[Wait for #2725][Layer] add "divide layer"

api/ccapi/include/layer.h

@@ -40,6 +40,7 @@ enum LayerType {
   LAYER_ADD = ML_TRAIN_LAYER_TYPE_ADD,           /**< Add Layer type */
   LAYER_SUBTRACT = ML_TRAIN_LAYER_TYPE_SUBTRACT, /**< Subtract Layer type */
   LAYER_MULTIPLY = ML_TRAIN_LAYER_TYPE_MULTIPLY, /**< Multiply Layer type */
+  LAYER_DIVIDE = ML_TRAIN_LAYER_TYPE_DIVIDE,     /**< Divide 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 */
@@ -332,6 +333,14 @@ MultiplyLayer(const std::vector<std::string> &properties = {}) {
   return createLayer(LayerType::LAYER_MULTIPLY, properties);
 }
 
+/**
+ * @brief Helper function to create divide layer
+ */
+inline std::unique_ptr<Layer>
+DivideLayer(const std::vector<std::string> &properties = {}) {
+  return createLayer(LayerType::LAYER_DIVIDE, properties);
+}
+
 /**
  * @brief Helper function to create fully connected layer
  */

api/nntrainer-api-common.h

@@ -68,6 +68,7 @@ typedef enum {
   ML_TRAIN_LAYER_TYPE_ADD = 32,      /**< Add Layer type (Since 9.0)*/
   ML_TRAIN_LAYER_TYPE_SUBTRACT = 33, /**< Subtract Layer type (Since 9.0)*/
   ML_TRAIN_LAYER_TYPE_MULTIPLY = 34, /**< Multiply Layer type (Since 9.0)*/
+  ML_TRAIN_LAYER_TYPE_DIVIDE = 35,   /**< Divide 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

@@ -42,6 +42,7 @@
 #include <conv2d_layer.h>
 #include <cross_entropy_sigmoid_loss_layer.h>
 #include <cross_entropy_softmax_loss_layer.h>
+#include <divide_layer.h>
 #include <dropout.h>
 #include <embedding.h>
 #include <fc_layer.h>
@@ -262,6 +263,8 @@ static void add_default_object(AppContext &ac) {
                      LayerType::LAYER_SUBTRACT);
   ac.registerFactory(nntrainer::createLayer<MultiplyLayer>, MultiplyLayer::type,
                      LayerType::LAYER_MULTIPLY);
+  ac.registerFactory(nntrainer::createLayer<DivideLayer>, DivideLayer::type,
+                     LayerType::LAYER_DIVIDE);
   ac.registerFactory(nntrainer::createLayer<FullyConnectedLayer>,
                      FullyConnectedLayer::type, LayerType::LAYER_FC);
   ac.registerFactory(nntrainer::createLayer<BatchNormalizationLayer>,

nntrainer/layers/divide_layer.cpp

@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Copyright (C) 2024 SeungBaek Hong <[email protected]>
+ *
+ * @file   divide_layer.cpp
+ * @date   10 Oct 2024
+ * @see    https://github.com/nnstreamer/nntrainer
+ * @author SeungBaek Hong <[email protected]>
+ * @bug    No known bugs except for NYI items
+ * @brief  This is div layer class (operation layer)
+ *
+ */
+
+#include <divide_layer.h>
+#include <nntrainer_error.h>
+#include <nntrainer_log.h>
+#include <node_exporter.h>
+#include <util_func.h>
+
+#include <layer_context.h>
+
+namespace nntrainer {
+
+void DivideLayer::finalize(InitLayerContext &context) {
+  context.setOutputDimensions({context.getInputDimensions()[0]});
+}
+
+void DivideLayer::forwarding_operation(const Tensor &input0,
+                                       const Tensor &input1, Tensor &hidden) {
+  input0.divide(input1, hidden);
+}
+
+void DivideLayer::calcDerivative(RunLayerContext &context) {
+  context.getOutgoingDerivative(0).copy(
+    context.getIncomingDerivative(SINGLE_INOUT_IDX)
+      .divide(context.getInput(1)));
+
+  context.getOutgoingDerivative(1).copy(
+    context.getIncomingDerivative(SINGLE_INOUT_IDX)
+      .multiply(context.getInput(0).multiply(-1))
+      .divide(context.getInput(1).pow(2)));
+}
+
+void DivideLayer::setProperty(const std::vector<std::string> &values) {
+  auto remain_props = loadProperties(values, divide_props);
+  if (!remain_props.empty()) {
+    std::string msg = "[DivideLayer] Unknown Layer Properties count " +
+                      std::to_string(values.size());
+    throw exception::not_supported(msg);
+  }
+}
+} /* namespace nntrainer */

nntrainer/layers/divide_layer.h

@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Copyright (C) 2024 SeungBaek Hong <[email protected]>
+ *
+ * @file   divide_layer.h
+ * @date   10 Oct 2024
+ * @see    https://github.com/nnstreamer/nntrainer
+ * @author SeungBaek Hong <[email protected]>
+ * @bug    No known bugs except for NYI items
+ * @brief  This is div layer class (operation layer)
+ *
+ */
+
+#ifndef __DIVIDE_LAYER_H__
+#define __DIVIDE_LAYER_H__
+#ifdef __cplusplus
+
+#include <common_properties.h>
+#include <layer_devel.h>
+#include <operation_layer.h>
+
+namespace nntrainer {
+
+/**
+ * @class Divide Layer
+ * @brief Divide Layer
+ */
+class DivideLayer : public BinaryOperationLayer {
+public:
+  /**
+   * @brief Constructor of Divide Layer
+   */
+  DivideLayer() : BinaryOperationLayer(), divide_props(props::Print()) {}
+
+  /**
+   * @brief Destructor of Divide Layer
+   */
+  ~DivideLayer(){};
+
+  /**
+   *  @brief  Move constructor of Divide Layer.
+   *  @param[in] DivideLayer &&
+   */
+  DivideLayer(DivideLayer &&rhs) noexcept = default;
+
+  /**
+   * @brief  Move assignment operator.
+   * @parma[in] rhs DivideLayer to be moved.
+   */
+  DivideLayer &operator=(DivideLayer &&rhs) = default;
+
+  /**
+   * @copydoc Layer::finalize(InitLayerContext &context)
+   */
+  void finalize(InitLayerContext &context) final;
+
+  /**
+   * @brief forwarding operation for add
+   *
+   * @param input0 input tensor 0
+   * @param input1 input tensor 1
+   * @param hidden tensor to store the result of addition
+   */
+  void forwarding_operation(const Tensor &input0, const Tensor &input1,
+                            Tensor &hidden) final;
+
+  /**
+   * @copydoc Layer::calcDerivative(RunLayerContext &context)
+   */
+  void calcDerivative(RunLayerContext &context) final;
+
+  /**
+   * @copydoc bool supportBackwarding() const
+   */
+  bool supportBackwarding() const final { return true; };
+
+  /**
+   * @copydoc Layer::exportTo(Exporter &exporter, ml::train::ExportMethods
+   * method)
+   */
+  void exportTo(Exporter &exporter,
+                const ml::train::ExportMethods &method) const final {}
+
+  /**
+   * @copydoc Layer::setProperty(const std::vector<std::string> &values)
+   */
+  void setProperty(const std::vector<std::string> &values) final;
+
+  /**
+   * @copydoc Layer::getType()
+   */
+  const std::string getType() const final { return DivideLayer::type; };
+
+  std::tuple<props::Print> divide_props;
+
+  inline static const std::string type = "divide";
+};
+
+} // namespace nntrainer
+
+#endif /* __cplusplus */
+#endif /* __DIVIDE_LAYER_H__ */

nntrainer/layers/meson.build

@@ -8,6 +8,7 @@ layer_sources = [
   'add_layer.cpp',
   'subtract_layer.cpp',
   'multiply_layer.cpp',
+  'divide_layer.cpp',
   'addition_layer.cpp',
   'attention_layer.cpp',
   'mol_attention_layer.cpp',

test/input_gen/genModelTests_v2.py

@@ -505,6 +505,19 @@ def forward(self, inputs, labels):
         return out, loss
 
 
+class DivideOperation(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.fc = torch.nn.Linear(2, 2)
+        self.loss = torch.nn.MSELoss()
+
+    def forward(self, inputs, labels):
+        out = self.fc(inputs[0])
+        out = inputs[0] / out
+        loss = self.loss(out, labels[0])
+        return out, loss
+
+
 if __name__ == "__main__":
     record_v2(
         ReduceMeanLast(),

test/unittest/layers/meson.build

@@ -50,6 +50,7 @@ test_target = [
   'unittest_layers_add.cpp',
   'unittest_layers_subtract.cpp',
   'unittest_layers_multiply.cpp',
+  'unittest_layers_divide.cpp',
   'unittest_layers_multiout.cpp',
   'unittest_layers_rnn.cpp',
   'unittest_layers_rnncell.cpp',

test/unittest/layers/unittest_layers_divide.cpp

@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: Apache-2.0
+/**
+ * Copyright (C) 2024 SeungBaek Hong <[email protected]>
+ *
+ * @file unittest_layers_divide.cpp
+ * @date 30 August 2024
+ * @brief Divide Layer Test
+ * @see	https://github.com/nnstreamer/nntrainer
+ * @author SeungBaek Hong <[email protected]>
+ * @bug No known bugs except for NYI items
+ */
+#include <tuple>
+
+#include <gtest/gtest.h>
+
+#include <divide_layer.h>
+#include <layers_common_tests.h>
+
+auto semantic_divide = LayerSemanticsParamType(
+  nntrainer::createLayer<nntrainer::DivideLayer>, nntrainer::DivideLayer::type,
+  {}, LayerCreateSetPropertyOptions::AVAILABLE_FROM_APP_CONTEXT, false, 1);
+
+auto semantic_divide_multi = LayerSemanticsParamType(
+  nntrainer::createLayer<nntrainer::DivideLayer>, nntrainer::DivideLayer::type,
+  {}, LayerCreateSetPropertyOptions::AVAILABLE_FROM_APP_CONTEXT, false, 2);
+
+GTEST_PARAMETER_TEST(Divide, LayerSemantics,
+                     ::testing::Values(semantic_divide, semantic_divide_multi));

test/unittest/models/unittest_models.cpp

@@ -929,6 +929,25 @@ static std::unique_ptr<NeuralNetwork> makeMultiplyOperation() {
   return nn;
 }
 
+static std::unique_ptr<NeuralNetwork> makeDivideOperation() {
+  std::unique_ptr<NeuralNetwork> nn(new NeuralNetwork());
+
+  auto outer_graph =
+    makeGraph({{"input", {"name=in", "input_shape=1:1:2"}},
+               {"fully_connected", {"name=fc", "unit=2", "input_layers=in"}},
+               {"divide", {"name=divide_layer", "input_layers=in,fc"}},
+               {"mse", {"name=loss", "input_layers=divide_layer"}}});
+
+  for (auto &node : outer_graph) {
+    nn->addLayer(node);
+  }
+
+  nn->setProperty({"batch_size=1"});
+  nn->setOptimizer(ml::train::createOptimizer("sgd", {"learning_rate=0.1"}));
+
+  return nn;
+}
+
 GTEST_PARAMETER_TEST(
   model, nntrainerModelTest,
   ::testing::ValuesIn({
@@ -1005,6 +1024,8 @@ GTEST_PARAMETER_TEST(
                  ModelTestOption::ALL_V2),
     mkModelTc_V2(makeMultiplyOperation, "multiply_operation",
                  ModelTestOption::ALL_V2),
+    mkModelTc_V2(makeDivideOperation, "divide_operation",
+                 ModelTestOption::ALL_V2),
   }),
   [](const testing::TestParamInfo<nntrainerModelTest::ParamType> &info)
     -> const auto & { return std::get<1>(info.param); });