[mlpack-git] master: A better attempt at batch training. (26ba0e2)

[gitdub at big.cc.gt.atl.ga.us]

Repository : https://github.com/mlpack/mlpack

On branch  : master
Link       : https://github.com/mlpack/mlpack/compare/de9cc4b05069e1fa4793d9355f2f595af5ff45d2...6070527af14296cd99739de6c62666cc5d2a2125

>---------------------------------------------------------------

commit 26ba0e2b009bd8a455a57db03c31d1638de88115
Author: Ryan Curtin <ryan at ratml.org>
Date:   Tue Nov 3 06:45:35 2015 -0800

    A better attempt at batch training.


>---------------------------------------------------------------

26ba0e2b009bd8a455a57db03c31d1638de88115
 src/mlpack/methods/hoeffding_trees/hoeffding_tree_impl.hpp | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/src/mlpack/methods/hoeffding_trees/hoeffding_tree_impl.hpp b/src/mlpack/methods/hoeffding_trees/hoeffding_tree_impl.hpp
index f7c48ab..d8193c8 100644
--- a/src/mlpack/methods/hoeffding_trees/hoeffding_tree_impl.hpp
+++ b/src/mlpack/methods/hoeffding_trees/hoeffding_tree_impl.hpp
@@ -187,8 +187,12 @@ void HoeffdingTree<
   {
     // Pass all the points through the nodes, and then split only after that.
     checkInterval = data.n_cols; // Only split on the last sample.
+    // Don't split if there are fewer than five points.
+    size_t oldMaxSamples = maxSamples;
+    maxSamples = std::max(size_t(data.n_cols - 1), size_t(5));
     for (size_t i = 0; i < data.n_cols; ++i)
       Train(data.col(i), labels[i]);
+    maxSamples = oldMaxSamples;
 
     // Now, if we did split, find out which points go to which child, and
     // perform the same batch training.
@@ -215,6 +219,11 @@ void HoeffdingTree<
       // batch-mode training.
       for (size_t i = 0; i < children.size(); ++i)
       {
+        // If we don't have any points that go to the child in question, don't
+        // train that child.
+        if (counts[i] == 0)
+          continue;
+
         // The submatrix here is non-contiguous, but I think this will be faster
         // than copying the points to an ordered state.  We still have to
         // assemble the labels vector, though.