[jvm-packages] use ML's para system to build the passed-in params to XGBoost (#2043)

* add back train method but mark as deprecated

* fix scalastyle error

* use ML's para system to build the passed-in params to XGBoost

* clean

jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/XGBoostEstimator.scala

@@ -32,7 +32,7 @@ import org.apache.spark.sql.{Dataset, Row}
  * XGBoost Estimator to produce a XGBoost model
  */
 class XGBoostEstimator private[spark](
-  override val uid: String, private[spark] var xgboostParams: Map[String, Any])
+  override val uid: String, xgboostParams: Map[String, Any])
   extends Predictor[MLVector, XGBoostEstimator, XGBoostModel]
   with LearningTaskParams with GeneralParams with BoosterParams {
 
@@ -41,7 +41,6 @@ class XGBoostEstimator private[spark](
 
   def this(uid: String) = this(uid, Map[String, Any]())
 
-
   // called in fromXGBParamMapToParams only when eval_metric is not defined
   private def setupDefaultEvalMetric(): String = {
     val objFunc = xgboostParams.getOrElse("objective", xgboostParams.getOrElse("obj_type", null))
@@ -93,16 +92,11 @@ class XGBoostEstimator private[spark](
 
   fromXGBParamMapToParams()
 
-  // only called when XGBParamMap is empty, i.e. in the constructor this(String)
-  // TODO: refactor to be functional
-  private def fromParamsToXGBParamMap(): Map[String, Any] = {
-    require(xgboostParams.isEmpty, "fromParamsToXGBParamMap can only be called when" +
-      " XGBParamMap is empty, i.e. in the constructor this(String)")
+  private[spark] def fromParamsToXGBParamMap: Map[String, Any] = {
     val xgbParamMap = new mutable.HashMap[String, Any]()
     for (param <- params) {
       xgbParamMap += param.name -> $(param)
     }
-    xgboostParams = xgbParamMap.toMap
     xgbParamMap.toMap
   }
 
@@ -116,8 +110,9 @@ class XGBoostEstimator private[spark](
         LabeledPoint(label, feature)
     }
     transformSchema(trainingSet.schema, logging = true)
-    val trainedModel = XGBoost.trainWithRDD(instances, xgboostParams, $(round), $(nWorkers),
-      $(customObj), $(customEval), $(useExternalMemory), $(missing)).setParent(this)
+    val trainedModel = XGBoost.trainWithRDD(instances, fromParamsToXGBParamMap,
+      $(round), $(nWorkers), $(customObj), $(customEval), $(useExternalMemory),
+      $(missing)).setParent(this)
     val returnedModel = copyValues(trainedModel)
     if (XGBoost.isClassificationTask(xgboostParams)) {
       val numClass = {
@@ -133,11 +128,6 @@ class XGBoostEstimator private[spark](
   }
 
   override def copy(extra: ParamMap): XGBoostEstimator = {
-    val est = defaultCopy(extra).asInstanceOf[XGBoostEstimator]
-    // we need to synchronize the params here instead of in the constructor
-    // because we cannot guarantee that params (default implementation) is initialized fully
-    // before the other params
-    est.fromParamsToXGBParamMap()
-    est
+    defaultCopy(extra).asInstanceOf[XGBoostEstimator]
   }
 }

jvm-packages/xgboost4j-spark/src/main/scala/ml/dmlc/xgboost4j/scala/spark/params/BoosterParams.scala

@@ -196,7 +196,7 @@ trait BoosterParams extends Params {
     minChildWeight -> 1, maxDeltaStep -> 0,
     subSample -> 1, colSampleByTree -> 1, colSampleByLevel -> 1,
     lambda -> 1, alpha -> 0, treeMethod -> "auto", sketchEps -> 0.03,
-    scalePosWeight -> 1, sampleType -> "uniform", normalizeType -> "tree",
+    scalePosWeight -> 1.0, sampleType -> "uniform", normalizeType -> "tree",
     rateDrop -> 0.0, skipDrop -> 0.0, lambdaBias -> 0)
 
   /**