fixed bug in the case of regression: it's all about the difference be…

…tween `a[indices] += 1` and `np.add.at(a,indices,1)` and the trick of using `np.bincount`, which is faster than `ufunc.at`. See: ml31415/numpy-groupies#1 and numpy/numpy#5922
kevin-keraudren · Feb 25, 2016 · 21116f0 · 21116f0
1 parent 462107b
commit 21116f0
Show file tree

Hide file tree

Showing 9 changed files with 3 additions and 2 deletions.
diff --git a/fern_AxisAligned_regression.png b/fern_AxisAligned_regression.png
diff --git a/fern_Conic_regression.png b/fern_Conic_regression.png
diff --git a/fern_Linear_regression.png b/fern_Linear_regression.png
diff --git a/fern_Parabola_regression.png b/fern_Parabola_regression.png
diff --git a/randomferns.py b/randomferns.py
@@ -15,10 +15,11 @@ def fit( self, points, responses ):
         self.tests = np.array( self.test_class.generate_all( points, self.depth ) )
         if self.regression:
             self.target_dim = responses.shape[1]
-            self.data = np.ones( (2**self.depth, self.target_dim), dtype='float64' )
+            self.data = np.zeros( (2**self.depth, self.target_dim), dtype='float64' )
             bins = self.apply_tests(points)
             bincount = np.bincount(bins, minlength=self.data.shape[0])
-            self.data[self.apply_tests(points)] += responses
+            for dim in range(self.target_dim):
+                self.data[:,dim] += np.bincount(bins, weights=responses[:,dim], minlength=self.data.shape[0])
             self.data[bincount>0] /= bincount[bincount>0][...,np.newaxis]
         else:
             self.n_classes = responses.max() + 1

diff --git a/randomferns_AxisAligned_regression.png b/randomferns_AxisAligned_regression.png
diff --git a/randomferns_Conic_regression.png b/randomferns_Conic_regression.png
diff --git a/randomferns_Linear_regression.png b/randomferns_Linear_regression.png
diff --git a/randomferns_Parabola_regression.png b/randomferns_Parabola_regression.png