[MERGE][master] updating from main

.gitlab/ci/build.gitlab-ci.yml

@@ -12,6 +12,7 @@ build:ubuntu_cpp:
     - make -j4 all install
 
   artifacts:
+    expire_in: 1 week
     paths:
       - build_cpp/
       - install_cpp/
@@ -29,6 +30,7 @@ build:ubuntu_python:
     - export AIDGE_INSTALL=`pwd`/install
     - python3 -m pip install .
   artifacts:
+    expire_in: 1 week
     paths:
       - venv/
 
@@ -57,6 +59,7 @@ build:windows_cpp:
     - cmake --install . --config Debug
 
   artifacts:
+    expire_in: 1 week
     paths:
       - build_cpp/
       - install_cpp/

CMakeLists.txt

@@ -52,9 +52,9 @@ target_include_directories(${module_name}
 )
 
 # PYTHON BINDING
-generate_python_binding(${project} ${module_name})
-
 if (PYBIND)
+    generate_python_binding(${project} ${module_name})
+
     # Handles Python + pybind11 headers dependencies
     target_link_libraries(${module_name}
         PUBLIC 
@@ -72,21 +72,13 @@ else()
 set(SANITIZE_FLAGS -fsanitize=address,leak,undefined,float-divide-by-zero -fno-omit-frame-pointer)
 endif()
 
-if(WERROR)
     target_compile_options(${module_name} PRIVATE
     $<$<OR:$<CXX_COMPILER_ID:Clang>,$<CXX_COMPILER_ID:AppleClang>,$<CXX_COMPILER_ID:GNU>>:
-    -Wall -Wextra -fPIC -Wold-style-cast -Winline -pedantic -Werror=narrowing -Wshadow -Werror ${SANITIZE_FLAGS}>)
+    -Wall -Wextra -fPIC -Wold-style-cast -Winline -pedantic -Werror=narrowing -Wshadow $<$<BOOL:${WERROR}>:-Werror> ${SANITIZE_FLAGS}>)
     target_compile_options(${module_name} PRIVATE
     $<$<CXX_COMPILER_ID:MSVC>:
     /W4>)
-else()
-    target_compile_options(${module_name} PRIVATE
-        $<$<OR:$<CXX_COMPILER_ID:Clang>,$<CXX_COMPILER_ID:AppleClang>,$<CXX_COMPILER_ID:GNU>>:
-        -Wall -Wextra -fPIC -Wold-style-cast -Winline -pedantic -Werror=narrowing -Wshadow -Wpedantic ${SANITIZE_FLAGS}>)
-        target_compile_options(${module_name} PRIVATE
-        $<$<CXX_COMPILER_ID:MSVC>:
-        /W4>)
-endif()
+
 target_link_options(${module_name} PUBLIC ${SANITIZE_FLAGS})
 
 if(CMAKE_COMPILER_IS_GNUCXX AND COVERAGE)

cmake/PybindModuleCreation.cmake

-function(generate_python_binding name target_to_bind) 
-    if (PYBIND)
-        add_definitions(-DPYBIND)
-        Include(FetchContent)
+function(generate_python_binding name target_to_bind)
+    add_definitions(-DPYBIND)
+    Include(FetchContent)
 
-        FetchContent_Declare(
-        PyBind11
-        GIT_REPOSITORY https://github.com/pybind/pybind11.git
-        GIT_TAG        v2.10.4 # or a later release
-        )
+    FetchContent_Declare(
+    PyBind11
+    GIT_REPOSITORY https://github.com/pybind/pybind11.git
+    GIT_TAG        v2.10.4 # or a later release
+    )
 
-        # Use the New FindPython mode, recommanded. Requires CMake 3.15+
-        find_package(Python COMPONENTS Interpreter Development)
-        FetchContent_MakeAvailable(PyBind11)
+    # Use the New FindPython mode, recommanded. Requires CMake 3.15+
+    find_package(Python COMPONENTS Interpreter Development)
+    FetchContent_MakeAvailable(PyBind11)
 
-        message(STATUS "Creating binding for module ${name}")
-        file(GLOB_RECURSE pybind_src_files "python_binding/*.cpp")
+    message(STATUS "Creating binding for module ${name}")
+    file(GLOB_RECURSE pybind_src_files "python_binding/*.cpp")
 
-        pybind11_add_module(${name} MODULE ${pybind_src_files} "NO_EXTRAS") # NO EXTRA recquired for pip install
-        target_include_directories(${name} PUBLIC "python_binding")
-        target_link_libraries(${name} PUBLIC ${target_to_bind})        
-    endif()
+    pybind11_add_module(${name} MODULE ${pybind_src_files} "NO_EXTRAS") # NO EXTRA recquired for pip install
+    target_include_directories(${name} PUBLIC "python_binding")
+    target_link_libraries(${name} PUBLIC ${target_to_bind})
 endfunction()

include/aidge/aidge.hpp

@@ -34,11 +34,13 @@
 #include "aidge/operator/FC.hpp"
 #include "aidge/operator/GenericOperator.hpp"
 #include "aidge/operator/Matmul.hpp"
-#include "aidge/operator/MetaOperator.hpp"
+#include "aidge/operator/MaxPooling.hpp"
+//#include "aidge/operator/MetaOperator.hpp"
 #include "aidge/operator/Operator.hpp"
 #include "aidge/operator/Producer.hpp"
 #include "aidge/operator/ReLU.hpp"
 #include "aidge/operator/Softmax.hpp"
+#include "aidge/operator/Scaling.hpp"
 #include "aidge/scheduler/Scheduler.hpp"
 #include "aidge/utils/CParameter.hpp"
 #include "aidge/utils/Parameter.hpp"

include/aidge/graph/Node.hpp

@@ -303,7 +303,7 @@ public:
    * @param inId Input index.
    * @return std::shared_ptr<Node>&
    */
-  inline NodePtr &getParents(const IOIndex_t inId) {
+  inline NodePtr &getParent(const IOIndex_t inId) {
     assert(inId != gk_IODefaultIndex);
     return mParents.at(inId);
   }

include/aidge/hook/execTime.hpp

+/**
+ * \file execTime.hpp
+ * \brief execTime structure
+ * \version file 1.0.0
+ * \date Creation 27 June 2023
+ * \date 27 June 2023
+ * \par ChangeLog
+ * \par
+ *  v1.0.0, 27 June 2023<br>
+ *  - Initial version.
+ * \author mn271187, ik243221
+ * \copyright
+ *  Copyright (c) 2023 CEA, LIST, Embedded Artificial Intelligence Laboratory. All
+ *  rights reserved.
+ */
+
+#ifndef execTime_H_
+#define execTime_H_
+
+#include "aidge/operator/Operator.hpp"
+#include "aidge/hook/hook.hpp"
+#include <memory>
+#include <chrono>
+#include <vector>
+
+namespace Aidge {
+
+class ExecTime : public Hook {
+private:
+    std::vector<std::chrono::high_resolution_clock::time_point> registeredTimes = std::vector<std::chrono::high_resolution_clock::time_point>();
+public:
+    ExecTime(const std::shared_ptr<Operator> op) : Hook(op) {}
+    ~ExecTime() = default;
+
+    void call() override final {
+        registeredTimes.push_back(std::chrono::high_resolution_clock::now());
+    }
+
+    static std::shared_ptr<ExecTime> create(const std::shared_ptr<Operator> op)
+    {
+        return std::make_shared<ExecTime>(op);
+    }
+
+    std::vector<std::chrono::high_resolution_clock::time_point> getTimes() {
+        return  registeredTimes;
+    }
+
+    std::chrono::high_resolution_clock::time_point getTime(size_t idx) {
+        return registeredTimes[idx];
+    }
+
+};
+
+namespace {
+    static Registrar<Hook> registrarHook_ExecTime({"execution_time"}, Aidge::ExecTime::create);
+}
+}
+
+#endif /* execTime_H_ */
\ No newline at end of file

include/aidge/hook/hook.hpp

+/**
+ * \file Hook.hpp
+ * \brief Hook structure
+ * \version file 1.0.0
+ * \date Creation 27 June 2023
+ * \date 27 June 2023
+ * \par ChangeLog
+ * \par
+ *  v1.0.0, 27 June 2023<br>
+ *  - Initial version.
+ * \author mn271187, ik243221
+ * \copyright
+ *  Copyright (c) 2023 CEA, LIST, Embedded Artificial Intelligence Laboratory. All
+ *  rights reserved.
+ */
+
+#ifndef Hook_H_
+#define Hook_H_
+
+#include "aidge/utils/Parameter.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include <memory>
+
+namespace Aidge {
+
+class Operator;
+class Hook : public Registrable<Hook, std::tuple<std::string>, std::shared_ptr<Hook>(const std::shared_ptr<Operator>)> {
+//class Hook : public Registrable<Hook, std::tuple<std::string>, std::shared_ptr<Hook>(const std::shared_ptr<Operator>)>{
+protected:
+    const std::shared_ptr<Operator> mOperator;
+
+public:
+    Hook(std::shared_ptr<Operator> op) : mOperator(op) {}
+    virtual ~Hook();
+
+    virtual void call() = 0;
+
+};
+}
+
+#endif /* Hook_H_ */
\ No newline at end of file

include/aidge/hook/outputRange.hpp

+/**
+ * \file execTime.hpp
+ * \brief execTime structure
+ * \version file 1.0.0
+ * \date Creation 27 June 2023
+ * \date 27 June 2023
+ * \par ChangeLog
+ * \par
+ *  v1.0.0, 27 June 2023<br>
+ *  - Initial version.
+ * \author ik243221
+ * \copyright
+ *  Copyright (c) 2023 CEA, LIST, Embedded Artificial Intelligence Laboratory. All
+ *  rights reserved.
+ */
+
+#ifndef AIDGE_CORE_HOOK_OUTPUTRANGE_H_
+#define AIDGE_CORE_HOOK_OUTPUTRANGE_H_
+
+#include "aidge/operator/Operator.hpp"
+#include "aidge/hook/hook.hpp"
+#include <memory>
+#include <chrono>
+#include <vector>
+#include <cmath>
+namespace Aidge {
+
+class OutputRange : public Hook {
+private:
+    std::vector<float> registeredOutputs = std::vector<float>();
+public:
+    OutputRange(const std::shared_ptr<Operator> op) : Hook(op) {}
+    ~OutputRange() = default;
+
+    void call() override final {
+        //std::cout << "call() outputRange hook " << std::endl;
+        //this assumes there is only 1 output possible
+        std::shared_ptr<Tensor> tensor = mOperator->getOutput(0);
+        //tensor->print();
+        //std::cout << "call() outputRange hook : tensor printed" << std::endl;
+        float max_value = 0.;
+        float * casted_tensor = static_cast<float *>(tensor->getImpl()->rawPtr());
+        //find the absolute max value in the tensor, save it to registered outputs
+        for(std::size_t i = 0; i < tensor->size(); ++i) {
+            //std::cout << "call() outputRange hook : casted_tensor[i] = " << casted_tensor[i] << std::endl;
+            if(std::abs(casted_tensor[i]) > max_value){
+                max_value = std::abs(casted_tensor[i]);
+            }
+        }
+        //std::cout << "call() outputRange hook : max_value = " << max_value << std::endl;
+        registeredOutputs.push_back(max_value);
+    }
+
+    static std::shared_ptr<OutputRange> create(const std::shared_ptr<Operator> op)
+    {
+        return std::make_shared<OutputRange>(op);
+    }
+
+    std::vector<float> getOutputs() {
+        return  registeredOutputs;
+    }
+
+    float getOutput(size_t idx) {
+        return registeredOutputs[idx];
+    }
+
+};
+
+namespace {
+    static Registrar<Hook> registrarHook_OutputRange({"output_range"}, Aidge::OutputRange::create);
+}
+}
+
+#endif /* outputRange_H_ */
\ No newline at end of file

include/aidge/operator/MaxPooling.hpp

+/********************************************************************************
+ * Copyright (c) 2023 CEA-List
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Eclipse Public License 2.0 which is available at
+ * http://www.eclipse.org/legal/epl-2.0.
+ *
+ * SPDX-License-Identifier: EPL-2.0
+ *
+ ********************************************************************************/
+
+#ifndef AIDGE_CORE_OPERATOR_MAXPOOLING_H_
+#define AIDGE_CORE_OPERATOR_MAXPOOLING_H_
+
+#include <array>
+#include <numeric>
+#include <vector>
+#include <cmath>
+
+#include "aidge/data/Tensor.hpp"
+#include "aidge/graph/Node.hpp"
+#include "aidge/operator/Operator.hpp"
+#include "aidge/operator/Producer.hpp"
+#include "aidge/utils/Parameter.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+
+namespace Aidge {
+enum class MaxPoolingParam { StrideDims, KernelDims, PaddingDims };
+
+template <DimIdx_t DIM>
+class MaxPooling_Op : public Operator,
+                public Registrable<MaxPooling_Op<DIM>, std::string, std::unique_ptr<OperatorImpl>(const MaxPooling_Op<DIM> &)>,
+                public Parameterizable<MaxPoolingParam,
+                                       std::array<DimSize_t, DIM>,
+                                       std::array<DimSize_t, DIM>,
+                                       std::array<DimSize_t, (DIM<<1) >> {
+private:
+    // FIXME: change accessibility
+    std::shared_ptr<Tensor> mInput = std::make_shared<Tensor>();
+    const std::shared_ptr<Tensor> mOutput = std::make_shared<Tensor>();
+
+public:
+    static constexpr const char *Type = "MaxPooling";
+
+    MaxPooling_Op() = delete;
+
+    using Parameterizable_ = Parameterizable<MaxPoolingParam,
+                                             std::array<DimSize_t, DIM>,
+                                             std::array<DimSize_t, DIM>,
+                                             std::array<DimSize_t, (DIM<<1)> >;
+    template <MaxPoolingParam e>
+    using param = typename Parameterizable_::template param<e>;
+
+    constexpr MaxPooling_Op(const std::array<DimSize_t, DIM> &kernel_dims,
+                            const std::array<DimSize_t, DIM> &stride_dims = create_array<DimSize_t,DIM>(1),
+                            const std::array<DimSize_t, (DIM<<1)> &padding_dims = create_array<DimSize_t,(DIM<<1)>(0))
+        : Operator(Type),
+          Parameterizable_(param<MaxPoolingParam::StrideDims>(stride_dims),
+                           param<MaxPoolingParam::KernelDims>(kernel_dims),
+                           param<MaxPoolingParam::PaddingDims>(padding_dims)),
+          mOutput(std::make_shared<Tensor>()) {
+        setDatatype(DataType::Float32);
+    }
+
+    constexpr void associateInput(const IOIndex_t inputIdx, std::shared_ptr<Data> data) override final {
+        assert(inputIdx < 1 && "operators supports only 3 inputs");
+        (void) inputIdx; // avoid unused warning
+        assert(strcmp(data->type(), Tensor::Type) == 0 && "input data must be of Tensor type");
+
+        mInput = std::dynamic_pointer_cast<Tensor>(data);
+    }
+
+    constexpr void computeOutputDims() override final {
+        if (!mInput->empty()) {
+            std::array<DimSize_t, DIM + 2> outputDims = {};
+
+            for (std::size_t dim = 0; dim < this->template get<MaxPoolingParam::KernelDims>().size() ; ++dim) {
+                outputDims[dim+2] = 1 + static_cast<DimSize_t>(
+                                            std::floor(static_cast<float>(mInput->dims()[dim+2] -
+                                                                    this->template get<MaxPoolingParam::KernelDims>()[dim] +
+                                                                    this->template get<MaxPoolingParam::PaddingDims>()[dim] +
+                                                                    this->template get<MaxPoolingParam::PaddingDims>()[dim+DIM]) /
+                                            static_cast<float>(this->template get<MaxPoolingParam::StrideDims>()[dim])));
+            }
+            outputDims[1] = mInput->dims()[1];
+            outputDims[0] = mInput->dims()[0];
+            mOutput->resize(outputDims);
+        }
+    }
+
+    bool outputDimsForwarded() const override final { return !(mOutput->empty()); }
+
+
+    inline Tensor& input(const IOIndex_t inputIdx) const override final {
+        assert(inputIdx == 0 && "operators supports only 1 inputs");
+        (void) inputIdx; // avoid unused warning
+        return *(mInput.get());
+    }
+    inline Tensor& output(const IOIndex_t /*outputIdx*/) const override final { return *(mOutput.get()); }
+
+
+    inline std::shared_ptr<Tensor> getInput(const IOIndex_t inputIdx) const override final {
+        assert(inputIdx == 0 && "MaxPooling Operators supports only 1 inputs");
+        (void) inputIdx; // avoid unused warning
+        return mInput;
+    }
+    inline std::shared_ptr<Tensor> getOutput(const IOIndex_t outputIdx) const override final {
+        assert(outputIdx == 0 && "MaxPooling Operators has only 1 outputs");
+        (void) outputIdx; // avoid unused warning
+        return mOutput;
+    }
+
+
+    std::shared_ptr<Data> getRawInput(const IOIndex_t inputIdx) const override final {
+        assert(inputIdx == 0 && "operators supports only 1 inputs");
+        (void) inputIdx; // avoid unused warning
+        return std::static_pointer_cast<Data>(mInput);
+    }
+    std::shared_ptr<Data> getRawOutput(const IOIndex_t outputIdx) const override final {
+        assert(outputIdx == 0 && "operator supports only 1 output");
+        (void) outputIdx; // avoid unused warning
+        return std::static_pointer_cast<Data>(mOutput);
+    }
+
+
+    void setBackend(const std::string &name) {
+        mImpl = Registrar<MaxPooling_Op<DIM>>::create(name)(*this);
+        mOutput->setBackend(name);
+
+        // FIXME: temporary workaround
+        mInput->setBackend(name);
+    }
+
+    void setDatatype(const DataType &datatype) {
+        mOutput->setDatatype(datatype);
+
+        // FIXME: temporary workaround
+        mInput->setDatatype(datatype);
+    }
+
+    inline IOIndex_t nbInputs() const noexcept override final { return 1; }
+    inline IOIndex_t nbDataInputs() const noexcept override final { return 1; }
+    inline IOIndex_t nbOutputs() const noexcept override final { return 1; }
+};
+
+template <std::array<DimSize_t, 1>::size_type DIM>
+inline std::shared_ptr<Node> MaxPooling(const std::array<DimSize_t, DIM> &kernel_dims,
+                                           const std::string& name = "",
+                                           const std::array<DimSize_t, DIM> &stride_dims = create_array<DimSize_t,DIM>(1),
+                                           const std::array<DimSize_t, (DIM<<1)> &padding_dims = create_array<DimSize_t,(DIM<<1)>(0)) {
+    // FIXME: properly handle default w&b initialization in every cases
+    static_assert(DIM<=MaxDim,"Too many kernel dimensions required by MaxPooling, not supported");
+    auto avgPool = std::make_shared<Node>(std::make_shared<MaxPooling_Op<static_cast<DimIdx_t>(DIM)>>(kernel_dims, stride_dims, padding_dims), name);
+    return avgPool;
+}
+
+template <DimSize_t DIM>
+inline std::shared_ptr<Node> MaxPooling(
+    DimSize_t const (&kernel_dims)[DIM],
+    const std::string& name = "",
+    const std::array<DimSize_t, DIM> &stride_dims = create_array<DimSize_t,DIM>(1),
+    const std::array<DimSize_t, (DIM<<1)> &padding_dims = create_array<DimSize_t,(DIM<<1)>(0)) {
+    static_assert(DIM<=MaxDim,"Too many kernel dimensions required by MaxPooling, not supported");
+    return MaxPooling(to_array(kernel_dims), name, stride_dims, padding_dims);
+}
+}  // namespace Aidge
+
+namespace {
+template <>
+const char *const EnumStrings<Aidge::MaxPoolingParam>::data[] = {"StrideDims", "KernelDims", "PaddingDims"};
+}
+
+#endif /* AIDGE_CORE_OPERATOR_MAXPOOLING_H_ */

include/aidge/operator/Operator.hpp

@@ -20,12 +20,14 @@
 #include "aidge/data/Data.hpp"
 #include "aidge/data/Tensor.hpp"
 #include "aidge/utils/Types.h"
+#include "aidge/hook/hook.hpp"
 
 namespace Aidge {
 
 class Operator : public std::enable_shared_from_this<Operator> {
 protected:
   std::unique_ptr<OperatorImpl> mImpl; // implementation of the operator
+  std::map<std::string, std::shared_ptr<Hook>> mHooks;
 
 private:
   std::string mType;
@@ -48,6 +50,15 @@ public:
     virtual std::shared_ptr<Tensor> getOutput(const IOIndex_t outputIdx) const = 0;
     virtual Tensor& output(const IOIndex_t /*outputIdx*/) const = 0;
 
+    std::shared_ptr<Hook> getHook(std::string hookName) {
+        return mHooks[hookName];
+    }
+    void addHook(std::string hookName) {
+        mHooks.insert(std::pair<std::string, std::shared_ptr<Hook>>(hookName,Registrar<Hook>::create({hookName})(shared_from_this())));
+    }
+
+    void runHooks() const;
+
 ///////////////////////////////////////////////////////
 //        IMPLEMENTATION
 ///////////////////////////////////////////////////////

include/aidge/operator/Scaling.hpp

+/********************************************************************************
+ * Copyright (c) 2023 CEA-List
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Eclipse Public License 2.0 which is available at
+ * http://www.eclipse.org/legal/epl-2.0.
+ *
+ * SPDX-License-Identifier: EPL-2.0
+ *
+ ********************************************************************************/
+
+#ifndef __AIDGE_CORE_OPERATOR_Scaling_H__
+#define __AIDGE_CORE_OPERATOR_Scaling_H__
+
+#include <vector>
+#include <memory>
+
+
+
+#include "aidge/utils/Parameter.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/operator/Operator.hpp"
+#include "aidge/backend/OperatorImpl.hpp"
+#include "aidge/data/Tensor.hpp"
+#include "aidge/data/Data.hpp"
+#include "aidge/graph/Node.hpp"
+#include "aidge/utils/Types.h"
+
+namespace Aidge {
+enum class ScalingParam {
+    scalingFactor
+};
+
+class Scaling_Op : public Operator,
+    public Registrable<Scaling_Op, std::string, std::unique_ptr<OperatorImpl>(const Scaling_Op&)>,
+    public Parameterizable<ScalingParam, float> {
+public:
+    // FIXME: change accessibility
+    std::shared_ptr<Tensor> mInput = std::make_shared<Tensor>();
+    const std::shared_ptr<Tensor> mOutput = std::make_shared<Tensor>();
+
+public:
+    static constexpr const char* Type = "Scaling";
+
+    Scaling_Op() = delete;
+
+    using Parameterizable_ = Parameterizable<ScalingParam, float>;
+    template <ScalingParam e> using param = typename Parameterizable_::template param<e>;
+
+    Scaling_Op(float scalingFactor)
+            : Operator(Type),
+            Parameterizable_(
+                param<ScalingParam::scalingFactor>(scalingFactor))
+    {
+        setDatatype(DataType::Float32);
+    }
+
+    void associateInput(const IOIndex_t inputIdx, std::shared_ptr<Data> data) override final {
+        assert(inputIdx == 0 && "operator supports only 1 input");
+        assert(strcmp(data->type(), Tensor::Type)==0 && "input data must be of Tensor type");
+        (void) inputIdx; //avoid unused warning
+        mInput = std::dynamic_pointer_cast<Tensor>(data);
+    }
+
+    void computeOutputDims() override final {
+        if (!mInput->empty())
+            mOutput->resize(mInput->dims());
+    }
+
+    bool outputDimsForwarded() const override final {
+        return !(mOutput->empty());
+    }
+
+
+    inline Tensor& input(const IOIndex_t inputIdx) const override final {
+        assert((inputIdx == 0) && "Scaling Operator has only 1 input");
+        (void) inputIdx; // avoid unused warning
+        return *(mInput.get());
+    }
+    inline Tensor& output(const IOIndex_t outputIdx) const override final {
+        assert((outputIdx == 0) && "Scaling Operator has only 1 output");
+        (void) outputIdx; // avoid unused warning
+        return *(mOutput.get());
+    }
+
+
+    inline std::shared_ptr<Tensor> getInput(const IOIndex_t inputIdx) const override final { 
+        assert((inputIdx == 0) && "Scaling Operator has only 1 input");
+        (void) inputIdx; // avoid unused warning
+        return mInput;
+    }
+    inline std::shared_ptr<Tensor> getOutput(const IOIndex_t outputIdx) const override final {
+        assert((outputIdx == 0) && "Scaling Operator has only 1 output");
+        (void) outputIdx; // avoid unused warning
+        return mOutput;
+    }
+
+
+    std::shared_ptr<Data> getRawInput(const IOIndex_t inputIdx) const override final {
+        assert(inputIdx == 0 && "operator supports only 1 input");
+        (void) inputIdx; // avoid unused warning
+        return std::static_pointer_cast<Data>(mInput);
+    }
+    std::shared_ptr<Data> getRawOutput(const IOIndex_t outputIdx) const override final {
+        assert(outputIdx == 0 && "operator supports only 1 output");
+        (void) outputIdx; // avoid unused warning;
+        return mOutput;
+    }
+
+
+    void setBackend(const std::string& name) {
+        mImpl = Registrar<Scaling_Op>::create(name)(*this);
+        mOutput->setBackend(name);
+        // FIXME: temporary workaround
+        mInput->setBackend(name);
+    }
+    void setDatatype(const DataType& datatype) {
+        mOutput->setDatatype(datatype);
+
+        // FIXME: temporary workaround
+        mInput->setDatatype(datatype);
+    }
+
+    inline IOIndex_t nbInputs() const noexcept override final { return 1; }
+    inline IOIndex_t nbDataInputs() const noexcept override final { return 1; }
+    inline IOIndex_t nbOutputs() const noexcept override final { return 1; }
+};
+
+inline std::shared_ptr<Node> Scaling(float scalingFactor = 1.0f, const std::string& name = "") {
+    return std::make_shared<Node>(std::make_shared<Scaling_Op>(scalingFactor), name);
+}
+}
+
+namespace {
+template <>
+const char* const EnumStrings<Aidge::ScalingParam>::data[]
+    = {"scalingFactor"};
+}
+
+#endif /* __AIDGE_CORE_OPERATOR_RELU_H__ */

python_binding/graph/pybind_Node.cpp

@@ -136,6 +136,16 @@ void init_Node(py::module& m) {
             :rtype: int
             )mydelimiter")
 
+            .def("get_parents", &Node::getParents,
+            R"mydelimiter(
+            Get parents.
+            )mydelimiter")
+
+            .def("get_children", (std::set<std::shared_ptr<Node>> (Node::*)() const) &Node::getChildren,
+            R"mydelimiter(
+            Get children.
+            )mydelimiter")
+
             .def("__call__", &Node::operator(), py::arg("connectors"));
 }
 }  // namespace Aidge

python_binding/operator/pybind_MaxPooling.cpp

+/********************************************************************************
+ * Copyright (c) 2023 CEA-List
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Eclipse Public License 2.0 which is available at
+ * http://www.eclipse.org/legal/epl-2.0.
+ *
+ * SPDX-License-Identifier: EPL-2.0
+ *
+ ********************************************************************************/
+#ifdef PYBIND
+#include <pybind11/pybind11.h>
+#include <pybind11/stl.h>
+
+#include <string>
+#include <vector>
+#include <array>
+
+#include "aidge/utils/Parameter.hpp"
+#include "aidge/backend/OperatorImpl.hpp"
+#include "aidge/operator/MaxPooling.hpp"
+#include "aidge/operator/Operator.hpp"
+#include "aidge/utils/Types.h"
+#include "aidge/data/Tensor.hpp"
+
+namespace py = pybind11;
+namespace Aidge {
+
+template <DimIdx_t DIM> void declare_MaxPoolingOp(py::module &m) {
+  py::class_<MaxPooling_Op<DIM>, std::shared_ptr<MaxPooling_Op<DIM>>, Operator, PyAbstractParametrizable>(
+    m, ("MaxPoolingOp" + std::to_string(DIM) + "D").c_str(),
+    py::multiple_inheritance())
+  .def(py::init<const std::array<DimSize_t, DIM> &,
+                const std::array<DimSize_t, DIM> &,
+                const std::array<DimSize_t, (DIM<<1)> &>(),
+        py::arg("kernel_dims"),
+        py::arg("stride_dims"),
+        py::arg("padding_dims"));
+  
+  m.def(("MaxPooling" + std::to_string(DIM) + "D").c_str(), [](const std::vector<DimSize_t>& kernel_dims, 
+                                                                  const std::string& name,
+                                                                  const std::vector<DimSize_t> &stride_dims,
+                                                                  const std::vector<DimSize_t> &padding_dims) {
+        // Lambda function wrapper because PyBind fails to convert const array.
+        // So we use a vector that we convert in this function to a const DimeSize_t [DIM] array. 
+        if (kernel_dims.size() != DIM) {
+            throw std::runtime_error("kernel_dims size [" + std::to_string(kernel_dims.size()) + "] does not match DIM [" + std::to_string(DIM) +"]");
+        }
+        if (stride_dims.size() != DIM) {
+            throw std::runtime_error("stride_dims size [" + std::to_string(stride_dims.size()) + "] does not match DIM [" + std::to_string(DIM) +"]");
+        }
+        if (padding_dims.size() != (DIM<<1)) {
+            throw std::runtime_error("padding_dims size [" + std::to_string(padding_dims.size()) + "] does not match DIM [" + std::to_string(DIM<<1) +"]");
+        }
+        DimSize_t tmp_kernel_dims_array[DIM];
+        for (size_t i = 0; i < DIM; ++i) {
+            tmp_kernel_dims_array[i] = kernel_dims[i];
+        }
+        DimSize_t tmp_stride_dims_array[DIM];
+        for (size_t i = 0; i < DIM; ++i) {
+            tmp_stride_dims_array[i] = stride_dims[i];
+        }
+        DimSize_t tmp_padding_dims_array[DIM<<1];
+        for (size_t i = 0; i < (DIM<<1); ++i) {
+            tmp_padding_dims_array[i] = padding_dims[i];
+        }
+        const DimSize_t (&kernel_dims_array)[DIM] = tmp_kernel_dims_array;
+        const DimSize_t (&stride_dims_array)[DIM] = tmp_stride_dims_array;
+        const DimSize_t (&padding_dims_array)[DIM<<1] = tmp_padding_dims_array;
+        return MaxPooling<DIM>(to_array(kernel_dims_array), name, to_array(stride_dims_array), to_array(padding_dims_array));
+    }, py::arg("kernel_dims"),
+       py::arg("name") = "",
+       py::arg("stride_dims") = std::vector<DimSize_t>(DIM,1),
+       py::arg("padding_dims") = std::vector<DimSize_t>(DIM<<1,0));
+  
+}
+
+
+void init_MaxPooling(py::module &m) {
+  declare_MaxPoolingOp<1>(m);
+  declare_MaxPoolingOp<2>(m);
+  declare_MaxPoolingOp<3>(m);
+ 
+  // FIXME:
+  // m.def("MaxPooling1D", static_cast<NodeAPI(*)(const char*, int, int, int const
+  // (&)[1])>(&MaxPooling));
+}
+} // namespace Aidge
+#endif
\ No newline at end of file

python_binding/pybind_core.cpp

@@ -29,6 +29,7 @@ void init_FC(py::module&);
 void init_GenericOperator(py::module&);
 void init_LeakyReLU(py::module&);
 void init_Matmul(py::module&);
+void init_MaxPooling(py::module&);
 void init_Producer(py::module&);
 void init_ReLU(py::module&);
 void init_Softmax(py::module&);
@@ -75,6 +76,7 @@ void init_Aidge(py::module& m){
     init_GenericOperator(m);
     init_LeakyReLU(m);
     init_Matmul(m);
+    init_MaxPooling(m);
     init_ReLU(m);
     init_Softmax(m);
 

src/graph/GraphView.cpp

@@ -326,7 +326,7 @@ void Aidge::GraphView::add(std::shared_ptr<Node> node, bool includeLearnablePara
   // add learnable parameters to the graph
   if (includeLearnableParam) {
     for (IOIndex_t i = node->nbDataInputs(); i < node->nbInputs(); ++i) {
-      std::shared_ptr<Node> parentNode = node->getParents(static_cast<IOIndex_t>(i));
+      std::shared_ptr<Node> parentNode = node->getParent(static_cast<IOIndex_t>(i));
       if (parentNode) {
           parentNode->addView(shared_from_this());
           mNodes.insert(parentNode);

src/graph/Node.cpp

@@ -226,7 +226,7 @@ void Aidge::Node::addChild(std::shared_ptr<GraphView> otherView, const IOIndex_t
 }
 
 void Aidge::Node::addParent(const std::shared_ptr<Node> other_node, const IOIndex_t inId) {
-    if (getParents(inId) != nullptr) {
+    if (getParent(inId) != nullptr) {
         printf("Warning, you're replacing a Parent.\n");
     }
     assert((inId != gk_IODefaultIndex) && (inId < nbInputs()) && "Input index out of bound.");

src/operator/Operator.cpp

@@ -42,6 +42,14 @@ void Aidge::Operator::updateConsummerProducer(){
     mImpl->updateConsummerProducer();
 }
 
-void Aidge::Operator::forward() { mImpl->forward(); }
+void Aidge::Operator::runHooks() const {
+    for (auto& hook : mHooks) {
+        hook.second->call();
+    }
+}
+void Aidge::Operator::forward() {
+    mImpl->forward();
+    runHooks();
+}
 
 void Aidge::Operator::backward() { mImpl->backward(); }

src/recipies/FuseMulAdd.cpp

@@ -59,12 +59,12 @@ void Aidge::fuseMulAdd(std::set<std::shared_ptr<Node>> nodes){
 
     // Step 2 : Branch existing producers & create the others
     // link weights & bias
-    if (matmul->getParents(1)==nullptr) {
-        matmul->getParents(0)->addChild(fc, 0, 1);
+    if (matmul->getParent(1)==nullptr) {
+        matmul->getParent(0)->addChild(fc, 0, 1);
     } else {
-        if (matmul->getParents(0)!=nullptr)
-            matmul->getParents(0)->addChild(fc, 0, 0);
-        matmul->getParents(1)->addChild(fc, 0, 1);
+        if (matmul->getParent(0)!=nullptr)
+            matmul->getParent(0)->addChild(fc, 0, 0);
+        matmul->getParent(1)->addChild(fc, 0, 1);
     }
     (producer_add_bias.first)->addChild(fc,0,2);