# Copyright (C) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License. See LICENSE in project root for information.
import sys
if sys.version >= '3':
basestring = str
from pyspark.ml.param.shared import *
from pyspark import keyword_only
from pyspark.ml.util import JavaMLReadable, JavaMLWritable
from pyspark.ml.wrapper import JavaTransformer, JavaEstimator, JavaModel
from pyspark.ml.common import inherit_doc
from mmlspark.Utils import *
from mmlspark.TypeConversionUtils import generateTypeConverter, complexTypeConverter
[docs]@inherit_doc
class TuneHyperparameters(ComplexParamsMixin, JavaMLReadable, JavaMLWritable, JavaEstimator):
"""
Tunes model hyperparameters
Supports distributed hyperparameter tuning through randomized grid search.
In the near future will support Nelder-Mead and kernel density optimization.
Args:
evaluationMetric (str): Metric to evaluate models with
models (object): Estimators to run
numFolds (int): Number of folds
numRuns (int): Termination criteria for randomized search
parallelism (int): The number of models to run in parallel
paramSpace (object): Parameter space for generating hyperparameters
seed (long): Random number generator seed (default: 0)
"""
@keyword_only
def __init__(self, evaluationMetric=None, models=None, numFolds=None, numRuns=None, parallelism=None, paramSpace=None, seed=0):
super(TuneHyperparameters, self).__init__()
self._java_obj = self._new_java_obj("com.microsoft.ml.spark.TuneHyperparameters")
self._cache = {}
self.evaluationMetric = Param(self, "evaluationMetric", "evaluationMetric: Metric to evaluate models with")
self.models = Param(self, "models", "models: Estimators to run", generateTypeConverter("models", self._cache, complexTypeConverter))
self.numFolds = Param(self, "numFolds", "numFolds: Number of folds")
self.numRuns = Param(self, "numRuns", "numRuns: Termination criteria for randomized search")
self.parallelism = Param(self, "parallelism", "parallelism: The number of models to run in parallel")
self.paramSpace = Param(self, "paramSpace", "paramSpace: Parameter space for generating hyperparameters")
self.seed = Param(self, "seed", "seed: Random number generator seed (default: 0)")
self._setDefault(seed=0)
if hasattr(self, "_input_kwargs"):
kwargs = self._input_kwargs
else:
kwargs = self.__init__._input_kwargs
self.setParams(**kwargs)
[docs] @keyword_only
def setParams(self, evaluationMetric=None, models=None, numFolds=None, numRuns=None, parallelism=None, paramSpace=None, seed=0):
"""
Set the (keyword only) parameters
Args:
evaluationMetric (str): Metric to evaluate models with
models (object): Estimators to run
numFolds (int): Number of folds
numRuns (int): Termination criteria for randomized search
parallelism (int): The number of models to run in parallel
paramSpace (object): Parameter space for generating hyperparameters
seed (long): Random number generator seed (default: 0)
"""
if hasattr(self, "_input_kwargs"):
kwargs = self._input_kwargs
else:
kwargs = self.__init__._input_kwargs
return self._set(**kwargs)
[docs] def setEvaluationMetric(self, value):
"""
Args:
evaluationMetric (str): Metric to evaluate models with
"""
self._set(evaluationMetric=value)
return self
[docs] def getEvaluationMetric(self):
"""
Returns:
str: Metric to evaluate models with
"""
return self.getOrDefault(self.evaluationMetric)
[docs] def setModels(self, value):
"""
Args:
models (object): Estimators to run
"""
self._set(models=value)
return self
[docs] def getModels(self):
"""
Returns:
object: Estimators to run
"""
return self._cache.get("models", None)
[docs] def setNumFolds(self, value):
"""
Args:
numFolds (int): Number of folds
"""
self._set(numFolds=value)
return self
[docs] def getNumFolds(self):
"""
Returns:
int: Number of folds
"""
return self.getOrDefault(self.numFolds)
[docs] def setNumRuns(self, value):
"""
Args:
numRuns (int): Termination criteria for randomized search
"""
self._set(numRuns=value)
return self
[docs] def getNumRuns(self):
"""
Returns:
int: Termination criteria for randomized search
"""
return self.getOrDefault(self.numRuns)
[docs] def setParallelism(self, value):
"""
Args:
parallelism (int): The number of models to run in parallel
"""
self._set(parallelism=value)
return self
[docs] def getParallelism(self):
"""
Returns:
int: The number of models to run in parallel
"""
return self.getOrDefault(self.parallelism)
[docs] def setParamSpace(self, value):
"""
Args:
paramSpace (object): Parameter space for generating hyperparameters
"""
self._set(paramSpace=value)
return self
[docs] def getParamSpace(self):
"""
Returns:
object: Parameter space for generating hyperparameters
"""
return self.getOrDefault(self.paramSpace)
[docs] def setSeed(self, value):
"""
Args:
seed (long): Random number generator seed (default: 0)
"""
self._set(seed=value)
return self
[docs] def getSeed(self):
"""
Returns:
long: Random number generator seed (default: 0)
"""
return self.getOrDefault(self.seed)
[docs] @classmethod
def read(cls):
""" Returns an MLReader instance for this class. """
return JavaMMLReader(cls)
[docs] @staticmethod
def getJavaPackage():
""" Returns package name String. """
return "com.microsoft.ml.spark.TuneHyperparameters"
@staticmethod
def _from_java(java_stage):
module_name=TuneHyperparameters.__module__
module_name=module_name.rsplit(".", 1)[0] + ".TuneHyperparameters"
return from_java(java_stage, module_name)
def _create_model(self, java_model):
return TuneHyperparametersModel(java_model)
[docs]class TuneHyperparametersModel(ComplexParamsMixin, JavaModel, JavaMLWritable, JavaMLReadable):
"""
Model fitted by :class:`TuneHyperparameters`.
This class is left empty on purpose.
All necessary methods are exposed through inheritance.
"""
[docs] @classmethod
def read(cls):
""" Returns an MLReader instance for this class. """
return JavaMMLReader(cls)
[docs] @staticmethod
def getJavaPackage():
""" Returns package name String. """
return "com.microsoft.ml.spark.TuneHyperparametersModel"
@staticmethod
def _from_java(java_stage):
module_name=TuneHyperparametersModel.__module__
module_name=module_name.rsplit(".", 1)[0] + ".TuneHyperparametersModel"
return from_java(java_stage, module_name)