Sparkit-learn

PySpark + Scikit-learn = Sparkit-learn

GitHub: https://github.com/lensacom/sparkit-learn

About

Sparkit-learn aims to provide scikit-learn functionality and API on PySpark. The main goal of the library is to create an API that stays close to sklearn's.

The driving principle was to "Think locally, execute distributively." To accomodate this concept, the basic data block is always an array or a (sparse) matrix and the operations are executed on block level.

Requirements

• Python 2.7.x or 3.4.x
• Spark[>=1.3.0]
• NumPy[>=1.9.0]
• SciPy[>=0.14.0]
• Scikit-learn[>=0.16]

Run IPython from notebooks directory

PYTHONPATH=${PYTHONPATH}:.. IPYTHON_OPTS="notebook" ${SPARK_HOME}/bin/pyspark --master local\[4\] --driver-memory 2G

Run tests with

./runtests.sh

Quick start

Sparkit-learn introduces three important distributed data format:

• ArrayRDD:

  A numpy.array like distributed array

  from splearn.rdd import ArrayRDD
  
  data = range(20)
  # PySpark RDD with 2 partitions
  rdd = sc.parallelize(data, 2) # each partition with 10 elements
  # ArrayRDD
  # each partition will contain blocks with 5 elements
  X = ArrayRDD(rdd, bsize=5) # 4 blocks, 2 in each partition

  Basic operations:

  len(X) # 20 - number of elements in the whole dataset
  X.blocks # 4 - number of blocks
  X.shape # (20,) - the shape of the whole dataset
  
  X # returns an ArrayRDD
  # <class 'splearn.rdd.ArrayRDD'> from PythonRDD...
  
  X.dtype # returns the type of the blocks
  # numpy.ndarray
  
  X.collect() # get the dataset
  # [array([0, 1, 2, 3, 4]),
  #  array([5, 6, 7, 8, 9]),
  #  array([10, 11, 12, 13, 14]),
  #  array([15, 16, 17, 18, 19])]
  
  X[1].collect() # indexing
  # [array([5, 6, 7, 8, 9])]
  
  X[1] # also returns an ArrayRDD!
  
  X[1::2].collect() # slicing
  # [array([5, 6, 7, 8, 9]),
  #  array([15, 16, 17, 18, 19])]
  
  X[1::2] # returns an ArrayRDD as well
  
  X.tolist() # returns the dataset as a list
  # [0, 1, 2, ... 17, 18, 19]
  X.toarray() # returns the dataset as a numpy.array
  # array([ 0,  1,  2, ... 17, 18, 19])
  
  # pyspark.rdd operations will still work
  X.getNumPartitions() # 2 - number of partitions

• SparseRDD:

  The sparse counterpart of the ArrayRDD, the main difference is that the blocks are sparse matrices. The reason behind this split is to follow the distinction between numpy.ndarray*s and *scipy.sparse matrices. Usually the SparseRDD is created by splearn's transformators, but one can instantiate too.

  # generate a SparseRDD from a text using SparkCountVectorizer
  from splearn.rdd import SparseRDD
  from sklearn.feature_extraction.tests.test_text import ALL_FOOD_DOCS
  ALL_FOOD_DOCS
  #(u'the pizza pizza beer copyright',
  # u'the pizza burger beer copyright',
  # u'the the pizza beer beer copyright',
  # u'the burger beer beer copyright',
  # u'the coke burger coke copyright',
  # u'the coke burger burger',
  # u'the salad celeri copyright',
  # u'the salad salad sparkling water copyright',
  # u'the the celeri celeri copyright',
  # u'the tomato tomato salad water',
  # u'the tomato salad water copyright')
  
  # ArrayRDD created from the raw data
  X = ArrayRDD(sc.parallelize(ALL_FOOD_DOCS, 4), 2)
  X.collect()
  # [array([u'the pizza pizza beer copyright',
  #         u'the pizza burger beer copyright'], dtype='<U31'),
  #  array([u'the the pizza beer beer copyright',
  #         u'the burger beer beer copyright'], dtype='<U33'),
  #  array([u'the coke burger coke copyright',
  #         u'the coke burger burger'], dtype='<U30'),
  #  array([u'the salad celeri copyright',
  #         u'the salad salad sparkling water copyright'], dtype='<U41'),
  #  array([u'the the celeri celeri copyright',
  #         u'the tomato tomato salad water'], dtype='<U31'),
  #  array([u'the tomato salad water copyright'], dtype='<U32')]
  
  # Feature extraction executed
  from splearn.feature_extraction.text import SparkCountVectorizer
  vect = SparkCountVectorizer()
  X = vect.fit_transform(X)
  # and we have a SparseRDD
  X
  # <class 'splearn.rdd.SparseRDD'> from PythonRDD...
  
  # it's type is the scipy.sparse's general parent
  X.dtype
  # scipy.sparse.base.spmatrix
  
  # slicing works just like in ArrayRDDs
  X[2:4].collect()
  # [<2x11 sparse matrix of type '<type 'numpy.int64'>'
  #   with 7 stored elements in Compressed Sparse Row format>,
  #  <2x11 sparse matrix of type '<type 'numpy.int64'>'
  #   with 9 stored elements in Compressed Sparse Row format>]
  
  # general mathematical operations are available
  X.sum(), X.mean(), X.max(), X.min()
  # (55, 0.45454545454545453, 2, 0)
  
  # even with axis parameters provided
  X.sum(axis=1)
  # matrix([[5],
  #         [5],
  #         [6],
  #         [5],
  #         [5],
  #         [4],
  #         [4],
  #         [6],
  #         [5],
  #         [5],
  #         [5]])
  
  # It can be transformed to dense ArrayRDD
  X.todense()
  # <class 'splearn.rdd.ArrayRDD'> from PythonRDD...
  X.todense().collect()
  # [array([[1, 0, 0, 0, 1, 2, 0, 0, 1, 0, 0],
  #         [1, 1, 0, 0, 1, 1, 0, 0, 1, 0, 0]]),
  #  array([[2, 0, 0, 0, 1, 1, 0, 0, 2, 0, 0],
  #         [2, 1, 0, 0, 1, 0, 0, 0, 1, 0, 0]]),
  #  array([[0, 1, 0, 2, 1, 0, 0, 0, 1, 0, 0],
  #         [0, 2, 0, 1, 0, 0, 0, 0, 1, 0, 0]]),
  #  array([[0, 0, 1, 0, 1, 0, 1, 0, 1, 0, 0],
  #         [0, 0, 0, 0, 1, 0, 2, 1, 1, 0, 1]]),
  #  array([[0, 0, 2, 0, 1, 0, 0, 0, 2, 0, 0],
  #         [0, 0, 0, 0, 0, 0, 1, 0, 1, 2, 1]]),
  #  array([[0, 0, 0, 0, 1, 0, 1, 0, 1, 1, 1]])]
  
  # One can instantiate SparseRDD manually too:
  sparse = sc.parallelize(np.array([sp.eye(2).tocsr()]*20), 2)
  sparse = SparseRDD(sparse, bsize=5)
  sparse
  # <class 'splearn.rdd.SparseRDD'> from PythonRDD...
  
  sparse.collect()
  # [<10x2 sparse matrix of type '<type 'numpy.float64'>'
  #   with 10 stored elements in Compressed Sparse Row format>,
  #  <10x2 sparse matrix of type '<type 'numpy.float64'>'
  #   with 10 stored elements in Compressed Sparse Row format>,
  #  <10x2 sparse matrix of type '<type 'numpy.float64'>'
  #   with 10 stored elements in Compressed Sparse Row format>,
  #  <10x2 sparse matrix of type '<type 'numpy.float64'>'
  #   with 10 stored elements in Compressed Sparse Row format>]

• DictRDD:

  A column based data format, each column with it's own type.

  from splearn.rdd import DictRDD
  
  X = range(20)
  y = list(range(2)) * 10
  # PySpark RDD with 2 partitions
  X_rdd = sc.parallelize(X, 2) # each partition with 10 elements
  y_rdd = sc.parallelize(y, 2) # each partition with 10 elements
  # DictRDD
  # each partition will contain blocks with 5 elements
  Z = DictRDD((X_rdd, y_rdd),
              columns=('X', 'y'),
              bsize=5,
              dtype=[np.ndarray, np.ndarray]) # 4 blocks, 2/partition
  # if no dtype is provided, the type of the blocks will be determined
  # automatically
  
  # or:
  import numpy as np
  
  data = np.array([range(20), list(range(2))*10]).T
  rdd = sc.parallelize(data, 2)
  Z = DictRDD(rdd,
              columns=('X', 'y'),
              bsize=5,
              dtype=[np.ndarray, np.ndarray])

  Basic operations:

  len(Z) # 8 - number of blocks
  Z.columns # returns ('X', 'y')
  Z.dtype # returns the types in correct order
  # [numpy.ndarray, numpy.ndarray]
  
  Z # returns a DictRDD
  #<class 'splearn.rdd.DictRDD'> from PythonRDD...
  
  Z.collect()
  # [(array([0, 1, 2, 3, 4]), array([0, 1, 0, 1, 0])),
  #  (array([5, 6, 7, 8, 9]), array([1, 0, 1, 0, 1])),
  #  (array([10, 11, 12, 13, 14]), array([0, 1, 0, 1, 0])),
  #  (array([15, 16, 17, 18, 19]), array([1, 0, 1, 0, 1]))]
  
  Z[:, 'y'] # column select - returns an ArrayRDD
  Z[:, 'y'].collect()
  # [array([0, 1, 0, 1, 0]),
  #  array([1, 0, 1, 0, 1]),
  #  array([0, 1, 0, 1, 0]),
  #  array([1, 0, 1, 0, 1])]
  
  Z[:-1, ['X', 'y']] # slicing - DictRDD
  Z[:-1, ['X', 'y']].collect()
  # [(array([0, 1, 2, 3, 4]), array([0, 1, 0, 1, 0])),
  #  (array([5, 6, 7, 8, 9]), array([1, 0, 1, 0, 1])),
  #  (array([10, 11, 12, 13, 14]), array([0, 1, 0, 1, 0]))]

Basic workflow

With the use of the described data structures, the basic workflow is almost identical to sklearn's.

Distributed vectorizing of texts

SparkCountVectorizer

from splearn.rdd import ArrayRDD
from splearn.feature_extraction.text import SparkCountVectorizer
from sklearn.feature_extraction.text import CountVectorizer

X = [...]  # list of texts
X_rdd = ArrayRDD(sc.parallelize(X, 4))  # sc is SparkContext

local = CountVectorizer()
dist = SparkCountVectorizer()

result_local = local.fit_transform(X)
result_dist = dist.fit_transform(X_rdd)  # SparseRDD

SparkHashingVectorizer

from splearn.rdd import ArrayRDD
from splearn.feature_extraction.text import SparkHashingVectorizer
from sklearn.feature_extraction.text import HashingVectorizer

X = [...]  # list of texts
X_rdd = ArrayRDD(sc.parallelize(X, 4))  # sc is SparkContext

local = HashingVectorizer()
dist = SparkHashingVectorizer()

result_local = local.fit_transform(X)
result_dist = dist.fit_transform(X_rdd)  # SparseRDD

SparkTfidfTransformer

from splearn.rdd import ArrayRDD
from splearn.feature_extraction.text import SparkHashingVectorizer
from splearn.feature_extraction.text import SparkTfidfTransformer
from splearn.pipeline import SparkPipeline

from sklearn.feature_extraction.text import HashingVectorizer
from sklearn.feature_extraction.text import TfidfTransformer
from sklearn.pipeline import Pipeline

X = [...]  # list of texts
X_rdd = ArrayRDD(sc.parallelize(X, 4))  # sc is SparkContext

local_pipeline = Pipeline((
    ('vect', HashingVectorizer()),
    ('tfidf', TfidfTransformer())
))
dist_pipeline = SparkPipeline((
    ('vect', SparkHashingVectorizer()),
    ('tfidf', SparkTfidfTransformer())
))

result_local = local_pipeline.fit_transform(X)
result_dist = dist_pipeline.fit_transform(X_rdd)  # SparseRDD

Distributed Classifiers

from splearn.rdd import DictRDD
from splearn.feature_extraction.text import SparkHashingVectorizer
from splearn.feature_extraction.text import SparkTfidfTransformer
from splearn.svm import SparkLinearSVC
from splearn.pipeline import SparkPipeline

from sklearn.feature_extraction.text import HashingVectorizer
from sklearn.feature_extraction.text import TfidfTransformer
from sklearn.svm import LinearSVC
from sklearn.pipeline import Pipeline

X = [...]  # list of texts
y = [...]  # list of labels
X_rdd = sc.parallelize(X, 4)
y_rdd = sc.parallelize(y, 4)
Z = DictRDD((X_rdd, y_rdd),
            columns=('X', 'y'),
            dtype=[np.ndarray, np.ndarray])

local_pipeline = Pipeline((
    ('vect', HashingVectorizer()),
    ('tfidf', TfidfTransformer()),
    ('clf', LinearSVC())
))
dist_pipeline = SparkPipeline((
    ('vect', SparkHashingVectorizer()),
    ('tfidf', SparkTfidfTransformer()),
    ('clf', SparkLinearSVC())
))

local_pipeline.fit(X, y)
dist_pipeline.fit(Z, clf__classes=np.unique(y))

y_pred_local = local_pipeline.predict(X)
y_pred_dist = dist_pipeline.predict(Z[:, 'X'])

Distributed Model Selection

from splearn.rdd import DictRDD
from splearn.grid_search import SparkGridSearchCV
from splearn.naive_bayes import SparkMultinomialNB

from sklearn.grid_search import GridSearchCV
from sklearn.naive_bayes import MultinomialNB

X = [...]
y = [...]
X_rdd = sc.parallelize(X, 4)
y_rdd = sc.parallelize(y, 4)
Z = DictRDD((X_rdd, y_rdd),
            columns=('X', 'y'),
            dtype=[np.ndarray, np.ndarray])

parameters = {'alpha': [0.1, 1, 10]}
fit_params = {'classes': np.unique(y)}

local_estimator = MultinomialNB()
local_grid = GridSearchCV(estimator=local_estimator,
                          param_grid=parameters)

estimator = SparkMultinomialNB()
grid = SparkGridSearchCV(estimator=estimator,
                         param_grid=parameters,
                         fit_params=fit_params)

local_grid.fit(X, y)
grid.fit(Z)

ROADMAP

• [ ] Transparent API to support plain numpy and scipy objects (partially done in the transparent_api branch)
• [ ] Update all dependencies
• [ ] Use Mllib and ML packages more extensively (since it becames more mature)
• [ ] Support Spark DataFrames

Special thanks

• scikit-learn community
• spylearn community
• pyspark community

Similar Projects

• Thunder
• Bolt