[SPARK-4586][MLLIB] Python API for ML pipeline and parameters

This PR adds Python API for ML pipeline and parameters. The design doc can be found on the JIRA page. It includes transformers and an estimator to demo the simple text classification example code.

TODO:
- [x] handle parameters in LRModel
- [x] unit tests
- [x] missing some docs

CC: davies jkbradley

Author: Xiangrui Meng <meng@databricks.com>
Author: Davies Liu <davies@databricks.com>

Closes #4151 from mengxr/SPARK-4586 and squashes the following commits:

415268e [Xiangrui Meng] remove inherit_doc from __init__
edbd6fe [Xiangrui Meng] move Identifiable to ml.util
44c2405 [Xiangrui Meng] Merge pull request #2 from davies/ml
dd1256b [Xiangrui Meng] Merge remote-tracking branch 'apache/master' into SPARK-4586
14ae7e2 [Davies Liu] fix docs
54ca7df [Davies Liu] fix tests
78638df [Davies Liu] Merge branch 'SPARK-4586' of github.com:mengxr/spark into ml
fc59a02 [Xiangrui Meng] Merge remote-tracking branch 'apache/master' into SPARK-4586
1dca16a [Davies Liu] refactor
090b3a3 [Davies Liu] Merge branch 'master' of github.com:apache/spark into ml
0882513 [Xiangrui Meng] update doc style
a4f4dbf [Xiangrui Meng] add unit test for LR
7521d1c [Xiangrui Meng] add unit tests to HashingTF and Tokenizer
ba0ba1e [Xiangrui Meng] add unit tests for pipeline
0586c7b [Xiangrui Meng] add more comments to the example
5153cff [Xiangrui Meng] simplify java models
036ca04 [Xiangrui Meng] gen numFeatures
46fa147 [Xiangrui Meng] update mllib/pom.xml to include python files in the assembly
1dcc17e [Xiangrui Meng] update code gen and make param appear in the doc
f66ba0c [Xiangrui Meng] make params a property
d5efd34 [Xiangrui Meng] update doc conf and move embedded param map to instance attribute
f4d0fe6 [Xiangrui Meng] use LabeledDocument and Document in example
05e3e40 [Xiangrui Meng] update example
d3e8dbe [Xiangrui Meng] more docs optimize pipeline.fit impl
56de571 [Xiangrui Meng] fix style
d0c5bb8 [Xiangrui Meng] a working copy
bce72f4 [Xiangrui Meng] Merge remote-tracking branch 'apache/master' into SPARK-4586
17ecfb9 [Xiangrui Meng] code gen for shared params
d9ea77c [Xiangrui Meng] update doc
c18dca1 [Xiangrui Meng] make the example working
dadd84e [Xiangrui Meng] add base classes and docs
a3015cf [Xiangrui Meng] add Estimator and Transformer
46eea43 [Xiangrui Meng] a pipeline in python
33b68e0 [Xiangrui Meng] a working LR

Author: mengxr

examples/src/main/python/ml/simple_text_classification_pipeline.py

@@ -0,0 +1,79 @@
+#
+# 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.
+#
+
+from pyspark import SparkContext
+from pyspark.sql import SQLContext, Row
+from pyspark.ml import Pipeline
+from pyspark.ml.feature import HashingTF, Tokenizer
+from pyspark.ml.classification import LogisticRegression
+
+
+"""
+A simple text classification pipeline that recognizes "spark" from
+input text. This is to show how to create and configure a Spark ML
+pipeline in Python. Run with:
+
+  bin/spark-submit examples/src/main/python/ml/simple_text_classification_pipeline.py
+"""
+
+
+if __name__ == "__main__":
+    sc = SparkContext(appName="SimpleTextClassificationPipeline")
+    sqlCtx = SQLContext(sc)
+
+    # Prepare training documents, which are labeled.
+    LabeledDocument = Row('id', 'text', 'label')
+    training = sqlCtx.inferSchema(
+        sc.parallelize([(0L, "a b c d e spark", 1.0),
+                        (1L, "b d", 0.0),
+                        (2L, "spark f g h", 1.0),
+                        (3L, "hadoop mapreduce", 0.0)])
+          .map(lambda x: LabeledDocument(*x)))
+
+    # Configure an ML pipeline, which consists of tree stages: tokenizer, hashingTF, and lr.
+    tokenizer = Tokenizer() \
+        .setInputCol("text") \
+        .setOutputCol("words")
+    hashingTF = HashingTF() \
+        .setInputCol(tokenizer.getOutputCol()) \
+        .setOutputCol("features")
+    lr = LogisticRegression() \
+        .setMaxIter(10) \
+        .setRegParam(0.01)
+    pipeline = Pipeline() \
+        .setStages([tokenizer, hashingTF, lr])
+
+    # Fit the pipeline to training documents.
+    model = pipeline.fit(training)
+
+    # Prepare test documents, which are unlabeled.
+    Document = Row('id', 'text')
+    test = sqlCtx.inferSchema(
+        sc.parallelize([(4L, "spark i j k"),
+                        (5L, "l m n"),
+                        (6L, "mapreduce spark"),
+                        (7L, "apache hadoop")])
+          .map(lambda x: Document(*x)))
+
+    # Make predictions on test documents and print columns of interest.
+    prediction = model.transform(test)
+    prediction.registerTempTable("prediction")
+    selected = sqlCtx.sql("SELECT id, text, prediction from prediction")
+    for row in selected.collect():
+        print row
+
+    sc.stop()

mllib/pom.xml

@@ -125,6 +125,8 @@
         <directory>../python</directory>
         <includes>
           <include>pyspark/mllib/*.py</include>
+          <include>pyspark/ml/*.py</include>
+          <include>pyspark/ml/param/*.py</include>
         </includes>
       </resource>
     </resources>

mllib/src/main/scala/org/apache/spark/ml/param/params.scala

@@ -164,6 +164,13 @@ trait Params extends Identifiable with Serializable {
     this
   }
 
+  /**
+   * Sets a parameter (by name) in the embedded param map.
+   */
+  private[ml] def set(param: String, value: Any): this.type = {
+    set(getParam(param), value)
+  }
+
   /**
    * Gets the value of a parameter in the embedded param map.
    */
@@ -286,7 +293,6 @@ class ParamMap private[ml] (private val map: mutable.Map[Param[Any], Any]) exten
     new ParamMap(this.map ++ other.map)
   }
 
-
   /**
    * Adds all parameters from the input param map into this param map.
    */

python/docs/conf.py

@@ -55,9 +55,9 @@
 # built documents.
 #
 # The short X.Y version.
-version = '1.2-SNAPSHOT'
+version = '1.3-SNAPSHOT'
 # The full version, including alpha/beta/rc tags.
-release = '1.2-SNAPSHOT'
+release = '1.3-SNAPSHOT'
 
 # The language for content autogenerated by Sphinx. Refer to documentation
 # for a list of supported languages.

python/docs/index.rst

@@ -14,6 +14,7 @@ Contents:
    pyspark
    pyspark.sql
    pyspark.streaming
+   pyspark.ml
    pyspark.mllib
 
 

python/docs/pyspark.ml.rst

@@ -0,0 +1,29 @@
+pyspark.ml package
+=====================
+
+Submodules
+----------
+
+pyspark.ml module
+-----------------
+
+.. automodule:: pyspark.ml
+    :members:
+    :undoc-members:
+    :inherited-members:
+
+pyspark.ml.feature module
+-------------------------
+
+.. automodule:: pyspark.ml.feature
+    :members:
+    :undoc-members:
+    :inherited-members:
+
+pyspark.ml.classification module
+--------------------------------
+
+.. automodule:: pyspark.ml.classification
+    :members:
+    :undoc-members:
+    :inherited-members:

python/docs/pyspark.rst

@@ -9,6 +9,7 @@ Subpackages
 
     pyspark.sql
     pyspark.streaming
+    pyspark.ml
     pyspark.mllib
 
 Contents

python/pyspark/ml/__init__.py

@@ -0,0 +1,21 @@
+#
+# 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.
+#
+
+from pyspark.ml.param import *
+from pyspark.ml.pipeline import *
+
+__all__ = ["Param", "Params", "Transformer", "Estimator", "Pipeline"]

python/pyspark/ml/classification.py

@@ -0,0 +1,76 @@
+#
+# 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.
+#
+
+from pyspark.ml.util import inherit_doc
+from pyspark.ml.wrapper import JavaEstimator, JavaModel
+from pyspark.ml.param.shared import HasFeaturesCol, HasLabelCol, HasPredictionCol, HasMaxIter,\
+    HasRegParam
+
+
+__all__ = ['LogisticRegression', 'LogisticRegressionModel']
+
+
+@inherit_doc
+class LogisticRegression(JavaEstimator, HasFeaturesCol, HasLabelCol, HasPredictionCol, HasMaxIter,
+                         HasRegParam):
+    """
+    Logistic regression.
+
+    >>> from pyspark.sql import Row
+    >>> from pyspark.mllib.linalg import Vectors
+    >>> dataset = sqlCtx.inferSchema(sc.parallelize([ \
+            Row(label=1.0, features=Vectors.dense(1.0)), \
+            Row(label=0.0, features=Vectors.sparse(1, [], []))]))
+    >>> lr = LogisticRegression() \
+            .setMaxIter(5) \
+            .setRegParam(0.01)
+    >>> model = lr.fit(dataset)
+    >>> test0 = sqlCtx.inferSchema(sc.parallelize([Row(features=Vectors.dense(-1.0))]))
+    >>> print model.transform(test0).head().prediction
+    0.0
+    >>> test1 = sqlCtx.inferSchema(sc.parallelize([Row(features=Vectors.sparse(1, [0], [1.0]))]))
+    >>> print model.transform(test1).head().prediction
+    1.0
+    """
+    _java_class = "org.apache.spark.ml.classification.LogisticRegression"
+
+    def _create_model(self, java_model):
+        return LogisticRegressionModel(java_model)
+
+
+class LogisticRegressionModel(JavaModel):
+    """
+    Model fitted by LogisticRegression.
+    """
+
+
+if __name__ == "__main__":
+    import doctest
+    from pyspark.context import SparkContext
+    from pyspark.sql import SQLContext
+    globs = globals().copy()
+    # The small batch size here ensures that we see multiple batches,
+    # even in these small test examples:
+    sc = SparkContext("local[2]", "ml.feature tests")
+    sqlCtx = SQLContext(sc)
+    globs['sc'] = sc
+    globs['sqlCtx'] = sqlCtx
+    (failure_count, test_count) = doctest.testmod(
+        globs=globs, optionflags=doctest.ELLIPSIS)
+    sc.stop()
+    if failure_count:
+        exit(-1)

python/pyspark/ml/feature.py

@@ -0,0 +1,82 @@
+#
+# 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.
+#
+
+from pyspark.ml.param.shared import HasInputCol, HasOutputCol, HasNumFeatures
+from pyspark.ml.util import inherit_doc
+from pyspark.ml.wrapper import JavaTransformer
+
+__all__ = ['Tokenizer', 'HashingTF']
+
+
+@inherit_doc
+class Tokenizer(JavaTransformer, HasInputCol, HasOutputCol):
+    """
+    A tokenizer that converts the input string to lowercase and then
+    splits it by white spaces.
+
+    >>> from pyspark.sql import Row
+    >>> dataset = sqlCtx.inferSchema(sc.parallelize([Row(text="a b c")]))
+    >>> tokenizer = Tokenizer() \
+            .setInputCol("text") \
+            .setOutputCol("words")
+    >>> print tokenizer.transform(dataset).head()
+    Row(text=u'a b c', words=[u'a', u'b', u'c'])
+    >>> print tokenizer.transform(dataset, {tokenizer.outputCol: "tokens"}).head()
+    Row(text=u'a b c', tokens=[u'a', u'b', u'c'])
+    """
+
+    _java_class = "org.apache.spark.ml.feature.Tokenizer"
+
+
+@inherit_doc
+class HashingTF(JavaTransformer, HasInputCol, HasOutputCol, HasNumFeatures):
+    """
+    Maps a sequence of terms to their term frequencies using the
+    hashing trick.
+
+    >>> from pyspark.sql import Row
+    >>> dataset = sqlCtx.inferSchema(sc.parallelize([Row(words=["a", "b", "c"])]))
+    >>> hashingTF = HashingTF() \
+            .setNumFeatures(10) \
+            .setInputCol("words") \
+            .setOutputCol("features")
+    >>> print hashingTF.transform(dataset).head().features
+    (10,[7,8,9],[1.0,1.0,1.0])
+    >>> params = {hashingTF.numFeatures: 5, hashingTF.outputCol: "vector"}
+    >>> print hashingTF.transform(dataset, params).head().vector
+    (5,[2,3,4],[1.0,1.0,1.0])
+    """
+
+    _java_class = "org.apache.spark.ml.feature.HashingTF"
+
+
+if __name__ == "__main__":
+    import doctest
+    from pyspark.context import SparkContext
+    from pyspark.sql import SQLContext
+    globs = globals().copy()
+    # The small batch size here ensures that we see multiple batches,
+    # even in these small test examples:
+    sc = SparkContext("local[2]", "ml.feature tests")
+    sqlCtx = SQLContext(sc)
+    globs['sc'] = sc
+    globs['sqlCtx'] = sqlCtx
+    (failure_count, test_count) = doctest.testmod(
+        globs=globs, optionflags=doctest.ELLIPSIS)
+    sc.stop()
+    if failure_count:
+        exit(-1)