Changes to Prob GP committee

- Improvements to GPC logging
- Changes to GP pipe API
- Bug fix to GP STP setState for Decomposable Kernels

Author: mandar2812

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/DynaMLPipe.scala

@@ -25,7 +25,7 @@ import breeze.numerics.sqrt
 import io.github.mandar2812.dynaml.evaluation.RegressionMetrics
 import io.github.mandar2812.dynaml.models.ParameterizedLearner
 import io.github.mandar2812.dynaml.models.gp.AbstractGPRegressionModel
-import io.github.mandar2812.dynaml.optimization.{CoupledSimulatedAnnealing, GloballyOptWithGrad, GradBasedGlobalOptimizer, GridSearch}
+import io.github.mandar2812.dynaml.optimization._
 import io.github.mandar2812.dynaml.pipes._
 import io.github.mandar2812.dynaml.utils.{GaussianScaler, MVGaussianScaler, MinMaxScaler}
 import io.github.mandar2812.dynaml.wavelets.{GroupedHaarWaveletFilter, HaarWaveletFilter, InvGroupedHaarWaveletFilter, InverseHaarWaveletFilter}
@@ -667,6 +667,43 @@ object DynaMLPipe {
         "maxIterations" -> grid.toString))
     })
 
+  def gpTuning[T, I:ClassTag](
+    startingState: Map[String, Double],
+    globalOpt: String = "GS",
+    grid: Int = 3, step: Double = 0.02,
+    maxIt: Int = 20, policy: String = "GS") =
+    DataPipe((model: AbstractGPRegressionModel[T, I]) => {
+      val gs = globalOpt match {
+        case "GS" => new GridSearch(model)
+          .setGridSize(grid)
+          .setStepSize(step)
+          .setLogScale(false)
+
+        case "ML" => new GradBasedGlobalOptimizer(model)
+
+        case "CSA" => new CoupledSimulatedAnnealing(model)
+          .setGridSize(grid)
+          .setStepSize(step)
+          .setLogScale(false)
+          .setMaxIterations(maxIt)
+          .setVariant(CoupledSimulatedAnnealing.MwVC)
+
+        case "GPC" => new ProbGPCommMachine(model)
+          .setPolicy(policy)
+          .setGridSize(grid)
+          .setStepSize(step)
+          .setMaxIterations(maxIt)
+      }
+
+      gs.optimize(
+        startingState,
+        Map(
+          "tolerance" -> "0.0001",
+          "step" -> step.toString,
+          "maxIterations" -> grid.toString,
+          "persist" -> "true"))
+    })
+
 
   def GPRegressionTest[T <: AbstractGPRegressionModel[
     Seq[(DenseVector[Double], Double)],

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

@@ -18,6 +18,7 @@ under the License.
 * */
 package io.github.mandar2812.dynaml.models
 
+import io.github.mandar2812.dynaml.DynaMLPipe
 import io.github.mandar2812.dynaml.kernels.LocalScalarKernel
 import io.github.mandar2812.dynaml.models.gp.AbstractGPRegressionModel
 import io.github.mandar2812.dynaml.pipes.DataPipe
@@ -27,20 +28,31 @@ import scala.reflect.ClassTag
 /**
   * Created by mandar on 15/6/16.
   */
-class GPRegressionPipe[
-M <: AbstractGPRegressionModel[Seq[(IndexSet, Double)], IndexSet],
-Source, IndexSet: ClassTag](
+class GPRegressionPipe[Source, IndexSet: ClassTag](
   pre: (Source) => Seq[(IndexSet, Double)],
   cov: LocalScalarKernel[IndexSet],
   n: LocalScalarKernel[IndexSet],
   order: Int = 0, ex: Int = 0,
   meanFunc: DataPipe[IndexSet, Double] = DataPipe((_: IndexSet) => 0.0))
-  extends ModelPipe[Source, Seq[(IndexSet, Double)], IndexSet, Double, M] {
+  extends ModelPipe[
+    Source, Seq[(IndexSet, Double)], IndexSet, Double,
+    AbstractGPRegressionModel[Seq[(IndexSet, Double)], IndexSet]] {
 
   override val preProcess: (Source) => Seq[(IndexSet, Double)] = pre
 
-  implicit val transform = DataPipe(preProcess)
+  implicit val transform = DynaMLPipe.identityPipe[Seq[(IndexSet, Double)]]
 
-  override def run(data: Source): M = AbstractGPRegressionModel(cov, n, meanFunc)(data, 0).asInstanceOf[M]
+  override def run(data: Source): AbstractGPRegressionModel[Seq[(IndexSet, Double)], IndexSet] =
+    AbstractGPRegressionModel(cov, n, meanFunc)(preProcess(data), 0)
 
 }
+
+
+object GPRegressionPipe {
+  def apply[Source, IndexSet: ClassTag](
+    pre: (Source) => Seq[(IndexSet, Double)],
+    cov: LocalScalarKernel[IndexSet], n: LocalScalarKernel[IndexSet],
+    order: Int = 0, ex: Int = 0,
+    meanFunc: DataPipe[IndexSet, Double] = DataPipe((_: IndexSet) => 0.0)) =
+    new GPRegressionPipe[Source, IndexSet](pre, cov, n, order, ex, meanFunc)
+}

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/gp/AbstractGPRegressionModel.scala

@@ -92,8 +92,14 @@ abstract class AbstractGPRegressionModel[T, I: ClassTag](
     * kernels.
     * */
   def setState(s: Map[String, Double]): this.type = {
-    covariance.setHyperParameters(s)
-    noiseModel.setHyperParameters(s)
+
+    val (covHyp, noiseHyp) = (
+      s.filterKeys(covariance.hyper_parameters.contains),
+      s.filterKeys(noiseModel.hyper_parameters.contains)
+    )
+
+    covariance.setHyperParameters(covHyp)
+    noiseModel.setHyperParameters(noiseHyp)
     current_state = covariance.state ++ noiseModel.state
     this
   }

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/gp/GPCommitteeRegression.scala

@@ -28,7 +28,7 @@ import scala.reflect.ClassTag
   */
 abstract class GPCommitteeRegression[D, I: ClassTag](
   num: Int, data: D,
-  networks: GPRegressionPipe[AbstractGPRegressionModel[Seq[(I, Double)], I], D, I]*) extends
+  networks: GPRegressionPipe[D, I]*) extends
   LinearModel[D, DenseVector[Double], I, Double, D] {
 
   override protected val g: D = data

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/models/stp/AbstractSTPRegressionModel.scala

@@ -89,8 +89,15 @@ abstract class AbstractSTPRegressionModel[T, I](
     * kernels.
     * */
   def setState(s: Map[String, Double]): this.type = {
-    covariance.setHyperParameters(s)
-    noiseModel.setHyperParameters(s)
+
+    val (covHyp, noiseHyp) = (
+      s.filterKeys(covariance.hyper_parameters.contains),
+      s.filterKeys(noiseModel.hyper_parameters.contains)
+    )
+
+    covariance.setHyperParameters(covHyp)
+    noiseModel.setHyperParameters(noiseHyp)
+
     current_state = covariance.state ++ noiseModel.state
     current_state += ("degrees_of_freedom" -> (s("degrees_of_freedom")+2.0))
     this

dynaml-core/src/main/scala-2.11/io/github/mandar2812/dynaml/optimization/CoupledSimulatedAnnealing.scala

@@ -344,8 +344,11 @@ class ProbGPCommMachine[T, I: ClassTag](
 
     logger.info("===============================================")
 
-    if(options.contains("persist") && (options("persist") == "true" || options("persist") == "1"))
+    if(options.contains("persist") && (options("persist") == "true" || options("persist") == "1")) {
+      logger.info("Persisting model state")
       committeeGP.persist(committeeGP._current_state)
+    }
+
 
     //Return the resultant model
     (committeeGP, committeeGP._current_state)