SVMClassifier/include/svm_classifier/multiclass_strategy.hpp

#pragma once

#include "types.hpp"
#include "kernel_parameters.hpp"
#include "data_converter.hpp"
#include <torch/torch.h>
#include <vector>
#include <memory>
#include <unordered_map>

// Forward declarations for external library structures
struct svm_model;
struct model;

namespace svm_classifier {

    /**
     * @brief Abstract base class for multiclass classification strategies
     */
    class MulticlassStrategyBase {
    public:
        /**
         * @brief Virtual destructor
         */
        virtual ~MulticlassStrategyBase() = default;

        /**
         * @brief Train the multiclass classifier
         * @param X Feature tensor of shape (n_samples, n_features)
         * @param y Target tensor of shape (n_samples,)
         * @param params Kernel parameters
         * @param converter Data converter instance
         * @return Training metrics
         */
        virtual TrainingMetrics fit(const torch::Tensor& X,
            const torch::Tensor& y,
            const KernelParameters& params,
            DataConverter& converter) = 0;

        /**
         * @brief Predict class labels
         * @param X Feature tensor of shape (n_samples, n_features)
         * @param converter Data converter instance
         * @return Predicted class labels
         */
        virtual std::vector<int> predict(const torch::Tensor& X,
            DataConverter& converter) = 0;

        /**
         * @brief Predict class probabilities
         * @param X Feature tensor of shape (n_samples, n_features)
         * @param converter Data converter instance
         * @return Class probabilities for each sample
         */
        virtual std::vector<std::vector<double>> predict_proba(const torch::Tensor& X,
            DataConverter& converter) = 0;

        /**
         * @brief Get decision function values
         * @param X Feature tensor of shape (n_samples, n_features)
         * @param converter Data converter instance
         * @return Decision function values
         */
        virtual std::vector<std::vector<double>> decision_function(const torch::Tensor& X,
            DataConverter& converter) = 0;

        /**
         * @brief Get unique class labels
         * @return Vector of unique class labels
         */
        virtual std::vector<int> get_classes() const = 0;

        /**
         * @brief Check if the model supports probability prediction
         * @return True if probabilities are supported
         */
        virtual bool supports_probability() const = 0;

        /**
         * @brief Get number of classes
         * @return Number of classes
         */
        virtual int get_n_classes() const = 0;

        /**
         * @brief Get strategy type
         * @return Multiclass strategy type
         */
        virtual MulticlassStrategy get_strategy_type() const = 0;

    protected:
        std::vector<int> classes_;  ///< Unique class labels
        bool is_trained_ = false;   ///< Whether the model is trained
    };

    /**
     * @brief One-vs-Rest (OvR) multiclass strategy
     */
    class OneVsRestStrategy : public MulticlassStrategyBase {
    public:
        /**
         * @brief Constructor
         */
        OneVsRestStrategy();

        /**
         * @brief Destructor
         */
        ~OneVsRestStrategy() override;

        TrainingMetrics fit(const torch::Tensor& X,
            const torch::Tensor& y,
            const KernelParameters& params,
            DataConverter& converter) override;

        std::vector<int> predict(const torch::Tensor& X,
            DataConverter& converter) override;

        std::vector<std::vector<double>> predict_proba(const torch::Tensor& X,
            DataConverter& converter) override;

        std::vector<std::vector<double>> decision_function(const torch::Tensor& X,
            DataConverter& converter) override;

        std::vector<int> get_classes() const override { return classes_; }

        bool supports_probability() const override;

        int get_n_classes() const override { return static_cast<int>(classes_.size()); }

        MulticlassStrategy get_strategy_type() const override { return MulticlassStrategy::ONE_VS_REST; }

    private:
        std::vector<std::unique_ptr<svm_model>> svm_models_;      ///< SVM models (one per class)
        std::vector<std::unique_ptr<model>> linear_models_;       ///< Linear models (one per class)
        KernelParameters params_;                                 ///< Stored parameters
        SVMLibrary library_type_;                                ///< Which library is being used

        /**
         * @brief Create binary labels for one-vs-rest
         * @param y Original labels
         * @param positive_class Positive class label
         * @return Binary labels (+1 for positive class, -1 for others)
         */
        torch::Tensor create_binary_labels(const torch::Tensor& y, int positive_class);

        /**
         * @brief Train a single binary classifier
         * @param X Feature tensor
         * @param y_binary Binary labels
         * @param params Kernel parameters
         * @param converter Data converter
         * @param class_idx Index of the class being trained
         * @return Training time for this classifier
         */
        double train_binary_classifier(const torch::Tensor& X,
            const torch::Tensor& y_binary,
            const KernelParameters& params,
            DataConverter& converter,
            int class_idx);

        /**
         * @brief Clean up all models
         */
        void cleanup_models();
    };

    /**
     * @brief One-vs-One (OvO) multiclass strategy
     */
    class OneVsOneStrategy : public MulticlassStrategyBase {
    public:
        /**
         * @brief Constructor
         */
        OneVsOneStrategy();

        /**
         * @brief Destructor
         */
        ~OneVsOneStrategy() override;

        TrainingMetrics fit(const torch::Tensor& X,
            const torch::Tensor& y,
            const KernelParameters& params,
            DataConverter& converter) override;

        std::vector<int> predict(const torch::Tensor& X,
            DataConverter& converter) override;

        std::vector<std::vector<double>> predict_proba(const torch::Tensor& X,
            DataConverter& converter) override;

        std::vector<std::vector<double>> decision_function(const torch::Tensor& X,
            DataConverter& converter) override;

        std::vector<int> get_classes() const override { return classes_; }

        bool supports_probability() const override;

        int get_n_classes() const override { return static_cast<int>(classes_.size()); }

        MulticlassStrategy get_strategy_type() const override { return MulticlassStrategy::ONE_VS_ONE; }

    private:
        std::vector<std::unique_ptr<svm_model>> svm_models_;      ///< SVM models (one per pair)
        std::vector<std::unique_ptr<model>> linear_models_;       ///< Linear models (one per pair)
        std::vector<std::pair<int, int>> class_pairs_;           ///< Class pairs for each model
        KernelParameters params_;                                 ///< Stored parameters
        SVMLibrary library_type_;                                ///< Which library is being used

        /**
         * @brief Extract samples for a specific class pair
         * @param X Feature tensor
         * @param y Label tensor
         * @param class1 First class
         * @param class2 Second class
         * @return Pair of (filtered_X, filtered_y)
         */
        std::pair<torch::Tensor, torch::Tensor> extract_binary_data(const torch::Tensor& X,
            const torch::Tensor& y,
            int class1,
            int class2);

        /**
         * @brief Train a single pairwise classifier
         * @param X Feature tensor
         * @param y Labels
         * @param class1 First class
         * @param class2 Second class
         * @param params Kernel parameters
         * @param converter Data converter
         * @param model_idx Index of the model being trained
         * @return Training time for this classifier
         */
        double train_pairwise_classifier(const torch::Tensor& X,
            const torch::Tensor& y,
            int class1,
            int class2,
            const KernelParameters& params,
            DataConverter& converter,
            int model_idx);

        /**
         * @brief Voting mechanism for OvO predictions
         * @param decisions Matrix of pairwise decisions
         * @return Predicted class for each sample
         */
        std::vector<int> vote_predictions(const std::vector<std::vector<double>>& decisions);

        /**
         * @brief Clean up all models
         */
        void cleanup_models();
    };

    /**
     * @brief Factory function to create multiclass strategy
     * @param strategy Strategy type
     * @return Unique pointer to multiclass strategy
     */
    std::unique_ptr<MulticlassStrategyBase> create_multiclass_strategy(MulticlassStrategy strategy);

} // namespace svm_classifier