eval_metric": "logloss", "min_child_weight": 1, "train_test_ratio": 0.8,
"objective": "binary:logistic"
}
scala_map = spark._jvm.PythonUtils.toScalaMap(xgb_params)
j = JavaWrapper._new_java_obj(
"ml.dmlc.xgboost4j.scala.spark.XGBoostClassifier", scala_map) \
.setFeaturesCol(FEATURES).setLabelCol(LABEL)
jmodel = j.train(train._jdf)
logger.info(jmodel.summary().toString())
# get validation metric
preds = jmodel.transform(test._jdf)
pred = DataFrame(preds, spark)
slogloss = pred.withColumn('log_loss', udf_logloss(LABEL, 'probability')) \
.agg({"log_loss": "mean"}).collect()[0]['avg(log_loss)']
logger.info('valid logloss: {}'.format(slogloss))
# save or update model
model_path = MODEL_PATH + '/model.bin'
if os.path.exists(model_path):
shutil.rmtree(model_path)
logger.info('model exist, rm old model')
jw = JavaWrapper._new_java_obj(
"ml.dmlc.xgboost4j.scala.spark.XGBoostClassificationModel.XGBoostClassificationModelWriter", jmodel)
jw.saveImpl(model_path)
logger.info('save model to {}'.format(model_path))
except Exception:
logger.error(traceback.print_exc())
finally:
# stop spark
spark.stop()
def fscore(self,fmap=""):
jxgb = JavaWrapper(self.getBooster())
return jxgb._call_java("getFeatureScore",fmap)
def empty(cls):
return JavaWrapper._new_java_obj("scala.Option.empty")
def _handleOverwrite(self, path):
from pyspark.ml.wrapper import JavaWrapper
_java_obj = JavaWrapper._new_java_obj("org.apache.spark.ml.util.FileSystemOverwrite")
wrapper = JavaWrapper(_java_obj)
wrapper._call_java("handleOverwrite", path, True, self.sc._jsc.sc())
def _get_single_metric(col, weightCol, metric):
col, weightCol = Summarizer._check_param(col, weightCol)
return Column(JavaWrapper._new_java_obj("org.apache.spark.ml.stat.Summarizer." + metric,
col._jc, weightCol._jc))
def get_dump(self,fmap="",with_stats=True,format= "text"):
jxgb = JavaWrapper(self.getBooster())
return jxgb._call_java("getModelDump",fmap,with_stats,format)
def main():
try:
# init spark
spark = get_spark(app_name="pyspark-xgb")
# get logger
logger = get_logger(spark, "app")
# load data
train = spark.read.csv(DATASET_PATH + "/iris_train.csv",
get_mtrain_schema(),
header=True)
test = spark.read.csv(DATASET_PATH + "/iris_test.csv",
get_mtrain_schema(),
header=True)
# preprocess
# get label encode result from csv, since StringIndexer will get different result
STR_LABEL = 'class'
LABEL = 'label'
FEATURES = 'features'
N_CLASS = 3
features = [c for c in train.columns if c not in [STR_LABEL, LABEL]]
assembler = VectorAssembler(inputCols=features, outputCol=FEATURES)
pipeline = Pipeline(stages=[assembler])
preprocess = pipeline.fit(train)
train = preprocess.transform(train).select(FEATURES, LABEL)
test = preprocess.transform(test).select(FEATURES, LABEL)
# set param map
xgb_params = {
"eta": 0.1, "eval_metric": "mlogloss",
"gamma": 0, "max_depth": 5, "min_child_weight": 1.0,
"objective": "multi:softprob", "seed": 0,
"num_class": N_CLASS,
# xgboost4j only
"num_round": 100, "num_early_stopping_rounds": 10,
"maximize_evaluation_metrics": False,
"num_workers": 1, "use_external_memory": False,
"missing": np.nan,
}
scala_map = spark._jvm.PythonUtils.toScalaMap(xgb_params)
# set evaluation set
eval_set = {'eval': test._jdf}
scala_eval_set = spark._jvm.PythonUtils.toScalaMap(eval_set)
logger.info('training')
j = JavaWrapper._new_java_obj(
"ml.dmlc.xgboost4j.scala.spark.XGBoostClassifier", scala_map) \
.setFeaturesCol(FEATURES).setLabelCol(LABEL) \
.setEvalSets(scala_eval_set)
jmodel = j.fit(train._jdf)
print_summary(jmodel)
# get validation metric
preds = jmodel.transform(test._jdf)
pred = DataFrame(preds, spark)
slogloss = pred.withColumn('log_loss', udf_logloss(LABEL, 'probability', N_CLASS)) \
.agg({"log_loss": "mean"}).collect()[0]['avg(log_loss)']
logger.info('[xgboost4j] valid logloss: {}'.format(slogloss))
# save model - using native booster for single node library to read
model_path = MODEL_PATH + '/model.bin'
logger.info('save model to {}'.format(model_path))
jbooster = jmodel.nativeBooster()
jbooster.saveModel(model_path)
# get feature score
imp_type = "gain"
feature_map_path = MODEL_PATH + '/feature.map'
create_feature_map(feature_map_path, features)
jfeatureMap = jbooster.getScore(feature_map_path, imp_type)
f_imp = dict()
for feature in features:
if not jfeatureMap.get(feature).isEmpty():
f_imp[feature] = jfeatureMap.get(feature).get()
feature_imp_path = MODEL_PATH + '/feature.imp'
create_feature_imp(feature_imp_path, f_imp)
# [Optional] load model training by xgboost, predict and get validation metric
local_model_path = LOCAL_MODEL_PATH + '/model.bin'
if os.path.exists(local_model_path):
logger.info('load model from {}'.format(local_model_path))
scala_xgb = spark.sparkContext._jvm.ml.dmlc.xgboost4j.scala.XGBoost
jbooster = scala_xgb.loadModel(local_model_path)
# uid, num_class, booster
xgb_cls_model = JavaWrapper._new_java_obj(
"ml.dmlc.xgboost4j.scala.spark.XGBoostClassificationModel",
"xgbc", N_CLASS, jbooster)
jpred = xgb_cls_model.transform(test._jdf)
pred = DataFrame(jpred, spark)
slogloss = pred.withColumn('log_loss', udf_logloss(LABEL, 'probability', N_CLASS)) \
.agg({"log_loss": "mean"}).collect()[0]['avg(log_loss)']
logger.info('[xgboost] valid logloss: {}'.format(slogloss))
else:
logger.info(
"local model is not exist, call python_xgb/train_multi.py to get the model "
"and compare logloss between xgboost and xgboost4j"
)
except Exception:
logger.error(traceback.print_exc())
finally:
# stop spark
spark.stop()
def saveBooster(self, save_path):
jxgb = JavaWrapper(self.getBooster())
jxgb._call_java("saveModel", save_path)
def main():
try:
# init spark
spark = get_spark(app_name="pyspark-xgb")
# get logger
logger = get_logger(spark, "app")
# load data
df = spark.read.csv(DATASET_PATH + "/iris.data", get_mtrain_schema())
# preprocess
LABEL = 'label'
FEATURES = 'features'
N_CLASS = 3
features = [c for c in df.columns if c != "class"]
assembler = VectorAssembler(inputCols=features, outputCol='features')
strIdxer = StringIndexer(inputCol="class", outputCol=LABEL)
pipeline = Pipeline(stages=[assembler, strIdxer])
df = pipeline.fit(df).transform(df).select(FEATURES, LABEL)
train, test = df.randomSplit([0.8, 0.2])
# training
logger.info('training')
xgb_params = {
"eta": 0.1, "gamma": 0, "max_depth": 4,
"num_round": 100, "num_early_stopping_rounds": 10,
"num_workers": 1, "use_external_memory": False, "missing": np.nan,
"num_class": 3, "eval_metric": "mlogloss",
"min_child_weight": 1, "train_test_ratio": 0.8,
"objective": "multi:softprob"
}
scala_map = spark._jvm.PythonUtils.toScalaMap(xgb_params)
j = JavaWrapper._new_java_obj(
"ml.dmlc.xgboost4j.scala.spark.XGBoostClassifier", scala_map) \
.setFeaturesCol(FEATURES).setLabelCol(LABEL)
jmodel = j.train(train._jdf)
logger.info(jmodel.summary().toString())
# get validation metric
preds = jmodel.transform(test._jdf)
pred = DataFrame(preds, spark)
slogloss = pred.withColumn('log_loss', udf_logloss('label', 'probability', N_CLASS)) \
.agg({"log_loss": "mean"}).collect()[0]['avg(log_loss)']
logger.info('valid logloss: {}'.format(slogloss))
# save or update model
model_path = MODEL_PATH + '/model.bin'
if os.path.exists(model_path):
shutil.rmtree(model_path)
logger.info('model exist, rm old model')
jmodel.save(model_path)
logger.info('save model to {}'.format(model_path))
except Exception:
logger.error(traceback.print_exc())
finally:
# stop spark
spark.stop()
def _getCreds(cls):
return JavaWrapper._new_java_obj(
"com.amazonaws.auth.DefaultAWSCredentialsProviderChain")
