PUBDEV-5294 - XGBoost model in Flow doesn't seem to converge

h2oai · Feb 15, 2018 · 9661331 · 9661331
1 parent 97f351b
commit 9661331
Showing 1 changed file with 40 additions and 16 deletions.
diff --git a/h2o-extensions/xgboost/src/main/java/hex/tree/xgboost/XGBoostModel.java b/h2o-extensions/xgboost/src/main/java/hex/tree/xgboost/XGBoostModel.java
@@ -13,6 +13,9 @@
 import water.fvec.Frame;
 import water.util.Log;
 import hex.ModelMetrics;
+
+import java.text.DecimalFormat;
+import java.text.NumberFormat;
 import java.util.HashMap;
 import java.util.Map;
 
@@ -21,6 +24,7 @@
 public class XGBoostModel extends Model<XGBoostModel, XGBoostModel.XGBoostParameters, XGBoostOutput> {
 
   private XGBoostModelInfo model_info;
+  private static final NumberFormat localizedNumberFormatter = DecimalFormat.getNumberInstance();
 
   XGBoostModelInfo model_info() { return model_info; }
 
@@ -290,9 +294,29 @@ public static HashMap<String, Object> createParams(XGBoostParameters p, XGBoostO
       Log.info(" " + s.getKey() + " = " + s.getValue());
     }
     Log.info("");
+
+    localizeDecimalParams(params);
     return params;
   }
 
+  /**
+   * Iterates over a set of parameters and applies locale-specific formatting
+   * to decimal ones (Floats and Doubles).
+   *
+   * @param params Parameters to localize
+   */
+  private static void localizeDecimalParams(final HashMap<String, Object> params) {
+
+    for (String key : params.keySet()) {
+      final Object value = params.get(key);
+      if (value instanceof Float || value instanceof Double) {
+        final String localizedValue = localizedNumberFormatter.format(value);
+        params.put(key, localizedValue);
+      }
+    }
+
+  }
+
   @Override
   protected double[] score0(double[] data, double[] preds) {
     return score0(data, preds, 0.0);
@@ -325,7 +349,7 @@ private ModelMetrics makeMetrics(Booster booster, Frame data, Frame originalData
     ModelMetrics[] mms = new ModelMetrics[1];
     Frame predictions = makePreds(booster, data, mms, true, predFrameKey, fs);
     if (predFrameKey == null) {
-        predictions.remove(fs);
+      predictions.remove(fs);
     } else {
       DKV.put(predictions, fs);
     }
@@ -345,17 +369,17 @@ private Frame makePredsOnly(Booster booster, Frame data, Key<Frame> destinationK
   }
 
   private Frame makePreds(Booster booster, Frame data, ModelMetrics[] mms, boolean computeMetrics, Key<Frame> destinationKey, Futures fs) throws XGBoostError {
-      assert (! computeMetrics) || (mms != null && mms.length == 1);
-
-      XGBoostScoreTask.XGBoostScoreTaskResult score = XGBoostScoreTask.runScoreTask(
-              model_info(), _output, _parms,
-              booster, destinationKey, data,
-              computeMetrics
-      );
-      if(computeMetrics) {
-        mms[0] = score.mm;
-      }
-      return score.preds;
+    assert (! computeMetrics) || (mms != null && mms.length == 1);
+
+    XGBoostScoreTask.XGBoostScoreTaskResult score = XGBoostScoreTask.runScoreTask(
+        model_info(), _output, _parms,
+        booster, destinationKey, data,
+        computeMetrics
+    );
+    if(computeMetrics) {
+      mms[0] = score.mm;
+    }
+    return score.preds;
   }
 
   /**
@@ -399,8 +423,8 @@ void computeVarImp(Map<String, Integer> varimp) {
   public double[] score0(double[] data, double[] preds, double offset) {
     DataInfo di = model_info._dataInfoKey.get();
     return XGBoostMojoModel.score0(data, offset, preds,
-            model_info.getBooster(), di._nums, di._cats, di._catOffsets, di._useAllFactorLevels,
-            _output.nclasses(), _output._priorClassDist, defaultThreshold(), _output._sparse);
+        model_info.getBooster(), di._nums, di._cats, di._catOffsets, di._useAllFactorLevels,
+        _output.nclasses(), _output._priorClassDist, defaultThreshold(), _output._sparse);
   }
 
   @Override
@@ -412,8 +436,8 @@ public double[][] score0( Chunk chks[], double[] offset, int[] rowsInChunk, doub
     }
     DataInfo di = model_info._dataInfoKey.get();
     double[][] scored = XGBoostMojoModel.bulkScore0(tmp, offset, preds,
-            model_info.getBooster(), di._nums, di._cats, di._catOffsets, di._useAllFactorLevels,
-            _output.nclasses(), _output._priorClassDist, defaultThreshold(), _output._sparse);
+        model_info.getBooster(), di._nums, di._cats, di._catOffsets, di._useAllFactorLevels,
+        _output.nclasses(), _output._priorClassDist, defaultThreshold(), _output._sparse);
 
     if(isSupervised()) {
       // Correct probabilities obtained from training on oversampled data back to original distribution