# frozen_string_literal: true

require 'lbfgsb'
require 'rumale/base/classifier'
require 'rumale/linear_model/base_sgd'
require 'rumale/preprocessing/label_binarizer'

module Rumale
  module LinearModel
    # LogisticRegression is a class that implements Logistic Regression.
    # In multiclass classification problem, it uses one-vs-the-rest strategy for the sgd solver
    # and multinomial logistic regression for the lbfgs solver.
    #
    # @note
    #   Rumale::SVM provides Logistic Regression based on LIBLINEAR.
    #   If you prefer execution speed, you should use Rumale::SVM::LogisticRegression.
    #   https://github.com/yoshoku/rumale-svm
    #
    # @example
    #   estimator =
    #     Rumale::LinearModel::LogisticRegression.new(reg_param: 1.0, random_seed: 1)
    #   estimator.fit(training_samples, traininig_labels)
    #   results = estimator.predict(testing_samples)
    #
    # *Reference*
    # - Shalev-Shwartz, S., Singer, Y., Srebro, N., and Cotter, A., "Pegasos: Primal Estimated sub-GrAdient SOlver for SVM," Mathematical Programming, vol. 127 (1), pp. 3--30, 2011.
    # - Tsuruoka, Y., Tsujii, J., and Ananiadou, S., "Stochastic Gradient Descent Training for L1-regularized Log-linear Models with Cumulative Penalty," Proc. ACL'09, pp. 477--485, 2009.
    # - Bottou, L., "Large-Scale Machine Learning with Stochastic Gradient Descent," Proc. COMPSTAT'10, pp. 177--186, 2010.
    class LogisticRegression < BaseSGD
      include Base::Classifier

      # Return the weight vector for Logistic Regression.
      # @return [Numo::DFloat] (shape: [n_classes, n_features])
      attr_reader :weight_vec

      # Return the bias term (a.k.a. intercept) for Logistic Regression.
      # @return [Numo::DFloat] (shape: [n_classes])
      attr_reader :bias_term

      # Return the class labels.
      # @return [Numo::Int32] (shape: [n_classes])
      attr_reader :classes

      # Return the random generator for performing random sampling.
      # @return [Random]
      attr_reader :rng

      # Create a new classifier with Logisitc Regression.
      #
      # @param learning_rate [Float] The initial value of learning rate.
      #   The learning rate decreases as the iteration proceeds according to the equation: learning_rate / (1 + decay * t).
      #   If solver = 'lbfgs', this parameter is ignored.
      # @param decay [Float] The smoothing parameter for decreasing learning rate as the iteration proceeds.
      #   If nil is given, the decay sets to 'reg_param * learning_rate'.
      #   If solver = 'lbfgs', this parameter is ignored.
      # @param momentum [Float] The momentum factor.
      #   If solver = 'lbfgs', this parameter is ignored.
      # @param penalty [String] The regularization type to be used ('l1', 'l2', and 'elasticnet').
      #   If solver = 'lbfgs', only 'l2' can be selected for this parameter.
      # @param l1_ratio [Float] The elastic-net type regularization mixing parameter.
      #   If penalty set to 'l2' or 'l1', this parameter is ignored.
      #   If l1_ratio = 1, the regularization is similar to Lasso.
      #   If l1_ratio = 0, the regularization is similar to Ridge.
      #   If 0 < l1_ratio < 1, the regularization is a combination of L1 and L2.
      #   If solver = 'lbfgs', this parameter is ignored.
      # @param reg_param [Float] The regularization parameter.
      # @param fit_bias [Boolean] The flag indicating whether to fit the bias term.
      # @param bias_scale [Float] The scale of the bias term.
      #   If fit_bias is true, the feature vector v becoms [v; bias_scale].
      # @param max_iter [Integer] The maximum number of epochs that indicates
      #   how many times the whole data is given to the training process.
      # @param batch_size [Integer] The size of the mini batches.
      #   If solver = 'lbfgs', this parameter is ignored.
      # @param tol [Float] The tolerance of loss for terminating optimization.
      #   If solver = 'lbfgs', this value is given as tol / Lbfgsb::DBL_EPSILON to the factr argument of Lbfgsb.minimize method.
      # @param solver [String] The algorithm for optimization. ('lbfgs' or 'sgd').
      #   'lbfgs' uses the L-BFGS with lbfgs.rb gem.
      #   'sgd' uses the stochastic gradient descent optimization.
      # @param n_jobs [Integer] The number of jobs for running the fit and predict methods in parallel.
      #   If nil is given, the methods do not execute in parallel.
      #   If zero or less is given, it becomes equal to the number of processors.
      #   This parameter is ignored if the Parallel gem is not loaded or the solver is 'lbfgs'.
      # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
      #   If solver = 'lbfgs' and true is given, 'iterate.dat' file is generated by lbfgsb.rb.
      # @param random_seed [Integer] The seed value using to initialize the random generator.
      def initialize(learning_rate: 0.01, decay: nil, momentum: 0.9,
                     penalty: 'l2', reg_param: 1.0, l1_ratio: 0.5,
                     fit_bias: true, bias_scale: 1.0,
                     max_iter: 1000, batch_size: 50, tol: 1e-4,
                     solver: 'lbfgs',
                     n_jobs: nil, verbose: false, random_seed: nil)
        check_params_numeric(learning_rate: learning_rate, momentum: momentum,
                             reg_param: reg_param, l1_ratio: l1_ratio, bias_scale: bias_scale,
                             max_iter: max_iter, batch_size: batch_size, tol: tol)
        check_params_boolean(fit_bias: fit_bias, verbose: verbose)
        check_params_string(solver: solver, penalty: penalty)
        check_params_numeric_or_nil(decay: decay, n_jobs: n_jobs, random_seed: random_seed)
        check_params_positive(learning_rate: learning_rate, reg_param: reg_param,
                              bias_scale: bias_scale, max_iter: max_iter, batch_size: batch_size)
        raise ArgumentError, "The 'lbfgs' solver supports only 'l2' penalties." if solver == 'lbfgs' && penalty != 'l2'

        super()
        @params.merge!(method(:initialize).parameters.map { |_t, arg| [arg, binding.local_variable_get(arg)] }.to_h)
        @params[:solver] = solver == 'sgd' ? 'sgd' : 'lbfgs'
        @params[:decay] ||= @params[:reg_param] * @params[:learning_rate]
        @params[:random_seed] ||= srand
        @rng = Random.new(@params[:random_seed])
        @penalty_type = @params[:penalty]
        @loss_func = LinearModel::Loss::LogLoss.new
        @weight_vec = nil
        @bias_term = nil
        @classes = nil
      end

      # Fit the model with given training data.
      #
      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The training data to be used for fitting the model.
      # @param y [Numo::Int32] (shape: [n_samples]) The labels to be used for fitting the model.
      # @return [LogisticRegression] The learned classifier itself.
      def fit(x, y)
        x = check_convert_sample_array(x)
        y = check_convert_label_array(y)
        check_sample_label_size(x, y)

        @classes = Numo::Int32[*y.to_a.uniq.sort]
        if @params[:solver] == 'sgd'
          fit_sgd(x, y)
        else
          fit_lbfgs(x, y)
        end

        self
      end

      # Calculate confidence scores for samples.
      #
      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to compute the scores.
      # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Confidence score per sample.
      def decision_function(x)
        x = check_convert_sample_array(x)
        x.dot(@weight_vec.transpose) + @bias_term
      end

      # Predict class labels for samples.
      #
      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the labels.
      # @return [Numo::Int32] (shape: [n_samples]) Predicted class label per sample.
      def predict(x)
        x = check_convert_sample_array(x)

        n_samples, = x.shape
        decision_values = predict_proba(x)
        predicted = if enable_parallel?
                      parallel_map(n_samples) { |n| @classes[decision_values[n, true].max_index] }
                    else
                      Array.new(n_samples) { |n| @classes[decision_values[n, true].max_index] }
                    end
        Numo::Int32.asarray(predicted)
      end

      # Predict probability for samples.
      #
      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the probailities.
      # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Predicted probability of each class per sample.
      def predict_proba(x)
        x = check_convert_sample_array(x)

        proba = 1.0 / (Numo::NMath.exp(-decision_function(x)) + 1.0)
        return (proba.transpose / proba.sum(axis: 1)).transpose.dup if multiclass_problem?

        n_samples, = x.shape
        probs = Numo::DFloat.zeros(n_samples, 2)
        probs[true, 1] = proba
        probs[true, 0] = 1.0 - proba
        probs
      end

      private

      def multiclass_problem?
        @classes.size > 2
      end

      def fit_lbfgs(base_x, base_y)
        if multiclass_problem?
          fnc = proc do |w, x, y, a|
            n_features = x.shape[1]
            n_classes = y.shape[1]
            z = x.dot(w.reshape(n_classes, n_features).transpose)
            # logsumexp and softmax
            z_max = z.max(-1).expand_dims(-1).dup
            z_max[~z_max.isfinite] = 0.0
            lgsexp = Numo::NMath.log(Numo::NMath.exp(z - z_max).sum(-1)).expand_dims(-1) + z_max
            t = z - lgsexp
            sftmax = Numo::NMath.exp(t)
            # loss and gradient
            loss = -(y * t).sum + 0.5 * a * w.dot(w)
            grad = (sftmax - y).transpose.dot(x).flatten.dup + a * w
            [loss, grad]
          end

          base_x = expand_feature(base_x) if fit_bias?
          encoder = Rumale::Preprocessing::LabelBinarizer.new
          onehot_y = encoder.fit_transform(base_y)
          n_classes = @classes.size
          n_features = base_x.shape[1]
          w_init = Numo::DFloat.zeros(n_classes * n_features)

          verbose = @params[:verbose] ? 1 : -1
          res = Lbfgsb.minimize(
            fnc: fnc, jcb: true, x_init: w_init, args: [base_x, onehot_y, @params[:reg_param]],
            maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON, verbose: verbose
          )

          if fit_bias?
            weight = res[:x].reshape(n_classes, n_features)
            @weight_vec = weight[true, 0...-1].dup
            @bias_term = weight[true, -1].dup
          else
            @weight_vec = res[:x].reshape(n_classes, n_features)
            @bias_term = Numo::DFloat.zeros(n_classes)
          end
        else
          fnc = proc do |w, x, y, a|
            z = 1 + Numo::NMath.exp(-y * x.dot(w))
            loss = Numo::NMath.log(z).sum + 0.5 * a * w.dot(w)
            grad = (y / z - y).dot(x) + a * w
            [loss, grad]
          end

          base_x = expand_feature(base_x) if fit_bias?
          negative_label = @classes[0]
          bin_y = Numo::Int32.cast(base_y.ne(negative_label)) * 2 - 1
          n_features = base_x.shape[1]
          w_init = Numo::DFloat.zeros(n_features)

          verbose = @params[:verbose] ? 1 : -1
          res = Lbfgsb.minimize(
            fnc: fnc, jcb: true, x_init: w_init, args: [base_x, bin_y, @params[:reg_param]],
            maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON, verbose: verbose
          )

          @weight_vec, @bias_term = split_weight(res[:x])
        end
      end

      def fit_sgd(x, y)
        if multiclass_problem?
          n_classes = @classes.size
          n_features = x.shape[1]
          @weight_vec = Numo::DFloat.zeros(n_classes, n_features)
          @bias_term = Numo::DFloat.zeros(n_classes)
          if enable_parallel?
            # :nocov:
            models = parallel_map(n_classes) do |n|
              bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
              partial_fit(x, bin_y)
            end
            # :nocov:
            n_classes.times { |n| @weight_vec[n, true], @bias_term[n] = models[n] }
          else
            n_classes.times do |n|
              bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
              @weight_vec[n, true], @bias_term[n] = partial_fit(x, bin_y)
            end
          end
        else
          negative_label = @classes[0]
          bin_y = Numo::Int32.cast(y.ne(negative_label)) * 2 - 1
          @weight_vec, @bias_term = partial_fit(x, bin_y)
        end
      end
    end
  end
end