[Wait for #2724][Layer] add "mul layer"

api/ccapi/include/layer.h

@@ -37,6 +37,9 @@ namespace train {
 enum LayerType {
   LAYER_IN = ML_TRAIN_LAYER_TYPE_INPUT,      /**< Input Layer type */
   LAYER_WEIGHT = ML_TRAIN_LAYER_TYPE_WEIGHT, /**< Weight Layer type */
+  LAYER_ADD = ML_TRAIN_LAYER_TYPE_ADD,       /**< Add Layer type */
+  LAYER_SUB = ML_TRAIN_LAYER_TYPE_SUB,       /**< Subtract Layer type */
+  LAYER_MUL = ML_TRAIN_LAYER_TYPE_MUL,       /**< Multiply Layer type */
   LAYER_FC = ML_TRAIN_LAYER_TYPE_FC,         /**< Fully Connected Layer type */
   LAYER_SWIGLU = ML_TRAIN_LAYER_TYPE_SWIGLU, /**< Swiglu Layer type */
   LAYER_BN = ML_TRAIN_LAYER_TYPE_BN, /**< Batch Normalization Layer type */
@@ -299,6 +302,30 @@ WeightLayer(const std::vector<std::string> &properties = {}) {
   return createLayer(LayerType::LAYER_WEIGHT, properties);
 }
 
+/**
+ * @brief Helper function to create add layer
+ */
+inline std::unique_ptr<Layer>
+AddLayer(const std::vector<std::string> &properties = {}) {
+  return createLayer(LayerType::LAYER_ADD, properties);
+}
+
+/**
+ * @brief Helper function to create sub layer
+ */
+inline std::unique_ptr<Layer>
+SubLayer(const std::vector<std::string> &properties = {}) {
+  return createLayer(LayerType::LAYER_SUB, properties);
+}
+
+/**
+ * @brief Helper function to create mul layer
+ */
+inline std::unique_ptr<Layer>
+MulLayer(const std::vector<std::string> &properties = {}) {
+  return createLayer(LayerType::LAYER_MUL, properties);
+}
+
 /**
  * @brief Helper function to create fully connected layer
  */

api/nntrainer-api-common.h

@@ -65,6 +65,9 @@ typedef enum {
   ML_TRAIN_LAYER_TYPE_IDENTITY = 29, /**< Identity Layer type (Since 8.0) */
   ML_TRAIN_LAYER_TYPE_SWIGLU = 30,   /**< Swiglu Layer type */
   ML_TRAIN_LAYER_TYPE_WEIGHT = 31,   /**< Weight Layer type (Since 9.0)*/
+  ML_TRAIN_LAYER_TYPE_ADD = 32,      /**< Add Layer type (Since 9.0)*/
+  ML_TRAIN_LAYER_TYPE_SUB = 33,      /**< Sub Layer type (Since 9.0)*/
+  ML_TRAIN_LAYER_TYPE_MUL = 34,      /**< Mul Layer type (Since 9.0)*/
   ML_TRAIN_LAYER_TYPE_PREPROCESS_FLIP =
     300, /**< Preprocess flip Layer (Since 6.5) */
   ML_TRAIN_LAYER_TYPE_PREPROCESS_TRANSLATE =

nntrainer/app_context.cpp

@@ -31,6 +31,7 @@
 #include <sgd.h>
 
 #include <activation_layer.h>
+#include <add_layer.h>
 #include <addition_layer.h>
 #include <attention_layer.h>
 #include <bn_layer.h>
@@ -57,6 +58,7 @@
 #include <lstmcell.h>
 #include <mol_attention_layer.h>
 #include <mse_loss_layer.h>
+#include <mul_layer.h>
 #include <multi_head_attention_layer.h>
 #include <multiout_layer.h>
 #include <nntrainer_error.h>
@@ -72,6 +74,7 @@
 #include <rnn.h>
 #include <rnncell.h>
 #include <split_layer.h>
+#include <sub_layer.h>
 #include <time_dist.h>
 #include <upsample2d_layer.h>
 #include <weight_layer.h>
@@ -248,6 +251,12 @@ static void add_default_object(AppContext &ac) {
                      LayerType::LAYER_IN);
   ac.registerFactory(nntrainer::createLayer<WeightLayer>, WeightLayer::type,
                      LayerType::LAYER_WEIGHT);
+  ac.registerFactory(nntrainer::createLayer<AddLayer>, AddLayer::type,
+                     LayerType::LAYER_ADD);
+  ac.registerFactory(nntrainer::createLayer<SubLayer>, SubLayer::type,
+                     LayerType::LAYER_SUB);
+  ac.registerFactory(nntrainer::createLayer<MulLayer>, MulLayer::type,
+                     LayerType::LAYER_MUL);
   ac.registerFactory(nntrainer::createLayer<FullyConnectedLayer>,
                      FullyConnectedLayer::type, LayerType::LAYER_FC);
   ac.registerFactory(nntrainer::createLayer<BatchNormalizationLayer>,

nntrainer/layers/add_layer.cpp

@@ -0,0 +1,94 @@
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Copyright (C) 2024 SeungBaek Hong <[email protected]>
+ *
+ * @file   add_layer.cpp
+ * @date   2 August 2024
+ * @see    https://github.com/nnstreamer/nntrainer
+ * @author SeungBaek Hong <[email protected]>
+ * @bug    No known bugs except for NYI items
+ * @brief  This is add layer class (operation layer)
+ *
+ */
+
+#include <add_layer.h>
+#include <nntrainer_error.h>
+#include <nntrainer_log.h>
+#include <node_exporter.h>
+#include <util_func.h>
+
+#include <layer_context.h>
+
+namespace nntrainer {
+
+static constexpr size_t SINGLE_INOUT_IDX = 0;
+
+void AddLayer::finalize(InitLayerContext &context) {
+  context.setOutputDimensions({context.getInputDimensions()[0]});
+}
+
+void AddLayer::forwarding(RunLayerContext &context, bool training) {
+  Tensor &hidden_ = context.getOutput(SINGLE_INOUT_IDX);
+
+  const Tensor &input0 = context.getInput(0);
+  const Tensor &input1 = context.getInput(1);
+
+  input0.add(input1, hidden_);
+}
+
+void AddLayer::incremental_forwarding(RunLayerContext &context,
+                                      unsigned int from, unsigned int to,
+                                      bool training) {
+  Tensor &hidden_ = context.getOutput(SINGLE_INOUT_IDX);
+  TensorDim hidden_dim = hidden_.getDim();
+  TensorDim hidden_step_dim = hidden_dim;
+
+  if (from) {
+    NNTR_THROW_IF(to - from != 1, std::invalid_argument)
+      << "incremental step size is not 1";
+    from = 0;
+    to = 1;
+  }
+
+  hidden_step_dim.batch(1);
+  hidden_step_dim.height(to - from);
+
+  for (unsigned int b = 0; b < hidden_.batch(); ++b) {
+    Tensor hidden_step = hidden_.getSharedDataTensor(
+      hidden_step_dim, b * hidden_dim.getFeatureLen(), true);
+
+    const Tensor &input0 = context.getInput(0);
+    const Tensor &input1 = context.getInput(1);
+
+    TensorDim input_dim = input0.getDim();
+    TensorDim input_step_dim = input_dim;
+    input_step_dim.batch(1);
+    input_step_dim.height(to - from);
+
+    Tensor input0_step = input0.getSharedDataTensor(
+      input_step_dim, b * input_dim.getFeatureLen(), true);
+
+    Tensor input1_step = input1.getSharedDataTensor(
+      input_step_dim, b * input_dim.getFeatureLen(), true);
+
+    input0_step.add(input1_step, hidden_step);
+  }
+}
+
+void AddLayer::calcDerivative(RunLayerContext &context) {
+  context.getOutgoingDerivative(0).copy(
+    context.getIncomingDerivative(SINGLE_INOUT_IDX));
+
+  context.getOutgoingDerivative(1).copy(
+    context.getIncomingDerivative(SINGLE_INOUT_IDX));
+}
+
+void AddLayer::setProperty(const std::vector<std::string> &values) {
+  auto remain_props = loadProperties(values, add_props);
+  if (!remain_props.empty()) {
+    std::string msg = "[AddLayer] Unknown Layer Properties count " +
+                      std::to_string(values.size());
+    throw exception::not_supported(msg);
+  }
+}
+} /* namespace nntrainer */

nntrainer/layers/add_layer.h

@@ -0,0 +1,103 @@
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Copyright (C) 2024 SeungBaek Hong <[email protected]>
+ *
+ * @file   add_layer.h
+ * @date   2 August 2024
+ * @see    https://github.com/nnstreamer/nntrainer
+ * @author SeungBaek Hong <[email protected]>
+ * @bug    No known bugs except for NYI items
+ * @brief  This is add layer class (operation layer)
+ *
+ */
+
+#ifndef __ADD_LAYER_H__
+#define __ADD_LAYER_H__
+#ifdef __cplusplus
+
+#include <common_properties.h>
+#include <layer_devel.h>
+
+namespace nntrainer {
+
+/**
+ * @class Add Layer
+ * @brief Add Layer
+ */
+class AddLayer : public Layer {
+public:
+  /**
+   * @brief Constructor of Add Layer
+   */
+  AddLayer() : Layer(), add_props(props::Print()) {}
+
+  /**
+   * @brief Destructor of Add Layer
+   */
+  ~AddLayer(){};
+
+  /**
+   *  @brief  Move constructor of Add Layer.
+   *  @param[in] AddLayer &&
+   */
+  AddLayer(AddLayer &&rhs) noexcept = default;
+
+  /**
+   * @brief  Move assignment operator.
+   * @parma[in] rhs AddLayer to be moved.
+   */
+  AddLayer &operator=(AddLayer &&rhs) = default;
+
+  /**
+   * @copydoc Layer::finalize(InitLayerContext &context)
+   */
+  void finalize(InitLayerContext &context) override;
+
+  /**
+   * @copydoc Layer::forwarding(RunLayerContext &context, bool training)
+   */
+  void forwarding(RunLayerContext &context, bool training) override;
+
+  /**
+   * @copydoc Layer::incremental_forwarding(RunLayerContext &context, unsigned
+   * int from, unsigned int to, bool training)
+   */
+  void incremental_forwarding(RunLayerContext &context, unsigned int from,
+                              unsigned int to, bool training) override;
+
+  /**
+   * @copydoc Layer::calcDerivative(RunLayerContext &context)
+   */
+  void calcDerivative(RunLayerContext &context) override;
+
+  /**
+   * @copydoc bool supportBackwarding() const
+   */
+  bool supportBackwarding() const override { return true; };
+
+  /**
+   * @copydoc Layer::exportTo(Exporter &exporter, ml::train::ExportMethods
+   * method)
+   */
+  void exportTo(Exporter &exporter,
+                const ml::train::ExportMethods &method) const override {}
+
+  /**
+   * @copydoc Layer::setProperty(const std::vector<std::string> &values)
+   */
+  void setProperty(const std::vector<std::string> &values) override;
+
+  /**
+   * @copydoc Layer::getType()
+   */
+  const std::string getType() const override { return AddLayer::type; };
+
+  std::tuple<props::Print> add_props;
+
+  inline static const std::string type = "add";
+};
+
+} // namespace nntrainer
+
+#endif /* __cplusplus */
+#endif /* __ADD_LAYER_H__ */

nntrainer/layers/meson.build

@@ -5,6 +5,9 @@ nntrainer_inc_abs += meson.current_source_dir() / 'loss'
 layer_sources = [
   'activation_layer.cpp',
   'weight_layer.cpp',
+  'add_layer.cpp',
+  'sub_layer.cpp',
+  'mul_layer.cpp',
   'addition_layer.cpp',
   'attention_layer.cpp',
   'mol_attention_layer.cpp',

nntrainer/layers/mul_layer.cpp

@@ -0,0 +1,96 @@
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Copyright (C) 2024 SeungBaek Hong <[email protected]>
+ *
+ * @file   mul_layer.cpp
+ * @date   30 August 2024
+ * @see    https://github.com/nnstreamer/nntrainer
+ * @author SeungBaek Hong <[email protected]>
+ * @bug    No known bugs except for NYI items
+ * @brief  This is mul layer class (operation layer)
+ *
+ */
+
+#include <mul_layer.h>
+#include <nntrainer_error.h>
+#include <nntrainer_log.h>
+#include <node_exporter.h>
+#include <util_func.h>
+
+#include <layer_context.h>
+
+namespace nntrainer {
+
+static constexpr size_t SINGLE_INOUT_IDX = 0;
+
+void MulLayer::finalize(InitLayerContext &context) {
+  context.setOutputDimensions({context.getInputDimensions()[0]});
+}
+
+void MulLayer::forwarding(RunLayerContext &context, bool training) {
+  Tensor &hidden_ = context.getOutput(SINGLE_INOUT_IDX);
+
+  const Tensor &input0 = context.getInput(0);
+  const Tensor &input1 = context.getInput(1);
+
+  input0.multiply(input1, hidden_);
+}
+
+void MulLayer::incremental_forwarding(RunLayerContext &context,
+                                      unsigned int from, unsigned int to,
+                                      bool training) {
+  Tensor &hidden_ = context.getOutput(SINGLE_INOUT_IDX);
+  TensorDim hidden_dim = hidden_.getDim();
+  TensorDim hidden_step_dim = hidden_dim;
+
+  if (from) {
+    NNTR_THROW_IF(to - from != 1, std::invalid_argument)
+      << "incremental step size is not 1";
+    from = 0;
+    to = 1;
+  }
+
+  hidden_step_dim.batch(1);
+  hidden_step_dim.height(to - from);
+
+  for (unsigned int b = 0; b < hidden_.batch(); ++b) {
+    Tensor hidden_step = hidden_.getSharedDataTensor(
+      hidden_step_dim, b * hidden_dim.getFeatureLen(), true);
+
+    const Tensor &input0 = context.getInput(0);
+    const Tensor &input1 = context.getInput(1);
+
+    TensorDim input_dim = input0.getDim();
+    TensorDim input_step_dim = input_dim;
+    input_step_dim.batch(1);
+    input_step_dim.height(to - from);
+
+    Tensor input0_step = input0.getSharedDataTensor(
+      input_step_dim, b * input_dim.getFeatureLen(), true);
+
+    Tensor input1_step = input1.getSharedDataTensor(
+      input_step_dim, b * input_dim.getFeatureLen(), true);
+
+    input0_step.multiply(input1_step, hidden_step);
+  }
+}
+
+void MulLayer::calcDerivative(RunLayerContext &context) {
+  context.getOutgoingDerivative(0).copy(
+    context.getIncomingDerivative(SINGLE_INOUT_IDX)
+      .multiply(context.getInput(1)));
+
+  context.getOutgoingDerivative(1).copy(
+    context.getIncomingDerivative(SINGLE_INOUT_IDX)
+      .multiply(context.getInput(0)));
+}
+
+void MulLayer::setProperty(const std::vector<std::string> &values) {
+  auto remain_props = loadProperties(values, mul_props);
+  if (!remain_props.empty()) {
+    std::string msg = "[MulLayer] Unknown Layer Properties count " +
+                      std::to_string(values.size());
+    throw exception::not_supported(msg);
+  }
+}
+} /* namespace nntrainer */