ConvDepthWise.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_CONVDEPTHWISE_H_
#define AIDGE_CORE_OPERATOR_CONVDEPTHWISE_H_

#include <array>
#include <cmath>
#include <numeric>
#include <vector>

#include "aidge/data/Tensor.hpp"
#include "aidge/graph/Node.hpp"
#include "aidge/operator/OperatorTensor.hpp"
#include "aidge/operator/Producer.hpp"
#include "aidge/utils/StaticAttributes.hpp"
#include "aidge/utils/Registrar.hpp"
#include "aidge/utils/Types.h"

namespace Aidge {
enum class ConvDepthWiseAttr { StrideDims, DilationDims, Channels, KernelDims };

template <DimIdx_t DIM>
class ConvDepthWise_Op : public OperatorTensor,
                public Registrable<ConvDepthWise_Op<DIM>, std::string, std::unique_ptr<OperatorImpl>(const ConvDepthWise_Op<DIM> &)>,
                public StaticAttributes<ConvDepthWiseAttr,
                                       std::array<DimSize_t, DIM>,
                                       std::array<DimSize_t, DIM>,
                                       DimSize_t,
                                       std::array<DimSize_t, DIM>> {
public:
    static constexpr const char *Type = "ConvDepthWise";

    ConvDepthWise_Op() = delete;

    using Attributes_ = StaticAttributes<ConvDepthWiseAttr,
                                             std::array<DimSize_t, DIM>,
                                             std::array<DimSize_t, DIM>,
                                             DimSize_t,
                                             std::array<DimSize_t, DIM>>;
    template <ConvDepthWiseAttr e>
    using attr = typename Attributes_::template attr<e>;

    constexpr ConvDepthWise_Op(const DimSize_t nbChannels,
                               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> &dilation_dims = create_array<DimSize_t,DIM>(1))
        : OperatorTensor(Type, 1, 2, 1),
          Attributes_(attr<ConvDepthWiseAttr::StrideDims>(stride_dims),
                      attr<ConvDepthWiseAttr::DilationDims>(dilation_dims),
                      attr<ConvDepthWiseAttr::Channels>(nbChannels),
                      attr<ConvDepthWiseAttr::KernelDims>(kernel_dims)) {}

    /**
     * @brief Copy-constructor. Copy the operator attributes and its output tensor(s), but not its input tensors (the new operator has no input associated).
     * @param op Operator to copy.
     */
    ConvDepthWise_Op(const ConvDepthWise_Op<DIM>& op)
        : OperatorTensor(op),
          Attributes_(op)
    {
        mImpl = op.mImpl ? Registrar<ConvDepthWise_Op<DIM>>::create(op.mOutputs[0]->getImpl()->backend())(*this) : nullptr;
    }

    /**
     * @brief Clone the operator using its copy-constructor.
     * @see Operator::ConvDepthWise_Op
     */
    std::shared_ptr<Operator> clone() const override {
        return std::make_shared<ConvDepthWise_Op<DIM>>(*this);
    }


    void computeOutputDims() override final {
        // check inputs have been associated
        // TODO : add a check of inputs dimensions ?
        bool associated = true;
        for (IOIndex_t i = 0; i < 3; ++i) {
            if (!getInput(i)) {
                AIDGE_THROW_OR_ABORT(std::runtime_error, "Every input should be associated with a Tensor");
            }
            associated &= !(getInput(i)->empty());
        }
        if (associated) {
            std::array<DimSize_t, DIM + 2> outputDims = {};
            const std::array<DimSize_t, DIM + 2> inputDims(getInput(0)->template dims<DIM+2>());

            for (std::size_t dim = 0; dim < this->template getAttr<ConvDepthWiseAttr::KernelDims>().size() ; ++dim) {
                const DimSize_t kernelExtent = this->template getAttr<ConvDepthWiseAttr::DilationDims>()[dim] *
                                                       (this->template getAttr<ConvDepthWiseAttr::KernelDims>()[dim] - 1) +
                                               1;

                outputDims[dim+2] = 1 + static_cast<DimSize_t>(
                        floor(static_cast<float>(inputDims[dim+2] - kernelExtent) /
                              static_cast<float>(this->template getAttr<ConvDepthWiseAttr::StrideDims>()[dim])));
            }
            // std::array<DimSize_t, DIM+2> weightDims = append(mInputs[0]->dims()[1],append(1, this->template getAttr<ConvDepthWiseAttr::KernelDims>()));
            // if (mInputs[1]->empty()) {
            //     mInputs[1]->resize(weightDims);
            // }
            // if (mInputs[2]->empty()) {
            //     mInputs[2]->resize({mInputs[0]->dims()[1]});
            // }
            outputDims[1] = inputDims[1];
            outputDims[0] = inputDims[0];
            mOutputs[0]->resize(outputDims);
        }
    }

    std::vector<std::pair<std::size_t, std::vector<DimSize_t>>> computeReceptiveField(const std::size_t firstIdx, const std::vector<DimSize_t>& outputDims, const IOIndex_t outputIdx = 0) const override {
        if (outputIdx != 0) {
            AIDGE_THROW_OR_ABORT(std::runtime_error, "Conv_Op Operator has got only one output Tensor.");
        }
        if ((outputDims.size() == (DIM+2)) && outputDimsForwarded()) {
            // Offset
            const auto outputIdxDims = mOutputs[0]->getCoord(firstIdx);
            auto inputIdxDims = outputIdxDims; // batch idx is the same

            for (DimIdx_t i = 0; i < (DIM+2); ++i) {
                if (((outputDims[i] + outputIdxDims[i]) > mOutputs[0]->template dims<DIM+2>()[i]) || (outputDims[i] == 0)) {
                    AIDGE_THROW_OR_ABORT(std::runtime_error, "Given outputDim out of range for dimension %lu (%lu + %lu)", static_cast<std::size_t>(i), outputIdxDims[i], outputDims[i]);
                }
            }

            // padding is not a parameter of Conv_Op. It is handled in Pad_Op Operator
            // Input
            // same batch value
            std::vector<DimSize_t> inputDims{outputDims[0], outputDims[1]};
            for (DimIdx_t i = 0; i < DIM; ++i) {
                inputDims.push_back((outputDims[2+static_cast<std::size_t>(i)] - 1)
                            * this->template getAttr<ConvDepthWiseAttr::StrideDims>()[static_cast<std::size_t>(i)]
                            + 1
                            + (this->template getAttr<ConvDepthWiseAttr::KernelDims>()[static_cast<std::size_t>(i)] - 1)
                            * this->template getAttr<ConvDepthWiseAttr::DilationDims>()[static_cast<std::size_t>(i)]);
                inputIdxDims[2+i] *= this->template getAttr<ConvDepthWiseAttr::StrideDims>()[static_cast<std::size_t>(i)];
            }

            // Weight
            std::vector<DimSize_t> weightDims{outputDims[1], 1};
            for (std::size_t i = 0; i < DIM; ++i) {
                weightDims.push_back(this->template getAttr<ConvDepthWiseAttr::KernelDims>()[i]);
            }
            std::vector<DimSize_t> weightIdxDims = std::vector<DimSize_t>(DIM+2, 0);
            weightIdxDims[0] = outputIdxDims[1];

            // Bias
            const std::vector<DimSize_t> biasDims{outputDims[1]}; // the number of output channel
            const std::vector<DimSize_t> biasIdxDims{outputIdxDims[1]};

            // Result
            std::vector<std::pair<std::size_t, std::vector<DimSize_t>>> res;
            res.push_back(std::pair<std::size_t, std::vector<DimSize_t>>(getInput(0)->getIdx(inputIdxDims), inputDims));
            res.push_back(std::pair<std::size_t, std::vector<DimSize_t>>(getInput(1)->getIdx(weightIdxDims), weightDims));
            res.push_back(std::pair<std::size_t, std::vector<DimSize_t>>(getInput(2)->getIdx(biasIdxDims), biasDims));
            return res;
        }
        AIDGE_THROW_OR_ABORT(std::runtime_error, "Given outputDim out of range or output dim not forwarded yet.");
    }

    void setBackend(const std::string &name, int device = 0) override {
        mImpl = Registrar<ConvDepthWise_Op<DIM>>::create(name)(*this);
        mOutputs[0]->setBackend(name, device);

        // By default, automatically set backend for weight and bias inputs
        getInput(1)->setBackend(name, device);
        getInput(2)->setBackend(name, device);
    }

    void setDataType(const DataType& dt) const override {
        mOutputs[0]->setDataType(dt);

        // By default, automatically set data type for weight and bias inputs
        getInput(1)->setDataType(dt);
        getInput(2)->setDataType(dt);
    }

    static const std::vector<std::string> getInputsName(){
        return {"data_input", "weight", "bias"};
    }
    static const std::vector<std::string> getOutputsName(){
        return {"data_output"};
    }
};

template <std::array<DimSize_t, 1>::size_type DIM>
inline std::shared_ptr<Node> ConvDepthWise(const DimSize_t nbChannels,
                                           const std::array<DimSize_t, DIM> &kernelDims,
                                           const std::string& name = "",
                                           const std::array<DimSize_t, DIM> &strideDims = create_array<DimSize_t,DIM>(1),
                                           const std::array<DimSize_t, DIM> &dilationDims = create_array<DimSize_t,DIM>(1)) {
    // FIXME: properly handle default w&b initialization in every cases
    static_assert(DIM<=MaxDim,"Too many kernel dimensions required by ConvDepthWise, not supported");
    auto convDW = std::make_shared<Node>(std::make_shared<ConvDepthWise_Op<static_cast<DimIdx_t>(DIM)>>(nbChannels, kernelDims, strideDims, dilationDims), name);
    addProducer(convDW, 1, append(nbChannels, append(DimSize_t(1), kernelDims)), "w");
    addProducer(convDW, 2, std::array<DimSize_t, 1>({nbChannels}), "b");
    return convDW;
}

// helper with C-style array instead of std::array for kernel_dims to allow automatic template DIM deduction
template <DimSize_t DIM>
inline std::shared_ptr<Node> ConvDepthWise(
    const DimSize_t nbChannels,
    DimSize_t const (&kernelDims)[DIM],
    const std::string& name = "",
    const std::array<DimSize_t, DIM> &strideDims = create_array<DimSize_t,DIM>(1),
    const std::array<DimSize_t, DIM> &dilationDims = create_array<DimSize_t,DIM>(1)) {
    static_assert(DIM<=MaxDim,"Too many kernel dimensions required by ConvDepthWise, not supported");
    return ConvDepthWise(nbChannels, to_array(kernelDims), name, strideDims, dilationDims);
}
}  // namespace Aidge

namespace {
template <>
const char *const EnumStrings<Aidge::ConvDepthWiseAttr>::data[] = {"StrideDims", "DilationDims", "Channels",
                                                          "KernelDims"};
}

#endif /* AIDGE_CORE_OPERATOR_CONVDEPTHWISE_H_ */