rumale-linear_model 0.24.0 → 0.25.0

data/lib/rumale/linear_model/sgd_regressor.rb

@@ -0,0 +1,138 @@
+# frozen_string_literal: true
+
+require 'rumale/base/regressor'
+require 'rumale/validation'
+
+require_relative 'sgd_estimator'
+
+module Rumale
+  module LinearModel
+    # SGDRegressor is a class that implements linear regressor with stochastic gradient descent optimization.
+    #
+    # @example
+    #   require 'rumale/linear_model/sgd_regressor'
+    #
+    #   estimator =
+    #     Rumale::LinearModel::SGDRegressor.new(loss: 'squared_error', reg_param: 1.0, max_iter: 1000, batch_size: 50, random_seed: 1)
+    #   estimator.fit(training_samples, traininig_target_values)
+    #   results = estimator.predict(testing_samples)
+    #
+    # *Reference*
+    # - Shalev-Shwartz, S., and Singer, Y., "Pegasos: Primal Estimated sub-GrAdient SOlver for SVM," Proc. ICML'07, pp. 807--814, 2007.
+    # - 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 SGDRegressor < Rumale::LinearModel::SGDEstimator
+      include Rumale::Base::Regressor
+
+      # Return the random generator for performing random sampling.
+      # @return [Random]
+      attr_reader :rng
+
+      # Create a new linear regressor with stochastic gradient descent optimization.
+      #
+      # @param loss [String] The loss function to be used ('squared_error' and 'epsilon_insensitive').
+      # @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).
+      # @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'.
+      # @param momentum [Float] The momentum factor.
+      # @param penalty [String] The regularization type to be used ('l1', 'l2', and 'elasticnet').
+      # @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.
+      # @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.
+      # @param epsilon [Float] The margin of tolerance. If loss set to 'squared_error', this parameter is ignored.
+      # @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.
+      # @param tol [Float] The tolerance of loss for terminating 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.
+      # @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(loss: 'squared_error', 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,
+                     epsilon: 0.1,
+                     max_iter: 1000, batch_size: 50, tol: 1e-4,
+                     n_jobs: nil, verbose: false, random_seed: nil)
+        super()
+        @params.merge!(
+          loss: loss,
+          learning_rate: learning_rate,
+          decay: decay,
+          momentum: momentum,
+          penalty: penalty,
+          reg_param: reg_param,
+          l1_ratio: l1_ratio,
+          fit_bias: fit_bias,
+          bias_scale: bias_scale,
+          epsilon: epsilon,
+          max_iter: max_iter,
+          batch_size: batch_size,
+          tol: tol,
+          n_jobs: n_jobs,
+          verbose: verbose,
+          random_seed: random_seed
+        )
+        @params[:decay] ||= @params[:reg_param] * @params[:learning_rate]
+        @params[:random_seed] ||= srand
+        @rng = Random.new(@params[:random_seed])
+        @penalty_type = @params[:penalty]
+        @loss_func = case @params[:loss]
+                     when Rumale::LinearModel::Loss::MeanSquaredError::NAME
+                       Rumale::LinearModel::Loss::MeanSquaredError.new
+                     when Rumale::LinearModel::Loss::EpsilonInsensitive::NAME
+                       Rumale::LinearModel::Loss::EpsilonInsensitive.new(epsilon: @params[:epsilon])
+                     else
+                       raise ArgumentError, "given loss '#{loss}' is not supported."
+                     end
+      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.
+      # @retu:rn [SGDRegressor] The learned regressor itself.
+      def fit(x, y)
+        x = Rumale::Validation.check_convert_sample_array(x)
+        y = Rumale::Validation.check_convert_target_value_array(y)
+        Rumale::Validation.check_sample_size(x, y)
+
+        n_outputs = y.shape[1].nil? ? 1 : y.shape[1]
+        n_features = x.shape[1]
+
+        if n_outputs > 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
+        else
+          @weight_vec, @bias_term = partial_fit(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 = Rumale::Validation.check_convert_sample_array(x)
+
+        x.dot(@weight_vec.transpose) + @bias_term
+      end
+    end
+  end
+end

data/lib/rumale/linear_model/svc.rb

@@ -1,15 +1,16 @@
 # frozen_string_literal: true
 
+require 'lbfgsb'
+
 require 'rumale/base/classifier'
-require 'rumale/linear_model/base_sgd'
 require 'rumale/probabilistic_output'
 require 'rumale/validation'
 
+require_relative 'base_estimator'
+
 module Rumale
-  # This module consists of the classes that implement generalized linear models.
   module LinearModel
-    # SVC is a class that implements Support Vector Classifier
-    # with stochastic gradient descent optimization.
+    # SVC is a class that implements Support Vector Classifier with the squared hinge loss.
     # For multiclass classification problem, it uses one-vs-the-rest strategy.
     #
     # @note
@@ -21,52 +22,23 @@ module Rumale
     #   require 'rumale/linear_model/svc'
     #
     #   estimator =
-    #     Rumale::LinearModel::SVC.new(reg_param: 1.0, max_iter: 1000, batch_size: 50, random_seed: 1)
+    #     Rumale::LinearModel::SVC.new(reg_param: 1.0)
     #   estimator.fit(training_samples, traininig_labels)
     #   results = estimator.predict(testing_samples)
-    #
-    # *Reference*
-    # - Shalev-Shwartz, S., and Singer, Y., "Pegasos: Primal Estimated sub-GrAdient SOlver for SVM," Proc. ICML'07, pp. 807--814, 2007.
-    # - 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 SVC < BaseSGD
-      include ::Rumale::Base::Classifier
-
-      # Return the weight vector for SVC.
-      # @return [Numo::DFloat] (shape: [n_classes, n_features])
-      attr_reader :weight_vec
-
-      # Return the bias term (a.k.a. intercept) for SVC.
-      # @return [Numo::DFloat] (shape: [n_classes])
-      attr_reader :bias_term
+    class SVC < Rumale::LinearModel::BaseEstimator
+      include Rumale::Base::Classifier
 
       # 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 Support Vector Machine by the SGD optimization.
+      # Create a new linear classifier with Support Vector Machine with the squared hinge loss.
       #
-      # @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).
-      # @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'.
-      # @param momentum [Float] The momentum factor.
-      # @param penalty [String] The regularization type to be used ('l1', 'l2', and 'elasticnet').
-      # @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.
       # @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.
       # @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.
       # @param tol [Float] The tolerance of loss for terminating optimization.
       # @param probability [Boolean] The flag indicating whether to perform probability estimation.
       # @param n_jobs [Integer] The number of jobs for running the fit and predict methods in parallel.
@@ -74,20 +46,20 @@ module Rumale
       #   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.
       # @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(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,
-                     probability: false,
-                     n_jobs: nil, verbose: false, random_seed: nil)
+      #   'iterate.dat' file is generated by lbfgsb.rb.
+      def initialize(reg_param: 1.0, fit_bias: true, bias_scale: 1.0, max_iter: 1000, tol: 1e-4, probability: false,
+                     n_jobs: nil, verbose: false)
         super()
-        @params.merge!(method(:initialize).parameters.to_h { |_t, arg| [arg, binding.local_variable_get(arg)] })
-        @params[:decay] ||= @params[:reg_param] * @params[:learning_rate]
-        @params[:random_seed] ||= srand
-        @rng = Random.new(@params[:random_seed])
-        @penalty_type = @params[:penalty]
-        @loss_func = ::Rumale::LinearModel::Loss::HingeLoss.new
+        @params = {
+          reg_param: reg_param,
+          fit_bias: fit_bias,
+          bias_scale: bias_scale,
+          max_iter: max_iter,
+          tol: tol,
+          probability: probability,
+          n_jobs: n_jobs,
+          verbose: verbose
+        }
       end
 
       # Fit the model with given training data.
@@ -96,20 +68,20 @@ module Rumale
       # @param y [Numo::Int32] (shape: [n_samples]) The labels to be used for fitting the model.
       # @return [SVC] The learned classifier itself.
       def fit(x, y)
-        x = ::Rumale::Validation.check_convert_sample_array(x)
-        y = ::Rumale::Validation.check_convert_label_array(y)
-        ::Rumale::Validation.check_sample_size(x, y)
+        x = Rumale::Validation.check_convert_sample_array(x)
+        y = Rumale::Validation.check_convert_label_array(y)
+        Rumale::Validation.check_sample_size(x, y)
 
         @classes = Numo::Int32[*y.to_a.uniq.sort]
+        x = expand_feature(x) if fit_bias?
 
         if multiclass_problem?
           n_classes = @classes.size
           n_features = x.shape[1]
-          # initialize model.
+          n_features -= 1 if fit_bias?
           @weight_vec = Numo::DFloat.zeros(n_classes, n_features)
           @bias_term = Numo::DFloat.zeros(n_classes)
           @prob_param = Numo::DFloat.zeros(n_classes, 2)
-          # fit model.
           models = if enable_parallel?
                      parallel_map(n_classes) do |n|
                        bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
@@ -121,7 +93,6 @@ module Rumale
                        partial_fit(x, bin_y)
                      end
                    end
-          # store model.
           models.each_with_index { |model, n| @weight_vec[n, true], @bias_term[n], @prob_param[n, true] = model }
         else
           negative_label = @classes[0]
@@ -137,7 +108,7 @@ module Rumale
       # @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 = ::Rumale::Validation.check_convert_sample_array(x)
+        x = Rumale::Validation.check_convert_sample_array(x)
 
         x.dot(@weight_vec.transpose) + @bias_term
       end
@@ -147,7 +118,7 @@ module Rumale
       # @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 = ::Rumale::Validation.check_convert_sample_array(x)
+        x = Rumale::Validation.check_convert_sample_array(x)
 
         n_samples = x.shape[0]
         predicted = if multiclass_problem?
@@ -169,13 +140,13 @@ module Rumale
       # @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 = ::Rumale::Validation.check_convert_sample_array(x)
+        x = Rumale::Validation.check_convert_sample_array(x)
 
         if multiclass_problem?
           probs = 1.0 / (Numo::NMath.exp(@prob_param[true, 0] * decision_function(x) + @prob_param[true, 1]) + 1.0)
           (probs.transpose / probs.sum(axis: 1)).transpose.dup
         else
-          n_samples, = x.shape
+          n_samples = x.shape[0]
           probs = Numo::DFloat.zeros(n_samples, 2)
           probs[true, 1] = 1.0 / (Numo::NMath.exp(@prob_param[0] * decision_function(x) + @prob_param[1]) + 1.0)
           probs[true, 0] = 1.0 - probs[true, 1]
@@ -185,14 +156,35 @@ module Rumale
 
       private
 
-      def partial_fit(x, bin_y)
-        w, b = super
-        p = if @params[:probability]
-              ::Rumale::ProbabilisticOutput.fit_sigmoid(x.dot(w.transpose) + b, bin_y)
-            else
-              Numo::DFloat[1, 0]
-            end
-        [w, b, p]
+      def partial_fit(base_x, bin_y)
+        fnc = proc do |w, x, y, reg_param|
+          n_samples = x.shape[0]
+          z = x.dot(w)
+          t = 1 - y * z
+          loss = 0.5 * reg_param * w.dot(w) + (x.class.maximum(0, t)**2).sum.fdiv(n_samples)
+          indices = t.gt(0)
+          grad = reg_param * w
+          if indices.count.positive?
+            sx = x[indices, true]
+            sy = y[indices]
+            grad += 2.fdiv(n_samples) * sx.transpose.dot((sx.dot(w) - sy))
+          end
+          [loss, grad]
+        end
+
+        n_features = base_x.shape[1]
+        w_init = Numo::DFloat.zeros(n_features)
+
+        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: @params[:verbose] ? 1 : -1
+        )
+
+        prb = @params[:probability] ? Rumale::ProbabilisticOutput.fit_sigmoid(base_x.dot(res[:x]), bin_y) : Numo::DFloat[1, 0]
+        w, b = split_weight(res[:x])
+
+        [w, b, prb]
       end
 
       def multiclass_problem?

data/lib/rumale/linear_model/svr.rb

@@ -1,13 +1,15 @@
 # frozen_string_literal: true
 
+require 'lbfgsb'
+
 require 'rumale/base/regressor'
 require 'rumale/validation'
-require 'rumale/linear_model/base_sgd'
+
+require_relative 'base_estimator'
 
 module Rumale
   module LinearModel
-    # SVR is a class that implements Support Vector Regressor
-    # with stochastic gradient descent optimization.
+    # SVR is a class that implements Support Vector Regressor with the squared epsilon-insensitive loss.
     #
     # @note
     #   Rumale::SVM provides linear and kernel support vector regressor based on LIBLINEAR and LIBSVM.
@@ -17,70 +19,39 @@ module Rumale
     # @example
     #   require 'rumale/linear_model/svr'
     #
-    #   estimator =
-    #     Rumale::LinearModel::SVR.new(reg_param: 1.0, epsilon: 0.1, max_iter: 1000, batch_size: 50, random_seed: 1)
+    #   estimator = Rumale::LinearModel::SVR.new(reg_param: 1.0, epsilon: 0.1)
     #   estimator.fit(training_samples, traininig_target_values)
     #   results = estimator.predict(testing_samples)
-    #
-    # *Reference*
-    # - Shalev-Shwartz, S., and Singer, Y., "Pegasos: Primal Estimated sub-GrAdient SOlver for SVM," Proc. ICML'07, pp. 807--814, 2007.
-    # - 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 SVR < BaseSGD
-      include ::Rumale::Base::Regressor
-
-      # Return the weight vector for SVR.
-      # @return [Numo::DFloat] (shape: [n_outputs, n_features])
-      attr_reader :weight_vec
-
-      # Return the bias term (a.k.a. intercept) for SVR.
-      # @return [Numo::DFloat] (shape: [n_outputs])
-      attr_reader :bias_term
-
-      # Return the random generator for performing random sampling.
-      # @return [Random]
-      attr_reader :rng
+    class SVR < Rumale::LinearModel::BaseEstimator
+      include Rumale::Base::Regressor
 
       # Create a new regressor with Support Vector Machine by the SGD optimization.
       #
-      # @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).
-      # @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'.
-      # @param momentum [Float] The momentum factor.
-      # @param penalty [String] The regularization type to be used ('l1', 'l2', and 'elasticnet').
-      # @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.
       # @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.
       # @param epsilon [Float] The margin of tolerance.
       # @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.
       # @param tol [Float] The tolerance of loss for terminating 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.
       # @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(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,
-                     epsilon: 0.1,
-                     max_iter: 1000, batch_size: 50, tol: 1e-4,
-                     n_jobs: nil, verbose: false, random_seed: nil)
+      def initialize(reg_param: 1.0, fit_bias: true, bias_scale: 1.0, epsilon: 0.1, max_iter: 1000, tol: 1e-4,
+                     n_jobs: nil, verbose: false)
         super()
-        @params.merge!(method(:initialize).parameters.to_h { |_t, arg| [arg, binding.local_variable_get(arg)] })
-        @params[:decay] ||= @params[:reg_param] * @params[:learning_rate]
-        @params[:random_seed] ||= srand
-        @rng = Random.new(@params[:random_seed])
-        @penalty_type = @params[:penalty]
-        @loss_func = ::Rumale::LinearModel::Loss::EpsilonInsensitive.new(epsilon: @params[:epsilon])
+        @params = {
+          reg_param: reg_param,
+          fit_bias: fit_bias,
+          bias_scale: bias_scale,
+          epsilon: epsilon,
+          max_iter: max_iter,
+          tol: tol,
+          n_jobs: n_jobs,
+          verbose: verbose
+        }
       end
 
       # Fit the model with given training data.
@@ -89,9 +60,9 @@ module Rumale
       # @param y [Numo::DFloat] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
       # @return [SVR] The learned regressor itself.
       def fit(x, y)
-        x = ::Rumale::Validation.check_convert_sample_array(x)
-        y = ::Rumale::Validation.check_convert_target_value_array(y)
-        ::Rumale::Validation.check_sample_size(x, y)
+        x = Rumale::Validation.check_convert_sample_array(x)
+        y = Rumale::Validation.check_convert_target_value_array(y)
+        Rumale::Validation.check_sample_size(x, y)
 
         n_outputs = y.shape[1].nil? ? 1 : y.shape[1]
         n_features = x.shape[1]
@@ -121,6 +92,37 @@ module Rumale
 
         x.dot(@weight_vec.transpose) + @bias_term
       end
+
+      private
+
+      def partial_fit(base_x, single_y)
+        fnc = proc do |w, x, y, eps, reg_param|
+          n_samples = x.shape[0]
+          z = x.dot(w)
+          d = y - z
+          loss = 0.5 * reg_param * w.dot(w) + (x.class.maximum(0, d.abs - eps)**2).sum.fdiv(n_samples)
+          c = x.class.zeros(n_samples)
+          indices = d.gt(eps)
+          c[indices] = -d[indices] + eps if indices.count.positive?
+          indices = d.lt(eps)
+          c[indices] = -d[indices] - eps if indices.count.positive?
+          grad = reg_param * w + 2.fdiv(n_samples) * x.transpose.dot(c)
+          [loss, grad]
+        end
+
+        base_x = expand_feature(base_x) if fit_bias?
+
+        n_features = base_x.shape[1]
+        w_init = Numo::DFloat.zeros(n_features)
+
+        res = Lbfgsb.minimize(
+          fnc: fnc, jcb: true, x_init: w_init, args: [base_x, single_y, @params[:epsilon], @params[:reg_param]],
+          maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON,
+          verbose: @params[:verbose] ? 1 : -1
+        )
+
+        split_weight(res[:x])
+      end
     end
   end
 end

data/lib/rumale/linear_model/version.rb

@@ -5,6 +5,6 @@ module Rumale
   # This module consists of the classes that implement generalized linear models.
   module LinearModel
     # @!visibility private
-    VERSION = '0.24.0'
+    VERSION = '0.25.0'
   end
 end

data/lib/rumale/linear_model.rb

@@ -2,7 +2,8 @@
 
 require 'numo/narray'
 
-require_relative 'linear_model/base_sgd'
+require_relative 'linear_model/sgd_classifier'
+require_relative 'linear_model/sgd_regressor'
 require_relative 'linear_model/elastic_net'
 require_relative 'linear_model/lasso'
 require_relative 'linear_model/linear_regression'

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: rumale-linear_model
 version: !ruby/object:Gem::Version
-  version: 0.24.0
+  version: 0.25.0
 platform: ruby
 authors:
 - yoshoku
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2022-12-31 00:00:00.000000000 Z
+date: 2023-01-18 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: lbfgsb
@@ -44,14 +44,14 @@ dependencies:
     requirements:
     - - "~>"
       - !ruby/object:Gem::Version
-        version: 0.24.0
+        version: 0.25.0
   type: :runtime
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - "~>"
       - !ruby/object:Gem::Version
-        version: 0.24.0
+        version: 0.25.0
 description: |
   Rumale::LinearModel provides linear model algorithms,
   such as Logistic Regression, Support Vector Machine, Lasso, and Ridge Regression
@@ -65,13 +65,16 @@ files:
 - LICENSE.txt
 - README.md
 - lib/rumale/linear_model.rb
-- lib/rumale/linear_model/base_sgd.rb
+- lib/rumale/linear_model/base_estimator.rb
 - lib/rumale/linear_model/elastic_net.rb
 - lib/rumale/linear_model/lasso.rb
 - lib/rumale/linear_model/linear_regression.rb
 - lib/rumale/linear_model/logistic_regression.rb
 - lib/rumale/linear_model/nnls.rb
 - lib/rumale/linear_model/ridge.rb
+- lib/rumale/linear_model/sgd_classifier.rb
+- lib/rumale/linear_model/sgd_estimator.rb
+- lib/rumale/linear_model/sgd_regressor.rb
 - lib/rumale/linear_model/svc.rb
 - lib/rumale/linear_model/svr.rb
 - lib/rumale/linear_model/version.rb