Merge branch 'master' of github.com:h2oai/h2o

R/h2o-package/R/ParseImport.R

@@ -494,7 +494,7 @@ h2o.loadModel <- function(object, path="") {
     # Read models from model_names meta file
     if(length(grep("hdfs://", path)) == 1) {
       fileConn = paste(path, "model_names", sep = "/")
-      model_names.hex = h2o.importFile(object = conn, path = fileConn)
+      model_names.hex = h2o.importFile(object = object, path = fileConn)
       model_names = as.matrix(model_names.hex)
       h2o.rm(object, model_names.hex@key)
     } else {

src/main/java/hex/deeplearning/Neurons.java

@@ -7,6 +7,7 @@
 import water.Iced;
 import water.MemoryManager;
 import water.api.Request.API;
+import water.util.Log;
 import water.util.Utils;
 
 import java.nio.ByteBuffer;
@@ -985,11 +986,11 @@ protected void fprop() {
         scale += _a.get(row);
       }
       for( int row = 0; row < rows; row++ ) {
+        _a.raw()[row] /= scale;
         if (Float.isNaN(_a.get(row))) {
           _minfo.set_unstable();
           throw new RuntimeException("Numerical instability, predicted NaN.");
         }
-        _a.raw()[row] /= scale;
       }
     }
 

src/test/java/hex/DeepLearningSpiralsTest.java

@@ -60,33 +60,23 @@ public class DeepLearningSpiralsTest extends TestUtil {
           p.score_training_samples = 1000;
           p.score_validation_samples = 10000;
           p.shuffle_training_data = false;
-          p.force_load_balance = true; //multi-threaded
+          p.force_load_balance = false;
           p.replicate_training_data = false;
           p.destination_key = dest;
           p.adaptive_rate = true;
+          p.reproducible = true;
           p.rho = 0.99;
           p.epsilon = 5e-3;
           p.invoke();
         }
 
         // score and check result
         {
-          DeepLearningModel mymodel = UKV.get(dest); //this actually *requires* frame to also still be in UKV (because of DataInfo...)
-          Frame pred = mymodel.score(frame);
-          water.api.ConfusionMatrix CM = new water.api.ConfusionMatrix();
-          CM.actual = frame;
-          CM.vactual = frame.lastVec();
-          CM.predict = pred;
-          CM.vpredict = pred.vecs()[0];
-          CM.invoke();
-          StringBuilder sb = new StringBuilder();
-          CM.toASCII(sb);
-          double error = new ConfusionMatrix(CM.cm).err();
-          Log.info(sb);
+          DeepLearningModel mymodel = UKV.get(dest);
+          double error = mymodel.error();
           if (error >= 0.025) {
             Assert.fail("Classification error is not less than 0.025, but " + error + ".");
           }
-          pred.delete();
           mymodel.delete();
           mymodel.delete_best_model();
         }