[ Tensor ] remove sscal to set zero.

We do need to remove the Nan or Inf value in Tensor by call setZero(). However, if we using sscal, then Nan or Inf values are remain still. This PR change the sscal to memset. Resolves: **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: jijoong.moon <[email protected]>
nnstreamer · Jun 9, 2024 · 8e06368 · 8e06368
1 parent b154146
commit 8e06368
Show file tree

Hide file tree

Showing 3 changed files with 9 additions and 6 deletions.
diff --git a/nntrainer/layers/layer_context.cpp b/nntrainer/layers/layer_context.cpp
@@ -52,7 +52,7 @@ InitLayerContext::InitLayerContext(
   prefix(prefix_),
   tensor_type(tensor_type_),
   loss_scale(loss_scale_),
-  mode(mode_){
+  mode(mode_) {
   NNTR_THROW_IF(!validate(), std::invalid_argument)
     << "Invalid init context name: " << name
     << " num inputs: " << getNumInputs();
@@ -292,7 +292,6 @@ const Tensor RunLayerContext::getIncomingDerivative(unsigned int idx) const {
   return getOutputGrad(idx);
 }
 
-
 /**
  * @brief Get the Input tensor object
  *

diff --git a/nntrainer/tensor/manager.cpp b/nntrainer/tensor/manager.cpp
@@ -455,8 +455,8 @@ std::vector<Weight *> Manager::requestWeights(
          * reduce the memory.
          */
         bool is_wgrad = true;
-        if (Weight::isGradientClipByGlobalNorm(clip_by_global_norm))
-          is_wgrad = false;
+        //        if (Weight::isGradientClipByGlobalNorm(clip_by_global_norm))
+        //          is_wgrad = false;
         grad = tensor_pool.request(name + Var_Grad::grad_suffix, dim_g,
                                    grad_exec_order, grad_ls,
                                    Tensor::Initializer::ZEROS, is_wgrad);

diff --git a/nntrainer/tensor/tensor.cpp b/nntrainer/tensor/tensor.cpp
@@ -3319,13 +3319,17 @@ void Tensor::setValue(float val) {
 void Tensor::setZero() {
   if (dim.getDataType() == ml::train::TensorDim::DataType::FP32) {
     if (contiguous)
-      sscal(size(), 0, getData<float>(), 1);
+      //      sscal(size(), 0, getData<float>(), 1);
+      /// @note we cannot use sscal, when we set zero. if the data is inf or
+      /// NaN, then the inf or NaN still remain.
+      memset(getData<_FP16>(), 0, sizeof(float) * size());
     else
       apply_i<float>([](float val) -> float { return 0; });
   } else if (dim.getDataType() == ml::train::TensorDim::DataType::FP16) {
 #ifdef ENABLE_FP16
     if (contiguous)
-      sscal(size(), 0, getData<_FP16>(), 1);
+      // sscal(size(), 0, getData<_FP16>(), 1);
+      memset(getData<_FP16>(), 0, sizeof(_FP16) * size());
     else
       apply_i<_FP16>([](_FP16 val) -> _FP16 { return 0; });
 #else