Fixed the SquareError value calculation and the documentation formulas.

The value was half of the expected result previously.
AltasK · Mar 29, 2016 · 34e3cc9 · 34e3cc9
1 parent fdf1138
commit 34e3cc9
Show file tree

Hide file tree

Showing 2 changed files with 7 additions and 7 deletions.
diff --git a/Documentation/CNTK-TechReport/lyx/CNTKBook_CN_Chapter.lyx b/Documentation/CNTK-TechReport/lyx/CNTKBook_CN_Chapter.lyx
@@ -8333,9 +8333,9 @@ SquareError
 \begin_layout Standard
 \begin_inset Formula 
 \begin{eqnarray}
-v\left(\mathbf{X},\mathbf{\mathbf{Y}}\right) & \leftarrow & \frac{1}{2}\mathrm{Tr}\left(\left(\mathbf{X}-\mathbf{Y}\right)\left(\mathbf{X}-\mathbf{Y}\right)^{T}\right)\\
-\nabla_{\mathbf{X}}^{J} & \leftarrow & \nabla_{\mathbf{X}}^{J}+\mathbf{\nabla_{n}^{\mathit{J}}}\left(\mathbf{X}-\mathbf{Y}\right)\\
-\nabla_{\mathbf{\mathbf{Y}}}^{J} & \leftarrow & \nabla_{\mathbf{\mathbf{Y}}}^{J}-\mathbf{\nabla_{n}^{\mathit{J}}}\left(\mathbf{X}-\mathbf{Y}\right).
+v\left(\mathbf{X},\mathbf{Y}\right) & \leftarrow & \mathrm{Tr}\left(\left(\mathbf{X}-\mathbf{Y}\right)\left(\mathbf{X}-\mathbf{Y}\right)^{T}\right)\\
+\nabla_{\mathbf{X}}^{J} & \leftarrow & \nabla_{\mathbf{X}}^{J}+2\mathbf{\nabla_{n}^{\mathit{J}}}\left(\mathbf{X}-\mathbf{Y}\right)\\
+\nabla_{\mathbf{Y}}^{J} & \leftarrow & \nabla_{\mathbf{Y}}^{J}-2\mathbf{\nabla_{n}^{\mathit{J}}}\left(\mathbf{X}-\mathbf{Y}\right).
 \end{eqnarray}
 
 \end_inset
@@ -8367,8 +8367,8 @@ Note that
 \color none
 \begin_inset Formula 
 \begin{eqnarray}
-\frac{\partial v}{\partial\mathbf{X}} & = & \mathbf{X}-\mathbf{Y}\\
-\frac{\partial v}{\partial\mathbf{Y}} & = & \mathbf{-\left(X-\mathbf{Y}\right)}.
+\frac{\partial v}{\partial\mathbf{X}} & = & +2\left(\mathbf{X}-\mathbf{Y}\right)\\
+\frac{\partial v}{\partial\mathbf{Y}} & = & -2\left(\mathbf{X}-\mathbf{Y}\right).
 \end{eqnarray}
 
 \end_inset

diff --git a/Source/ComputationNetworkLib/TrainingNodes.h b/Source/ComputationNetworkLib/TrainingNodes.h
@@ -44,7 +44,7 @@ class SquareErrorNode : public ComputationNodeNonLooping /*ComputationNode*/<Ele
     {
         FrameRange fr(Input(0)->GetMBLayout());
         auto gradient = Input(inputIndex)->GradientFor(fr);
-        Matrix<ElemType>::Multiply1x1AndWeightedAdd(inputIndex == 0 ? 1.0f : -1.0f, Gradient() /*1x1*/, *m_leftMinusRight, 1.0f, gradient);
+        Matrix<ElemType>::Multiply1x1AndWeightedAdd(inputIndex == 0 ? 2.0f : -2.0f, Gradient() /*1x1*/, *m_leftMinusRight, 1.0f, gradient);
     }
 
     virtual bool OutputUsedInComputingInputNodesGradients() const override
@@ -68,7 +68,7 @@ class SquareErrorNode : public ComputationNodeNonLooping /*ComputationNode*/<Ele
         MaskMissingColumnsToZero(*m_leftMinusRight, Input(0)->GetMBLayout(), fr); // we are fine since it will only be called with full minibatch.
         ElemType v = m_leftMinusRight->FrobeniusNorm();
         Value().VerifySize(1, 1);
-        Value().SetValue(v * v / 2);
+        Value().SetValue(v * v);
 #if NANCHECK
         Value().HasNan("SquareError");
 #endif