rumale 0.23.3 → 0.24.0

data/lib/rumale/linear_model/linear_regression.rb

@@ -1,201 +0,0 @@
-# frozen_string_literal: true
-
-require 'lbfgsb'
-
-require 'rumale/linear_model/base_sgd'
-require 'rumale/base/regressor'
-
-module Rumale
-  module LinearModel
-    # LinearRegression is a class that implements ordinary least square linear regression
-    # with stochastic gradient descent (SGD) optimization,
-    # singular value decomposition (SVD), or L-BFGS optimization.
-    #
-    # @example
-    #   estimator =
-    #     Rumale::LinearModel::LinearRegression.new(max_iter: 1000, batch_size: 20, random_seed: 1)
-    #   estimator.fit(training_samples, traininig_values)
-    #   results = estimator.predict(testing_samples)
-    #
-    #   # If Numo::Linalg is installed, you can specify 'svd' for the solver option.
-    #   require 'numo/linalg/autoloader'
-    #   estimator = Rumale::LinearModel::LinearRegression.new(solver: 'svd')
-    #   estimator.fit(training_samples, traininig_values)
-    #   results = estimator.predict(testing_samples)
-    #
-    # *Reference*
-    # - Bottou, L., "Large-Scale Machine Learning with Stochastic Gradient Descent," Proc. COMPSTAT'10, pp. 177--186, 2010.
-    class LinearRegression < BaseSGD
-      include Base::Regressor
-
-      # Return the weight vector.
-      # @return [Numo::DFloat] (shape: [n_outputs, n_features])
-      attr_reader :weight_vec
-
-      # Return the bias term (a.k.a. intercept).
-      # @return [Numo::DFloat] (shape: [n_outputs])
-      attr_reader :bias_term
-
-      # Return the random generator for random sampling.
-      # @return [Random]
-      attr_reader :rng
-
-      # Create a new ordinary least square linear regressor.
-      #
-      # @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 is not 'sgd', 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 'learning_rate'.
-      #   If solver is not 'sgd', this parameter is ignored.
-      # @param momentum [Float] The momentum factor.
-      #   If solver is not 'sgd', this parameter is ignored.
-      # @param fit_bias [Boolean] The flag indicating whether to fit the bias term.
-      # @param bias_scale [Float] The scale of the bias term.
-      # @param max_iter [Integer] The maximum number of epochs that indicates
-      #   how many times the whole data is given to the training process.
-      #   If solver is 'svd', this parameter is ignored.
-      # @param batch_size [Integer] The size of the mini batches.
-      #   If solver is not 'sgd', this parameter is ignored.
-      # @param tol [Float] The tolerance of loss for terminating optimization.
-      #   If solver is 'svd', this parameter is ignored.
-      # @param solver [String] The algorithm to calculate weights. ('auto', 'sgd', 'svd' or 'lbfgs').
-      #   'auto' chooses the 'svd' solver if Numo::Linalg is loaded. Otherwise, it chooses the 'lbfgs' solver.
-      #   'sgd' uses the stochastic gradient descent optimization.
-      #   'svd' performs singular value decomposition of samples.
-      #   'lbfgs' uses the L-BFGS method for optimization.
-      # @param n_jobs [Integer] The number of jobs for running the fit method in parallel.
-      #   If nil is given, the method does 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 solver is not 'sgd'.
-      # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
-      #   If solver is 'svd', this parameter is ignored.
-      # @param random_seed [Integer] The seed value using to initialize the random generator.
-      def initialize(learning_rate: 0.01, decay: nil, momentum: 0.9,
-                     fit_bias: true, bias_scale: 1.0, max_iter: 1000, batch_size: 50, tol: 1e-4,
-                     solver: 'auto',
-                     n_jobs: nil, verbose: false, random_seed: nil)
-        check_params_numeric(learning_rate: learning_rate, momentum: momentum,
-                             bias_scale: bias_scale, max_iter: max_iter, batch_size: batch_size)
-        check_params_boolean(fit_bias: fit_bias, verbose: verbose)
-        check_params_string(solver: solver)
-        check_params_numeric_or_nil(decay: decay, n_jobs: n_jobs, random_seed: random_seed)
-        check_params_positive(learning_rate: learning_rate, max_iter: max_iter, batch_size: batch_size)
-        super()
-        @params.merge!(method(:initialize).parameters.map { |_t, arg| [arg, binding.local_variable_get(arg)] }.to_h)
-        @params[:solver] = if solver == 'auto'
-                             enable_linalg?(warning: false) ? 'svd' : 'lbfgs'
-                           else
-                             solver.match?(/^svd$|^sgd$|^lbfgs$/) ? solver : 'lbfgs'
-                           end
-        @params[:decay] ||= @params[:learning_rate]
-        @params[:random_seed] ||= srand
-        @rng = Random.new(@params[:random_seed])
-        @loss_func = LinearModel::Loss::MeanSquaredError.new
-        @weight_vec = nil
-        @bias_term = 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::DFloat] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
-      # @return [LinearRegression] The learned regressor itself.
-      def fit(x, y)
-        x = check_convert_sample_array(x)
-        y = check_convert_tvalue_array(y)
-        check_sample_tvalue_size(x, y)
-
-        if @params[:solver] == 'svd' && enable_linalg?(warning: false)
-          fit_svd(x, y)
-        elsif @params[:solver] == 'lbfgs'
-          fit_lbfgs(x, y)
-        else
-          fit_sgd(x, y)
-        end
-
-        self
-      end
-
-      # Predict values for samples.
-      #
-      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the values.
-      # @return [Numo::DFloat] (shape: [n_samples, n_outputs]) Predicted values per sample.
-      def predict(x)
-        x = check_convert_sample_array(x)
-        x.dot(@weight_vec.transpose) + @bias_term
-      end
-
-      private
-
-      def fit_svd(x, y)
-        x = expand_feature(x) if fit_bias?
-        w = Numo::Linalg.pinv(x, driver: 'svd').dot(y)
-        @weight_vec, @bias_term = single_target?(y) ? split_weight(w) : split_weight_mult(w)
-      end
-
-      def fit_lbfgs(x, y)
-        fnc = proc do |w, x, y| # rubocop:disable Lint/ShadowingOuterLocalVariable
-          n_samples, n_features = x.shape
-          w = w.reshape(y.shape[1], n_features) unless y.shape[1].nil?
-          z = x.dot(w.transpose)
-          d = z - y
-          loss = (d**2).sum.fdiv(n_samples)
-          gradient = 2.fdiv(n_samples) * d.transpose.dot(x)
-          [loss, gradient.flatten.dup]
-        end
-
-        x = expand_feature(x) if fit_bias?
-
-        n_features = x.shape[1]
-        n_outputs = single_target?(y) ? 1 : y.shape[1]
-
-        res = Lbfgsb.minimize(
-          fnc: fnc, jcb: true, x_init: init_weight(n_features, n_outputs), args: [x, y],
-          maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON,
-          verbose: @params[:verbose] ? 1 : -1
-        )
-
-        @weight_vec, @bias_term =
-          if single_target?(y)
-            split_weight(res[:x])
-          else
-            split_weight_mult(res[:x].reshape(n_outputs, n_features).transpose)
-          end
-      end
-
-      def fit_sgd(x, y)
-        if single_target?(y)
-          @weight_vec, @bias_term = partial_fit(x, y)
-        else
-          n_outputs = y.shape[1]
-          n_features = x.shape[1]
-          @weight_vec = Numo::DFloat.zeros(n_outputs, n_features)
-          @bias_term = Numo::DFloat.zeros(n_outputs)
-          if enable_parallel?
-            models = parallel_map(n_outputs) { |n| partial_fit(x, y[true, n]) }
-            n_outputs.times { |n| @weight_vec[n, true], @bias_term[n] = models[n] }
-          else
-            n_outputs.times { |n| @weight_vec[n, true], @bias_term[n] = partial_fit(x, y[true, n]) }
-          end
-        end
-      end
-
-      def single_target?(y)
-        y.ndim == 1
-      end
-
-      def init_weight(n_features, n_outputs)
-        Rumale::Utils.rand_normal([n_outputs, n_features], @rng.dup).flatten.dup
-      end
-
-      def split_weight_mult(w)
-        if fit_bias?
-          [w[0...-1, true].dup, w[-1, true].dup]
-        else
-          [w.dup, Numo::DFloat.zeros(w.shape[1])]
-        end
-      end
-    end
-  end
-end

data/lib/rumale/linear_model/logistic_regression.rb

@@ -1,275 +0,0 @@
-# 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) # rubocop:disable Metrics/AbcSize, Metrics/MethodLength
-        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

data/lib/rumale/linear_model/nnls.rb

@@ -1,137 +0,0 @@
-# frozen_string_literal: true
-
-require 'lbfgsb'
-
-require 'rumale/base/base_estimator'
-require 'rumale/base/regressor'
-
-module Rumale
-  module LinearModel
-    # NNLS is a class that implements non-negative least squares regression.
-    # NNLS solves least squares problem under non-negative constraints on the coefficient using L-BFGS-B method.
-    #
-    # @example
-    #   estimator = Rumale::LinearModel::NNLS.new(reg_param: 0.01, random_seed: 1)
-    #   estimator.fit(training_samples, traininig_values)
-    #   results = estimator.predict(testing_samples)
-    #
-    class NNLS
-      include Base::BaseEstimator
-      include Base::Regressor
-
-      # Return the weight vector.
-      # @return [Numo::DFloat] (shape: [n_outputs, n_features])
-      attr_reader :weight_vec
-
-      # Return the bias term (a.k.a. intercept).
-      # @return [Numo::DFloat] (shape: [n_outputs])
-      attr_reader :bias_term
-
-      # Returns the number of iterations when converged.
-      # @return [Integer]
-      attr_reader :n_iter
-
-      # Return the random generator for initializing weight.
-      # @return [Random]
-      attr_reader :rng
-
-      # Create a new regressor with non-negative least squares method.
-      #
-      # @param reg_param [Float] The regularization parameter for L2 regularization term.
-      # @param fit_bias [Boolean] The flag indicating whether to fit the bias term.
-      # @param bias_scale [Float] The scale of the bias term.
-      # @param max_iter [Integer] The maximum number of epochs that indicates
-      #   how many times the whole data is given to the training process.
-      # @param tol [Float] The tolerance of loss for terminating optimization.
-      #   If solver = 'svd', this parameter is ignored.
-      # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
-      # @param random_seed [Integer] The seed value using to initialize the random generator.
-      def initialize(reg_param: 1.0, fit_bias: true, bias_scale: 1.0,
-                     max_iter: 1000, tol: 1e-4, verbose: false, random_seed: nil)
-        check_params_numeric(reg_param: reg_param, bias_scale: bias_scale, max_iter: max_iter, tol: tol)
-        check_params_boolean(fit_bias: fit_bias, verbose: verbose)
-        check_params_numeric_or_nil(random_seed: random_seed)
-        check_params_positive(reg_param: reg_param, max_iter: max_iter)
-        @params = method(:initialize).parameters.each_with_object({}) { |(_, prm), obj| obj[prm] = binding.local_variable_get(prm) }
-        @params[:random_seed] ||= srand
-        @n_iter = nil
-        @weight_vec = nil
-        @bias_term = nil
-        @rng = Random.new(@params[:random_seed])
-      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::DFloat] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
-      # @return [NonneagtiveLeastSquare] The learned regressor itself.
-      def fit(x, y)
-        x = check_convert_sample_array(x)
-        y = check_convert_tvalue_array(y)
-        check_sample_tvalue_size(x, y)
-
-        x = expand_feature(x) if fit_bias?
-
-        n_features = x.shape[1]
-        n_outputs = single_target?(y) ? 1 : y.shape[1]
-
-        w_init = Rumale::Utils.rand_normal([n_outputs, n_features], @rng.dup).flatten.dup
-        w_init[w_init.lt(0)] = 0
-        bounds = Numo::DFloat.zeros(n_outputs * n_features, 2)
-        bounds.shape[0].times { |n| bounds[n, 1] = Float::INFINITY }
-
-        res = Lbfgsb.minimize(
-          fnc: method(:nnls_fnc), jcb: true, x_init: w_init, args: [x, y, @params[:reg_param]], bounds: bounds,
-          maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON, verbose: @params[:verbose] ? 1 : -1
-        )
-
-        @n_iter = res[:n_iter]
-        w = single_target?(y) ? res[:x] : res[:x].reshape(n_outputs, n_features).transpose
-
-        if fit_bias?
-          @weight_vec = single_target?(y) ? w[0...-1].dup : w[0...-1, true].dup
-          @bias_term = single_target?(y) ? w[-1] : w[-1, true].dup
-        else
-          @weight_vec = w.dup
-          @bias_term = single_target?(y) ? 0 : Numo::DFloat.zeros(y.shape[1])
-        end
-
-        self
-      end
-
-      # Predict values for samples.
-      #
-      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the values.
-      # @return [Numo::DFloat] (shape: [n_samples, n_outputs]) Predicted values per sample.
-      def predict(x)
-        x = check_convert_sample_array(x)
-        x.dot(@weight_vec.transpose) + @bias_term
-      end
-
-      private
-
-      def nnls_fnc(w, x, y, alpha)
-        n_samples, n_features = x.shape
-        w = w.reshape(y.shape[1], n_features) unless y.shape[1].nil?
-        z = x.dot(w.transpose)
-        d = z - y
-        loss = (d**2).sum.fdiv(n_samples) + alpha * (w * w).sum
-        gradient = 2.fdiv(n_samples) * d.transpose.dot(x) + 2.0 * alpha * w
-        [loss, gradient.flatten.dup]
-      end
-
-      def expand_feature(x)
-        n_samples = x.shape[0]
-        Numo::NArray.hstack([x, Numo::DFloat.ones([n_samples, 1]) * @params[:bias_scale]])
-      end
-
-      def fit_bias?
-        @params[:fit_bias] == true
-      end
-
-      def single_target?(y)
-        y.ndim == 1
-      end
-    end
-  end
-end