FC.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_FC_H_
#define AIDGE_CORE_OPERATOR_FC_H_

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

#include "aidge/utils/Types.h"
#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"

namespace Aidge {
enum class FCAttr { OutChannels, NoBias };

class FC_Op : public OperatorTensor,
              public Registrable<FC_Op,
                                 std::string,
                                 std::unique_ptr<OperatorImpl>(const FC_Op &)>,
              public StaticAttributes<FCAttr, DimSize_t, bool> {
public:
    static constexpr const char* Type = "FC";

    FC_Op() = delete;

    using Attributes_ = StaticAttributes<FCAttr, DimSize_t, bool>;
    template <FCAttr e> using attr = typename Attributes_::template attr<e>;

    FC_Op(DimSize_t out_channels, bool noBias)
    : OperatorTensor(Type, 1, 2, 1),
      Attributes_(
        attr<FCAttr::OutChannels>(out_channels),
        attr<FCAttr::NoBias>(noBias))
    {}

    /**
     * @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.
     */
    FC_Op(const FC_Op& op)
        : OperatorTensor(op),
          Attributes_(op)
    {
        mImpl = op.mImpl ? Registrar<FC_Op>::create(op.mOutputs[0]->getImpl()->backend())(*this) : nullptr;
    }

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

    void associateInput(const IOIndex_t inputIdx, const std::shared_ptr<Data>& data) override final {
        assert(inputIdx < 3 && "operators supports only 3 inputs");
        assert(strcmp(data->type(), Tensor::Type)==0 && "input data must be of Tensor type");
        if (inputIdx == 2) {
            assert(std::dynamic_pointer_cast<Tensor>(data)->size() == ((this->template getAttr<FCAttr::NoBias>()) == false ? static_cast<std::size_t>(this->template getAttr<FCAttr::OutChannels>()) : 0));
            assert(std::dynamic_pointer_cast<Tensor>(data)->nbDims() == 1);
        }
        mInputs[inputIdx] = std::dynamic_pointer_cast<Tensor>(data);
        if (inputIdx == 0 && getInput(0)->nbDims() == 1)
            mInputs[inputIdx]->resize(std::array<DimSize_t, 2>({1, getInput(inputIdx)->size()}));
    }

    void computeOutputDims() override final {
        bool associated = true;
        for (IOIndex_t i = 0; i < nbInputs(); ++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) {
            // <batch, OutChannels>
            mOutputs[0]->resize({getInput(0)->dims()[0], this->template getAttr<FCAttr::OutChannels>()});
        }
    }


    void setBackend(const std::string& name, int device = 0) override {
        mImpl = Registrar<FC_Op>::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"};
    }
};

inline std::shared_ptr<Node> FC(DimSize_t inChannels, DimSize_t outChannels, bool noBias = false, const std::string& name = "") {
    // FIXME: properly handle default w&b initialization in every cases
    auto fc = std::make_shared<Node>(std::make_shared<FC_Op>(outChannels, noBias), name);
    addProducer(fc, 1, std::array<DimSize_t, 2>({outChannels, inChannels}), "w");
    addProducer(fc, 2, (noBias ? std::array<DimSize_t, 1>({0}) : std::array<DimSize_t, 1>({outChannels})), "b"); // already sets bias dims
    return fc;
}
} // namespace Aidge

namespace {
template <>
const char *const EnumStrings<Aidge::FCAttr>::data[] = {"OutChannels",
                                                        "NoBias"};
}

#endif /* AIDGE_CORE_OPERATOR_FC_H_ */