1. Corrected typo in Backpropagation class name 2. Added composition operations for Scaler objects 3. Added Autoencoder class

Author: mandar2812

README.md

@@ -88,10 +88,25 @@ Steps
 * Clone this repository
 * Run the following.
 ```shell
-  sbt console
+  sbt
 ```
 
-  You should get the following prompt.
+The sbt shell will open
+
+```shell
+ [info] Loading project definition from ~/DynaML/project
+ [info] Set current project to DynaML (in build file:~/Development/DynaML/)
+ >
+```
+
+Now enter the following commands
+
+```shell
+>stage
+>console
+```
+
+After the project builds, you should get the following prompt.
 
 ```
       ___           ___           ___           ___           ___           ___

build.sbt

@@ -12,7 +12,7 @@ packageDescription := "DynaML is a scala library/repl for implementing and worki
   "which can be extended easily to implement advanced models for small and large scale applications.\n\n"+
   "But the library can also be used as an educational/research tool for data analysis."
 
-val mainVersion = "v1.4-beta.14"
+val mainVersion = "v1.4-beta.15"
 
 val dataDirectory = settingKey[File]("The directory holding the data files for running example scripts")
 

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/neuralnets/AutoEncoder.scala

@@ -19,13 +19,45 @@ under the License.
 package io.github.mandar2812.dynaml.models.neuralnets
 
 import breeze.linalg.DenseVector
-import io.github.mandar2812.dynaml.models.ModelPipe
+import io.github.mandar2812.dynaml.graph.FFNeuralGraph
+import io.github.mandar2812.dynaml.graph.utils.Neuron
+import io.github.mandar2812.dynaml.optimization.BackPropagation
+import io.github.mandar2812.dynaml.pipes.{ReversibleScaler, Scaler}
 
 /**
-  * Created by mandar on 23/2/16.
+  * @author mandar2812 22/6/16.
+  *
+  * Base implementation of a Sparse Autoencoder
+  *
+  * It is represented as a [[ReversibleScaler]] transforming
+  * a breeze [[DenseVector]] into another breeze Dense Vector.
   */
-trait AutoEncoder[Source, Data] extends
-ModelPipe[Source, Data, DenseVector[Double],
-  DenseVector[Double], FeedForwardNetwork[Data]]{
+class AutoEncoder(inDim: Int, outDim: Int) extends ReversibleScaler[DenseVector[Double]]{
+
+  def initialize() =
+    FFNeuralGraph(
+      inDim, inDim, 1,
+      List("logsig", "logsig"),
+      List(outDim))
+
+  var graph = initialize()
+
+  val optimizer = new BackPropagation
+
+  val i = Scaler((xhat: DenseVector[Double]) => {
+    graph.getLayer(1)
+      .filter(_.getNeuronType() == "perceptron")
+      .foreach(n => n.setValue(xhat(n.getNID())))
+    DenseVector(graph.getLayer(2).map(n => Neuron.getLocalField(n)._1).toArray)
+  })
+
+  def learn(data: Stream[(DenseVector[Double], DenseVector[Double])]) = {
+    graph = optimizer.optimize(data.length.toLong, data, initialize())
+  }
+
+  override def run(x: DenseVector[Double]) = {
+    graph.forwardPass(x)
+    DenseVector(graph.getLayer(1).filter(_.getNeuronType() == "perceptron").map(_.getValue()).toArray)
+  }
 
 }

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/neuralnets/CommitteeNetwork.scala

@@ -21,7 +21,7 @@ package io.github.mandar2812.dynaml.models.neuralnets
 import breeze.linalg.DenseVector
 import io.github.mandar2812.dynaml.graph.FFNeuralGraph
 import io.github.mandar2812.dynaml.models.LinearModel
-import io.github.mandar2812.dynaml.optimization.{BackPropogation, CommitteeModelSolver, RegularizedOptimizer}
+import io.github.mandar2812.dynaml.optimization.{BackPropagation, CommitteeModelSolver, RegularizedOptimizer}
 import io.github.mandar2812.dynaml.pipes.DataPipe
 
 /**
@@ -40,7 +40,7 @@ LinearModel[D, DenseVector[Double], DenseVector[Double],
 
   val baseNetworks: List[FFNeuralGraph] = networks.toList
 
-  val baseOptimizer = new BackPropogation()
+  val baseOptimizer = new BackPropagation()
     .setMomentum(0.01).setRegParam(0.001)
     .setMiniBatchFraction(1.0/baseNetworks.length)
     .setNumIterations(20)

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/neuralnets/FeedForwardNetwork.scala

@@ -22,7 +22,7 @@ import breeze.linalg.DenseVector
 import com.tinkerpop.blueprints.Graph
 import com.tinkerpop.frames.FramedGraph
 import io.github.mandar2812.dynaml.graph.FFNeuralGraph
-import io.github.mandar2812.dynaml.optimization.BackPropogation
+import io.github.mandar2812.dynaml.optimization.BackPropagation
 import io.github.mandar2812.dynaml.pipes.DataPipe
 
 
@@ -32,12 +32,9 @@ import io.github.mandar2812.dynaml.pipes.DataPipe
   * backed by an underlying graph.
   *
   * @tparam D The type of the underlying training data structure.
-  *
   * @param data The training data
-  *
   * @param netgraph The [[FFNeuralGraph]] object which represents the
   *                 network.
-  *
   * @param transform A [[DataPipe]] which takes input of type [[D]] and
   *                  returns a [[Stream]] of input, output tuples.
   *
@@ -74,14 +71,14 @@ class FeedForwardNetwork[D](
 
   /**
     * Model optimizer set to
-    * [[BackPropogation]] which
+    * [[BackPropagation]] which
     * is an implementation of
     * gradient based Back-propogation
     * with a momentum term.
     *
     * */
   override protected val optimizer =
-    new BackPropogation()
+    new BackPropagation()
       .setNumIterations(100)
       .setStepSize(0.01)
 
@@ -90,7 +87,7 @@ class FeedForwardNetwork[D](
     this
   }
 
-  override def dataAsStream(d: D) = transform.run(d)
+  override def dataAsStream(d: D) = transform(d)
 
   /**
    * Learn the parameters

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/optimization/BackPropogation.scala renamed to dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/optimization/BackPropagation.scala

@@ -29,7 +29,7 @@ import scala.util.Random
   * Implementation of the standard back pro-pogation with momentum
   * using the "generalized delta rule".
   */
-class BackPropogation extends RegularizedOptimizer[FFNeuralGraph,
+class BackPropagation extends RegularizedOptimizer[FFNeuralGraph,
   DenseVector[Double], DenseVector[Double],
   Stream[(DenseVector[Double], DenseVector[Double])]] {
 

dynaml-pipes/src/main/scala-2.11/io/github/mandar2812/dynaml/pipes/Scaler.scala

@@ -17,6 +17,15 @@ trait Scaler[S] extends DataPipe[S, S]{
     }
   }
 
+  def >(otherScaler: Scaler[S]) = {
+
+    val firstRun = this.run _
+
+    new Scaler[S] {
+      def run(data: S) = otherScaler.run(firstRun(data))
+    }
+  }
+
 }
 
 object Scaler {
@@ -52,4 +61,16 @@ trait ReversibleScaler[S] extends Scaler[S] {
       override def run(data: (S, T)): (S, T) = (firstRun(data._1), that(data._2))
     }
   }
+
+  def >(otherRevScaler: ReversibleScaler[S]): ReversibleScaler[S] = {
+
+    val firstInv = this.i
+
+    val firstRun = this.run _
+
+    new ReversibleScaler[S] {
+      val i: Scaler[S] = otherRevScaler.i > firstInv
+      def run(data: S) = otherRevScaler.run(firstRun(data))
+    }
+  }
 }