ml_model.hpp

/*
  Copyright (c) 2016 Robert W. Rose
  Copyright (c) 2018 Paul Maevskikh
  Copyright (c) 2024 NORCE
 
  Permission is hereby granted, free of charge, to any person obtaining a copy
  of this software and associated documentation files (the "Software"), to deal
  in the Software without restriction, including without limitation the rights
  to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  copies of the Software, and to permit persons to whom the Software is
  furnished to do so, subject to the following conditions:
 
  The above copyright notice and this permission notice shall be included in all
  copies or substantial portions of the Software.
 
  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  SOFTWARE.
 
  Note: This file is based on kerasify/keras_model.hh
*/
 
#ifndef ML_MODEL_H_
#define ML_MODEL_H_
 
#include <fmt/format.h>
 
#include <algorithm>
#include <chrono>
#include <cmath>
#include <cstdio>
#include <numeric>
#include <opm/common/ErrorMacros.hpp>
#include <opm/material/densead/Math.hpp>
#include <string>
#include <vector>
 
namespace Opm
{
 
namespace ML
{
 
    // NN layer
    // ---------------------
    template <class T>
    class Tensor
    {
    public:
        Tensor()
        {
        }
 
        explicit Tensor(int i)
        {
            resizeI<std::vector<int>>({i});
        }
 
        Tensor(int i, int j)
        {
            resizeI<std::vector<int>>({i, j});
        }
 
        Tensor(int i, int j, int k)
        {
            resizeI<std::vector<int>>({i, j, k});
        }
 
        Tensor(int i, int j, int k, int l)
        {
            resizeI<std::vector<int>>({i, j, k, l});
        }
 
        template <typename Sizes>
        void resizeI(const Sizes& sizes)
        {
            if (sizes.size() == 1)
                dims_ = {(int)sizes[0]};
            if (sizes.size() == 2)
                dims_ = {(int)sizes[0], (int)sizes[1]};
            if (sizes.size() == 3)
                dims_ = {(int)sizes[0], (int)sizes[1], (int)sizes[2]};
            if (sizes.size() == 4)
                dims_ = {(int)sizes[0], (int)sizes[1], (int)sizes[2], (int)sizes[3]};
 
            data_.resize(std::accumulate(begin(dims_), end(dims_), 1.0, std::multiplies<>()));
        }
 
        void flatten()
        {
            OPM_ERROR_IF(dims_.size() == 0, "Invalid tensor");
 
            int elements = dims_[0];
            for (unsigned int i = 1; i < dims_.size(); i++) {
                elements *= dims_[i];
            }
            dims_ = {elements};
        }
 
        T& operator()(int i)
        {
            OPM_ERROR_IF(dims_.size() != 1, "Invalid indexing for tensor");
 
            OPM_ERROR_IF(!(i < dims_[0] && i >= 0),
                         fmt::format(" Invalid i:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     i,
                                     dims_[0]));
 
            return data_[i];
        }
 
        T& operator()(int i, int j)
        {
            OPM_ERROR_IF(dims_.size() != 2, "Invalid indexing for tensor");
            OPM_ERROR_IF(!(i < dims_[0] && i >= 0),
                         fmt::format(" Invalid i:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     i,
                                     dims_[0]));
            OPM_ERROR_IF(!(j < dims_[1] && j >= 0),
                         fmt::format(" Invalid j:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     j,
                                     dims_[1]));
 
            return data_[dims_[1] * i + j];
        }
 
        const T& operator()(int i, int j) const
        {
            OPM_ERROR_IF(dims_.size() != 2, "Invalid indexing for tensor");
            OPM_ERROR_IF(!(i < dims_[0] && i >= 0),
                         fmt::format(" Invalid i:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     i,
                                     dims_[0]));
            OPM_ERROR_IF(!(j < dims_[1] && j >= 0),
                         fmt::format(" Invalid j:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     j,
                                     dims_[1]));
            return data_[dims_[1] * i + j];
        }
 
        T& operator()(int i, int j, int k)
        {
            OPM_ERROR_IF(dims_.size() != 3, "Invalid indexing for tensor");
            OPM_ERROR_IF(!(i < dims_[0] && i >= 0),
                         fmt::format(" Invalid i:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     i,
                                     dims_[0]));
            OPM_ERROR_IF(!(j < dims_[1] && j >= 0),
                         fmt::format(" Invalid j:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     j,
                                     dims_[1]));
            OPM_ERROR_IF(!(k < dims_[2] && k >= 0),
                         fmt::format(" Invalid k:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     k,
                                     dims_[2]));
 
            return data_[dims_[2] * (dims_[1] * i + j) + k];
        }
 
        const T& operator()(int i, int j, int k) const
        {
            OPM_ERROR_IF(dims_.size() != 3, "Invalid indexing for tensor");
            OPM_ERROR_IF(!(i < dims_[0] && i >= 0),
                         fmt::format(" Invalid i:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     i,
                                     dims_[0]));
            OPM_ERROR_IF(!(j < dims_[1] && j >= 0),
                         fmt::format(" Invalid j:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     j,
                                     dims_[1]));
            OPM_ERROR_IF(!(k < dims_[2] && k >= 0),
                         fmt::format(" Invalid k:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     k,
                                     dims_[2]));
 
            return data_[dims_[2] * (dims_[1] * i + j) + k];
        }
 
        T& operator()(int i, int j, int k, int l)
        {
            OPM_ERROR_IF(dims_.size() != 4, "Invalid indexing for tensor");
            OPM_ERROR_IF(!(i < dims_[0] && i >= 0),
                         fmt::format(" Invalid i:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     i,
                                     dims_[0]));
            OPM_ERROR_IF(!(j < dims_[1] && j >= 0),
                         fmt::format(" Invalid j:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     j,
                                     dims_[1]));
            OPM_ERROR_IF(!(k < dims_[2] && k >= 0),
                         fmt::format(" Invalid k:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     k,
                                     dims_[2]));
            OPM_ERROR_IF(!(l < dims_[3] && l >= 0),
                         fmt::format(" Invalid l:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     l,
                                     dims_[3]));
 
            return data_[dims_[3] * (dims_[2] * (dims_[1] * i + j) + k) + l];
        }
 
        const T& operator()(int i, int j, int k, int l) const
        {
            OPM_ERROR_IF(dims_.size() != 4, "Invalid indexing for tensor");
            OPM_ERROR_IF(!(i < dims_[0] && i >= 0),
                         fmt::format(" Invalid i:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     i,
                                     dims_[0]));
            OPM_ERROR_IF(!(j < dims_[1] && j >= 0),
                         fmt::format(" Invalid j:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     j,
                                     dims_[1]));
            OPM_ERROR_IF(!(k < dims_[2] && k >= 0),
                         fmt::format(" Invalid k:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     k,
                                     dims_[2]));
            OPM_ERROR_IF(!(l < dims_[3] && l >= 0),
                         fmt::format(" Invalid l:  "
                                     "{}"
                                     "  max: "
                                     "{}",
                                     l,
                                     dims_[3]));
 
            return data_[dims_[3] * (dims_[2] * (dims_[1] * i + j) + k) + l];
        }
 
        void fill(const T& value)
        {
            std::fill(data_.begin(), data_.end(), value);
        }
 
        // Tensor addition
        Tensor operator+(const Tensor& other)
        {
            OPM_ERROR_IF(dims_.size() != other.dims_.size(),
                         "Cannot add tensors with different dimensions");
            Tensor result;
            result.dims_ = dims_;
            result.data_.resize(data_.size());
 
            std::transform(data_.begin(),
                           data_.end(),
                           other.data_.begin(),
                           result.data_.begin(),
                           [](const T& x, const T& y) { return x + y; });
 
            return result;
        }
 
        // Tensor multiplication
        Tensor multiply(const Tensor& other)
        {
            OPM_ERROR_IF(dims_.size() != other.dims_.size(),
                         "Cannot multiply elements with different dimensions");
 
            Tensor result;
            result.dims_ = dims_;
            result.data_.resize(data_.size());
 
            std::transform(data_.begin(),
                           data_.end(),
                           other.data_.begin(),
                           result.data_.begin(),
                           [](const T& x, const T& y) { return x * y; });
 
            return result;
        }
 
        // Tensor dot for 2d tensor
        Tensor dot(const Tensor& other)
        {
            OPM_ERROR_IF(dims_.size() != 2, "Invalid tensor dimensions");
            OPM_ERROR_IF(other.dims_.size() != 2, "Invalid tensor dimensions");
 
            OPM_ERROR_IF(dims_[1] != other.dims_[0],
                         "Cannot multiply with different inner dimensions");
 
            Tensor tmp(dims_[0], other.dims_[1]);
 
            for (int i = 0; i < dims_[0]; i++) {
                for (int j = 0; j < other.dims_[1]; j++) {
                    for (int k = 0; k < dims_[1]; k++) {
                        tmp(i, j) += (*this)(i, k) * other(k, j);
                    }
                }
            }
 
            return tmp;
        }
 
        void swap(Tensor& other)
        {
            dims_.swap(other.dims_);
            data_.swap(other.data_);
        }
 
        std::vector<int> dims_;
        std::vector<T> data_;
    };
 
    // NN layer
    // ---------------------
    template <class Evaluation>
    class NNLayer
    {
    public:
        NNLayer()
        {
        }
 
        virtual ~NNLayer()
        {
        }
 
        // Loads the ML trained file, returns true if the file exists
        virtual bool loadLayer(std::ifstream& file) = 0;
        // Apply the NN layers
        virtual bool apply(const Tensor<Evaluation>& in, Tensor<Evaluation>& out) = 0;
    };

    template <class Evaluation>
    class NNLayerActivation : public NNLayer<Evaluation>
    {
    public:
        enum class ActivationType {
            kLinear = 1,
            kRelu = 2,
            kSoftPlus = 3,
            kSigmoid = 4,
            kTanh = 5,
            kHardSigmoid = 6
        };
 
        NNLayerActivation()
            : activation_type_(ActivationType::kLinear)
        {
        }
 
        bool loadLayer(std::ifstream& file) override;
 
        bool apply(const Tensor<Evaluation>& in, Tensor<Evaluation>& out) override;
 
    private:
        ActivationType activation_type_;
    };

    template <class Evaluation>
    class NNLayerScaling : public NNLayer<Evaluation>
    {
    public:
        NNLayerScaling()
            : data_min(1.0f)
            , data_max(1.0f)
            , feat_inf(1.0f)
            , feat_sup(1.0f)
        {
        }
 
        bool loadLayer(std::ifstream& file) override;
 
        bool apply(const Tensor<Evaluation>& in, Tensor<Evaluation>& out) override;
 
    private:
        Tensor<float> weights_;
        Tensor<float> biases_;
        float data_min;
        float data_max;
        float feat_inf;
        float feat_sup;
    };

    template <class Evaluation>
    class NNLayerUnScaling : public NNLayer<Evaluation>
    {
    public:
        NNLayerUnScaling()
            : data_min(1.0f)
            , data_max(1.0f)
            , feat_inf(1.0f)
            , feat_sup(1.0f)
        {
        }
 
        bool loadLayer(std::ifstream& file) override;
 
        bool apply(const Tensor<Evaluation>& in, Tensor<Evaluation>& out) override;
 
    private:
        Tensor<float> weights_;
        Tensor<float> biases_;
        float data_min;
        float data_max;
        float feat_inf;
        float feat_sup;
    };

    template <class Evaluation>
    class NNLayerDense : public NNLayer<Evaluation>
    {
    public:
        bool loadLayer(std::ifstream& file) override;
 
        bool apply(const Tensor<Evaluation>& in, Tensor<Evaluation>& out) override;
 
    private:
        Tensor<float> weights_;
        Tensor<float> biases_;
 
        NNLayerActivation<Evaluation> activation_;
    };

    template <class Evaluation>
    class NNLayerEmbedding : public NNLayer<Evaluation>
    {
    public:
        bool loadLayer(std::ifstream& file) override;
 
        bool apply(const Tensor<Evaluation>& in, Tensor<Evaluation>& out) override;
 
    private:
        Tensor<float> weights_;
    };

    template <class Evaluation>
    class NNModel
    {
    public:
        enum class LayerType { kScaling = 1, kUnScaling = 2, kDense = 3, kActivation = 4 };
 
        virtual ~NNModel() = default;
 
        // loads models (.model files) generated by Kerasify
        virtual bool loadModel(const std::string& filename);
 
        virtual bool apply(const Tensor<Evaluation>& in, Tensor<Evaluation>& out);
 
    private:
        std::vector<std::unique_ptr<NNLayer<Evaluation>>> layers_;
    };

    class NNTimer
    {
    public:
        void start()
        {
            start_ = std::chrono::high_resolution_clock::now();
        }
 
        float stop()
        {
            std::chrono::time_point<std::chrono::high_resolution_clock> now
                = std::chrono::high_resolution_clock::now();
 
            std::chrono::duration<double> diff = now - start_;
 
            return diff.count();
        }
 
    private:
        std::chrono::time_point<std::chrono::high_resolution_clock> start_;
    };
 
} // namespace ML
 
} // namespace Opm
 
#endif // ML_MODEL_H_