Coverage for pyspark/mllib/regression.py : 94%

# 

# Licensed to the Apache Software Foundation (ASF) under one or more 

# contributor license agreements. See the NOTICE file distributed with 

# this work for additional information regarding copyright ownership. 

# The ASF licenses this file to You under the Apache License, Version 2.0 

# (the "License"); you may not use this file except in compliance with 

# the License. You may obtain a copy of the License at 

# 

# http://www.apache.org/licenses/LICENSE-2.0 

# 

# Unless required by applicable law or agreed to in writing, software 

# distributed under the License is distributed on an "AS IS" BASIS, 

# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 

# See the License for the specific language governing permissions and 

# limitations under the License. 

# 

import sys 

import warnings 

import numpy as np 

from pyspark import RDD, since 

from pyspark.streaming.dstream import DStream 

from pyspark.mllib.common import callMLlibFunc, _py2java, _java2py, inherit_doc 

from pyspark.mllib.linalg import _convert_to_vector 

from pyspark.mllib.util import Saveable, Loader 

__all__ = ['LabeledPoint', 'LinearModel', 

'LinearRegressionModel', 'LinearRegressionWithSGD', 

'RidgeRegressionModel', 'RidgeRegressionWithSGD', 

'LassoModel', 'LassoWithSGD', 'IsotonicRegressionModel', 

'IsotonicRegression', 'StreamingLinearAlgorithm', 

'StreamingLinearRegressionWithSGD'] 

class LabeledPoint(object): 

""" 

Class that represents the features and labels of a data point. 

.. versionadded:: 1.0.0 

Parameters 

---------- 

label : int 

Label for this data point. 

features : :py:class:`pyspark.mllib.linalg.Vector` or convertible 

Vector of features for this point (NumPy array, list, 

pyspark.mllib.linalg.SparseVector, or scipy.sparse column matrix). 

Notes 

----- 

'label' and 'features' are accessible as class attributes. 

""" 

def __init__(self, label, features): 

self.label = float(label) 

self.features = _convert_to_vector(features) 

def __reduce__(self): 

return (LabeledPoint, (self.label, self.features)) 

def __str__(self): 

return "(" + ",".join((str(self.label), str(self.features))) + ")" 

def __repr__(self): 

return "LabeledPoint(%s, %s)" % (self.label, self.features) 

class LinearModel(object): 

""" 

A linear model that has a vector of coefficients and an intercept. 

.. versionadded:: 0.9.0 

Parameters 

---------- 

weights : :py:class:`pyspark.mllib.linalg.Vector` 

Weights computed for every feature. 

intercept : float 

Intercept computed for this model. 

""" 

def __init__(self, weights, intercept): 

self._coeff = _convert_to_vector(weights) 

self._intercept = float(intercept) 

@property 

@since("1.0.0") 

def weights(self): 

"""Weights computed for every feature.""" 

return self._coeff 

@property 

@since("1.0.0") 

def intercept(self): 

"""Intercept computed for this model.""" 

return self._intercept 

def __repr__(self): 

return "(weights=%s, intercept=%r)" % (self._coeff, self._intercept) 

@inherit_doc 

class LinearRegressionModelBase(LinearModel): 

"""A linear regression model. 

.. versionadded:: 0.9.0 

Examples 

-------- 

>>> from pyspark.mllib.linalg import SparseVector 

>>> lrmb = LinearRegressionModelBase(np.array([1.0, 2.0]), 0.1) 

>>> abs(lrmb.predict(np.array([-1.03, 7.777])) - 14.624) < 1e-6 

True 

>>> abs(lrmb.predict(SparseVector(2, {0: -1.03, 1: 7.777})) - 14.624) < 1e-6 

True 

""" 

@since("0.9.0") 

def predict(self, x): 

""" 

Predict the value of the dependent variable given a vector or 

an RDD of vectors containing values for the independent variables. 

""" 

if isinstance(x, RDD): 

return x.map(self.predict) 

x = _convert_to_vector(x) 

return self.weights.dot(x) + self.intercept 

@inherit_doc 

class LinearRegressionModel(LinearRegressionModelBase): 

"""A linear regression model derived from a least-squares fit. 

.. versionadded:: 0.9.0 

Examples 

-------- 

>>> from pyspark.mllib.linalg import SparseVector 

>>> from pyspark.mllib.regression import LabeledPoint 

>>> data = [ 

... LabeledPoint(0.0, [0.0]), 

... LabeledPoint(1.0, [1.0]), 

... LabeledPoint(3.0, [2.0]), 

... LabeledPoint(2.0, [3.0]) 

... ] 

>>> lrm = LinearRegressionWithSGD.train(sc.parallelize(data), iterations=10, 

... initialWeights=np.array([1.0])) 

>>> abs(lrm.predict(np.array([0.0])) - 0) < 0.5 

True 

>>> abs(lrm.predict(np.array([1.0])) - 1) < 0.5 

True 

>>> abs(lrm.predict(SparseVector(1, {0: 1.0})) - 1) < 0.5 

True 

>>> abs(lrm.predict(sc.parallelize([[1.0]])).collect()[0] - 1) < 0.5 

True 

>>> import os, tempfile 

>>> path = tempfile.mkdtemp() 

>>> lrm.save(sc, path) 

>>> sameModel = LinearRegressionModel.load(sc, path) 

>>> abs(sameModel.predict(np.array([0.0])) - 0) < 0.5 

True 

>>> abs(sameModel.predict(np.array([1.0])) - 1) < 0.5 

True 

>>> abs(sameModel.predict(SparseVector(1, {0: 1.0})) - 1) < 0.5 

True 

>>> from shutil import rmtree 

>>> try: 

... rmtree(path) 

... except: 

... pass 

>>> data = [ 

... LabeledPoint(0.0, SparseVector(1, {0: 0.0})), 

... LabeledPoint(1.0, SparseVector(1, {0: 1.0})), 

... LabeledPoint(3.0, SparseVector(1, {0: 2.0})), 

... LabeledPoint(2.0, SparseVector(1, {0: 3.0})) 

... ] 

>>> lrm = LinearRegressionWithSGD.train(sc.parallelize(data), iterations=10, 

... initialWeights=np.array([1.0])) 

>>> abs(lrm.predict(np.array([0.0])) - 0) < 0.5 

True 

>>> abs(lrm.predict(SparseVector(1, {0: 1.0})) - 1) < 0.5 

True 

>>> lrm = LinearRegressionWithSGD.train(sc.parallelize(data), iterations=10, step=1.0, 

... miniBatchFraction=1.0, initialWeights=np.array([1.0]), regParam=0.1, regType="l2", 

... intercept=True, validateData=True) 

>>> abs(lrm.predict(np.array([0.0])) - 0) < 0.5 

True 

>>> abs(lrm.predict(SparseVector(1, {0: 1.0})) - 1) < 0.5 

True 

""" 

@since("1.4.0") 

def save(self, sc, path): 

"""Save a LinearRegressionModel.""" 

java_model = sc._jvm.org.apache.spark.mllib.regression.LinearRegressionModel( 

_py2java(sc, self._coeff), self.intercept) 

java_model.save(sc._jsc.sc(), path) 

@classmethod 

@since("1.4.0") 

def load(cls, sc, path): 

"""Load a LinearRegressionModel.""" 

java_model = sc._jvm.org.apache.spark.mllib.regression.LinearRegressionModel.load( 

sc._jsc.sc(), path) 

weights = _java2py(sc, java_model.weights()) 

intercept = java_model.intercept() 

model = LinearRegressionModel(weights, intercept) 

return model 

# train_func should take two parameters, namely data and initial_weights, and 

# return the result of a call to the appropriate JVM stub. 

# _regression_train_wrapper is responsible for setup and error checking. 

def _regression_train_wrapper(train_func, modelClass, data, initial_weights): 

from pyspark.mllib.classification import LogisticRegressionModel 

first = data.first() 

222 ↛ 223line 222 didn't jump to line 223, because the condition on line 222 was never true if not isinstance(first, LabeledPoint): 

raise TypeError("data should be an RDD of LabeledPoint, but got %s" % type(first)) 

if initial_weights is None: 

initial_weights = [0.0] * len(data.first().features) 

if (modelClass == LogisticRegressionModel): 

weights, intercept, numFeatures, numClasses = train_func( 

data, _convert_to_vector(initial_weights)) 

return modelClass(weights, intercept, numFeatures, numClasses) 

else: 

weights, intercept = train_func(data, _convert_to_vector(initial_weights)) 

return modelClass(weights, intercept) 

class LinearRegressionWithSGD(object): 

""" 

Train a linear regression model with no regularization using Stochastic Gradient Descent. 

.. versionadded:: 0.9.0 

.. deprecated:: 2.0.0 

Use :py:class:`pyspark.ml.regression.LinearRegression`. 

""" 

@classmethod 

def train(cls, data, iterations=100, step=1.0, miniBatchFraction=1.0, 

initialWeights=None, regParam=0.0, regType=None, intercept=False, 

validateData=True, convergenceTol=0.001): 

""" 

Train a linear regression model using Stochastic Gradient 

Descent (SGD). This solves the least squares regression 

formulation 

f(weights) = 1/(2n) ||A weights - y||^2 

which is the mean squared error. Here the data matrix has n rows, 

and the input RDD holds the set of rows of A, each with its 

corresponding right hand side label y. 

See also the documentation for the precise formulation. 

.. versionadded:: 0.9.0 

Parameters 

---------- 

data : :py:class:`pyspark.RDD` 

The training data, an RDD of LabeledPoint. 

iterations : int, optional 

The number of iterations. 

(default: 100) 

step : float, optional 

The step parameter used in SGD. 

(default: 1.0) 

miniBatchFraction : float, optional 

Fraction of data to be used for each SGD iteration. 

(default: 1.0) 

initialWeights : :py:class:`pyspark.mllib.linalg.Vector` or convertible, optional 

The initial weights. 

(default: None) 

regParam : float, optional 

The regularizer parameter. 

(default: 0.0) 

regType : str, optional 

The type of regularizer used for training our model. 

Supported values: 

- "l1" for using L1 regularization 

- "l2" for using L2 regularization 

- None for no regularization (default) 

intercept : bool, optional 

Boolean parameter which indicates the use or not of the 

augmented representation for training data (i.e., whether bias 

features are activated or not). 

(default: False) 

validateData : bool, optional 

Boolean parameter which indicates if the algorithm should 

validate data before training. 

(default: True) 

convergenceTol : float, optional 

A condition which decides iteration termination. 

(default: 0.001) 

""" 

warnings.warn( 

"Deprecated in 2.0.0. Use ml.regression.LinearRegression.", FutureWarning) 

def train(rdd, i): 

return callMLlibFunc("trainLinearRegressionModelWithSGD", rdd, int(iterations), 

float(step), float(miniBatchFraction), i, float(regParam), 

regType, bool(intercept), bool(validateData), 

float(convergenceTol)) 

return _regression_train_wrapper(train, LinearRegressionModel, data, initialWeights) 

@inherit_doc 

class LassoModel(LinearRegressionModelBase): 

"""A linear regression model derived from a least-squares fit with 

an l_1 penalty term. 

.. versionadded:: 0.9.0 

Examples 

-------- 

>>> from pyspark.mllib.linalg import SparseVector 

>>> from pyspark.mllib.regression import LabeledPoint 

>>> data = [ 

... LabeledPoint(0.0, [0.0]), 

... LabeledPoint(1.0, [1.0]), 

... LabeledPoint(3.0, [2.0]), 

... LabeledPoint(2.0, [3.0]) 

... ] 

>>> lrm = LassoWithSGD.train( 

... sc.parallelize(data), iterations=10, initialWeights=np.array([1.0])) 

>>> abs(lrm.predict(np.array([0.0])) - 0) < 0.5 

True 

>>> abs(lrm.predict(np.array([1.0])) - 1) < 0.5 

True 

>>> abs(lrm.predict(SparseVector(1, {0: 1.0})) - 1) < 0.5 

True 

>>> abs(lrm.predict(sc.parallelize([[1.0]])).collect()[0] - 1) < 0.5 

True 

>>> import os, tempfile 

>>> path = tempfile.mkdtemp() 

>>> lrm.save(sc, path) 

>>> sameModel = LassoModel.load(sc, path) 

>>> abs(sameModel.predict(np.array([0.0])) - 0) < 0.5 

True 

>>> abs(sameModel.predict(np.array([1.0])) - 1) < 0.5 

True 

>>> abs(sameModel.predict(SparseVector(1, {0: 1.0})) - 1) < 0.5 

True 

>>> from shutil import rmtree 

>>> try: 

... rmtree(path) 

... except: 

... pass 

>>> data = [ 

... LabeledPoint(0.0, SparseVector(1, {0: 0.0})), 

... LabeledPoint(1.0, SparseVector(1, {0: 1.0})), 

... LabeledPoint(3.0, SparseVector(1, {0: 2.0})), 

... LabeledPoint(2.0, SparseVector(1, {0: 3.0})) 

... ] 

>>> lrm = LinearRegressionWithSGD.train(sc.parallelize(data), iterations=10, 

... initialWeights=np.array([1.0])) 

>>> abs(lrm.predict(np.array([0.0])) - 0) < 0.5 

True 

>>> abs(lrm.predict(SparseVector(1, {0: 1.0})) - 1) < 0.5 

True 

>>> lrm = LassoWithSGD.train(sc.parallelize(data), iterations=10, step=1.0, 

... regParam=0.01, miniBatchFraction=1.0, initialWeights=np.array([1.0]), intercept=True, 

... validateData=True) 

>>> abs(lrm.predict(np.array([0.0])) - 0) < 0.5 

True 

>>> abs(lrm.predict(SparseVector(1, {0: 1.0})) - 1) < 0.5 

True 

""" 

@since("1.4.0") 

def save(self, sc, path): 

"""Save a LassoModel.""" 

java_model = sc._jvm.org.apache.spark.mllib.regression.LassoModel( 

_py2java(sc, self._coeff), self.intercept) 

java_model.save(sc._jsc.sc(), path) 

@classmethod 

@since("1.4.0") 

def load(cls, sc, path): 

"""Load a LassoModel.""" 

java_model = sc._jvm.org.apache.spark.mllib.regression.LassoModel.load( 

sc._jsc.sc(), path) 

weights = _java2py(sc, java_model.weights()) 

intercept = java_model.intercept() 

model = LassoModel(weights, intercept) 

return model 

class LassoWithSGD(object): 

""" 

Train a regression model with L1-regularization using Stochastic Gradient Descent. 

.. versionadded:: 0.9.0 

.. deprecated:: 2.0.0 

Use :py:class:`pyspark.ml.regression.LinearRegression` with elasticNetParam = 1.0. 

Note the default regParam is 0.01 for LassoWithSGD, but is 0.0 for LinearRegression. 

""" 

@classmethod 

def train(cls, data, iterations=100, step=1.0, regParam=0.01, 

miniBatchFraction=1.0, initialWeights=None, intercept=False, 

validateData=True, convergenceTol=0.001): 

""" 

Train a regression model with L1-regularization using Stochastic 

Gradient Descent. This solves the l1-regularized least squares 

regression formulation 

f(weights) = 1/(2n) ||A weights - y||^2 + regParam ||weights||_1 

Here the data matrix has n rows, and the input RDD holds the set 

of rows of A, each with its corresponding right hand side label y. 

See also the documentation for the precise formulation. 

.. versionadded:: 0.9.0 

Parameters 

---------- 

data : :py:class:`pyspark.RDD` 

The training data, an RDD of LabeledPoint. 

iterations : int, optional 

The number of iterations. 

(default: 100) 

step : float, optional 

The step parameter used in SGD. 

(default: 1.0) 

regParam : float, optional 

The regularizer parameter. 

(default: 0.01) 

miniBatchFraction : float, optional 

Fraction of data to be used for each SGD iteration. 

(default: 1.0) 

initialWeights : :py:class:`pyspark.mllib.linalg.Vector` or convertible, optional 

The initial weights. 

(default: None) 

intercept : bool, optional 

Boolean parameter which indicates the use or not of the 

augmented representation for training data (i.e. whether bias 

features are activated or not). 

(default: False) 

validateData : bool, optional 

Boolean parameter which indicates if the algorithm should 

validate data before training. 

(default: True) 

convergenceTol : float, optional 

A condition which decides iteration termination. 

(default: 0.001) 

""" 

warnings.warn( 

"Deprecated in 2.0.0. Use ml.regression.LinearRegression with elasticNetParam = 1.0. " 

"Note the default regParam is 0.01 for LassoWithSGD, but is 0.0 for LinearRegression.", 

FutureWarning 

) 

def train(rdd, i): 

return callMLlibFunc("trainLassoModelWithSGD", rdd, int(iterations), float(step), 

float(regParam), float(miniBatchFraction), i, bool(intercept), 

bool(validateData), float(convergenceTol)) 

return _regression_train_wrapper(train, LassoModel, data, initialWeights) 

@inherit_doc 

class RidgeRegressionModel(LinearRegressionModelBase): 

"""A linear regression model derived from a least-squares fit with 

an l_2 penalty term. 

.. versionadded:: 0.9.0 

Examples 

-------- 

>>> from pyspark.mllib.linalg import SparseVector 

>>> from pyspark.mllib.regression import LabeledPoint 

>>> data = [ 

... LabeledPoint(0.0, [0.0]), 

... LabeledPoint(1.0, [1.0]), 

... LabeledPoint(3.0, [2.0]), 

... LabeledPoint(2.0, [3.0]) 

... ] 

>>> lrm = RidgeRegressionWithSGD.train(sc.parallelize(data), iterations=10, 

... initialWeights=np.array([1.0])) 

>>> abs(lrm.predict(np.array([0.0])) - 0) < 0.5 

True 

>>> abs(lrm.predict(np.array([1.0])) - 1) < 0.5 

True 

>>> abs(lrm.predict(SparseVector(1, {0: 1.0})) - 1) < 0.5 

True 

>>> abs(lrm.predict(sc.parallelize([[1.0]])).collect()[0] - 1) < 0.5 

True 

>>> import os, tempfile 

>>> path = tempfile.mkdtemp() 

>>> lrm.save(sc, path) 

>>> sameModel = RidgeRegressionModel.load(sc, path) 

>>> abs(sameModel.predict(np.array([0.0])) - 0) < 0.5 

True 

>>> abs(sameModel.predict(np.array([1.0])) - 1) < 0.5 

True 

>>> abs(sameModel.predict(SparseVector(1, {0: 1.0})) - 1) < 0.5 

True 

>>> from shutil import rmtree 

>>> try: 

... rmtree(path) 

... except: 

... pass 

>>> data = [ 

... LabeledPoint(0.0, SparseVector(1, {0: 0.0})), 

... LabeledPoint(1.0, SparseVector(1, {0: 1.0})), 

... LabeledPoint(3.0, SparseVector(1, {0: 2.0})), 

... LabeledPoint(2.0, SparseVector(1, {0: 3.0})) 

... ] 

>>> lrm = LinearRegressionWithSGD.train(sc.parallelize(data), iterations=10, 

... initialWeights=np.array([1.0])) 

>>> abs(lrm.predict(np.array([0.0])) - 0) < 0.5 

True 

>>> abs(lrm.predict(SparseVector(1, {0: 1.0})) - 1) < 0.5 

True 

>>> lrm = RidgeRegressionWithSGD.train(sc.parallelize(data), iterations=10, step=1.0, 

... regParam=0.01, miniBatchFraction=1.0, initialWeights=np.array([1.0]), intercept=True, 

... validateData=True) 

>>> abs(lrm.predict(np.array([0.0])) - 0) < 0.5 

True 

>>> abs(lrm.predict(SparseVector(1, {0: 1.0})) - 1) < 0.5 

True 

""" 

@since("1.4.0") 

def save(self, sc, path): 

"""Save a RidgeRegressionMode.""" 

java_model = sc._jvm.org.apache.spark.mllib.regression.RidgeRegressionModel( 

_py2java(sc, self._coeff), self.intercept) 

java_model.save(sc._jsc.sc(), path) 

@classmethod 

@since("1.4.0") 

def load(cls, sc, path): 

"""Load a RidgeRegressionMode.""" 

java_model = sc._jvm.org.apache.spark.mllib.regression.RidgeRegressionModel.load( 

sc._jsc.sc(), path) 

weights = _java2py(sc, java_model.weights()) 

intercept = java_model.intercept() 

model = RidgeRegressionModel(weights, intercept) 

return model 

class RidgeRegressionWithSGD(object): 

""" 

Train a regression model with L2-regularization using Stochastic Gradient Descent. 

.. versionadded:: 0.9.0 

.. deprecated:: 2.0.0 

Use :py:class:`pyspark.ml.regression.LinearRegression` with elasticNetParam = 0.0. 

Note the default regParam is 0.01 for RidgeRegressionWithSGD, but is 0.0 for 

LinearRegression. 

""" 

@classmethod 

def train(cls, data, iterations=100, step=1.0, regParam=0.01, 

miniBatchFraction=1.0, initialWeights=None, intercept=False, 

validateData=True, convergenceTol=0.001): 

""" 

Train a regression model with L2-regularization using Stochastic 

Gradient Descent. This solves the l2-regularized least squares 

regression formulation 

f(weights) = 1/(2n) ||A weights - y||^2 + regParam/2 ||weights||^2 

Here the data matrix has n rows, and the input RDD holds the set 

of rows of A, each with its corresponding right hand side label y. 

See also the documentation for the precise formulation. 

.. versionadded:: 0.9.0 

Parameters 

---------- 

data : :py:class:`pyspark.RDD` 

The training data, an RDD of LabeledPoint. 

iterations : int, optional 

The number of iterations. 

(default: 100) 

step : float, optional 

The step parameter used in SGD. 

(default: 1.0) 

regParam : float, optional 

The regularizer parameter. 

(default: 0.01) 

miniBatchFraction : float, optional 

Fraction of data to be used for each SGD iteration. 

(default: 1.0) 

initialWeights : :py:class:`pyspark.mllib.linalg.Vector` or convertible, optional 

The initial weights. 

(default: None) 

intercept : bool, optional 

Boolean parameter which indicates the use or not of the 

augmented representation for training data (i.e. whether bias 

features are activated or not). 

(default: False) 

validateData : bool, optional 

Boolean parameter which indicates if the algorithm should 

validate data before training. 

(default: True) 

convergenceTol : float, optional 

A condition which decides iteration termination. 

(default: 0.001) 

""" 

warnings.warn( 

"Deprecated in 2.0.0. Use ml.regression.LinearRegression with elasticNetParam = 0.0. " 

"Note the default regParam is 0.01 for RidgeRegressionWithSGD, but is 0.0 for " 

"LinearRegression.", FutureWarning) 

def train(rdd, i): 

return callMLlibFunc("trainRidgeModelWithSGD", rdd, int(iterations), float(step), 

float(regParam), float(miniBatchFraction), i, bool(intercept), 

bool(validateData), float(convergenceTol)) 

return _regression_train_wrapper(train, RidgeRegressionModel, data, initialWeights) 

class IsotonicRegressionModel(Saveable, Loader): 

""" 

Regression model for isotonic regression. 

.. versionadded:: 1.4.0 

Parameters 

---------- 

boundaries : ndarray 

Array of boundaries for which predictions are known. Boundaries 

must be sorted in increasing order. 

predictions : ndarray 

Array of predictions associated to the boundaries at the same 

index. Results of isotonic regression and therefore monotone. 

isotonic : true 

Indicates whether this is isotonic or antitonic. 

Examples 

-------- 

>>> data = [(1, 0, 1), (2, 1, 1), (3, 2, 1), (1, 3, 1), (6, 4, 1), (17, 5, 1), (16, 6, 1)] 

>>> irm = IsotonicRegression.train(sc.parallelize(data)) 

>>> irm.predict(3) 

2.0 

>>> irm.predict(5) 

16.5 

>>> irm.predict(sc.parallelize([3, 5])).collect() 

[2.0, 16.5] 

>>> import os, tempfile 

>>> path = tempfile.mkdtemp() 

>>> irm.save(sc, path) 

>>> sameModel = IsotonicRegressionModel.load(sc, path) 

>>> sameModel.predict(3) 

2.0 

>>> sameModel.predict(5) 

16.5 

>>> from shutil import rmtree 

>>> try: 

... rmtree(path) 

... except OSError: 

... pass 

""" 

def __init__(self, boundaries, predictions, isotonic): 

self.boundaries = boundaries 

self.predictions = predictions 

self.isotonic = isotonic 

def predict(self, x): 

""" 

Predict labels for provided features. 

Using a piecewise linear function. 

1) If x exactly matches a boundary then associated prediction 

is returned. In case there are multiple predictions with the 

same boundary then one of them is returned. Which one is 

undefined (same as java.util.Arrays.binarySearch). 

2) If x is lower or higher than all boundaries then first or 

last prediction is returned respectively. In case there are 

multiple predictions with the same boundary then the lowest 

or highest is returned respectively. 

3) If x falls between two values in boundary array then 

prediction is treated as piecewise linear function and 

interpolated value is returned. In case there are multiple 

values with the same boundary then the same rules as in 2) 

are used. 

.. versionadded:: 1.4.0 

Parameters 

---------- 

x : :py:class:`pyspark.mllib.linalg.Vector` or :py:class:`pyspark.RDD` 

Feature or RDD of Features to be labeled. 

""" 

if isinstance(x, RDD): 

return x.map(lambda v: self.predict(v)) 

return np.interp(x, self.boundaries, self.predictions) 

@since("1.4.0") 

def save(self, sc, path): 

"""Save an IsotonicRegressionModel.""" 

java_boundaries = _py2java(sc, self.boundaries.tolist()) 

java_predictions = _py2java(sc, self.predictions.tolist()) 

java_model = sc._jvm.org.apache.spark.mllib.regression.IsotonicRegressionModel( 

java_boundaries, java_predictions, self.isotonic) 

java_model.save(sc._jsc.sc(), path) 

@classmethod 

@since("1.4.0") 

def load(cls, sc, path): 

"""Load an IsotonicRegressionModel.""" 

java_model = sc._jvm.org.apache.spark.mllib.regression.IsotonicRegressionModel.load( 

sc._jsc.sc(), path) 

py_boundaries = _java2py(sc, java_model.boundaryVector()).toArray() 

py_predictions = _java2py(sc, java_model.predictionVector()).toArray() 

return IsotonicRegressionModel(py_boundaries, py_predictions, java_model.isotonic) 

class IsotonicRegression(object): 

""" 

Isotonic regression. 

Currently implemented using parallelized pool adjacent violators 

algorithm. Only univariate (single feature) algorithm supported. 

.. versionadded:: 1.4.0 

Notes 

----- 

Sequential PAV implementation based on 

Tibshirani, Ryan J., Holger Hoefling, and Robert Tibshirani (2011) [1]_ 

Sequential PAV parallelization based on 

Kearsley, Anthony J., Richard A. Tapia, and Michael W. Trosset (1996) [2]_ 

See also 

`Isotonic regression (Wikipedia) <http://en.wikipedia.org/wiki/Isotonic_regression>`_. 

.. [1] Tibshirani, Ryan J., Holger Hoefling, and Robert Tibshirani. 

"Nearly-isotonic regression." Technometrics 53.1 (2011): 54-61. 

Available from http://www.stat.cmu.edu/~ryantibs/papers/neariso.pdf 

.. [2] Kearsley, Anthony J., Richard A. Tapia, and Michael W. Trosset 

"An approach to parallelizing isotonic regression." 

Applied Mathematics and Parallel Computing. Physica-Verlag HD, 1996. 141-147. 

Available from http://softlib.rice.edu/pub/CRPC-TRs/reports/CRPC-TR96640.pdf 

""" 

@classmethod 

def train(cls, data, isotonic=True): 

""" 

Train an isotonic regression model on the given data. 

.. versionadded:: 1.4.0 

Parameters 

---------- 

data : :py:class:`pyspark.RDD` 

RDD of (label, feature, weight) tuples. 

isotonic : bool, optional 

Whether this is isotonic (which is default) or antitonic. 

(default: True) 

""" 

boundaries, predictions = callMLlibFunc("trainIsotonicRegressionModel", 

data.map(_convert_to_vector), bool(isotonic)) 

return IsotonicRegressionModel(boundaries.toArray(), predictions.toArray(), isotonic) 

class StreamingLinearAlgorithm(object): 

""" 

Base class that has to be inherited by any StreamingLinearAlgorithm. 

Prevents reimplementation of methods predictOn and predictOnValues. 

.. versionadded:: 1.5.0 

""" 

def __init__(self, model): 

self._model = model 

@since("1.5.0") 

def latestModel(self): 

""" 

Returns the latest model. 

""" 

return self._model 

def _validate(self, dstream): 

786 ↛ 787line 786 didn't jump to line 787, because the condition on line 786 was never true if not isinstance(dstream, DStream): 

raise TypeError( 

"dstream should be a DStream object, got %s" % type(dstream)) 

789 ↛ 790line 789 didn't jump to line 790, because the condition on line 789 was never true if not self._model: 

raise ValueError( 

"Model must be initialized using setInitialWeights") 

def predictOn(self, dstream): 

""" 

Use the model to make predictions on batches of data from a 

DStream. 

.. versionadded:: 1.5.0 

Returns 

------- 

:py:class:`pyspark.streaming.DStream` 

DStream containing predictions. 

""" 

self._validate(dstream) 

return dstream.map(lambda x: self._model.predict(x)) 

def predictOnValues(self, dstream): 

""" 

Use the model to make predictions on the values of a DStream and 

carry over its keys. 

.. versionadded:: 1.5.0 

Returns 

------- 

:py:class:`pyspark.streaming.DStream` 

DStream containing predictions. 

""" 

self._validate(dstream) 

return dstream.mapValues(lambda x: self._model.predict(x)) 

@inherit_doc 

class StreamingLinearRegressionWithSGD(StreamingLinearAlgorithm): 

""" 

Train or predict a linear regression model on streaming data. 

Training uses Stochastic Gradient Descent to update the model 

based on each new batch of incoming data from a DStream 

(see `LinearRegressionWithSGD` for model equation). 

Each batch of data is assumed to be an RDD of LabeledPoints. 

The number of data points per batch can vary, but the number 

of features must be constant. An initial weight vector must 

be provided. 

.. versionadded:: 1.5.0 

Parameters 

---------- 

stepSize : float, optional 

Step size for each iteration of gradient descent. 

(default: 0.1) 

numIterations : int, optional 

Number of iterations run for each batch of data. 

(default: 50) 

miniBatchFraction : float, optional 

Fraction of each batch of data to use for updates. 

(default: 1.0) 

convergenceTol : float, optional 

Value used to determine when to terminate iterations. 

(default: 0.001) 

""" 

def __init__(self, stepSize=0.1, numIterations=50, miniBatchFraction=1.0, convergenceTol=0.001): 

self.stepSize = stepSize 

self.numIterations = numIterations 

self.miniBatchFraction = miniBatchFraction 

self.convergenceTol = convergenceTol 

self._model = None 

super(StreamingLinearRegressionWithSGD, self).__init__( 

model=self._model) 

@since("1.5.0") 

def setInitialWeights(self, initialWeights): 

""" 

Set the initial value of weights. 

This must be set before running trainOn and predictOn 

""" 

initialWeights = _convert_to_vector(initialWeights) 

self._model = LinearRegressionModel(initialWeights, 0) 

return self 

@since("1.5.0") 

def trainOn(self, dstream): 

"""Train the model on the incoming dstream.""" 

self._validate(dstream) 

def update(rdd): 

# LinearRegressionWithSGD.train raises an error for an empty RDD. 

if not rdd.isEmpty(): 

self._model = LinearRegressionWithSGD.train( 

rdd, self.numIterations, self.stepSize, 

self.miniBatchFraction, self._model.weights, 

intercept=self._model.intercept, convergenceTol=self.convergenceTol) 

dstream.foreachRDD(update) 

def _test(): 

import doctest 

from pyspark.sql import SparkSession 

import pyspark.mllib.regression 

globs = pyspark.mllib.regression.__dict__.copy() 

spark = SparkSession.builder\ 

.master("local[2]")\ 

.appName("mllib.regression tests")\ 

.getOrCreate() 

globs['sc'] = spark.sparkContext 

(failure_count, test_count) = doctest.testmod(globs=globs, optionflags=doctest.ELLIPSIS) 

spark.stop() 

902 ↛ 903line 902 didn't jump to line 903, because the condition on line 902 was never true if failure_count: 

sys.exit(-1) 

if __name__ == "__main__": 

_test()