[WIP] SPARK-1430: Support sparse data in Python MLlib

This PR adds a SparseVector class in PySpark and updates all the regression, classification and clustering algorithms and models to support sparse data, similar to MLlib. I chose to add this class because SciPy is quite difficult to install in many environments (more so than NumPy), but I plan to add support for SciPy sparse vectors later too, and make the methods work transparently on objects of either type.

On the Scala side, we keep Python sparse vectors sparse and pass them to MLlib. We always return dense vectors from our models.

Some to-do items left:
- [x] Support SciPy's scipy.sparse matrix objects when SciPy is available. We can easily add a function to convert these to our own SparseVector.
- [x] MLlib currently uses a vector with one extra column on the left to represent what we call LabeledPoint in Scala. Do we really want this? It may get annoying once you deal with sparse data since you must add/subtract 1 to each feature index when training. We can remove this API in 1.0 and use tuples for labeling.
- [x] Explain how to use these in the Python MLlib docs.

CC @mengxr, @joshrosen

Author: Matei Zaharia <matei@databricks.com>

Closes #341 from mateiz/py-ml-update and squashes the following commits:

d52e763 [Matei Zaharia] Remove no-longer-needed slice code and handle review comments
ea5a25a [Matei Zaharia] Fix remaining uses of copyto() after merge
b9f97a3 [Matei Zaharia] Fix test
1e1bd0f [Matei Zaharia] Add MLlib logistic regression example in Python
88bc01f [Matei Zaharia] Clean up inheritance of LinearModel in Python, and expose its parametrs
37ab747 [Matei Zaharia] Fix some examples and docs due to changes in MLlib API
da0f27e [Matei Zaharia] Added a MLlib K-means example and updated docs to discuss sparse data
c48e85a [Matei Zaharia] Added some tests for passing lists as input, and added mllib/tests.py to run-tests script.
a07ba10 [Matei Zaharia] Fix some typos and calculation of initial weights
74eefe7 [Matei Zaharia] Added LabeledPoint class in Python
889dde8 [Matei Zaharia] Support scipy.sparse matrices in all our algorithms and models
ab244d1 [Matei Zaharia] Allow SparseVectors to be initialized using a dict
a5d6426 [Matei Zaharia] Add linalg.py to run-tests script
0e7a3d8 [Matei Zaharia] Keep vectors sparse in Java when reading LabeledPoints
eaee759 [Matei Zaharia] Update regression, classification and clustering models for sparse data
2abbb44 [Matei Zaharia] Further work to get linear models working with sparse data
154f45d [Matei Zaharia] Update docs, name some magic values
881fef7 [Matei Zaharia] Added a sparse vector in Python and made Java-Python format more compact

3 files changed, 132 insertions, 50 deletions

diff --git a/mllib/src/main/scala/org/apache/spark/mllib/api/python/PythonMLLibAPI.scala b/mllib/src/main/scala/org/apache/spark/mllib/api/python/PythonMLLibAPI.scala
index a6c049e517..7c65b0d475 100644
--- a/mllib/src/main/scala/org/apache/spark/mllib/api/python/PythonMLLibAPI.scala
+++ b/mllib/src/main/scala/org/apache/spark/mllib/api/python/PythonMLLibAPI.scala
@@ -23,7 +23,7 @@ import org.apache.spark.annotation.DeveloperApi
 import org.apache.spark.api.java.JavaRDD
 import org.apache.spark.mllib.classification._
 import org.apache.spark.mllib.clustering._
-import org.apache.spark.mllib.linalg.Vectors
+import org.apache.spark.mllib.linalg.{SparseVector, Vector, Vectors}
 import org.apache.spark.mllib.recommendation._
 import org.apache.spark.mllib.regression._
 import org.apache.spark.rdd.RDD
@@ -31,56 +31,112 @@ import org.apache.spark.rdd.RDD
 /**
  * :: DeveloperApi ::
  * The Java stubs necessary for the Python mllib bindings.
+ *
+ * See python/pyspark/mllib/_common.py for the mutually agreed upon data format.
  */
 @DeveloperApi
 class PythonMLLibAPI extends Serializable {
-  private def deserializeDoubleVector(bytes: Array[Byte]): Array[Double] = {
-    val packetLength = bytes.length
-    if (packetLength < 16) {
-      throw new IllegalArgumentException("Byte array too short.")
-    }
-    val bb = ByteBuffer.wrap(bytes)
-    bb.order(ByteOrder.nativeOrder())
-    val magic = bb.getLong()
-    if (magic != 1) {
+  private val DENSE_VECTOR_MAGIC: Byte = 1
+  private val SPARSE_VECTOR_MAGIC: Byte = 2
+  private val DENSE_MATRIX_MAGIC: Byte = 3
+  private val LABELED_POINT_MAGIC: Byte = 4
+
+  private def deserializeDoubleVector(bytes: Array[Byte], offset: Int = 0): Vector = {
+    require(bytes.length - offset >= 5, "Byte array too short")
+    val magic = bytes(offset)
+    if (magic == DENSE_VECTOR_MAGIC) {
+      deserializeDenseVector(bytes, offset)
+    } else if (magic == SPARSE_VECTOR_MAGIC) {
+      deserializeSparseVector(bytes, offset)
+    } else {
       throw new IllegalArgumentException("Magic " + magic + " is wrong.")
     }
-    val length = bb.getLong()
-    if (packetLength != 16 + 8 * length) {
-      throw new IllegalArgumentException("Length " + length + " is wrong.")
-    }
+  }
+
+  private def deserializeDenseVector(bytes: Array[Byte], offset: Int = 0): Vector = {
+    val packetLength = bytes.length - offset
+    require(packetLength >= 5, "Byte array too short")
+    val bb = ByteBuffer.wrap(bytes, offset, bytes.length - offset)
+    bb.order(ByteOrder.nativeOrder())
+    val magic = bb.get()
+    require(magic == DENSE_VECTOR_MAGIC, "Invalid magic: " + magic)
+    val length = bb.getInt()
+    require (packetLength == 5 + 8 * length, "Invalid packet length: " + packetLength)
     val db = bb.asDoubleBuffer()
     val ans = new Array[Double](length.toInt)
     db.get(ans)
-    ans
+    Vectors.dense(ans)
   }
 
-  private def serializeDoubleVector(doubles: Array[Double]): Array[Byte] = {
+  private def deserializeSparseVector(bytes: Array[Byte], offset: Int = 0): Vector = {
+    val packetLength = bytes.length - offset
+    require(packetLength >= 9, "Byte array too short")
+    val bb = ByteBuffer.wrap(bytes, offset, bytes.length - offset)
+    bb.order(ByteOrder.nativeOrder())
+    val magic = bb.get()
+    require(magic == SPARSE_VECTOR_MAGIC, "Invalid magic: " + magic)
+    val size = bb.getInt()
+    val nonZeros = bb.getInt()
+    require (packetLength == 9 + 12 * nonZeros, "Invalid packet length: " + packetLength)
+    val ib = bb.asIntBuffer()
+    val indices = new Array[Int](nonZeros)
+    ib.get(indices)
+    bb.position(bb.position() + 4 * nonZeros)
+    val db = bb.asDoubleBuffer()
+    val values = new Array[Double](nonZeros)
+    db.get(values)
+    Vectors.sparse(size, indices, values)
+  }
+
+  private def serializeDenseVector(doubles: Array[Double]): Array[Byte] = {
     val len = doubles.length
-    val bytes = new Array[Byte](16 + 8 * len)
+    val bytes = new Array[Byte](5 + 8 * len)
     val bb = ByteBuffer.wrap(bytes)
     bb.order(ByteOrder.nativeOrder())
-    bb.putLong(1)
-    bb.putLong(len)
+    bb.put(DENSE_VECTOR_MAGIC)
+    bb.putInt(len)
     val db = bb.asDoubleBuffer()
     db.put(doubles)
     bytes
   }
 
+  private def serializeSparseVector(vector: SparseVector): Array[Byte] = {
+    val nonZeros = vector.indices.length
+    val bytes = new Array[Byte](9 + 12 * nonZeros)
+    val bb = ByteBuffer.wrap(bytes)
+    bb.order(ByteOrder.nativeOrder())
+    bb.put(SPARSE_VECTOR_MAGIC)
+    bb.putInt(vector.size)
+    bb.putInt(nonZeros)
+    val ib = bb.asIntBuffer()
+    ib.put(vector.indices)
+    bb.position(bb.position() + 4 * nonZeros)
+    val db = bb.asDoubleBuffer()
+    db.put(vector.values)
+    bytes
+  }
+
+  private def serializeDoubleVector(vector: Vector): Array[Byte] = vector match {
+    case s: SparseVector =>
+      serializeSparseVector(s)
+    case _ =>
+      serializeDenseVector(vector.toArray)
+  }
+
   private def deserializeDoubleMatrix(bytes: Array[Byte]): Array[Array[Double]] = {
     val packetLength = bytes.length
-    if (packetLength < 24) {
+    if (packetLength < 9) {
       throw new IllegalArgumentException("Byte array too short.")
     }
     val bb = ByteBuffer.wrap(bytes)
     bb.order(ByteOrder.nativeOrder())
-    val magic = bb.getLong()
-    if (magic != 2) {
+    val magic = bb.get()
+    if (magic != DENSE_MATRIX_MAGIC) {
       throw new IllegalArgumentException("Magic " + magic + " is wrong.")
     }
-    val rows = bb.getLong()
-    val cols = bb.getLong()
-    if (packetLength != 24 + 8 * rows * cols) {
+    val rows = bb.getInt()
+    val cols = bb.getInt()
+    if (packetLength != 9 + 8 * rows * cols) {
       throw new IllegalArgumentException("Size " + rows + "x" + cols + " is wrong.")
     }
     val db = bb.asDoubleBuffer()
@@ -98,12 +154,12 @@ class PythonMLLibAPI extends Serializable {
     if (rows > 0) {
       cols = doubles(0).length
     }
-    val bytes = new Array[Byte](24 + 8 * rows * cols)
+    val bytes = new Array[Byte](9 + 8 * rows * cols)
     val bb = ByteBuffer.wrap(bytes)
     bb.order(ByteOrder.nativeOrder())
-    bb.putLong(2)
-    bb.putLong(rows)
-    bb.putLong(cols)
+    bb.put(DENSE_MATRIX_MAGIC)
+    bb.putInt(rows)
+    bb.putInt(cols)
     val db = bb.asDoubleBuffer()
     for (i <- 0 until rows) {
       db.put(doubles(i))
@@ -111,18 +167,27 @@ class PythonMLLibAPI extends Serializable {
     bytes
   }
 
+  private def deserializeLabeledPoint(bytes: Array[Byte]): LabeledPoint = {
+    require(bytes.length >= 9, "Byte array too short")
+    val magic = bytes(0)
+    if (magic != LABELED_POINT_MAGIC) {
+      throw new IllegalArgumentException("Magic " + magic + " is wrong.")
+    }
+    val labelBytes = ByteBuffer.wrap(bytes, 1, 8)
+    labelBytes.order(ByteOrder.nativeOrder())
+    val label = labelBytes.asDoubleBuffer().get(0)
+    LabeledPoint(label, deserializeDoubleVector(bytes, 9))
+  }
+
   private def trainRegressionModel(
-      trainFunc: (RDD[LabeledPoint], Array[Double]) => GeneralizedLinearModel,
+      trainFunc: (RDD[LabeledPoint], Vector) => GeneralizedLinearModel,
       dataBytesJRDD: JavaRDD[Array[Byte]],
       initialWeightsBA: Array[Byte]): java.util.LinkedList[java.lang.Object] = {
-    val data = dataBytesJRDD.rdd.map(xBytes => {
-        val x = deserializeDoubleVector(xBytes)
-        LabeledPoint(x(0), Vectors.dense(x.slice(1, x.length)))
-    })
+    val data = dataBytesJRDD.rdd.map(deserializeLabeledPoint)
     val initialWeights = deserializeDoubleVector(initialWeightsBA)
     val model = trainFunc(data, initialWeights)
     val ret = new java.util.LinkedList[java.lang.Object]()
-    ret.add(serializeDoubleVector(model.weights.toArray))
+    ret.add(serializeDoubleVector(model.weights))
     ret.add(model.intercept: java.lang.Double)
     ret
   }
@@ -143,7 +208,7 @@ class PythonMLLibAPI extends Serializable {
           numIterations,
           stepSize,
           miniBatchFraction,
-          Vectors.dense(initialWeights)),
+          initialWeights),
       dataBytesJRDD,
       initialWeightsBA)
   }
@@ -166,7 +231,7 @@ class PythonMLLibAPI extends Serializable {
           stepSize,
           regParam,
           miniBatchFraction,
-          Vectors.dense(initialWeights)),
+          initialWeights),
       dataBytesJRDD,
       initialWeightsBA)
   }
@@ -189,7 +254,7 @@ class PythonMLLibAPI extends Serializable {
           stepSize,
           regParam,
           miniBatchFraction,
-          Vectors.dense(initialWeights)),
+          initialWeights),
       dataBytesJRDD,
       initialWeightsBA)
   }
@@ -212,7 +277,7 @@ class PythonMLLibAPI extends Serializable {
           stepSize,
           regParam,
           miniBatchFraction,
-          Vectors.dense(initialWeights)),
+          initialWeights),
       dataBytesJRDD,
       initialWeightsBA)
   }
@@ -233,7 +298,7 @@ class PythonMLLibAPI extends Serializable {
           numIterations,
           stepSize,
           miniBatchFraction,
-          Vectors.dense(initialWeights)),
+          initialWeights),
       dataBytesJRDD,
       initialWeightsBA)
   }
@@ -244,14 +309,11 @@ class PythonMLLibAPI extends Serializable {
   def trainNaiveBayes(
       dataBytesJRDD: JavaRDD[Array[Byte]],
       lambda: Double): java.util.List[java.lang.Object] = {
-    val data = dataBytesJRDD.rdd.map(xBytes => {
-      val x = deserializeDoubleVector(xBytes)
-      LabeledPoint(x(0), Vectors.dense(x.slice(1, x.length)))
-    })
+    val data = dataBytesJRDD.rdd.map(deserializeLabeledPoint)
     val model = NaiveBayes.train(data, lambda)
     val ret = new java.util.LinkedList[java.lang.Object]()
-    ret.add(serializeDoubleVector(model.labels))
-    ret.add(serializeDoubleVector(model.pi))
+    ret.add(serializeDoubleVector(Vectors.dense(model.labels)))
+    ret.add(serializeDoubleVector(Vectors.dense(model.pi)))
     ret.add(serializeDoubleMatrix(model.theta))
     ret
   }
@@ -265,7 +327,7 @@ class PythonMLLibAPI extends Serializable {
       maxIterations: Int,
       runs: Int,
       initializationMode: String): java.util.List[java.lang.Object] = {
-    val data = dataBytesJRDD.rdd.map(xBytes => Vectors.dense(deserializeDoubleVector(xBytes)))
+    val data = dataBytesJRDD.rdd.map(bytes => deserializeDoubleVector(bytes))
     val model = KMeans.train(data, k, maxIterations, runs, initializationMode)
     val ret = new java.util.LinkedList[java.lang.Object]()
     ret.add(serializeDoubleMatrix(model.clusterCenters.map(_.toArray)))
diff --git a/mllib/src/main/scala/org/apache/spark/mllib/linalg/Vectors.scala b/mllib/src/main/scala/org/apache/spark/mllib/linalg/Vectors.scala
index 99a849f1c6..7cdf6bd56a 100644
--- a/mllib/src/main/scala/org/apache/spark/mllib/linalg/Vectors.scala
+++ b/mllib/src/main/scala/org/apache/spark/mllib/linalg/Vectors.scala
@@ -130,9 +130,11 @@ object Vectors {
   private[mllib] def fromBreeze(breezeVector: BV[Double]): Vector = {
     breezeVector match {
       case v: BDV[Double] =>
-        require(v.offset == 0, s"Do not support non-zero offset ${v.offset}.")
-        require(v.stride == 1, s"Do not support stride other than 1, but got ${v.stride}.")
-        new DenseVector(v.data)
+        if (v.offset == 0 && v.stride == 1) {
+          new DenseVector(v.data)
+        } else {
+          new DenseVector(v.toArray)  // Can't use underlying array directly, so make a new one
+        }
       case v: BSV[Double] =>
         new SparseVector(v.length, v.index, v.data)
       case v: BV[_] =>
diff --git a/mllib/src/test/scala/org/apache/spark/mllib/linalg/VectorsSuite.scala b/mllib/src/test/scala/org/apache/spark/mllib/linalg/VectorsSuite.scala
index 8a200310e0..cfe8a27fcb 100644
--- a/mllib/src/test/scala/org/apache/spark/mllib/linalg/VectorsSuite.scala
+++ b/mllib/src/test/scala/org/apache/spark/mllib/linalg/VectorsSuite.scala
@@ -82,4 +82,22 @@ class VectorsSuite extends FunSuite {
       assert(v.## != another.##)
     }
   }
+
+  test("indexing dense vectors") {
+    val vec = Vectors.dense(1.0, 2.0, 3.0, 4.0)
+    assert(vec(0) === 1.0)
+    assert(vec(3) === 4.0)
+  }
+
+  test("indexing sparse vectors") {
+    val vec = Vectors.sparse(7, Array(0, 2, 4, 6), Array(1.0, 2.0, 3.0, 4.0))
+    assert(vec(0) === 1.0)
+    assert(vec(1) === 0.0)
+    assert(vec(2) === 2.0)
+    assert(vec(3) === 0.0)
+    assert(vec(6) === 4.0)
+    val vec2 = Vectors.sparse(8, Array(0, 2, 4, 6), Array(1.0, 2.0, 3.0, 4.0))
+    assert(vec2(6) === 4.0)
+    assert(vec2(7) === 0.0)
+  }
 }