[addition] Bug fix addition layer calcDerivative

Bug fix for addition layer calcDerivative().
addition layer assign the same tensor derivative memory back to its
other layers. If the other layers connecting with addition layers are
inplace, then this can lead to wrong results.

Signed-off-by: Parichay Kapoor <pk.kapoor@samsung.com>

diff --git a/nntrainer/layers/addition_layer.cpp b/nntrainer/layers/addition_layer.cpp
index cd8c590..953d6e8 100644 (file)
--- a/nntrainer/layers/addition_layer.cpp
+++ b/nntrainer/layers/addition_layer.cpp
@@ -59,7 +59,12 @@ void AdditionLayer::forwarding(bool training) {
 void AdditionLayer::calcDerivative() {
 
   for (unsigned int i = 0; i < getNumInputs(); ++i) {
-    net_input[i]->getGradientRef() = net_hidden[0]->getGradientRef();
+    /**
+     * TODO: replace this with tensor assignment during optimization.
+     * Tensor assignement needs to make sure that the previous connected layers
+     * are not inplace
+     */
+    net_input[i]->getGradientRef().copy(net_hidden[0]->getGradientRef());
   }
 }
 

diff --git a/nntrainer/layers/concat_layer.cpp b/nntrainer/layers/concat_layer.cpp
index 0ab0bf1..1c0f2bd 100644 (file)
--- a/nntrainer/layers/concat_layer.cpp
+++ b/nntrainer/layers/concat_layer.cpp
@@ -99,6 +99,7 @@ void ConcatLayer::calcDerivative() {
     TensorDim in_dim = input_dim[idx];
 
     for (unsigned int b = 0; b < in_dim.batch(); ++b) {
+      // TODO: replace with tensor::copy/fill
       memcpy(
         net_input[idx]->getGradient().getAddress(b * in_dim.getFeatureLen()),
         net_hidden[0]->getGradient().getAddress(b * d.getFeatureLen() +