Refactor SVD method name and add svd_method option to R package

h2o-algos/src/main/java/hex/schemas/SVDV99.java

@@ -29,7 +29,7 @@ public static final class SVDParametersV99 extends ModelParametersSchema<SVDMode
     @API(help = "Transformation of training data", values = { "NONE", "STANDARDIZE", "NORMALIZE", "DEMEAN", "DESCALE" })  // TODO: pull out of enum class
     public DataInfo.TransformType transform;
 
-    @API(help = "Method for computing SVD (Caution: Power and Probablistic are currently experimental and unstable)", values = { "GramSVD", "Power", "Probabilistic" })   // TODO: pull out of enum class
+    @API(help = "Method for computing SVD (Caution: Power and Probablistic are currently experimental and unstable)", values = { "GramSVD", "Power", "Randomized" })   // TODO: pull out of enum class
     public SVDParameters.Method svd_method;
 
     @API(help = "Number of right singular vectors")

h2o-algos/src/main/java/hex/svd/SVD.java

@@ -1,6 +1,5 @@
 package hex.svd;
 
-import Jama.CholeskyDecomposition;
 import Jama.Matrix;
 import Jama.QRDecomposition;
 import Jama.SingularValueDecomposition;
@@ -27,7 +26,7 @@
  * Singular Value Decomposition
  * <a href = "http://www.cs.yale.edu/homes/el327/datamining2013aFiles/07_singular_value_decomposition.pdf">SVD via Power Method Algorithm</a>
  * <a href = "https://www.cs.cmu.edu/~venkatg/teaching/CStheory-infoage/book-chapter-4.pdf">Proof of Convergence for Power Method</a>
- * <a href = "http://arxiv.org/pdf/0909.4061.pdf">Probabilistic Algorithms for Matrix Approximation</a>
+ * <a href = "http://arxiv.org/pdf/0909.4061.pdf">Randomized Algorithms for Matrix Approximation</a>
  * @author anqi_fu
  */
 public class SVD extends ModelBuilder<SVDModel,SVDModel.SVDParameters,SVDModel.SVDOutput> {
@@ -55,7 +54,7 @@ public long progressUnits() {
         return 2;
       case Power:
         return 1 + _parms._nv;
-      case Probabilistic:
+      case Randomized:
         return 5 + _parms._max_iterations;
       default: return _parms._nv;
     }
@@ -486,7 +485,7 @@ private Frame directSVD(DataInfo dinfo, Frame qfrm, SVDModel model) {
             DivideU utsk = new DivideU(model._output._d);
             utsk.doAll(u);
           }
-        } else if(_parms._svd_method == SVDParameters.Method.Probabilistic) {
+        } else if(_parms._svd_method == SVDParameters.Method.Randomized) {
           qfrm = randSubIter(dinfo, _parms._max_iterations, _parms._seed);
           u = directSVD(dinfo, qfrm, model);
         } else

h2o-algos/src/main/java/hex/svd/SVDModel.java

@@ -24,7 +24,7 @@ public static class SVDParameters extends Model.Parameters {
     public boolean _impute_missing = false;   // Should missing numeric values be imputed with the column mean?
 
     public enum Method {
-      GramSVD, Power, Probabilistic
+      GramSVD, Power, Randomized
     }
   }
 

h2o-algos/src/test/java/hex/svd/SVDTest.java

@@ -2,10 +2,7 @@
 
 import hex.DataInfo;
 import hex.SplitFrame;
-import hex.gram.Gram;
 import hex.svd.SVDModel.SVDParameters;
-import org.apache.commons.math3.analysis.function.Pow;
-import org.apache.commons.math3.analysis.function.Power;
 import org.junit.Assert;
 import org.junit.BeforeClass;
 import org.junit.Test;
@@ -15,8 +12,6 @@
 import water.Scope;
 import water.TestUtil;
 import water.fvec.Frame;
-import water.rapids.Exec;
-import water.util.ArrayUtils;
 import water.util.FrameUtils;
 import water.util.Log;
 
@@ -227,7 +222,7 @@ public class SVDTest extends TestUtil {
       parms._nv = 4;
       parms._keep_u = true;
       parms._transform = DataInfo.TransformType.STANDARDIZE;
-      parms._svd_method = SVDParameters.Method.Probabilistic;
+      parms._svd_method = SVDParameters.Method.Randomized;
       parms._max_iterations = 4;
 
       SVD job = new SVD(parms);
@@ -422,7 +417,7 @@ public class SVDTest extends TestUtil {
       parms._use_all_factor_levels = true;
       parms._keep_u = false;
       parms._transform = DataInfo.TransformType.NONE;
-      parms._svd_method = SVDParameters.Method.Probabilistic;
+      parms._svd_method = SVDParameters.Method.Randomized;
       parms._max_iterations = 7;
 
       SVD job = new SVD(parms);
@@ -464,7 +459,7 @@ public class SVDTest extends TestUtil {
       parms._nv = 5;
       parms._keep_u = true;
       parms._transform = DataInfo.TransformType.DEMEAN;
-      parms._svd_method = SVDParameters.Method.Probabilistic;
+      parms._svd_method = SVDParameters.Method.Randomized;
       parms._impute_missing = true;
       parms._max_iterations = 20;
 
@@ -513,7 +508,7 @@ public class SVDTest extends TestUtil {
       parms._nv = 8;
       parms._only_v = false;
       parms._keep_u = true;
-      parms._svd_method = SVDParameters.Method.Probabilistic;
+      parms._svd_method = SVDParameters.Method.Randomized;
       parms._impute_missing = true;
       parms._max_iterations = 20;
 

h2o-r/h2o-package/R/svd.R

@@ -19,13 +19,19 @@
 #'        column; "DESCALE" for dividing by the standard deviation of each
 #'        column; "STANDARDIZE" for demeaning and descaling; and "NORMALIZE"
 #'        for demeaning and dividing each column by its range (max - min).
+#' @param svd_method A character string that indicates how SVD should be calculated.
+#'        Possible values are "GramSVD": distributed computation of the Gram matrix
+#'        followed by a local SVD using the JAMA package, "Power": computation of
+#'        the SVD using the power iteration method, "Randomized": approximate SVD
+#'        by projecting onto a random subspace (see references).
 #' @param seed (Optional) Random seed used to initialize the right singular vectors
 #'        at the beginning of each power method iteration.
 #' @param use_all_factor_levels (Optional) A logical value indicating whether all
 #'        factor levels should be included in each categorical column expansion.
 #'        If FALSE, the indicator column corresponding to the first factor level
 #'        of every categorical variable will be dropped. Defaults to TRUE.
 #' @return Returns an object of class \linkS4class{H2ODimReductionModel}.
+#' @references N. Halko, P.G. Martinsson, J.A. Tropp. {Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions}[http://arxiv.org/abs/0909.4061]. SIAM Rev., Survey and Review section, Vol. 53, num. 2, pp. 217-288, June 2011.
 #' @examples
 #' library(h2o)
 #' localH2O <- h2o.init()
@@ -38,6 +44,7 @@ h2o.svd <- function(training_frame, x, nv,
                     destination_key,                   # h2o generates its own default parameters
                     max_iterations = 1000,
                     transform = "NONE",
+                    svd_method = c("GramSVD", "Power", "Randomized"),
                     seed,
                     use_all_factor_levels)
 {
@@ -69,6 +76,8 @@ h2o.svd <- function(training_frame, x, nv,
     parms$max_iterations <- max_iterations
   if(!missing(transform))
     parms$transform <- transform
+  if(!missing(svd_method))
+    parms$svd_method <- svd_method
   if(!missing(seed))
     parms$seed <- seed
   if(!missing(use_all_factor_levels))