[layercontext] Add unsafe methods

getGradient methods always perform checks about memory being allocated
for the gradients. However for labels, we donot want to allocate memory
as label is allocated by the dataset. Yet, we still want to able to set
it at the output gradient. So, a getGradientUnsafe is added which
allows returing empty tensor reference so that gradient can be set.

This new interface does not provide new functionality for the layer
developer but rather provides an unsafe method with the existing
functionality.

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

diff --git a/nntrainer/layers/activation_layer.cpp b/nntrainer/layers/activation_layer.cpp
index e7819bc..a798851 100644 (file)
--- a/nntrainer/layers/activation_layer.cpp
+++ b/nntrainer/layers/activation_layer.cpp
@@ -47,7 +47,7 @@ void ActivationLayer::calcDerivative(RunLayerContext &context) {
   Tensor &ret = context.getOutgoingDerivative(SINGLE_INOUT_IDX);
   Tensor &out = context.getOutput(SINGLE_INOUT_IDX);
 
-  ret = acti_func.run_prime_fn(out, ret, deriv);
+  acti_func.run_prime_fn(out, ret, deriv);
 }
 
 void ActivationLayer::setProperty(const std::vector<std::string> &values) {

diff --git a/nntrainer/layers/layer_context.h b/nntrainer/layers/layer_context.h
index cddd855..4270653 100644 (file)
--- a/nntrainer/layers/layer_context.h
+++ b/nntrainer/layers/layer_context.h
@@ -325,6 +325,19 @@ public:
     if (!outputs[idx]->hasGradient())
       throw std::invalid_argument(
         "Requesting gradient for a non-trainable tensor.");
+    return getOutputGradUnsafe(idx);
+  }
+
+  /**
+   * @brief Get the Output Grad tensor object
+   *
+   * @param idx Identifier of the output
+   * @return Tensor& Reference to the output grad tensor
+   *
+   * @note recommended to NOT use this function as a layer developer but rather
+   * use getOutputGrad().
+   */
+  Tensor &getOutputGradUnsafe(unsigned int idx) {
     return outputs[idx]->getGradientRef();
   }
 

diff --git a/nntrainer/layers/layer_node.cpp b/nntrainer/layers/layer_node.cpp
index 991a34e..1a5e03e 100644 (file)
--- a/nntrainer/layers/layer_node.cpp
+++ b/nntrainer/layers/layer_node.cpp
@@ -427,8 +427,6 @@ void LayerNode::setBatch(unsigned int batch) {
       run_context.setBatch(batch);
       layer->setBatch(run_context, batch);
     } else {
-      for (auto &dim : input_dim)
-        dim.batch(batch);
       init_context.setBatch(batch);
       layer->setBatch(init_context, batch);
     }

diff --git a/nntrainer/layers/layer_node.h b/nntrainer/layers/layer_node.h
index f89ef7f..f58b139 100644 (file)
--- a/nntrainer/layers/layer_node.h
+++ b/nntrainer/layers/layer_node.h
@@ -328,7 +328,7 @@ public:
    * @brief     Get number of inputs
    * @retval    number of inputs
    */
-  unsigned int getNumInputs() const { return input_dim.size(); }
+  unsigned int getNumInputs() const { return init_context.getNumInputs(); }
 
   /**
    * @brief     Get number of outputs
@@ -462,7 +462,7 @@ public:
    * @brief Get the Weight object
    *
    * @param idx Identifier of the weight
-   * @return Tensor& Reference to the weight tensor
+   * @return Weight& Reference to the weight
    */
   Weight getWeightWrapper(unsigned int idx) {
     if (layerv1 == nullptr) {
@@ -586,6 +586,20 @@ public:
   }
 
   /**
+   * @brief Get the Output Grad unsafe
+   *
+   * @param idx Identifier of the output
+   * @return Tensor& Reference to the output grad tensor
+   */
+  Tensor &getOutputGradUnsafe(unsigned int idx) {
+    if (layerv1 == nullptr) {
+      return run_context.getOutputGradUnsafe(idx);
+    } else {
+      return getLayer()->getOutputRef()[idx]->getGradientRef();
+    }
+  }
+
+  /**
    * @brief     read layer Weight & Bias data from file
    * @param file input file stream
    */

diff --git a/nntrainer/models/neuralnet.cpp b/nntrainer/models/neuralnet.cpp
index 92736f4..e606e4d 100644 (file)
--- a/nntrainer/models/neuralnet.cpp
+++ b/nntrainer/models/neuralnet.cpp
@@ -237,13 +237,13 @@ sharedConstTensors NeuralNetwork::forwarding(sharedConstTensors input,
       << " requirements size: " << layer_node->getNumOutputs();
 
     for (unsigned int i = 0; i < layer_node->getNumOutputs(); i++) {
-      layer_node->getOutputGrad(i) = *label[i];
+      layer_node->getOutputGradUnsafe(i) = *label[i];
     }
   };
 
   auto clear_label = [](auto const &layer_node) {
     for (unsigned int i = 0; i < layer_node->getNumOutputs(); i++) {
-      layer_node->getOutputGrad(i) = Tensor();
+      layer_node->getOutputGradUnsafe(i) = Tensor();
     }
   };
 

diff --git a/nntrainer/tensor/var_grad.h b/nntrainer/tensor/var_grad.h
index 0359c74..c4fa320 100644 (file)
--- a/nntrainer/tensor/var_grad.h
+++ b/nntrainer/tensor/var_grad.h
@@ -84,7 +84,7 @@ public:
     dim(v.getDim()),
     var(std::make_shared<Tensor>(v.getSharedDataTensor(dim, 0, false))),
     grad(std::make_shared<Tensor>(g.getSharedDataTensor(dim, 0, false))),
-    trainable(!g.uninitialized()),
+    need_gradient(!g.uninitialized()),
     alloc_now(v.isAllocated()),
     name(n) {}
 
@@ -131,6 +131,8 @@ public:
     initializeVariable(var_preallocated);
     if (gtrain)
       initializeGradient(grad_preallocated);
+    else
+      grad = std::make_shared<Tensor>();
   }
 
   /**